Multiple-instance learning for whole-slide images

Multiple-instance learning (MIL) is the dominant way to train a whole-slide image (WSI) model when the available target belongs to a slide, case, or patient rather than to every tile. It is attractive because it replaces dense annotation with weak labels. It is difficult because the weak label does not say where the evidence is, how much evidence is required, or whether several tissue patterns must interact.

Latest literature discovery audit: 2 October 2026. Existing paper reviews retain their individual source-check dates.

Chapter coverage: a formal WSI-MIL model, an architecture taxonomy, primary readings organized by scientific mechanism, detailed routes through CLAM/DSMIL/TransMIL/DTFD and their successors, spatial and multiscale models, long-sequence/state-space models, survival and biomarker adaptations, sampling and end-to-end learning, a leakage-safe evaluation protocol, reproducibility records, and research exercises.

Maintenance policy: this is a curated research route, not an automatically generated leaderboard. New papers enter only after their identity, status, method, evaluation unit, and scientific role have been checked.

Scope, evidence, and publication policy

This chapter covers weakly supervised learning from pathology WSIs when tiles or regions are treated as instances and the learning target is attached to a bag. It includes related set, graph, hierarchical, long-context, and end-to-end models when they answer the same slide-to-label problem.

It does not claim that one architecture is universally best. Results in this literature change with:

• the patient population, specimen and endpoint;
• the patch encoder and its pretraining overlap;
• microns per pixel (MPP), tile size, tissue mask, and sampling policy;
• whether coordinates and multiple magnifications are available;
• the patient split, external site, tuning budget, and reported metric.

The year in each reading table is the peer-reviewed version-of-record year unless the row explicitly says preprint. Workshop papers are marked. An implementation link is called official only when the paper or an author organization identifies it as such. Absence of a code link is not evidence that no implementation exists.

The self-supervised-learning chapter explains how tile representations are learned. The foundation-encoder guide explains checkpoint identity, preprocessing, and fair encoder comparison. This chapter starts where a set of tile representations and a slide-, case-, or patient-level target meet.

Prerequisites and the shortest route

Level	Prerequisite	You are ready when you can…
L0 · WSI orientation	supervised classification, CNNs/ViTs, train/validation/test splits	trace patient → case → slide → tile and state the unit of every label
L1 · MIL foundation	softmax, cross-entropy, attention, permutation invariance	distinguish instance scoring from embedding aggregation and derive mean, max, and attention pooling
L2 · pathology research	MPP, tissue masking, frozen encoders, class imbalance	state the biological meaning of a positive bag and audit whether a sampled bag can contain the evidence
L3 · research synthesis	transformers, graphs, survival analysis, calibration, causal confounding	design a leakage-safe comparison and explain which model assumption matches the endpoint

Take this minimum 14-step route before designing a new WSI aggregator:

1. Dietterich et al. (1997) for the classical bag/instance problem.
2. Deep Sets (2017) for permutation-invariant set functions.
3. Attention-based Deep MIL (2018) for trainable pooling.
4. Campanella et al. (2019) for large-scale weakly supervised WSI training.
5. CLAM (2021) for attention plus instance-level clustering.
6. DSMIL (2021) for instance and bag streams.
7. TransMIL (2021) for correlated instances and approximate self-attention.
8. DTFD-MIL (2022) for pseudo-bags and two-tier distillation.
9. HIPT (2022) or ZoomMIL (2022) for hierarchy and scale.
10. MHIM-MIL (2023) for hard-instance masking.
11. SlideGCD (2024) for slide-graph collaborative online distillation.
12. RRT-MIL (2024) and LongMIL (2024) for re-embedding and long context.
13. S4MIL (2023) then MambaMIL (2024) for state-space sequence models.
14. xMIL (2024) for explanation faithfulness.
15. FEATHER (2025) for the question of whether a supervised MIL aggregator transfers.

Then choose the endpoint-specific route in survival and biomarker learning.

The formal problem: a slide is not automatically a classical MIL bag

Let patient p have one or more slides. For slide s, tissue processing produces instances and coordinates

B_s = {(x_i, c_i, m_i)} for i = 1,...,n_s

where x_i is a tile, c_i is its coordinate, and m_i records MPP, magnification, or scale. A common two-stage system is

h_i = f_encoder(x_i)
u_i = f_reembed(h_i, c_i, neighbours)
z_s = A({u_i}, {c_i})
ŷ_s = g(z_s)

f_encoder may be frozen or trainable. f_reembed is optional. A may be a fixed pool, learned attention, transformer, graph network, prototype model, or state-space model. A case or patient model needs one more explicitly defined aggregation over slides.

Three different label assumptions

1. Standard or existential MIL: a positive bag contains at least one positive instance and a negative bag contains none. In binary notation, Y = max_i(y_i). Small metastasis detection is close to this assumption.
2. Collective MIL: the class depends on the distribution or proportion of several instance phenotypes. Grade, subtype, inflammation, and many molecular targets are often closer to this setting.
3. Relational MIL: the label depends on spatial organization or interactions among instances. Invasion, tissue architecture, and tumour–immune relations may require coordinates, graphs, multiple scales, or contextual sequence models.

Do not justify an existential top-1 model merely by calling the dataset “MIL.” Write the target assumption before selecting the pooling operator.

Instance-space, embedding-space, and hybrid models

Paradigm	Computation	Strength	Failure to test
Instance-score	classify each instance, then pool scores	direct ranking and efficient rare-evidence detection	weak labels can turn normal tiles in positive slides into false positives
Embedding-space	aggregate features, then classify the bag	can represent distributed evidence	bag prediction need not imply meaningful instance scores
Hybrid	combine an instance scorer with a learned bag representation	preserves local and global signals	two branches may exploit different shortcuts or be poorly calibrated

Architecture taxonomy

These rows are overlapping design families, not mutually exclusive model classes.

Family	Representative systems	What is added	Invariance / complexity question
Fixed pooling	mean, max, log-sum-exp, top-k, noisy-OR	a known symmetric reduction	what fraction of tissue should determine the label?
Learned attention pooling	ABMIL, CLAM, MAD-MIL	trainable non-negative instance weights	are weights stable, faithful, and calibrated as localization?
Instance–bag hybrid	DSMIL	a critical-instance stream plus a bag stream	does one selected instance define a useful query for diffuse disease?
Clustering / prototypes	CLAM, DTFD, TPMIL, Flow-MIL	pseudo-labels, pseudo-bags, or latent prototypes	do clusters reflect morphology, class, site, stain, or sampling frequency?
Set / transformer	Set Transformer, TransMIL	pairwise or induced interactions	is the position encoding meaningful and is attention approximation adequate?
Graph / topology	FS-GCN-MIL, Patch-GCN, SlideGraph+, WSI-HGNN, WiKG, TopoSlide	explicit spatial, cell, semantic, or topological relations	how are nodes and edges constructed, and what true tissue relation do they encode?
Hierarchical / multiscale	HIPT, ZoomMIL, HACT-Net	tile → region → slide or cell → tissue structure	are levels aligned, registered, and sampled at physical rather than nominal scale?
Long-context local/global	RRT-MIL, LongMIL	re-embedding or sparse/local-global attention	what information is lost by neighbourhood windows, chunks, or low-rank attention?
State-space / Mamba	S4MIL, MambaMIL, M3amba, GMMamba	linear-complexity sequence state	which scan order is used and how sensitive is the result to permutation?
Transfer / distillation	DTFD, MHIM, SlideGCD, FEATHER, MKD, HistoMILKD	tiers, online graph peers, pretrained slide aggregators, multimodal branches, or multiple encoder teachers	what transfers: instance evidence, graph neighbourhoods, encoder information, aggregator weights, modality-specific knowledge, task taxonomy, or site bias?
End-to-end pixels	EPL, ZoomMIL, Pixel-Mamba, E2E-ViT	slide supervision reaches image features	what pixels receive gradients, and is “end to end” genuinely raw-pixel training?
Outcome / multimodal	DeepAttnMISL, MCAT, PORPOISE, MOTCat, SurvPath	censor-aware loss and/or omics–histology fusion	is the prediction available at the stated clinical time and does each modality add value?

Part I — foundations of bags, sets, pooling, and attention

Read these papers to recover the assumptions that pathology implementations often hide behind a model name.

ID	Primary paper and year	Status / official resource	What to recover
F1	Solving the Multiple Instance Problem with Axis-Parallel Rectangles (1997)	Artificial Intelligence, peer reviewed	bags, latent instance identity, and the existential positive-bag assumption
F2	A Framework for Multiple-Instance Learning (1998; NeurIPS 1997 proceedings)	NeurIPS, peer reviewed	Diverse Density and the idea that positive bags intersect around an unknown concept
F3	Deep Multiple Instance Learning for Image Classification and Auto-Annotation (2015)	CVPR, peer reviewed; author project	max, mean, and log-sum-exp pooling inside a deep weakly supervised system
F4	Deep Sets (2017)	NeurIPS, peer reviewed	why rho(sum_i phi(x_i)) is a basic permutation-invariant construction
F5	Attention-based Deep Multiple Instance Learning (2018)	ICML, peer reviewed; author code	gated versus ungated attention, normalized weights, and embedding-space aggregation
F6	Set Transformer (2019)	ICML, peer reviewed; code linked by PMLR	interactions within a set and inducing-point attention for reduced cost
F7	Attention is not Explanation (2019)	NAACL, peer reviewed	why a normalized attention weight is not automatically a faithful causal contribution

Derive the baseline before reading variants

For encoded instances h_i, gated attention pooling computes

a_i = softmax_i(wᵀ[tanh(Vh_i) ⊙ sigmoid(Uh_i)])
z   = Σ_i a_i h_i
ŷ   = g(z)

Check four facts:

1. Permuting the input instances leaves z unchanged if no coordinate- or order-dependent operation is inserted.
2. Attention weights compete through the softmax. Adding many irrelevant tiles can change every weight even when their content is unchanged.
3. A high a_i means the learned aggregator uses that embedding strongly relative to the other embeddings in this bag; it does not prove that the tile is malignant or that removing it causes the prediction to change.
4. The representation h_i is part of the model. Comparing aggregators while silently changing the encoder is not an aggregator comparison.

Checkpoint: implement mean, max, log-sum-exp, ABMIL, and a linear instance-score pooler on the same frozen synthetic bags. Test bag-size shift, instance permutation, one-positive-instance bags, and proportion-defined bags.

Part II — the core WSI-MIL lineage

Early weak supervision at slide scale

ID	Primary paper and year	Status / official resource	Contribution and question
C1	Classification and Disease Localization in Histopathology Using Only Global Labels (2018)	arXiv preprint; earlier workshop version	Chowder-style top positive and negative evidence; ask when negative evidence is diagnostically meaningful
C2	Clinical-grade computational pathology using weakly supervised deep learning on whole slide images (2019)	Nature Medicine, peer reviewed	large-scale iterative tile ranking, recurrent aggregation, and the effect of data scale
C3	Processing Megapixel Images with Deep Attention-Sampling Models (2019)	ICML, peer reviewed; code linked by PMLR	differentiable attention sampling and the compute–coverage trade-off
C4	Multi-scale Domain-adversarial Multiple-instance CNN for Cancer Subtype Classification (2020)	CVPR, peer reviewed	multi-scale weak supervision plus domain-adversarial learning
C5	Beyond Classification: Whole Slide Tissue Histopathology Analysis by End-to-End Part Learning (2020)	MIDL, peer reviewed	cluster tiles into parts and backpropagate slide loss through representative pixels

Four anchors: CLAM, DSMIL, TransMIL, and DTFD-MIL

ID	Primary paper and year	Status / official implementation	What changes
C6	Data-efficient and weakly supervised computational pathology on whole-slide images (CLAM) (2021)	Nature Biomedical Engineering, peer reviewed; author code	class-specific gated attention plus instance-level clustering constraints
C7	Dual-stream Multiple Instance Learning Network for Whole Slide Image Classification (DSMIL) (2021)	CVPR, peer reviewed; author code	combines an instance classifier with a similarity-weighted bag classifier
C8	TransMIL: Transformer based Correlated Multiple Instance Learning (2021)	NeurIPS, peer reviewed; author code	Nyström self-attention and a pyramid positional-encoding generator
C9	DTFD-MIL: Double-Tier Feature Distillation Multiple Instance Learning (2022)	CVPR, peer reviewed; author code	pseudo-bags enlarge the effective bag count; tier-one features are distilled for tier-two MIL

CLAM teardown

• Input: typically all retained tile features from a frozen encoder.
• Aggregator: gated, class-specific attention; CLAM-SB uses a single attention branch and CLAM-MB uses multiple class branches.
• Auxiliary signal: high- and low-attention instances receive pseudo cluster targets; the paper uses an instance-level smooth SVM objective in addition to slide classification.
• Value: it regularizes the feature space using representative evidence and produces inspectable heatmaps with a complete WSI-processing pipeline.
• Failure tests: instability across folds/seeds; tissue or scanner artifacts in high-attention regions; false localization when the slide label is collective; sensitivity to the number of sampled positive/negative instances.

DSMIL teardown

• Instance stream: scores every instance and selects a highest-scoring critical instance for a class.
• Bag stream: uses similarity to the critical instance to weight value embeddings and construct a bag representation.
• Value: instance evidence and bag evidence are trained together rather than assuming one fixed pool is sufficient.
• Failure tests: the critical instance can be an artifact; one query can be a poor summary of heterogeneous or diffuse evidence; self-supervised pretraining in the original system is a separate contributor from the MIL head.

TransMIL teardown

• Input: a sequence of tile features plus a class token.
• Context: Nyström attention approximates full self-attention; a pyramid position-encoding generator injects two-dimensional neighbourhood information.
• Value: explicitly models correlated instances rather than treating the slide only as independent evidence.
• Failure tests: padding and square-grid reconstruction; approximate-attention rank; coordinate scaling; long-bag memory; whether shuffling coordinates changes the result in the expected way.

DTFD-MIL teardown

• Pseudo-bags: one WSI bag is partitioned into smaller bags that inherit the slide label.
• Tier one: an attention MIL model processes each pseudo-bag and distils a representation using attention features or selected high/low evidence.
• Tier two: the distilled pseudo-bag representations form a new bag for final prediction.
• Value: increases the number of optimization units when slides are few and tiles are many.
• Failure tests: a positive slide’s pseudo-bag can contain no positive evidence; random partitioning can break spatial structures; performance can depend on the number of pseudo-bags and distillation variant.

Direct successors: selection, confounding, and hard instances

ID	Primary paper and year	Status / official resource	Read it for
C10	Interventional Bag Multi-Instance Learning on Whole-Slide Pathological Images (IBMIL) (2023)	CVPR, peer reviewed; author code	a causal-intervention framing for bag-level confounding; audit how the confounder dictionary is estimated
C11	Multiple Instance Learning with Masked Hard Instance Mining (MHIM-MIL) (2023)	ICCV, peer reviewed; author code	an EMA teacher identifies high-attention/easy evidence to mask so the student must learn harder instances
C12	Dynamic Policy-Driven Adaptive Multi-Instance Learning (PAMIL) (2024)	CVPR, peer reviewed	reinforcement-learning-driven dynamic instance sampling and pseudo-bag bias correction
C13	SlideGCD: Slide-based Graph Collaborative Training with Knowledge Distillation for WSI Classification (2024)	MICCAI, peer reviewed; author code	turns slides into graph-linked collaborators and uses online knowledge distillation; audit graph construction, queue/state leakage, and whether gains survive patient-disjoint external cohorts
C14	Attention Entropy Maximization (AEM) (2025)	MICCAI, peer reviewed; author code	regularizes softmax-normalized ABMIL attention against early over-concentration with cosine weight annealing; audit whether entropy improves generalization rather than merely flattening evidence

Required synthesis: recreate all four anchor models with the same features, folds, optimization budget, and evaluation code. Then add one successor at a time. If the ordering changes when the encoder changes, report an encoder–aggregator interaction rather than a universal ranking.

Part III — space, graphs, hierarchy, and multiple magnifications

A coordinate-free set model intentionally discards topology. That is a useful baseline, not proof that topology is irrelevant. Spatial models differ in what they call a node and what they believe an edge means.

Graph and hierarchy reading route

ID	Primary paper and year	Status / official resource	Representation to inspect
X1	Predicting Lymph Node Metastasis Using Histopathological Images Based on MIL with Deep Graph Convolution (FS-GCN-MIL) (2020)	CVPR, peer reviewed	feature selection followed by graph convolution when standard MIL is insufficient
X2	Whole Slide Images are 2D Point Clouds (Patch-GCN) (2021)	MICCAI, peer reviewed; author code	a spatially resolved patch graph and hierarchical message passing for survival
X3	Hierarchical Graph Representations in Digital Pathology (HACT-Net) (2022)	Medical Image Analysis, peer reviewed; author code/data	cell graphs, tissue graphs, and explicit cell-to-tissue assignment
X4	SlideGraph+: Whole Slide Image Level Graphs to Predict HER2 Status (2022)	Medical Image Analysis, peer reviewed; author code	WSI graphs combining deep and cellular morphology features
X5	Scaling Vision Transformers to Gigapixel Images via Hierarchical Self-Supervised Learning (HIPT) (2022)	CVPR, peer reviewed; author code	nested tile/region/slide representations and hierarchical pretraining
X6	Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images (ZoomMIL) (2022)	ECCV, peer reviewed; author code	differentiable selection across magnifications and its compute savings
X7	Histopathology WSI Analysis with Heterogeneous Graph Representation Learning (2023)	CVPR, peer reviewed; author code	nucleus types, heterogeneous edge attributes, semantic pooling, and causal attribution
X8	Dynamic Graph Representation with Knowledge-aware Attention (WiKG) (2024)	CVPR, peer reviewed; author code	directed, dynamically chosen semantic neighbours rather than only physical adjacency
X9	Feature Re-Embedding: Towards Foundation Model-Level Performance (RRT-MIL) (2024)	CVPR, peer reviewed; author code	a portable local/global regional transformer that adapts frozen tile features online
X10	TopoSlide: Topologically-Informed Histopathology WSI Representation Learning (2026)	CVPR, peer reviewed	persistent-homology targets for global tissue organization and topology-conditioned retrieval
X11	Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation (HDMIL) (2025)	CVPR, peer reviewed	high-to-low-resolution hierarchical distillation; report which evidence survives resolution transfer and the resulting compute
X12	LRMIL: Efficient Low-Resolution MIL via High-Resolution Knowledge Distillation (2026)	MICCAI 2026, peer reviewed; official proceedings verified; official code	align one 5× field with sixteen 20× patches, then distil bag logits and instance attention so deployment retains only the low-resolution encoder and MIL head
X13	Dual-Level Hypergraph Generation for Addressing Feature Scarcity in WSI Classification (2026)	CVPR, peer reviewed	instance- and bag-level hypergraph structure under feature scarcity; test the graph construction against simpler spatial and semantic neighbours
X14	A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification (2024)	IEEE TMI, peer reviewed; official code	graph context within and between 5×/10×/20× streams plus prediction self-distillation; all scale branches remain at inference
X15	Reducing Histopathology Slide Magnification Improves the Accuracy and Speed of Ovarian Cancer Subtyping (2024)	ISBI, peer reviewed	direct low-magnification ABMIL control; separates the benefit of a lower-scale view from any distillation mechanism
X16	CEMIL (2025)	WACV, peer reviewed	frozen instructor modules score every pre-extracted 20× embedding, then transfer representations and predictions into a subset learner; feature extraction is not shown removed
X17	PatHME (2025)	IEEE Access, peer reviewed; official code	prompt-guided cross-scale and multi-expert foundation-model transfer followed by hierarchical multiscale attention
X18	MAG-GLTrans (2026)	Cancer Research, peer reviewed; official code	registered magnification alignment trains a low-resolution global–local slide model; high-resolution evidence is training-only in the low-resolution deployment path
X19	LaGuadia (2026)	MICCAI 2026, peer reviewed; official code	a language-guided meta-teacher adapts several pathology foundation-model teachers into a compact student used across MIL and multimodal tasks

LRMIL and HDMIL answer different deployment questions. HDMIL uses a coarse branch to choose evidence and still evaluates selected high-resolution regions; LRMIL pays the high-resolution cost during training, transfers that information at patch and slide levels, and runs only the 5× student at inference. The latter therefore tests whether high-resolution semantics can survive resolution transfer—not whether a selector can find every rare high-resolution focus. Its reported “survival” experiments classify alive/dead status and must not be read as censoring-aware time-to-event analysis.

A project-ready 20× teacher → 5× deployment comparison matrix

Here “20× → 5×” names the knowledge-transfer direction: a deployed 5× model aims to recover useful 20× teacher information. It does not claim that a 5× image can reconstruct detail that was never sampled.

Do not begin by choosing a loss. First choose the deployment contract.

Family	Teacher / privileged evidence	Deployed evidence	Transfer interface	Essential control
Resolution KD	higher-resolution patch classifier	lower-resolution patch classifier	logits plus spatial feature maps	native low-resolution student with the same architecture and optimization
XMAG	frozen 20× UNI2 over sixteen aligned fields	one compact 5× encoder	mean teacher-tile CLS ↔ student CLS, plus sixteen teacher CLS ↔ sixteen pooled student regions	5× student without KD and raw per-WSI time rather than prose speed ratio
MAG-GLTrans	registered 20× representations	5× global–local WSI model	multilayer feature alignment	identical 5× architecture without alignment and a communication/storage ledger
LRMIL	20× CONCH encoder and 20× MIL teacher	5× CONCH-architecture encoder and 5× MIL head	multi-layer representation alignment between frozen 20× CLS embeddings and pooled same-field 5× spatial tokens; bag logit, soft-attention, and hard-instance losses	5× supervised-only encoder/MIL; same patches, split, and aggregation budget
Graph multi-scale KD / PatHME	multiple active scale or expert streams	the same multiscale system	prediction, feature, prompt, or expert alignment	report that high-resolution inference was not removed
ZoomMIL / HDMIL	coarse selector	chosen high-resolution subset	adaptive zoom or hierarchical feature transfer	rare-lesion coverage at fixed acquisition and wall-clock budget
CEMIL	all pre-extracted 20× embeddings plus frozen instructor	selected 20× embedding subset in the learner	instructor representation and prediction transfer	include full extraction and instructor-scoring cost, not learner FLOPs alone

Distillation targets: what to test before adding another loss

The cross-resolution diagnostic study (2026, preprint; ICLR 2027 under review, not accepted) separates three questions: does an observed 20× target improve the task, can a low-power predictor recover it, and does the slide model benefit from the prediction? Its direct predictor has lower MSE than its residual predictor, yet a constant training-mean target wins on several cohorts. Lower feature error is therefore not sufficient evidence of better slide prediction. Preserve native-feature and untrained-MLP controls.

Neighboring KD methods can provide useful controls without answering the same acquisition question:

Read next	What transfers	What it does not establish
MIL pretraining from TITAN and CARE (2026, preprint)	reusable aggregator initialization, with teacher-specific projections and dispersion-normalized alignment	that more teachers always help, or that the high-resolution tile encoder can be removed
MoPE (2026, preprint)	training-only RNA pathway supervision through shared-memory distributions	that missing molecular information and missing 20× pixels are equivalent problems
Efficient pathology segmentation KD (2026, preprint)	pixel predictions and features from an adapted Virchow2 teacher to smaller segmentation networks	end-to-end WSI speed from a patch-throughput measurement
PAGET (2026, MIDL)	hierarchical pseudo-labels from specialist teachers to a tissue-segmentation student	that every variant removes all teacher components: PAGET-H still invokes HoverNet
ConStruct (2026, MIDL)	SegFormer structural affinities into adapted CONCH features for prototype-guided segmentation	residual-error KD merely because residual adapters appear in the model

Start from matched patient splits, MPP/field of view, encoder identity, teacher pretraining overlap, training budget and inference inputs. Keep proxy feature metrics separate from the final slide endpoint. The detailed topic tables below and the general distillation route provide the corresponding mechanisms and primary sources.

For an LRMIL-like project, run ablations in this order: 5× supervised baseline; patch-level cross-resolution transfer; slide-logit transfer; soft-attention transfer; hard-instance transfer; then the complete system. Use patient-disjoint splits and freeze the teacher before student comparisons. Report extraction and cache generation for both scales, peak memory, wall time, storage, communication, calibration, and rare-evidence recall. A 5× model matching one 20× task does not prove that every cellular endpoint is recoverable from the coarser field.

Edge construction is a hypothesis

For every graph model, record:

node:
  entity: tile | nucleus | tissue-region | prototype
  feature_source: exact encoder or cell/tissue model
edge:
  rule: radius | k-nearest-coordinate | k-nearest-feature | learned-directed
  physical_units: micrometres | pixels | none
  attributes: []
hierarchy:
  levels: []
  assignment_rule: ""
pooling:
  operator: ""

A Euclidean k-nearest-neighbour graph encodes physical adjacency. A feature-space graph encodes phenotypic similarity. WiKG’s dynamic semantic edges can connect distant but similar tiles. Those are different biological hypotheses and should not be described simply as “using spatial context.”

Multiscale design questions

• Are scales specified in MPP, or only as scanner “10×/20×” labels?
• Does a high-magnification tile lie inside its low-magnification parent?
• Is the tissue registered across stains or serial sections?
• Are low- and high-scale views assumed equivalent, complementary, or predictive?
• Does coarse selection prevent rare high-resolution evidence from being visited?
• Is compute reported per sampled tile, slide, and patient?

Checkpoint: take one trained coordinate-free ABMIL model. Compare it with shuffled coordinates, true coordinates, feature-neighbour graphs, and coordinate-neighbour graphs. The comparison isolates whether gain comes from context, extra parameters, or an encoder change.

Foundation hierarchy, adaptive fusion, and acquired evidence

• MRPT (2026, preprint) learns a cell→region→WSI hierarchy with consecutive cross-resolution attention and semantic consistency. Compare each level with scale-matched mean/ABMIL pooling before attributing gains to a foundation-model hierarchy.
• Adaptive fusion of pathology foundation models (2026, preprint; code) asks whether several frozen encoders carry complementary instance/bag evidence. A fair test must match aggregate dimension, trainable parameters, task labels, and inference cost against concatenation, averaging, and the strongest single encoder.
• BEACON (2026, preprint; code) replaces relevance-only WSI retrieval with expected information gain under a diagnostic belief. Its unit is a sequentially acquired evidence set, so patch proposal recall, acquisition rounds, stopping, calibration, and final reasoning are separate denominators.

Part IV — long bags, local–global context, and state-space models

A tissue-rich WSI can yield tens of thousands of tile tokens. Full self-attention is quadratic in token count. Approximate attention, local windows, graph sparsity, hierarchy, prototypes, and state-space models solve different versions of this cost problem.

ID	Primary paper and year	Status / official resource	What to test
X13	Structured State Space Models for Multiple Instance Learning in Digital Pathology (S4MIL) (2023)	MICCAI, peer reviewed; author code	an S4 sequence learner for long WSI bags; permutation and ordering sensitivity
X14	MambaMIL: Enhancing Long Sequence Modeling with Sequence Reordering (2024)	MICCAI, peer reviewed; author code	selective state-space modeling, bidirectionality, and sequence reordering
X15	Rethinking Transformer for Long Contextual Histopathology WSI Analysis (LongMIL) (2024)	NeurIPS, peer reviewed; author code	chunked local attention plus global interaction, linear cost in a local bandwidth, and position extrapolation
X16	M3amba: Memory Mamba is All You Need for WSI Classification (2025)	CVPR, peer reviewed	dynamic memory across groups and bidirectional within-group modeling
X17	GMMamba: Group Masking Mamba for WSI Classification (2025)	ICCV, peer reviewed	location-based groups, instance masking, and cross-group super-feature sampling
X18	Bridging Local Inductive Bias and Long-Range Dependencies with Pixel-Mamba (2025)	ICCV, peer reviewed	hierarchical raw-pixel end-to-end processing rather than a conventional frozen-feature MIL head
X19	2DMamba: Efficient State Space Model for Image Representation with Applications on Gigapixel Images (2025)	CVPR, peer reviewed	a two-dimensional state-space construction; compare its ordering and spatial inductive bias with one-dimensional scans
X20	FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot WSI Classification (2025)	CVPR, peer reviewed	language-guided token compression; report retained visual evidence, language dependence, and compute
X21	Cracking Instance Jigsaw Puzzles: An Alternative to Multiple Instance Learning (2025)	ICCV, peer reviewed	order-aware instance reconstruction as an alternative to conventional bag aggregation; test whether the imposed order represents tissue structure

Sequence order is part of the method

Sets are permutation invariant; S4 and Mamba models are sequence models. A WSI does not arrive with one canonical one-dimensional order. Raster order, Hilbert curves, coordinate clustering, random order, attention order, and learned reordering expose different neighbours and histories.

Every state-space experiment should include:

1. the exact order-construction algorithm;
2. at least one fixed alternative order;
3. repeated random permutations at inference;
4. forward, reverse, and bidirectional ablations;
5. memory, throughput, and performance against bag length;
6. an all-instance ABMIL baseline and a spatial baseline using the same encoder.

Linear computational complexity does not guarantee complete spatial context, order invariance, or better clinical generalization.

Part V — prototypes, feature transformation, expert routing, case aggregation, explanations, transfer, and continual learning

ID	Primary paper and year	Status / official resource	Research question
N1	Trainable Prototype Enhanced MIL (TPMIL) (2024)	MIDL, peer reviewed	can softly assigned trainable prototypes refine the feature space and provide a different inspection target from attention?
N2	xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology (2024)	NeurIPS, peer reviewed; author code	explanation with instance interactions and quantitative faithfulness tests
N3	Do Multiple Instance Learning Models Transfer? (FEATHER) (2025)	ICML, peer reviewed; MIL-Lab and weights	transfer of supervised pan-cancer slide aggregators across organs and tasks
N4	Advancing MIL with Continual Learning for Whole Slide Imaging (2025)	CVPR, peer reviewed	attention knowledge distillation and pseudo-bag memory under class-incremental learning
N5	Flow-MIL: Constructing Highly-expressive Latent Feature Space Using Normalizing Flow (2025)	ICCV, peer reviewed	invertible feature transformation, latent semantic prototypes, and pseudo-instance labels
N6	SMMILe enables accurate spatial quantification in digital pathology using MIL (2025 online)	Nature Cancer, peer reviewed; author code linked by paper	superpatch-based measurable MIL and the distinction between slide classification and spatial quantification
N7	HistoMILKD: Multi-Teacher Knowledge Distillation for WSI Classification (2026)	WACV, peer reviewed; author code	distilling complementary pathology foundation encoders into one student-adapter/MIL system
N8	MoA: Mixture of Aggregators Improves Slide-Level Diagnosis (2026)	MIDL, peer reviewed	whether different pooling operators capture complementary disease-specific instance distributions
N9	MAMMOTH (2026)	ICLR, peer reviewed; author code	a mixture of low-rank mini-experts replaces the usual linear task projection; compare feature transformation before attributing gains to aggregation
N10	Learning Heterogeneous Tissues with Mixture of Experts for Gigapixel WSI Classification (PAMoE) (2025)	CVPR, peer reviewed	pathology-aware experts for heterogeneous tissues; audit expert collapse, routing stability, and organ/site specialization
N11	Continual Multiple Instance Learning with Enhanced Localization (CoMEL) (2025)	ICCV, peer reviewed	continual adaptation with localization; evaluate forgetting and localization independently
N12	Universal-to-Specific Dynamic Knowledge-Guided MIL for Few-Shot WSI Classification (2026)	CVPR, peer reviewed	few-shot transfer from universal to task-specific knowledge; compare against parameter- and label-budget-matched transfer baselines
N13	Multi-modal Knowledge Decomposition based Online Distillation for Biomarker Prediction in Breast Cancer Histopathology (2025)	MICCAI, peer reviewed; author code	decomposes histology and genomic knowledge for online distillation into a deployable biomarker predictor; audit missing-modality handling, patient alignment, and whether each branch contributes beyond matched unimodal baselines
N14	RACR-MIL (2026 accepted manuscript)	Medical Image Analysis acceptance is reported by the complete arXiv record; final DOI/issue pending at the cutoff	combines local and non-local WSI graph context with rank-aware ordinal attention for squamous-cell-carcinoma grading; audit grade ordering, anatomy transfer, graph construction, and reader-study denominators separately

MOOZY: patient/case-level multi-slide learning

MOOZY is accepted at ECCV 2026 according to the author project; proceedings were still forthcoming at the cutoff. It learns a slide encoder with masked self-distillation and then a case transformer over all slides assigned to a patient, supervised across 333 tasks from 56 datasets. Code and weights are public.

This is not ordinary single-slide MIL and not a VLM. Compare it only after matching the case definition, slide set, missing-slide policy, patient split, tile encoder, and probe. Report slide-level and case-level ablations separately.

Treat explanation as a separate evaluated output

An explanation claim needs its own target and metric.

Claim	Required reference	Suitable tests
“Finds tumour”	pixel/region tumour annotation	Dice/IoU, lesion FROC, point localization, pathologist review on held-out slides
“Identifies sufficient evidence”	model prediction under controlled interventions	deletion/insertion, patch flipping, minimal sufficient subsets, retraining sanity checks
“Discovers a phenotype”	blinded morphology or molecular validation	enrichment, reproducibility across folds/sites, independent assay or pathologist scoring
“Explains a patient risk”	a clearly defined explanation semantics	faithfulness, stability, subgroup analysis, counterfactual or perturbation tests

Attention heatmaps are useful exploratory objects. They are not tumour masks, causal maps, uncertainty maps, or proof of a new biomarker without these validations.

Transfer changes the baseline

FEATHER asks whether the aggregator, not only the patch encoder, can be pretrained. Therefore a current low-data study should compare:

• a randomly initialized simple ABMIL head;
• a pretrained transferable MIL head with the same patch encoder;
• a randomly initialized stronger aggregator;
• a slide foundation encoder, if available;
• full fine-tuning and a frozen slide-embedding probe under equal label/compute budgets.

Record the pretraining label taxonomy and patient/site overlap. Supervised pan-cancer pretraining can transfer useful morphology, but it can also transfer institution, sampling, and label-definition biases.

Part VI — survival, biomarker, and multimodal adaptations

Classification, survival, regression, and multimodal fusion are not interchangeable heads on an otherwise identical scientific problem.

Endpoint-aware reading route

ID	Primary paper and year	Status / official resource	What it adds
A1	WSISA: Making Survival Prediction from Whole Slide Histopathological Images (2017)	CVPR, peer reviewed	early weakly supervised WSI survival learning through patch grouping and multiple survival models
A2	Classification and Mutation Prediction from NSCLC Histopathology Images (2018)	Nature Medicine, peer reviewed; author code	morphology and mutation prediction from H&E; separate slide aggregation from molecular validity
A3	Deep learning can predict microsatellite instability directly from histology (2019)	Nature Medicine, peer reviewed	MSI prediction and the need for independent, population-specific validation
A4	Whole Slide Images Based Cancer Survival Prediction Using Attention Guided Deep MIL (DeepAttnMISL) (2020)	Medical Image Analysis, peer reviewed; author code	patient-level attention over WSI patterns for censored survival
A5	A deep learning model to predict RNA-Seq expression of tumours from WSIs (HE2RNA) (2020)	Nature Communications, peer reviewed	transcriptome prediction and spatialized tile scores from bulk expression targets
A6	Pan-cancer computational histopathology reveals mutations, tumour composition and prognosis (2020)	Nature Cancer, peer reviewed	broad image–genomic associations and external validation across cancer types
A7	Pan-cancer image-based detection of clinically actionable genetic alterations (2020)	Nature Cancer, peer reviewed; author workflow	multi-target weak supervision and task-specific external generalization
A8	Development and validation of a weakly supervised framework for molecular pathways and mutations in colorectal cancer (IDaRS) (2021)	Lancet Digital Health, peer reviewed; official TIAToolbox implementation	iterative random-plus-top tile sampling and external MSI validation
A9	Patch-GCN (2021)	MICCAI, peer reviewed; author code	spatial patch graphs for patient survival
A10	Multimodal Co-Attention Transformer for Survival Prediction (MCAT) (2021)	ICCV, peer reviewed; author code	genomics-guided co-attention that selects histology and reduces bag size
A11	Pan-cancer integrative histology-genomic analysis via multimodal deep learning (PORPOISE) (2022)	Cancer Cell, peer reviewed; author code	attention MIL, molecular features, Kronecker fusion, and pan-cancer risk attribution
A12	Multimodal Optimal Transport-based Co-Attention Transformer (MOTCat) (2023)	ICCV, peer reviewed; author code	unbalanced mini-batch optimal transport for global histology–genomic matching
A13	Modeling Dense Multimodal Interactions Between Biological Pathways and Histology (SurvPath) (2024)	CVPR, peer reviewed; author code	pathway tokens, histology tokens, memory-efficient fusion, and disease-specific survival
A14	ModalTune: Fine-Tuning Slide-Level Foundation Models with Multi-Modal Information for Multi-task Learning (2025)	ICCV, peer reviewed	multimodal adapters for slide foundation models; record every modality required during training and inference and include unimodal counterfactuals
A15	PANDA: Prototype-Anchored Alignment (2026 preprint)	arXiv; code is promised only upon acceptance	aligns partially paired WSI and RNA representations to class prototypes, then deploys without RNA; compare against histology-only, complete-case, missing-modality, and shuffled-pair controls, and retain the paper’s uncertainty/significance boundary
A16	Native-language prostate pathology report generation (2026 preprint)	arXiv; no public training code, weights, or paired clinical corpus verified	trains an end-to-end slide-to-German-report system from 17,344 core–report pairs across 2,402 cases; preserve case-level splitting, finding-level factuality, vocabulary coverage, and clinician-review requirements

Survival is a censored time-to-event problem

For each patient, define:

t0      = clinically valid prediction time
T       = observed event or censoring time measured from t0
δ       = 1 if the event is observed, 0 if censored
endpoint = overall | disease-specific | progression-free | other

The model may optimize Cox partial likelihood, a discrete-time hazard likelihood, a ranking loss, or another censor-aware objective. Those objectives answer different questions. Report:

• the time origin, event definition, censoring rule, and competing-event handling;
• one prediction per independent patient;
• Harrell or Uno concordance with confidence intervals and the estimator named;
• time-dependent AUROC and integrated Brier score at prespecified horizons when appropriate;
• calibration, not only discrimination;
• a clinical-covariate baseline and the incremental value of histology;
• how multiple slides and modalities are combined;
• missing-modality handling and whether missingness itself leaks outcome.

Do not choose a high/low-risk cut point on the test outcomes. Fit it in training or validation, lock it, and then plot test Kaplan–Meier curves. A log-rank p value is not a calibration assessment.

Biomarker prediction requires a causal and clinical audit

For mutation, expression, receptor, pathway, or treatment-response targets, ask:

1. Was the molecular assay performed on the same block, a neighbouring section, another slide, or only the same patient?
2. Is the label binary because of a clinical assay threshold, or because a continuous quantity was dichotomized for modelling convenience?
3. Is tumour content required and, if so, how was it selected without leaking the target?
4. Is prevalence correlated with submitting site, scanner, ancestry, stage, or treatment?
5. Does the external cohort use the same assay and label definition?
6. Would the model replace an assay, triage an assay, or generate a research hypothesis? These require different sensitivity, calibration, and validation.

The 2026 study Confounding factors and biases abound when predicting molecular biomarkers from histological images is a useful endpoint-specific audit: strong cross-validation does not establish that the learned signal is causal morphology rather than a cohort or site correlate.

Multimodal fusion needs unimodal counterfactuals

MCAT, PORPOISE, MOTCat, and SurvPath should be compared with:

• histology-only and omics-only models;
• clinical-covariate-only and clinical-plus-each-modality models;
• parameter- and tuning-budget-matched early, intermediate, and late fusion;
• missing and corrupted modality tests;
• permutation of one modality across patients as a leakage sanity check.

Cross-attention can visualize associations between modalities. It does not prove that a histologic region expresses a gene or pathway without spatially matched ground truth.

Part VII — sampling, pseudo-bags, and end-to-end learning

Sampling is an estimator, not preprocessing trivia

If a slide has N tissue tiles, r tiles contain rare diagnostic evidence, and k tiles are sampled without replacement, the probability of missing all evidence is

P(miss) = choose(N-r, k) / choose(N, k)
        ≈ (1 - r/N)^k when k is small relative to N.

This makes fixed-size random bags especially fragile for micrometastases or rare patterns. Report the empirical bag-size and tissue-area distributions, not only the chosen k.

Sampling strategy	Benefit	Bias / failure
all tissue tiles	maximum deterministic coverage	storage and long-context cost; large slides dominate
uniform random	simple stochastic regularization	misses rare evidence and underrepresents small compartments
class- or slide-balanced	reduces label-frequency imbalance	does not correct within-slide tissue imbalance
top-k / iterative rank	focuses compute on current evidence	self-reinforcing errors and artifact mining
attention sampling	can be differentiable and compute-aware	early attention can collapse exploration
clustering / prototypes	preserves diversity with fewer representatives	clustering may follow stain, site, or prevalence
pseudo-bags	creates more optimization units	inherited labels are noisy when evidence is sparse
contrastive cross-bag augmentation	expands bag-level variation and shared evidence	C2Aug can mix cohort, site, or label shortcuts unless cross-bag construction is audited
hard-instance masking	forces alternatives beyond easy evidence	may remove the only valid evidence or emphasize label noise
coarse-to-fine zoom	efficient multi-resolution search	an incorrect coarse decision prevents high-resolution recovery
language-guided coarse-to-fine selection	uses task concepts to rank low-resolution regions before selective 20× encoding	concepts can be incomplete or language-model-dependent; retained 20× encoding remains part of the cost ledger

LanGuSTE instantiates the last row: it builds task concepts with an LLM, adapts cross-scale visual prompts, scores the 5× view, and encodes only selected 20× patches. It is therefore a selective high-resolution acquisition method—not a 5×-only student and not conventional knowledge distillation. Compare it with low-only LRMIL/XMAG, retained-multiscale models, ZoomMIL-style acquisition, and post-extraction subset selection using the same physical field, retained-area budget, encoder, and complete I/O/encoding time.

Train/inference mismatch

Many systems train on sampled bags but infer on all tiles. Record both paths. Attention normalization, class token context, and state-space history can change with bag length; the all-tile prediction is not merely a lower-variance version of the sampled prediction.

Perform:

• repeated inference with different samples and report prediction variance;
• performance as a function of tile budget;
• a fixed deterministic evaluation bag plus all-tile evaluation;
• rare-lesion coverage analysis where region annotations exist;
• bag-size stress tests beyond the training distribution.

“End to end” has at least four meanings

1. Frozen two-stage: cache f_encoder(x_i); train only re-embedding, aggregation, and prediction.
2. Feature end-to-end: train all modules after cached tile embeddings, as in online feature re-embedding. This does not update pixels-to-feature encoding.
3. Selected-tile end-to-end: choose a tractable subset and backpropagate slide loss through those image tiles.
4. Whole-tissue end-to-end: stream, compress, or hierarchically process the tissue pixels so slide supervision updates image features across the modeled field.

EPL, ZoomMIL, Pixel-Mamba, and E2E-ViT explore different points in this design space. A frozen foundation encoder plus RRT-MIL remains feature-level adaptation, not raw-pixel end-to-end training.

Practical end-to-end reading

Method	Status / official resource	End-to-end mechanism and audit
C5 EPL (2020)	MIDL, peer reviewed	representative-part backpropagation; record which parts receive pixel gradients
X6 ZoomMIL (2022)	CVPR, peer reviewed	differentiable coarse-to-fine selection; measure rare-evidence coverage after coarse selection
X18 Pixel-Mamba (2025)	ICCV, peer reviewed	hierarchical pixel modeling rather than a conventional frozen-feature MIL head
E2E-ViT (2026)	CVPR, peer reviewed	parameter-neutral conversion of a pretrained ViT for end-to-end slide processing and survival prediction
FBTA (2026)	CVPR, peer reviewed	feature bridging for single-GPU end-to-end WSI classification; verify which image features actually receive slide-level gradients

Compare each method with a frozen-feature baseline because joint training cost and small-cohort overfitting can outweigh added adaptability.

Record the gradient-support set: exactly which slides, regions, tiles, pixels, and encoder blocks receive gradients in one optimization step.

Part VIII — exact evaluation and leakage protocol

1. Split the highest linked unit before any learning

patient → case → specimen → block → slide → region → tile

Split at the patient level unless a higher linked unit exists. Keep every derived tile, cached embedding, stain variant, serial section, and repeated scan with its parent. A tile-random split is invalid for slide or patient generalization.

AI slipping on tiles (2019) showed that putting tiles from the same subject in train and validation can markedly inflate digital pathology performance. This remains possible even when cross-validation code is otherwise correct.

2. Resolve the label unit

Label unit	Required model/evaluation unit
slide diagnosis	slide prediction; patient-disjoint split
case diagnosis from several slides	aggregate slides into one prespecified case prediction
molecular assay per patient	one patient prediction; document slide/block–assay matching
survival	one patient risk with one time origin and censoring record
lesion localization	slide prediction plus independent region/pixel reference

Do not count several slides from one patient as independent test observations when the target is patient-level. Confidence intervals must resample the independent unit.

3. Isolate validation from test

• Select encoder, MPP, tissue threshold, tile budget, architecture, learning rate, epoch, and decision threshold using training/validation data only.
• Use nested cross-validation when the dataset is too small for a fixed test set and extensive model selection is unavoidable.
• Do not report the best test result across learning rates, folds, seeds, or model variants.
• Lock the statistical analysis plan before opening an external test cohort.

4. Audit site and acquisition shortcuts

Howard et al. (2021) demonstrated that TCGA tissue-submitting sites are highly learnable from histology and can bias survival, genomic, and stage prediction. Stain normalization alone did not remove site identification.

At minimum:

• tabulate target prevalence by hospital, scanner, year, preparation, stain, and demographic group;
• train a site-prediction probe on the same features;
• evaluate leave-site-out or external-site performance;
• report site-stratified metrics and calibration;
• inspect slide labels, pen marks, tissue borders, blur, folds, coverslips, and background in high-attribution tiles.

5. Audit pretraining overlap

For every encoder and slide model, classify overlap as:

none | possible | known | unknown

Check patient, slide, institution, and public-cohort overlap. TCGA, CPTAC, CAMELYON, PANDA, and other public cohorts recur in pretraining and downstream benchmarks. “Frozen,” “zero-shot,” and “external” do not remove overlap.

6. Use endpoint-complete metrics

For classification, report:

• AUROC and AUPRC for imbalanced binary targets;
• class-wise and macro metrics for multiclass targets;
• sensitivity/specificity or predictive values at a validation-chosen threshold;
• calibration intercept/slope, reliability plots, and Brier score;
• patient-clustered confidence intervals and paired comparisons;
• coverage and selective risk if the model can abstain.

For localization, report region-level metrics against spatial labels. For survival, use the censor-aware protocol above. Accuracy alone is not sufficient for class imbalance, and AUROC alone does not define a clinical operating point.

7. Compare systems, not paper headlines

Use one locked matrix:

fixed:
  patients_and_folds: true
  tissue_mask_and_coordinates: true
  tile_mpp_size_and_transforms: true
  encoder_checkpoint: true
  sampled_or_all_tile_policy: true
  tuning_budget_and_early_stopping: true
  metric_code_and_bootstrap_units: true
varied:
  aggregator: [mean, max, ABMIL, candidate]

Then run a second matrix varying only the encoder. If resources allow, fit an encoder × aggregator factorial experiment. This reveals interaction effects that a single leaderboard column hides.

Reproducibility record

Store one record per run, including failed and aborted runs:

identity:
  experiment_id: ""
  commit: ""
  environment_lock: ""
  seed: null
cohort:
  eligibility: ""
  exclusions: {}
  patients_cases_slides: {}
  sites_scanners_years: {}
  target_definition: ""
  target_unit: patient | case | slide
split:
  strategy: ""
  grouping_key: patient_id
  train_validation_test_counts: {}
  duplicate_and_related-slide_audit: ""
wsi:
  reader_version: ""
  native_mpp_policy: ""
  analysis_mpp: null
  tile_pixels: []
  overlap_stride: []
  tissue_mask: ""
  quality_control: ""
  coordinates_preserved: true
sampling:
  train_policy: ""
  inference_policy: ""
  tiles_per_slide_distribution: {}
encoder:
  name_checkpoint_hash: ""
  official_transforms: ""
  frozen_blocks: []
  embedding_dimension: null
  pretraining_overlap: none | possible | known | unknown
mil:
  implementation_source: ""
  aggregator: ""
  position_or_graph_rule: ""
  bag_batching_padding: ""
  parameters_and_flops: {}
optimization:
  loss_and_weights: ""
  optimizer_schedule: ""
  epochs_and_early_stop: ""
  search_space_and_selection_metric: ""
evaluation:
  prediction_unit: ""
  metrics: []
  threshold_source: validation
  uncertainty_method: ""
  external_sites: []
  subgroup_plan: []
interpretation:
  method: ""
  reference_annotations: ""
  faithfulness_and_stability_tests: []
artifacts:
  checkpoints_features_predictions: []
  license_and_access_notes: ""

Also save patient-level predictions, fold manifests, sampled tile coordinates, feature-checkpoint hashes, and the exact code that produces every table. A mean and standard deviation without fold membership and predictions is not an auditable result.

A defensible experiment sequence

1. Define the target. Decide whether evidence is existential, collective, relational, temporal, or multimodal.
2. Freeze the cohort. Resolve patient/case/slide identifiers, site, assay, time origin, and external cohort before tiling.
3. Build transparent baselines. Mean, max, and ABMIL with one frozen encoder.
4. Measure sampling. Plot evidence coverage, bag length, prediction variance, and compute against tile budget.
5. Add the missing inductive bias. Use a graph only for a spatial hypothesis, a state-space model only for a long-order hypothesis, or prototypes only for a mixture hypothesis.
6. Run negative controls. Shuffle coordinates, sites, modalities, or attention maps as appropriate; permute sequence order; evaluate artifact-only regions.
7. Test external generalization. Lock preprocessing and decision threshold before the external cohort.
8. Evaluate explanation separately. Use spatial labels, perturbation faithfulness, stability, and blinded expert review.
9. Report compute and access. Include extraction time, training time, peak memory, feature storage, code, weights, and license.

Advanced research exercises

Exercise 1 — assumptions before architecture

Choose metastasis detection, tumour subtype, MSI, and survival. For each, state the plausible label assumption, required spatial scale, positive evidence prevalence, and a failure case for max, mean, and attention pooling.

Exercise 2 — reproduce the four anchors

Using one locked feature set, reproduce CLAM, DSMIL, TransMIL, and DTFD-MIL. Report patient-stratified folds, five seeds, bootstrap confidence intervals, runtime, memory, and paired differences. Do not copy reported scores from incompatible papers.

Exercise 3 — bag-size shift

Train at k ∈ {256, 1024, 4096} sampled tiles and evaluate at those sizes plus all tiles. Plot accuracy, calibration, prediction variance, and latency. Explain which models are sensitive to attention normalization or context length.

Exercise 4 — explanation falsification

For ABMIL, CLAM, TransMIL, and xMIL, perform high-to-low and low-to-high patch deletion, random deletion, seed stability, and coordinate perturbation. If tumour masks exist, compare model attribution with localization independently.

Exercise 5 — spatial hypothesis

Compare no coordinates, coordinate positional encoding, a coordinate-neighbour graph, a feature-neighbour graph, and shuffled edges. State which intervention supports a claim about tissue topology.

Exercise 6 — ordering in state-space MIL

Train S4MIL or MambaMIL under raster, Hilbert, random-fixed, and feature-clustered orders. Evaluate reversed and randomly permuted test sequences. A model that changes substantially under biologically equivalent orders needs that limitation reported.

Exercise 7 — biomarker confounding

For one TCGA biomarker, tabulate prevalence by submitting site, train a site probe, compare random patient folds with leave-site-out validation, and inspect whether site-aware performance changes the biological claim.

Exercise 8 — survival protocol

Reproduce a histology-only survival baseline with prespecified time origin, censoring, clinical covariates, C-index, time-dependent AUROC, integrated Brier score, and calibration. Fit any risk-group threshold without test outcomes.

Exercise 9 — frozen versus end to end

Compare cached features, online feature re-embedding, selected-tile encoder fine-tuning, and a true end-to-end method. Match label budget, count trainable parameters and gradient-supported pixels, and report total compute including feature extraction.

Exercise 10 — a useful negative result

Pre-register one plausible architecture improvement and one external cohort. Publish the result even if a mean or ABMIL baseline matches it. Analyze whether encoder, sampling, label noise, or cohort size limits the aggregator.

How to read a new MIL paper

Before accepting the paper’s main comparison, answer:

1. What is the independent unit and how was it split?
2. What makes a bag positive, and does that match the endpoint?
3. Which exact encoder, checkpoint, MPP, and transforms produced the instances?
4. Are all tiles used, or how are they sampled at train and inference?
5. Are coordinates used? If yes, in what physical coordinate system?
6. What is the actual complexity in bag length and peak memory?
7. Were baseline implementations, tuning budgets, and encoders matched?
8. Was the test set used to choose a model, epoch, threshold, or learning rate?
9. Are results patient-level with uncertainty and an external site?
10. Is a heatmap evaluated as an explanation, or merely shown?
11. Could pretraining or site overlap contaminate the claim?
12. Is code sufficient to recreate folds, features, bags, and predictions?

Maintenance notes

• Monthly: screen the Daily arXiv Radar for multiple instance learning, whole slide, weak supervision, slide representation, survival, state space, and graph pathology. A preprint stays marked as such.
• Quarterly: verify publication status, official code/weight links, licenses, retractions/corrections, and whether a newer version changes the method.
• On every addition: assign one scientific role, state prerequisites, record the evaluation unit, and add a reason to read it. Do not add a paper only because it reports a larger AUROC.
• On model updates: keep checkpoint versions separate. An updated encoder can change a historical MIL comparison without changing the aggregator.
• On benchmark updates: preserve old folds and add a leakage-safe external/site split rather than silently replacing the result.
• Promotion rule: a radar paper moves into the core route only after metadata, publication status, assumptions, and evidence have been reviewed against the primary paper.

The durable research question is not “Which MIL model is number one?” It is: which weak-label assumption, spatial representation, sampling estimator, and evaluation design are justified for this patient-level claim?

Risk-specific multi-teacher survival adapters

SurvMILKD combines frozen encoder features through trainable MIL adapters, teacher-space feature translation and soft ordinal risk distributions. It removes teacher branches at deployment while retaining one student feature pipeline.

Primary paper · year · evidence	Core idea and why to read it	Official resources
SurvMILKD: A Weakly Supervised Survival Analysis Framework for Multi-Teacher Knowledge Distillation using Pathology Foundation ModelsMayur Mallya, Ali Khajegili Mirabadi, Hossein Farahani, and Ali Bashashati2026 · ECCV · acceptedThe official ECCV 2026 poster and conference-hosted manuscript verify acceptance; a Springer version of record was not verified. The official repository was a code-coming-soon placeholder on 2026-09-12. All foundation encoders remain frozen: task-specific MIL adapters and translation modules learn, not the UNI backbone.	Risk-aware multi-teacher distillation combines complementary slide-level survival signals while deploying only a student branch. SurvMILKD uses frozen pathology foundation encoders, trainable attention-MIL adapters, teacher-specific feature translators, and a four-bin representation of scalar risk. Its contribution is task-specific survival transfer rather than another general-purpose encoder pretraining recipe.	Official ECCV 2026 Poster · Official repository — code pending

Reading question: How does distillation represent censored survival risk, and which teacher and feature costs remain in training versus inference?

Evidence boundary: Not every cohort benefits from more teachers; frozen encoders, external risk pooling and code-release status must remain explicit.

Spatial-token retention for slide pretraining

TokRet uses greedy feature diversity to retain many tile-level representatives; PathTokScale adds sparse slide attention and DINOv2-style pretraining. Compare token selection alone against the complete additional training recipe.

Primary paper · year · evidence	Core idea and why to read it	Official resources
Scaling Whole-Slide Pathology Foundation Model Pretraining with Billions Off-the-Shelf TokensHonglin Li, Zhongyi Shui, Chenglu Zhu, and Lin Yang2026 · ECCV · acceptedECCV 2026 acceptance and the manuscript are verified from the official program. No Springer version of record or author-owned public code/checkpoint was verified. Section 4.1 explicitly combines HistAI with about 10,000 private WSIs; do not describe pretraining as public-only. PathTokScale and PathTokenScaling both occur in the manuscript.	TokRet retains diverse spatial tokens from a frozen tile encoder instead of reducing every tile to its CLS token. PathTokScale then pretrains a sparse slide transformer over these richer token sets. The scientific question is how much useful tissue information can be retained per unit of slide-model computation, rather than how to enlarge the tile backbone.	Official ECCV 2026 Poster

Reading question: Do retained tokens preserve rare tissue evidence, and how much computation, private data and pretraining are actually required?

Evidence boundary: The corpus includes private WSIs; token-similarity coverage is not a guarantee of diagnostic fidelity or low-magnification equivalence.

Taxonomy-conditioned diagnostic label generation

TaxoMIL generates coarse and fine diagnostic strings using a shared decoder, branch-specific prefixes and hierarchy-aware losses. Its target is a known label pair, so compare it with hierarchical MIL and exact-match diagnosis rather than free-text report generation.

Primary paper · year · evidence	Core idea and why to read it	Official resources
TaxoMIL: Taxonomy-Constrained Learning for Hierarchical Whole Slide Image AnalysisChaeyeon Lee, Khang Nguyen Quoc, Jinsol Song, Yosep Chong, Kwangil Yim, and Jin Tae Kwak2026 · ECCV · acceptedOfficial ECCV 2026 poster and conference-hosted manuscript verify acceptance; no Springer version of record was verified. Author spelling follows the paper rather than the abbreviated program entry. Output is coarse/fine diagnostic label text, not full clinical reports. GastWSI is private.	TaxoMIL learns to generate a coherent pair of coarse and fine diagnostic labels from a whole slide. It combines MIL aggregation, two granularity-specific visual prefixes, a shared GPT-2 decoder with separate heads, and taxonomy-aware training losses. The key teaching idea is to structure both the label space and its alignment with visual evidence before judging hierarchical predictions.	Official ECCV 2026 Poster · Official implementation · arXiv manuscript

Reading question: Does each prediction satisfy both granularity-specific correctness and the supplied parent–child taxonomy?

Evidence boundary: Taxonomy-aware training is not a formal guarantee for arbitrary generated strings; private cohorts and predefined labels limit transfer claims.

Cross-magnification transfer: utility, predictability and task benefit

After XMaG and LRMIL, test whether observed teacher targets help, whether native low-power features predict them, and whether MIL benefits from the predictions. Direct versus residual regression and native-feature retention are separate controls.

Primary paper · year · evidence	Core idea and why to read it	Official resources
What Makes High-Magnification Knowledge Transferable? A Study of Cross-Resolution Distillation in Whole-Slide ImagingZhiyuan Yang, Jiahao Cheng, and Mahdi S. Hosseini2026 · arXiv · preprintPublic arXiv v2 (2026-10-01) verified 2026-10-02. The authors state ICLR 2027 under review; no acceptance inferred. Only public source material was used. A diagnostic study of target utility, predictability and downstream benefit, not a new universally superior distillation architecture.	Separate teacher usefulness, student predictability and downstream benefit. Observed 20× regional means can help slide prediction when supplied with native 5× features, yet a 5× predictor may fail to reconstruct them usefully. Lower reconstruction error does not consistently translate into better slide-level performance.	Reviewed manuscript revision

Reading question: Can the proposed target beat a constant training-mean baseline and improve the final slide endpoint?

Evidence boundary: Observed 20× access is a diagnostic reference, not low-power-only deployment. Lower reconstruction MSE is not guaranteed downstream improvement; the study is an ICLR 2027 submission-stage preprint.

Distilling reusable MIL initializations

TITAN and CARE supervise nine lightweight MIL aggregators over regular ROI bags. Teacher-specific projections and dispersion-normalized alignment precede linear probing or supervised fine-tuning.

Primary paper · year · evidence	Core idea and why to read it	Official resources
Pretraining Multiple Instance Learning Networks with Multi-Teacher Distillation from Pathology Slide Foundation ModelsMingxi Fu, Jiawen Li, Renao Yan, Jiali Hu, Qiehe Sun, Tian Guan, and Yonghong He2026 · arXiv · preprintarXiv v1 (2026-07-16) checked 2026-10-02. Six cohorts cover 15 tasks, not 15 independent datasets. Official code and pretrained aggregator checkpoints are listed; experiments were not rerun. This transfers slide embeddings into MIL initialization, not 20×-to-5× acquisition.	Pretrain the MIL aggregator, not only its tile encoder. A small, initially headless MIL network learns to reproduce complementary TITAN and CARE region embeddings. The resulting aggregator weights initialize multiple downstream slide tasks; this is reusable representation transfer, not a 20×-to-5× acquisition method.	Versioned manuscript · Official implementation and pretrained MIL weights

Reading question: What improves because of pretraining, and does that gain survive a matched architecture, encoder and tuning budget?

Evidence boundary: Six cohorts contain 15 tasks; more teachers do not always win. Main-text angular loss and appendix pseudocode differ, and aggregator pretraining does not itself reduce acquisition magnification.

Pathway-structured privileged supervision

MoPE uses training-only RNA to supervise WSI pathway experts through shared-memory distributions. It aligns a structured view of molecular information without requiring RNA at inference.

Primary paper · year · evidence	Core idea and why to read it	Official resources
Pathway-Structured Privileged Distillation for Deployable Computational PathologyYongxin Guo, Hao Lu, Onur C. Koyun, Zhengjie Zhu, Muhammet F. Demir, and Metin N. Gurcan2026 · arXiv · preprintarXiv v1 (2026-06-01) checked 2026-10-02; no conference acceptance verified. Paired RNA is training-only, while WSI is used at deployment. Pathway-memory readouts are not independently validated molecular measurements.	MoPE learns pathway-structured histology predictors using paired bulk RNA during training, but deploys with WSI alone. Instead of forcing RNA and tissue embeddings to be identical, it aligns their distributions over a shared memory. This is privileged-modality distillation, not magnification reduction or direct RNA reconstruction.	Versioned manuscript

Reading question: Which part of the molecular target is transferable from morphology, and what independent evidence supports biological interpretation?

Evidence boundary: Paired RNA is still needed for training. Model attention is not measured pathway activation; this is neither optical cross-magnification transfer nor prospective clinical validation.

Case-level composites and diagnostic text transfer

WILSON combines actual multi-magnification tissue views into composite images and studies patient-level embedding and text-generation variants. Separate image-embedding, report-distillation and generative branches.

Primary paper · year · evidence	Core idea and why to read it	Official resources
WILSON - a pathology foundation model framework for patient-level analysis and diagnostic text generationSaghir Alfasly, Wataru Uegami, Sobhan Hemati, Wenchao Han, Xiaojia Tang, Kevin Thompson, Daniel Stone, Ghazal Alabtah, Saba Yasir, Michael R. Lucas, Eric W. Klee, Cheryl L. Willman, Judy C. Boughey, Matthew P. Goetz, Krishna R. Kalari, and H. R. Tizhoosh2026 · arXiv · preprintarXiv v1 submitted 2026-09-20. No peer-reviewed record was verified. The manuscript says model release is planned after publication and institutional clearance; Mayo training WSIs/reports are not publicly downloadable.	WILSON makes a slide or a multi-slide patient case into a compact multi-magnification composite image, then learns one visual representation aligned with pathology-report supervision. It combines an organ-aware region-selection pipeline, a convolutional encoder, and alternative dense-text, keyword or CLIP alignment branches; a CoCa-style decoder adds caption generation. The central idea is to change the computational unit from thousands of separately encoded tiles to a structured image summary.	Full manuscript and supplementary material

Reading question: What images and reports are needed at training and deployment, and which compute path was measured?

Evidence boundary: The system still reads high-power tissue; analytic embedding TFLOPs are not end-to-end speed. It is a preprint with restricted training data and a planned model release.

Selecting encoders before expensive MIL sweeps

SAM-Cluster and SAM-LP compare feature extractors using structure-aware regions and proxy evaluations before full MIL training. They provide a teacher-selection starting point, not a guaranteed downstream ranking.

Primary paper · year · evidence	Core idea and why to read it	Official resources
SAM-Cluster and SAM-LP: Structure-Aware Evaluation Metrics for Selecting WSI Feature Extractors Without MIL TrainingJuhyeon Kim, Paul Ssemakula, Chanjae Song, and Mun Yong Yi2026 · MICCAI · peer-reviewedOfficial MICCAI 2026 open-access proceedings version. Springer DOI and final page range were not listed at the audit date.	SAM-Cluster and SAM-LP rank frozen WSI feature extractors before committing to full multiple-instance learning (MIL) training. SAM supplies structural regions, not tissue-class truth. The label-free score asks whether embeddings respect those regions; the supervised score learns which region prototypes support the slide label. The product is a candidate ranking for a particular cohort, not a new diagnostic classifier.	Official accepted-version PDF · Official implementation and benchmark results

Reading question: Does a proxy ranking agree with the target dataset, endpoint and aggregation protocol?

Evidence boundary: Correlation is not identical rank or clinical validation. Keep SAM segmentation, extractor inference, label requirements and final MIL confirmation in the resource account.

From fixed tissue graphs to evolving concept-guided fields

TMEvolve groups frozen CONCHv1.5 patch features into soft semantic regions, exchanges messages on a spatial graph, and lets regional and local states repeatedly update each other before slide prediction. Concept guidance and task supervision do not make its latent regions measured cell types or biological dynamics.

Primary paper · year · evidence	Core idea and why to read it	Official resources
Modeling Whole-Slide Images as Dynamic Tumor Microenvironment FieldsLei Wu, Jiashuai Liu, Di Zhang, Zhangpeng Gong, Yingkang Zhan, Yi Niu, Jiusong Ge, Chunze Yang, Kai Yi, Mireia Crispin-Ortuzar, Chen Li, and Zeyu Gao2026 · NeurIPS · acceptedAccepted as a NeurIPS 2026 poster: official conference program event 149764 matches the exact title and 12 authors, checked 2026-10-08. Final proceedings were not verified; review uses arXiv v1 (2026-09-28). Official method code is available under a noncommercial research license, but does not include the WSI feature-extraction pipeline. Tissue-tile pixel size, extraction magnification/MPP, and image preprocessing remain unreported in the inspected paper and release.	TMEvolve replaces a static bag of patch features with adaptive tissue regions whose states are repeatedly refined. Similar-region neighbors exchange visual information; heterogeneous boundaries carry directional visual and pathological-concept messages. The final regions are pooled for slide-level prediction. This is reaction-diffusion-inspired representation learning, not a simulation of tumor progression or a generative diffusion model.	Reviewed manuscript and supplement (v1) · Author publication and acceptance statement · Official NeurIPS 2026 poster · Official method implementation (noncommercial research)

Reading question: Does adaptive region–patch interaction add value beyond the same encoder with set pooling or a fixed spatial graph?

Evidence boundary: Tile pixel size, extraction magnification and MPP are not reported in the inspected paper or method release. The 512 setting is a hidden feature width, not pixels; offline feature extraction is excluded from the reported method cost.

Condense a slide distribution, not its acquisition magnification

NICER fits adaptive nonparametric prototype sets to each slide’s UNI features using distribution matching. It studies task-agnostic bag condensation, with separate experiments for condensed-training/full-bag testing and condensed training plus testing.

Primary paper · year · evidence	Core idea and why to read it	Official resources
Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image CondensationDuong M. Nguyen, Trong Nghia Hoang, Hang Thi Nguyen, Thanh Trung Huynh, Phi Le Nguyen, and Minh N. Do2026 · NeurIPS · acceptedAcceptance independently verified by NeurIPS 2026 poster 153235 and author repository. Review uses arXiv:2610.00678v1, submitted 2026-09-30. Repository contained only a README on 2026-10-08. Feature condensation at 20× is not low-magnification encoder distillation; Table 1 tests on full feature bags.	NICER condenses a whole slide's large frozen-feature bag into a smaller, slide-adaptive set of learned prototypes. Its organizing principle is preserving the distribution seen by the downstream encoder representation, not preserving arbitrary pixel similarity or forcing every slide into the same fixed prototype count. Local pattern exploration preserves diverse features; nonparametric concept condensation then merges redundancy. This is feature/data condensation, not distilling a large encoder into a smaller student.	Official NeurIPS 2026 program · Versioned full manuscript · Author repository (README only at audit)

Reading question: Which information survives prototype condensation, and does the measured saving include prototype optimization and feature extraction?

Evidence boundary: The acquisition remains nonoverlapping 256×256 tiles at 20× (0.5 µm/pixel). Table 1 is not a compressed-inference benchmark; feature prototypes are not low-power image tiles or distilled encoder weights.

Sparse features for slide prototypes and weakly supervised localization

ProtoMIL uses SAE-derived patch prototypes in slide learning; SAE-1N selects a discriminative latent using slide supervision. These methods connect sparse features to WSI tasks, but localization quality and slide-level predictive performance can move in different directions.

Primary paper · year · evidence	Core idea and why to read it	Official resources
Prototype-Based Multiple Instance Learning for Gigapixel Whole Slide Image ClassificationSusu Sun, Dominique van Midden, Geert Litjens, and Christian F. Baumgartner2025 · MICCAI · peer-reviewedOfficial MICCAI 2025 record and 11-page accepted manuscript checked. PANDA is evaluated as binary tumor classification in this study, not the original six-class ISUP grading task. Artifact concepts are masked before MIL retraining; this is not a universal AUC improvement.	ProtoMIL discovers sparse visual concepts from frozen CONCH patch embeddings, then expresses a WSI class logit as a sum of weighted concept contributions. Pathologists can inspect maximally activating patches and disable unwanted concepts before retraining the slide classifier. The point is an inspectable decision decomposition, not a claim that attention alone explains a prediction.	Official MICCAI open-access record · Official accepted manuscript · Author implementation · arXiv manuscript
Patch-level phenotype identification via weakly supervised neuron selection in sparse autoencoders for CLIP-derived pathology embeddingsKeita Tamura, Yao-zhong Zhang, Yohei Okubo, and Seiya Imoto2026 · Pacific Symposium on Biocomputing · peer-reviewedPSB 2026 volume 31, pages 708–721, with source supplement checked. Neuron discovery uses slide labels, but the reported localization thresholds use patch-level validation annotations. Single-neuron localization does not preserve all multi-feature slide-classification performance.	SAE-1N selects one sparse feature using only slide labels, then uses its patch activation as a tumor-localization signal. A BatchTopK SAE decomposes frozen image embeddings; each feature is max-pooled over a slide and ranked by its tumor-versus-normal association. The selected coordinate is not normalized separately within each slide, unlike softmax attention. The reported thresholded localization still uses annotated validation patches.	Official PSB proceedings PDF · Author implementation and supplement

Reading question: Which labels select the feature, tune its threshold, and evaluate localization or slide prediction?

Evidence boundary: Do not hide patch-annotation use in threshold selection or assume improved localization implies better slide AUC. Distinguish retraining after an intervention from a fixed-classifier causal test.

Encoder selection: connect cheap proxies to the intended endpoint

Tile-to-slide transferability compares the same frozen encoders on local morphology and slide tasks. Strong aggregate correlations support shortlisting, while learned aggregation and context-heavy tasks can change which model wins. Use this alongside structure-aware feature-selection metrics, not as a replacement for final WSI validation.

Primary paper · year · evidence	Core idea and why to read it	Official resources
From Patches to Patients: A Study of the Tile-to-Slide Performance Transferability in Digital PathologySofiène Boutaj, Leo Fillioux, Maria Vakalopoulou, Stergios Christodoulidis, and Pierre Marza2026 · MICCAI · peer-reviewedMICCAI 2026 chapter, LNCS 16891, pages 149–159; publisher first-online date is in September 2026, although the volume copyright/bibliographic imprint says 2027. Review uses the inspected publisher PDF, not only arXiv:2606.10778. Across-encoder rank correlation supports shortlisting, not guaranteed selection of the best model for a new clinical cohort.	Tile-level linear probing can help shortlist pathology encoders before expensive whole-slide experiments. Across 19 frozen models, aggregate tile performance correlates strongly with aggregate slide performance, but learned attention changes the ranking. The contribution is a benchmark of when a cheaper proxy transfers, not a new encoder, distillation method, or promise that the best tile model wins on every clinical task.	Publisher full chapter · Official MICCAI open-access record · Official accepted manuscript · Earlier arXiv manuscript

Reading question: Does the ranking survive the actual slide aggregator and held-out clinical task?

Evidence boundary: Across-model correlation is not guaranteed top-one selection, patient-level validity, or evidence that a distilled student preserves cross-magnification information.

Question-conditioned evidence acquisition across magnifications

PathPilot separates question-agnostic diagnostic localization from question-dependent evidence accumulation, movement, zoom and stopping. The agent accesses real multi-scale ROIs; it is neither low-only distillation nor a new MIL pooling operator. Its MIL experiment evaluates whether selected regions remain useful.

Primary paper · year · evidence	Core idea and why to read it	Official resources
PathPilot: An Adaptive Question-Conditioned Multi-Magnification Navigation Agent for Whole-Slide Image AnalysisZhengyang Xu, Han Li, Jingsong Liu, Ayako Ura, Shihui Zu, Shaohua Kevin Zhou, Nassir Navab, and Peter Schüffler2026 · arXiv · preprintarXiv:2610.16734v1, submitted 12 October 2026; manuscript says under review, with no independently verified acceptance. Full 24-page public manuscript inspected; referenced supplementary material was not attached or independently obtained. Code is promised on publication, not a verified release. Adaptive thumbnail/1.25×/2.5×/5×/10× exploration is not knowledge distillation or low-resolution-only inference. QA scores use a language-model judge; deterministic repeated runs do not establish population uncertainty.	PathPilot separates where tissue looks diagnostically promising from what evidence a particular question still needs. CrossMagLoc learns question-agnostic region priorities across magnifications; EviRegionSearch uses a VLM, language-model memory and explicit evidence gaps to move, zoom or stop. This is adaptive acquisition of real multi-magnification evidence, not distillation into a low-resolution-only student.	Inspected arXiv v1 manuscript

Reading question: What evidence is missing, which region and scale can supply it, and what justifies stopping?

Evidence boundary: Small morphology subsets and a language-model judge limit inference. A same-institution cohort called external is not independent-site validation; deterministic repeatability is not patient uncertainty.

Complete maintained paper route

81 canonical readings · 1997–2026· route reviewed through October 2, 2026.

This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.

Show all 81 readings in maintained order

1. #11997Artificial IntelligencePeer reviewedL1

   Solving the Multiple Instance Problem with Axis-Parallel Rectangles

   Thomas G. Dietterich et al.

   Metadata checked 2026-07-27

2. #22017NeurIPSPeer reviewedL1–L2

   Deep Sets

   Manzil Zaheer et al.

   Metadata checked 2026-07-27

3. #32018ICMLPeer reviewedL1–L2

   Attention-based Deep Multiple Instance Learning

   Maximilian Ilse, Jakub Tomczak, and Max Welling

   Metadata checked 2026-07-27

4. #42019Nature MedicinePeer reviewedL2

   Clinical-grade computational pathology using weakly supervised deep learning on whole slide images

   Gabriele Campanella et al.

   Metadata checked 2026-07-27

5. #52021Nature Biomedical EngineeringPeer reviewedL2

   Data-efficient and weakly supervised computational pathology on whole-slide images

   Ming Y. Lu et al.

   Metadata checked 2026-07-27

6. #62021CVPRPeer reviewedL2–L3

   Dual-Stream Multiple Instance Learning Network for Whole Slide Image Classification With Self-Supervised Contrastive Learning

   Bin Li et al.

   Metadata checked 2026-07-27

7. #72021NeurIPSPeer reviewedL2–L3

   TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification

   Zhuchen Shao et al.

   Metadata checked 2026-07-27

8. #82022CVPRPeer reviewedL2–L3

   DTFD-MIL: Double-Tier Feature Distillation Multiple Instance Learning for Histopathology Whole Slide Image Classification

   Hongrun Zhang et al.

   Metadata checked 2026-07-27

9. #92022CVPRPeer reviewedL2–L3

   Scaling Vision Transformers to Gigapixel Images via Hierarchical Self-Supervised Learning

   Richard J. Chen et al.

   Metadata checked 2026-07-27

10. #102023ICCVPeer reviewedL3

    Multiple Instance Learning Framework with Masked Hard Instance Mining for Whole Slide Image Classification

    Wenhao Tang et al.

    Metadata checked 2026-07-27

11. #112024CVPRPeer reviewedL3

    Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology

    Wenhao Tang et al.

    Metadata checked 2026-07-27

12. #122024NeurIPSPeer reviewedL3

    Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis

    Honglin Li et al.

    Metadata checked 2026-07-27

13. #132023MICCAIPeer reviewedL3

    Structured State Space Models for Multiple Instance Learning in Digital Pathology

    Leo Fillioux et al.

    Metadata checked 2026-07-27

14. #142024MICCAIPeer reviewedL3

    MambaMIL: Enhancing Long Sequence Modeling with Sequence Reordering in Computational Pathology

    Shu Yang et al.

    Metadata checked 2026-07-27

15. #152024NeurIPSPeer reviewedL3

    xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology

    Julius Hense et al.

    Metadata checked 2026-07-27

16. #162025ICMLPeer reviewedL3

    Do Multiple Instance Learning Models Transfer?

    Daniel Shao et al.

    Metadata checked 2026-07-27

17. #172026CVPRPeer reviewedL2–L3

    TopoSlide: Topologically-Informed Histopathology Whole Slide Image Representation Learning

    Shahira Abousamra, Asmita Sood, and Sylvia Plevritis

    Official artifacts: Code and models

    Metadata checked 2026-07-27

18. #182026CVPRPeer reviewedL3

    Turning Pre-Trained Vision Transformers into End-to-End Histopathology Whole Slide Image Models for Survival Prediction

    Jiawen Li, Jiali Hu, Xitong Ling, Renao Yan, Yuxuan Chen, Tian Guan, and Yonghong He

    Metadata checked 2026-07-27

19. #192026CVPRPeer reviewedL3

    CARE: A Molecular-Guided Foundation Model with Adaptive Region Modeling for Whole Slide Image Analysis

    Di Zhang et al.

    Metadata checked 2026-07-27

20. #202026ECCVAccepted · proceedings pendingL3

    MOOZY: A Patient-First Foundation Model for Computational Pathology

    Yousef Kotp, Vincent Quoc-Huy Trinh, Christopher Pal, and Mahdi S. Hosseini

    ECCV 2026 poster presentation is now verified by the official conference program (poster 4955), superseding author-only acceptance evidence. A Springer version of record was not verified on 2026-09-12. The program lists Yousef Hassan whereas the retained arXiv record lists Yousef Kotp; preserve the declared manuscript authorship rather than silently replacing it.

    Official artifacts: Project · Code · Model · Official ECCV 2026 poster · Conference manuscript

    Metadata checked 2026-09-12

21. #212026WACVPeer reviewedL2–L3

    HistoMILKD: A Multiple Instance Learning based Multi-Teacher Knowledge Distillation Framework for Whole Slide Image Classification

    Mayur Mallya, Ali Khajegili Mirabadi, Hossein Farahani, and Ali Bashashati

    Official artifacts: Code

    Metadata checked 2026-07-27

22. #222026CVPRPeer reviewedL3

    Act Like a Pathologist: Tissue-Aware Whole Slide Image Reasoning

    Wentao Huang et al.

    Metadata checked 2026-07-27

23. #232026CVPRPeer reviewedL3

    MUSE: Harnessing Precise and Diverse Semantics for Few-Shot Whole Slide Image Classification

    Jiahao Xu et al.

    Official artifacts: Code

    Metadata checked 2026-07-27

24. #242025CVPRPeer reviewedL3

    Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning

    Tim Lenz, Peter Neidlinger, Marta Ligero, Georg Wölflein, Marko van Treeck, and Jakob N. Kather

    Official artifacts: Code

    Metadata checked 2026-07-27

25. #252026ICLRPeer reviewedL3

    Mixture of Mini Experts: Overcoming the Linear Layer Bottleneck in Multiple Instance Learning

    Daniel Shao, Joel Runevic, Richard J. Chen, Drew F. K. Williamson, Ahrong Kim, Andrew H. Song, and Faisal Mahmood

    Official artifacts: Code

    Metadata checked 2026-07-27

26. #262026arXivPreprintL3

    GigaPath-Flash and GigaTIME-Flash: Efficient Pathology Foundation Models for Whole-Slide and Tumor Microenvironment Analysis

    Naoto Usuyama et al.

    Version 2 dated 2026-07-22; author-reported performance remains provisional.

    Metadata checked 2026-07-27

27. #272025arXivPreprintL3

    EXAONE Path 2.0: Pathology Foundation Model with End-to-End Supervision

    Myeongjang Pyeon et al.

    Metadata checked 2026-07-27

28. #282025CVPRPeer reviewedL3

    2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification

    Jingwei Zhang, Anh Tien Nguyen, Xi Han, Vincent Quoc-Huy Trinh, Hong Qin, Dimitris Samaras, and Mahdi S. Hosseini

    Official artifacts: Code

    Metadata checked 2026-07-27

29. #292025CVPRPeer reviewedL3

    Learning Heterogeneous Tissues with Mixture of Experts for Gigapixel Whole Slide Images

    Wu et al.

    Metadata checked 2026-07-27

30. #302025CVPRPeer reviewedL3

    Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning

    Jiuyang Dong et al.

    Official artifacts: Code

    Metadata checked 2026-07-27

31. #312025CVPRPeer reviewedL3

    FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification

    Guo et al.

    Metadata checked 2026-07-27

32. #322025ICCVPeer reviewedL3

    Continual Multiple Instance Learning with Enhanced Localization for Histopathological Whole Slide Image Analysis

    Lee et al.

    Metadata checked 2026-07-27

33. #332025ICCVPeer reviewedL3

    Cracking Instance Jigsaw Puzzles: An Alternative to Multiple Instance Learning

    Chen et al.

    Metadata checked 2026-07-27

34. #342025ICCVPeer reviewedL3

    ModalTune: Fine-Tuning Slide-Level Foundation Models with Multi-Modal Information for Multi-task Adaptation

    Ramanathan et al.

    Metadata checked 2026-07-27

35. #352026CVPRPeer reviewedL3

    FBTA: Enabling Single-GPU End-to-End Gigapixel WSI Classification with Feature Bridging

    Dong et al.

    Metadata checked 2026-07-27

36. #362026CVPRPeer reviewedL3

    Contrastive Cross-Bag Augmentation for Multiple Instance Learning-based Whole Slide Image Classification

    Zhang et al.

    Metadata checked 2026-07-27

37. #372026CVPRPeer reviewedL3

    Universal-to-Specific: Dynamic Knowledge-Guided Multiple Instance Learning for Few-Shot Whole Slide Image Classification

    Li et al.

    Metadata checked 2026-07-27

38. #382026CVPRPeer reviewedL3

    Dual-Level Hypergraph Generation for Addressing Feature Scarcity in Whole-Slide Image Classification

    Yao et al.

    Metadata checked 2026-07-27

39. #392026MICCAIPeer reviewedL3

    LRMIL: Efficient Low-Resolution Multiple Instance Learning via High-Resolution Knowledge Distillation for Whole Slide Image Classification

    Yonghan Shin and Won-Ki Jeong

    Official MICCAI 2026 proceedings record and accepted-version open-access PDF verified 2026-10-02. Springer DOI and pagination remain pending; do not invent either.

    Official artifacts: Official code · Official MICCAI 2026 proceedings · Official accepted-version PDF

    Metadata checked 2026-10-02

40. #402026arXivPreprintL3

    Beyond Relevance: Bayesian Evidence Acquisition for Agentic Whole-Slide Image Reasoning

    Bryan Wong, Xun Xu, Huazhu Fu, Nancy F. Chen, and Mun Yong Yi

    Official artifacts: Official code

    Metadata checked 2026-08-08

41. #412026ECCVAccepted · proceedings pendingL3

    From Multi-Resolution Cells to Gigapixel Whole Slide Images Foundation Model for Computational Pathology

    Basit Alawode, Moshira Ali Abdalla, Dwarikanath Mahapatra, Muhammad Muzammal Naseer, and Sajid Javed

    ECCV 2026 acceptance is now verified by official poster 4422, superseding the preprint-only catalog status. A Springer version of record was not verified on 2026-09-12; the existing review remains tied to its declared manuscript revision.

    Official artifacts: Official ECCV 2026 poster · Conference manuscript

    Metadata checked 2026-09-12

42. #422026arXivPreprintL3

    Understanding Synergistic Interactions among Pathology Foundation Models via Adaptive Fusion

    Yuxiang Xiao, Yang Hu, Bin Li, Tianyang Zhang, Zexi Li, Huazhu Fu, Jens Rittscher, and Kaixiang Yang

    Official artifacts: Official code

    Metadata checked 2026-08-08

43. #432026arXivPreprintL3

    A Multi-Cohort Validation of Censoring-Aware Conformal Lower Predictive Bounds for Pathology Survival Models

    Mingi Hong

    Metadata checked 2026-08-08

44. #442026arXivPreprintL3

    Spatial proteomics guided by H&E-based AI reveals recurrence-risk niches in triple-negative breast cancer

    Yesung Cho et al.

    Metadata checked 2026-08-08

45. #452024MICCAIPeer reviewedL3

    SlideGCD: Slide-based Graph Collaborative Training with Knowledge Distillation for Whole Slide Image Classification

    Tong Shu, Jun Shi, Dongdong Sun, Zhiguo Jiang, and Yushan Zheng

    Official artifacts: Official code

    Metadata checked 2026-08-10

46. #462025MICCAIPeer reviewedL3

    Multi-modal Knowledge Decomposition based Online Distillation for Biomarker Prediction in Breast Cancer Histopathology

    Qibin Zhang, Xinyu Hao, Qiao Chen, Rui Xu, Fengyu Cong, Cheng Lu, and Hongming Xu

    Official artifacts: Official code

    Metadata checked 2026-08-10

47. #472025MICCAIPeer reviewedL3

    AEM: Attention Entropy Maximization for Multiple Instance Learning based Whole Slide Image Classification

    Yunlong Zhang, Honglin Li, Yuxuan Sun, Zhongyi Shui, Jingxiong Li, Chenglu Zhu, and Lin Yang

    Official artifacts: Official code

    Metadata checked 2026-08-10

48. #482022Medical Image AnalysisPeer reviewedL3

    Federated learning for computational pathology on gigapixel whole slide images

    Ming Y. Lu, Richard J. Chen, Dehan Kong, Jana Lipkova, Rajendra Singh, Drew F. K. Williamson, Tiffany Y. Chen, and Faisal Mahmood

    Peer-reviewed Medical Image Analysis article 76:102298; DOI 10.1016/j.media.2021.102298.

    Official artifacts: DOI · Author manuscript · Official code

    Metadata checked 2026-08-20

49. #492023MICCAIPeer reviewedL3

    Detecting Domain Shift in Multiple Instance Learning for Digital Pathology Using Fréchet Domain Distance

    Milda Pocevičiūtė, Gabriel Eilertsen, Stina Garvin, and Claes Lundström

    Peer-reviewed MICCAI 2023 proceedings paper, pages 157–167; its arXiv upload appeared in 2024 and does not change the venue year.

    Official artifacts: MICCAI paper record · Author manuscript

    Metadata checked 2026-08-20

50. #502025CVPRPeer reviewedL3

    HistoFS: Non-IID Histopathologic Whole Slide Image Classification via Federated Style Transfer with RoI-Preserving

    Farchan Hakim Raswa, Chun-Shien Lu, and Jia-Ching Wang

    Peer-reviewed CVPR 2025 proceedings paper, pages 30251–30260.

    Official artifacts: Official project · Official code

    Metadata checked 2026-08-20

51. #512022ECCVPeer reviewedL3

    Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images

    Kevin Thandiackal, Boqi Chen, Pushpak Pati, Guillaume Jaume, Drew F. K. Williamson, Maria Gabrani, and Orcun Goksel

    Peer-reviewed ECCV 2022 proceedings paper.

    Official artifacts: Official code

    Metadata checked 2026-08-23

52. #522024IEEE Transactions on Medical ImagingPeer reviewedL3

    A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification

    Gianpaolo Bontempo, Federico Bolelli, Angelo Porrello, Simone Calderara, and Elisa Ficarra

    Peer-reviewed journal article 43(4):1412–1421; journal extension of DAS-MIL.

    Official artifacts: Official code

    Metadata checked 2026-08-23

53. #532024ISBIPeer reviewedL3

    Reducing Histopathology Slide Magnification Improves the Accuracy and Speed of Ovarian Cancer Subtyping

    Jack Breen, Katie Allen, Kieran Zucker, Nicolas M. Orsi, and Nishant Ravikumar

    Peer-reviewed ISBI 2024 proceedings paper.

    Official artifacts: Author manuscript

    Metadata checked 2026-08-23

54. #542025IEEE AccessPeer reviewedL3

    PatHME: Hierarchical Multi-Expert Knowledge Distillation for Whole Slide Image Analysis

    Jing Wei Tan, Gangsu Kim, and Won-Ki Jeong

    Peer-reviewed journal article 13:163246–163257.

    Official artifacts: Official code

    Metadata checked 2026-08-23

55. #552025WACVPeer reviewedL3

    CEMIL: Contextual Attention Based Efficient Weakly Supervised Approach for Histopathology Image Classification

    Tawsifur Rahman, Alexander S. Baras, and Rama Chellappa

    Peer-reviewed WACV 2025 proceedings paper, pages 4248–4257.

    Metadata checked 2026-08-23

56. #562026Cancer ResearchPeer reviewedL3

    A Magnification Alignment Framework Enables Computation- and Communication-Efficient Computational Pathology

    Chu Han, Bingchao Zhao, Tianpeng Deng, Jingqi Huang, Fangfang Liu, Jiatai Lin, Shanshan Lyu, Longfei Wang, Cheng Lu, Changhong Liang, Hannah Y. Wen, Zhenwei Shi, Xiaojing Guo, and Zaiyi Liu

    Peer-reviewed journal article 86(12):3060–3073.

    Official artifacts: Author manuscript · Official code

    Metadata checked 2026-08-23

57. #572026MICCAIPeer reviewedL3

    LaGuadia: Language-Guided Adaptive Distillation from Pathology Foundation Models

    Gangsu Kim and Won-Ki Jeong

    Official MICCAI 2026 proceedings record and accepted-version open-access PDF verified 2026-10-02; this supersedes the prior preprint classification. Springer DOI remains pending.

    Official artifacts: Official code · Official MICCAI 2026 proceedings · Official accepted-version PDF

    Metadata checked 2026-10-02

58. #582026Medical Image AnalysisAccepted · proceedings pendingL3

    RACR-MIL: Rank-aware contextual reasoning for weakly supervised grading of squamous cell carcinoma using whole slide images

    Anirudh Choudhary, Mosbah Aouad, Krishnakant Saboo, Angelina Hwang, Jacob Kechter, Blake Bordeaux, Puneet Bhullar, David DiCaudo, Steven Nelson, Nneka Comfere, Emma Johnson, Olayemi Sokumbi, Jason Sluzevich, Leah Swanson, Dennis Murphree, Aaron Mangold, and Ravishankar Iyer

    The complete arXiv v3 manuscript identifies Medical Image Analysis as the journal and the abstract record reports acceptance; a final DOI or journal issue record was not yet available when checked on 30 August 2026.

    Metadata checked 2026-08-30

59. #592026arXivPreprintL3

    LanGuSTE: Language-Guided Coarse-to-Fine Patch Selection for Efficient Whole Slide Image Analysis

    Yonghan Shin, Gangsu Kim, and Won-Ki Jeong

    arXiv preprint first submitted in 2025 and substantially revised as v2 on 24 August 2026; no peer-reviewed version or official implementation was verified by the review cutoff.

    Metadata checked 2026-08-30

60. #602026arXivPreprintL3

    An end-to-end-trained vision-language model for native-language prostate pathology report generation

    Christian Grashei, Fabian Gülhan, Maximilian Legnar, Fabian Stögbauer, Cleo-Aron Weis, Carolin Mogler, and Peter Schüffler

    arXiv v1 preprint submitted 24 August 2026; no peer-reviewed record, public training code, model weights, or paired clinical dataset release was verified by the review cutoff.

    Metadata checked 2026-08-30

61. #612026arXivPreprintL3

    PANDA: Prototype-Anchored Alignment for Partially Unpaired Multimodal Learning, with Applications to Alzheimer's MRI and TCGA Pathology

    Sheethal Bhat, Mahfuzur Rahman Chowdhury, Paula Andrea Perez-Toro, Stephan Wunderlich, Rose Dawn Bharat, Siming Bayer, and Andreas Maier

    arXiv v1 preprint submitted 26 August 2026. The manuscript promises code upon acceptance, but no public implementation was available at the review cutoff.

    Metadata checked 2026-08-30

62. #622026arXivPreprintL3

    Benchmarking Vision-Language Models for Automated Pathology Diagnosis and Report Generation

    Yumi Lee, Harim Oh, Hyoryung Kim, Minji Kim, Eunsu Kim, Hyeseong Lee, Junya Fukuoka, Andrey Bychkov, Jijgee Munkhdelger, Rajiv Kumar Kaushal, Ayushi Sahay, Rajni Yadav, Bharathi Prabakaran, Sulen Sarioglu, Serdar Balcı, Ilknur Turkmen, Yuri Tolkach, Christian Harder, Julian Westerdorf, Reinhard Buettner, Audun Ljone Henriksen, Sepp De Raedt, Byung Hyun Lee, Sungjin Lim, Joohoon Lee, Gwanghyun Kim, Se Young Chun, Suryakant Singh, Saarthak Kapse, Prateek Prasanna, Kyung A Kim, Yousun Kang, Sehwan Yoo, Sungman Hong, Shubham Innani, Michael Feldman, Spyridon Bakas, Ujjwal Baid, Prasad Dutande, Suhas Gajare, Bhakti Baheti, Serkan Sökmen, Ece Tuğba Cebeci, Ahmet Halıcı, Musa Balcı, Kardelen Peçenek, Srividhya Sainath, Kyongseok Jang, Messi H. J. Lee, Noorul Wahab, Bodong Du, Jiaming Zhang, Qixiang Zhang, Jang-Hwan Choi, and Sangjeong Ahn

    arXiv v1 submitted 1 September 2026 reports the results of the MICCAI REG 2025 challenge. No final peer-reviewed proceedings or journal record was verified on 2 September 2026. The approximately 10,500 institutional WSI-report pairs are not a generally downloadable public dataset; the released repositories expose benchmark information and team artifacts.

    Official artifacts: Official challenge repository · Official winning-system repository · Official challenge site

    Metadata checked 2026-09-02

63. #632026Engineering Applications of Artificial IntelligencePeer reviewedL3

    Cross-Stain Knowledge Distillation for Low-Cost Lung Cancer Programmed Death Ligand-1 Assessment With Multi-Granularity Multiple Instance Learning

    Yi Shi, Chong Ge, Fang Zhao, Anli Zhang, Ao Li, Haibo Wu, and Minghui Wang

    Peer-reviewed journal article, volume 167, article 113762. IHC is privileged teacher information during training; deployment uses the H&E student, but its macro- and micro-granularity branches remain. No official code or model release was verified on 2026-09-02.

    Metadata checked 2026-09-02

64. #642026Computers & Electrical EngineeringPeer reviewedL3

    CMA-MIL: A Cross-Magnification Attention-Based Multiple Instance Learning Framework for Renal Cell Carcinoma Subtyping Using Whole-Slide Images

    Rashika Bagri and Ankit Rajpal

    Peer-reviewed journal article, volume 135, article 111204. The weakly supervised model uses 5×, 10×, and 20× streams and retains their cross-attention fusion at inference; it is not a low-magnification-only student. No official artifact was verified on 2026-09-02.

    Metadata checked 2026-09-02

65. #652026Medical Image AnalysisPeer reviewedL3

    M2OTCA: Multiple-Magnification Optimal Transport-Based Cross-Attention Learning for Whole Slide Image Classification

    Zhonghang Zhu, Erik Meijering, and Liansheng Wang

    Peer-reviewed journal article, volume 112, article 104082. Region structural prototypes reduce the cost of optimal transport across 5×, 10×, and 20×; all three magnifications feed magnification-specific predictions and remain part of inference. No official artifact was verified on 2026-09-02.

    Official artifacts: PubMed record

    Metadata checked 2026-09-02

66. #662026ECCVAccepted · proceedings pendingL3

    SurvMILKD: A Weakly Supervised Survival Analysis Framework for Multi-Teacher Knowledge Distillation using Pathology Foundation Models

    Mayur Mallya, Ali Khajegili Mirabadi, Hossein Farahani, and Ali Bashashati

    The official ECCV 2026 poster and conference-hosted manuscript verify acceptance; a Springer version of record was not verified. The official repository was a code-coming-soon placeholder on 2026-09-12. All foundation encoders remain frozen: task-specific MIL adapters and translation modules learn, not the UNI backbone.

    Official artifacts: Official ECCV 2026 Poster · Official repository — code pending

    Metadata checked 2026-09-12

67. #672026ECCVAccepted · proceedings pendingL3

    Scaling Whole-Slide Pathology Foundation Model Pretraining with Billions Off-the-Shelf Tokens

    Honglin Li, Zhongyi Shui, Chenglu Zhu, and Lin Yang

    ECCV 2026 acceptance and the manuscript are verified from the official program. No Springer version of record or author-owned public code/checkpoint was verified. Section 4.1 explicitly combines HistAI with about 10,000 private WSIs; do not describe pretraining as public-only. PathTokScale and PathTokenScaling both occur in the manuscript.

    Official artifacts: Official ECCV 2026 Poster

    Metadata checked 2026-09-12

68. #682026ECCVAccepted · proceedings pendingL3

    TaxoMIL: Taxonomy-Constrained Learning for Hierarchical Whole Slide Image Analysis

    Chaeyeon Lee, Khang Nguyen Quoc, Jinsol Song, Yosep Chong, Kwangil Yim, and Jin Tae Kwak

    Official ECCV 2026 poster and conference-hosted manuscript verify acceptance; no Springer version of record was verified. Author spelling follows the paper rather than the abbreviated program entry. Output is coarse/fine diagnostic label text, not full clinical reports. GastWSI is private.

    Official artifacts: Official ECCV 2026 Poster · Official implementation · arXiv manuscript

    Metadata checked 2026-09-12

69. #692026MICCAIPeer reviewedL3

    SAM-Cluster and SAM-LP: Structure-Aware Evaluation Metrics for Selecting WSI Feature Extractors Without MIL Training

    Juhyeon Kim, Paul Ssemakula, Chanjae Song, and Mun Yong Yi

    Official MICCAI 2026 open-access proceedings version. Springer DOI and final page range were not listed at the audit date.

    Official artifacts: Official accepted-version PDF · Official implementation and benchmark results

    Metadata checked 2026-10-02

70. #702026arXivPreprintL3

    WILSON - a pathology foundation model framework for patient-level analysis and diagnostic text generation

    Saghir Alfasly, Wataru Uegami, Sobhan Hemati, Wenchao Han, Xiaojia Tang, Kevin Thompson, Daniel Stone, Ghazal Alabtah, Saba Yasir, Michael R. Lucas, Eric W. Klee, Cheryl L. Willman, Judy C. Boughey, Matthew P. Goetz, Krishna R. Kalari, and H. R. Tizhoosh

    arXiv v1 submitted 2026-09-20. No peer-reviewed record was verified. The manuscript says model release is planned after publication and institutional clearance; Mayo training WSIs/reports are not publicly downloadable.

    Official artifacts: Full manuscript and supplementary material

    Metadata checked 2026-10-02

71. #712026arXivPreprintL3

    Pretraining Multiple Instance Learning Networks with Multi-Teacher Distillation from Pathology Slide Foundation Models

    Mingxi Fu, Jiawen Li, Renao Yan, Jiali Hu, Qiehe Sun, Tian Guan, and Yonghong He

    arXiv v1 (2026-07-16) checked 2026-10-02. Six cohorts cover 15 tasks, not 15 independent datasets. Official code and pretrained aggregator checkpoints are listed; experiments were not rerun. This transfers slide embeddings into MIL initialization, not 20×-to-5× acquisition.

    Official artifacts: Versioned manuscript · Official implementation and pretrained MIL weights

    Metadata checked 2026-10-02

72. #722026arXivPreprintL3

    What Makes High-Magnification Knowledge Transferable? A Study of Cross-Resolution Distillation in Whole-Slide Imaging

    Zhiyuan Yang, Jiahao Cheng, and Mahdi S. Hosseini

    Public arXiv v2 (2026-10-01) verified 2026-10-02. The authors state ICLR 2027 under review; no acceptance inferred. Only public source material was used. A diagnostic study of target utility, predictability and downstream benefit, not a new universally superior distillation architecture.

    Official artifacts: Reviewed manuscript revision

    Metadata checked 2026-10-02

73. #732026arXivPreprintL3

    Pathway-Structured Privileged Distillation for Deployable Computational Pathology

    Yongxin Guo, Hao Lu, Onur C. Koyun, Zhengjie Zhu, Muhammet F. Demir, and Metin N. Gurcan

    arXiv v1 (2026-06-01) checked 2026-10-02; no conference acceptance verified. Paired RNA is training-only, while WSI is used at deployment. Pathway-memory readouts are not independently validated molecular measurements.

    Official artifacts: Versioned manuscript

    Metadata checked 2026-10-02

74. #742026NeurIPSAccepted · proceedings pendingL3

    Modeling Whole-Slide Images as Dynamic Tumor Microenvironment Fields

    Lei Wu, Jiashuai Liu, Di Zhang, Zhangpeng Gong, Yingkang Zhan, Yi Niu, Jiusong Ge, Chunze Yang, Kai Yi, Mireia Crispin-Ortuzar, Chen Li, and Zeyu Gao

    Accepted as a NeurIPS 2026 poster: official conference program event 149764 matches the exact title and 12 authors, checked 2026-10-08. Final proceedings were not verified; review uses arXiv v1 (2026-09-28). Official method code is available under a noncommercial research license, but does not include the WSI feature-extraction pipeline. Tissue-tile pixel size, extraction magnification/MPP, and image preprocessing remain unreported in the inspected paper and release.

    Official artifacts: Reviewed manuscript and supplement (v1) · Author publication and acceptance statement · Official NeurIPS 2026 poster · Official method implementation (noncommercial research)

    Metadata checked 2026-10-08

75. #752026NeurIPSAccepted · proceedings pendingL3

    Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation

    Duong M. Nguyen, Trong Nghia Hoang, Hang Thi Nguyen, Thanh Trung Huynh, Phi Le Nguyen, and Minh N. Do

    Acceptance independently verified by NeurIPS 2026 poster 153235 and author repository. Review uses arXiv:2610.00678v1, submitted 2026-09-30. Repository contained only a README on 2026-10-08. Feature condensation at 20× is not low-magnification encoder distillation; Table 1 tests on full feature bags.

    Official artifacts: Official NeurIPS 2026 program · Versioned full manuscript · Author repository (README only at audit)

    Metadata checked 2026-10-08

76. #762025MICCAIPeer reviewedL3

    Prototype-Based Multiple Instance Learning for Gigapixel Whole Slide Image Classification

    Susu Sun, Dominique van Midden, Geert Litjens, and Christian F. Baumgartner

    Official MICCAI 2025 record and 11-page accepted manuscript checked. PANDA is evaluated as binary tumor classification in this study, not the original six-class ISUP grading task. Artifact concepts are masked before MIL retraining; this is not a universal AUC improvement.

    Official artifacts: Official MICCAI open-access record · Official accepted manuscript · Author implementation · arXiv manuscript

    Metadata checked 2026-10-09

77. #772026Pacific Symposium on BiocomputingPeer reviewedL3

    Patch-level phenotype identification via weakly supervised neuron selection in sparse autoencoders for CLIP-derived pathology embeddings

    Keita Tamura, Yao-zhong Zhang, Yohei Okubo, and Seiya Imoto

    PSB 2026 volume 31, pages 708–721, with source supplement checked. Neuron discovery uses slide labels, but the reported localization thresholds use patch-level validation annotations. Single-neuron localization does not preserve all multi-feature slide-classification performance.

    Official artifacts: Official PSB proceedings PDF · Author implementation and supplement

    Metadata checked 2026-10-09

78. #782026MICCAIPeer reviewedL3

    From Patches to Patients: A Study of the Tile-to-Slide Performance Transferability in Digital Pathology

    Sofiène Boutaj, Leo Fillioux, Maria Vakalopoulou, Stergios Christodoulidis, and Pierre Marza

    MICCAI 2026 chapter, LNCS 16891, pages 149–159; publisher first-online date is in September 2026, although the volume copyright/bibliographic imprint says 2027. Review uses the inspected publisher PDF, not only arXiv:2606.10778. Across-encoder rank correlation supports shortlisting, not guaranteed selection of the best model for a new clinical cohort.

    Official artifacts: Publisher full chapter · Official MICCAI open-access record · Official accepted manuscript · Earlier arXiv manuscript

    Metadata checked 2026-10-12

79. #792026arXivPreprintL3

    PathPilot: An Adaptive Question-Conditioned Multi-Magnification Navigation Agent for Whole-Slide Image Analysis

    Zhengyang Xu, Han Li, Jingsong Liu, Ayako Ura, Shihui Zu, Shaohua Kevin Zhou, Nassir Navab, and Peter Schüffler

    arXiv:2610.16734v1, submitted 12 October 2026; manuscript says under review, with no independently verified acceptance. Full 24-page public manuscript inspected; referenced supplementary material was not attached or independently obtained. Code is promised on publication, not a verified release. Adaptive thumbnail/1.25×/2.5×/5×/10× exploration is not knowledge distillation or low-resolution-only inference. QA scores use a language-model judge; deterministic repeated runs do not establish population uncertainty.

    Official artifacts: Inspected arXiv v1 manuscript

    Metadata checked 2026-10-26

80. #802024CVPRPeer reviewedL3

    Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology

    Andrew H. Song et al.

    Metadata checked 2026-07-27

81. #812024ECCVPeer reviewedL3

    Multistain Pretraining for Slide Representation Learning in Pathology

    Jaume et al.

    Metadata checked 2026-07-27