Skip to content
Computational Pathology Reading Roadmap

Foundation encoders

A pathology foundation encoder converts a region, tile, whole slide, or image–text pair into a representation that can be reused for another task. “Foundation” is a claim about intended breadth, not proof of universal transfer, clinical validity, open access, or superiority.

Discovery update: 26 October 2026. Targeted new readings and publication-status checks were added; existing paper reviews retain their individual source-check dates. This is not a fresh full-text audit of every model.

This guide provides a staged reading route, checkpoint-level taxonomy, tile/region and slide/case model registries, exact tile/slide dependencies, native preprocessing recipes, access and publication-status checks, benchmark guidance, and a reproducible comparison protocol.

Scope, evidence, and status

Section titled “Scope, evidence, and status”

This is a curated research map, not an exhaustive catalogue or global ranking. Models enter the core tables when they introduce an important scale, objective, modality, representation level, or reproducible checkpoint. A newer, larger model is not assumed to be better.

Every entry links to a primary paper, official proceedings or publisher page, or an author-controlled repository/model card. Publication status was checked on the review date:

  • peer reviewed means a version of record in a journal or conference proceedings;
  • abstract only means a peer-reviewed conference abstract, not a full methods article;
  • preprint means no peer-reviewed version of record was located;
  • model-card/software release means the checkpoint exists, but a dedicated peer-reviewed paper does not establish its claims;
  • no public checkpoint located means exactly that as of the review date, not that no internal or future artifact exists.

Paper, code, weights, and data are four separate artifacts. A public paper does not make its checkpoint reproducible; a downloadable checkpoint does not make its training data or intended use open.

First classify what is being encoded

Section titled “First classify what is being encoded”
WSI pyramid
  │
  ├─ tissue detection, QC, MPP selection, tile coordinates
  │
  ▼
tile pixels ──► tile / region encoder ──► token or vector per tile
                                              │
                                              ├─► linear probe, retrieval, segmentation
                                              │
                         coordinates + tokens ▼
                                      slide encoder / MIL
                                              │
                                              ▼
                                         slide vector
                                              │
                              slides + patient metadata
                                              ▼
                                      case/patient model


text ──► text encoder ──► shared image–text space
  └────────────────────► multimodal slide/report model
Representation Input and output What it does not provide by itself
Tile / region vision encoder pixels at a defined field of view → class token, pooled token, patch-token map, or vector whole-slide context, patient reasoning, or text semantics
Tile vision–language encoder tile and text → aligned vectors, sometimes captioning tokens a slide representation unless a pooling/slide method is added
Slide encoder many tile vectors, often with coordinates → contextualized tile tokens or one slide vector a new pixel encoder unless one is explicitly bundled
Slide vision–language model tile grid plus report/caption/text → aligned slide/text vectors or generated text interchangeable features with its tile dependency
Case/patient encoder several slides and possibly clinical context → patient representation evidence that slide-level labels generalize to another workflow
General medical encoder multiple medical image modalities, possibly text → shared representation pathology specialization merely because histology was included

The minimum reproducible identity is:

(paper, exact checkpoint + revision, input transform, physical field of view,
 token extraction, feature normalization, tile coordinates, slide aggregator)

Changing any item can change the result while the displayed “model name” stays the same.

Prerequisites and shortest reading route

Section titled “Prerequisites and shortest reading route”
Level Prerequisite You are ready when you can…
L0 · orientation CNNs, ViTs, transfer learning distinguish pretraining objective, encoder, checkpoint, feature, and downstream head
L1 · WSI practice pyramid levels, magnification, MPP, tissue masking calculate physical field of view and reconstruct a tile extraction recipe
L2 · research SSL, contrastive learning, masked modeling, MIL identify the data unit, supervision, collapse control, and aggregation level
L3 · audit external validation, calibration, licensing, cohort provenance design a patient-level overlap audit and a fair multi-model comparison

Read the self-supervised-learning chapter for objective mechanics and the WSI engineering chapter for slide handling. Then take this minimum encoder route:

  1. ImageNet ResNet (2016) as a control and CTransPath (2022) as an early pathology-specific transfer baseline;
  2. HIPT (2022) to separate tile, region, and slide hierarchy;
  3. Kang et al. / Lunit-DINO (2023) to learn why data and evaluation breadth can matter more than an algorithm label;
  4. UNI, Virchow, and Prov-GigaPath (2024) to compare large tile pretraining with a bundled slide encoder;
  5. PLIP (2023), CONCH (2024), and MUSK (2025) to separate visual features from image–text alignment;
  6. CHIEF (2024), PRISM (preprint, 2024), TITAN (2025), and PRISM2 (preprint, 2025) to study slide encoders and their tile dependencies;
  7. Campanella et al. (2025), Neidlinger et al. (2025), PathoROB (2026), and Bareja et al. (2026) before selecting a model from a leaderboard.
  8. COBRA and GECKO (2025), TopoSlide, CARE, MOOZY, DaX, and GigaPath-Flash (2026) to compare current slide/case objectives, adaptive regions, scale consistency, and efficient distillation.

Part I — historical anchors

Section titled “Part I — historical anchors”

These are not obsolete simply because later checkpoints are larger. They expose the design decisions that are hidden inside current systems.

Order Primary reading and year Status Released artifact Read it for
1 Deep Residual Learning for Image Recognition (2016) CVPR, peer reviewed Author implementation a natural-image control; domain-specific gains must beat a controlled baseline, not a historical number
2 CTransPath (2022) Medical Image Analysis, peer reviewed Author code and checkpoint a convolutional stem plus multiscale Swin Transformer and semantically relevant contrastive learning
3 HIPT (2022) CVPR, peer reviewed Author code and checkpoints hierarchical DINO over 104 million 256-pixel tiles, 408,218 regions, and 10,678 TCGA WSIs
4 RetCCL (2023) Medical Image Analysis, peer reviewed Author code and checkpoint clustering-guided contrastive learning and WSI retrieval rather than only classification
5 Benchmarking SSL on diverse pathology datasets (2023) CVPR, peer reviewed Author project and Lunit-DINO weights controlled MoCo-v2, SwAV, Barlow Twins, and DINO comparisons using up to approximately 33 million tiles
6 Phikon scaling study (2023) medRxiv preprint Author code, checkpoint, and features iBOT scale, public-cohort overlap, and the difference between tile count and independent biology

Checkpoint: for each anchor, record the pixel input, physical field of view, objective, training-data unit, output token, and downstream aggregator. If a paper does not report an item, write not reported.

Part II — tile and region vision encoders

Section titled “Part II — tile and region vision encoders”

Established public checkpoints

Section titled “Established public checkpoints”

Counts below preserve the unit used by the source. Do not convert tiles to slides or slides to patients. Parameter and output dimensions refer to the named public checkpoint, not every model in the associated paper.

Model Year and evidence Checkpoint role Architecture and official downstream feature Pretraining evidence Artifact / central caveat
Lunit-DINO / DinoSSLPath Kang et al. (2023), peer reviewed tile vision encoder public ViT-S/8 DINO checkpoint; preserve its official transform and token rule up to 19 million TCGA plus 13.6 million private TULIP tiles in the study Official release; a benchmark paper and checkpoint family, not one universal recipe
Phikon Scaling paper (2023), preprint tile vision encoder ViT-B/16, about 86M parameters; 224-pixel input; 768-dimensional class token iBOT on 40 million TCGA tiles from approximately 6,000 WSIs Official model card; public-cohort overlap and Owkin non-commercial terms require an explicit audit
Path Foundation Technical report (2023), preprint tile vision encoder ViT-S; 224-pixel input; 384-dimensional embedding Masked Siamese Networks on 60 million TCGA patches sampled at approximately 5×, 10×, and 20× Official model card; click-through Health AI terms and TCGA overlap
UNI Nature Medicine paper (2024), peer reviewed tile / ROI vision encoder ViT-L/16, 307M parameters; 224-pixel input; 1,024-dimensional class token DINOv2 losses on exactly 100,130,900 tiles from 100,402 H&E WSIs in Mass-100K Official model card and code; gated, academic non-commercial
Virchow Nature Medicine paper (2024), peer reviewed tile vision encoder ViT-H/14, 632M parameters; 1,280-dimensional class token; official WSI recipe concatenates class-token and mean patch-token vectors to 2,560 dimensions DINOv2 on 1.5 million internal H&E WSIs Official model card; 224 pixels at 0.5 µm/px, gated access, and limited public overlap auditing
Prov-GigaPath tile Nature paper (2024), peer reviewed tile component of a tile-plus-slide system ViT-g/16, about 1.1B parameters; 1,536-dimensional tile vector DINOv2 on exactly 1,384,860,229 256-pixel tiles from 171,189 H&E/IHC slides, over 30,000 patients and 31 tissues Official code and weights; do not call its tile checkpoint “the GigaPath slide encoder”
Phikon-v2 arXiv paper (2024), preprint tile vision encoder ViT-L/16, about 307M parameters; 224-pixel input; 1,024-dimensional class token DINO, iBOT, and KoLeo losses on exactly 456,060,584 20× H&E tiles from approximately 60,000 public WSIs Official model card; its headline rounds this to 450M, so preserve the exact training-section count
Virchow2 arXiv paper (2024), preprint mixed-magnification tile vision encoder ViT-H/14, 632M parameters, four register tokens; official class-plus-mean feature is 2,560 dimensions modified DINOv2 on 3.1 million internal WSIs with 2.0, 1.0, 0.5, and 0.25 µm/px sampling Official model card; exclude register tokens from mean pooling and retain preprint status

Current public model families

Section titled “Current public model families”

This table answers “which version?” before “which model?” It does not imply that a lineage successor dominates its predecessor.

Family Public version and status on 2 September 2026 Scale / objective disclosed by the official source What must be recorded
UNI UNI is peer reviewed; UNI2-h is a model-card release without a dedicated peer-reviewed paper UNI2-h is a custom ViT-H/14 with 681M parameters, eight register tokens, and a 1,536-dimensional output, trained with DINOv2 losses on more than 200M tiles from more than 350,000 H&E/IHC slides do not cite the original UNI paper as an evaluation of UNI2-h; pin the card/checkpoint revision and register-token handling
Virchow Virchow is peer reviewed; Virchow2 remains a 2024 preprint Virchow2 expands to 3.1M WSIs, mixed MPP, KDE regularization in place of KoLeo, and extended-context translation; the paper also studies mini and 1.9B Virchow2G variants state whether the released Virchow2 checkpoint, an unreleased paper variant, class token, or class-plus-mean feature was used
H-optimus H-optimus-0 is a 2024 model-card/software release; H-optimus-1 has a 2026 AACR abstract, not a full methods article both releases use a roughly 1.1B-parameter ViT-g family and 1,536-dimensional output; H-optimus-0 reports more than 500,000 proprietary WSIs, while H-optimus-1 reports billions of tiles from more than 1M slides and 800,000 patients 224 pixels at 0.5 µm/px and the card’s custom stain normalization; distinguish Apache-2.0 H-optimus-0 terms from gated CC-BY-NC-ND H-optimus-1
Hibou Hibou-B and Hibou-L, 2024 preprint DINOv2 with registers on a proprietary corpus of more than 1M WSIs; cards report 512M clean patches for B and 1.2B for L B versus L checkpoint, gated click-through access, Apache-2.0 license, exact official transform, and token selection
Kaiko five DINO/DINOv2 checkpoints, 2024 preprint; Midnight-12k, 2025 preprint the first family samples online from about 29,000 TCGA FF/FFPE WSIs across 5×/10×/20×/40×; Midnight-12k is the public member of a high-resolution post-training family architecture/patch size, magnification, online sampling policy, and Kaiko license; do not invent a distinct-tile count for an online sampler
PathOrchestra npj Digital Medicine paper (2025), peer reviewed DINOv2 pretraining on 287,424 WSIs across 21 tissue types and three medical centers; evaluation spans 112 tasks official gated checkpoint, version, CC-BY-NC-ND terms, and overlap with its many evaluation datasets

Objective, scale, and stain-specialized branches

Section titled “Objective, scale, and stain-specialized branches”

Scale-only comparisons miss models that change the supervision source.

Model Year and status What changes Official artifact
GPFM Nature Biomedical Engineering (2025), peer reviewed combines multiple expert-model distillation, self-distillation, and local/global alignment; it is not purely label-free visual SSL Author code and model
mSTAR tile encoder Nature Communications (2025), peer reviewed learns a slide aggregator from WSI–report–gene-expression relations, then injects whole-slide knowledge into a tile encoder; 26,169 slide-level modality pairs from 10,275 patients and more than 116M tiles Author code and gated model
BEPH Nature Communications (2025), peer reviewed masked-image modeling with BEiT on 11M unlabeled histopathology images, followed by patch and WSI transfer Author code and checkpoint
StainNet (foundation encoder; not the 2021 normalization network) MIDL proceedings (2026), peer reviewed ViT-S and ViT-B self-distillation dedicated to IHC and special stains: more than 1.4M patches from 20,231 public HISTAI WSIs Author code and model links
ALICE arXiv (2026), preprint agglomerative distillation from eight vision, VLM, and slide teachers into a unified backbone; 24,985,184 tile images plus 155,604 high-resolution images Author code/model repository; too new for independent validation at this review
MRPT arXiv (2026), preprint consecutive cross-resolution attention and semantic consistency across cell, region, and WSI levels; pretraining reports 624M patches, 2.4M regions, and 36K WSIs the manuscript says code/models are available via a placeholder link; no working release is claimed here
SmartStu Official proceedings (2026), MICCAI, peer reviewed; official proceedings verified multi-teacher distillation into compact breast-specific students, with an adversarial nuisance model and artifact-injection SSL to suppress non-biological shortcuts Author code; compare tiny backbones, nuisance targets, and external cohorts separately
DaX arXiv (2026), preprint continuous-magnification and cross-scale-consistency SSL; test whether those mechanisms transfer beyond the authors’ 161-task benchmark paper is the primary artifact; independent evaluation remains needed
GenBio-PathFM Company technical report (2026), not peer reviewed morphology-aware curation followed by DINO-to-JEPA training record the exact released checkpoint and its non-OSI community license rather than treating report-only evidence as peer review
EXAONEPath family 1.0, 2.0, and 2.5 (2024–2025), preprints progresses from stain-normalized SSL to slide-label supervision and multi-omics SigLIP alignment treat the three objectives and checkpoints as separate systems, not interchangeable versions of one SSL model

Distilled and deployment-oriented encoder branch

Section titled “Distilled and deployment-oriented encoder branch”

These models reduce different bottlenecks. A smaller parameter count, a shorter embedding, and a lower WSI magnification are three separate interventions.

Model Year and evidence Teacher → student contract Deployed artifact and audit
H0-mini MICCAI paper (2025), peer reviewed distils a large pathology foundation model into a much smaller encoder; validation includes HEST/EVA transfer plus scanner/stain robustness on PLISM-WSI and a private breast cohort official model and PLISM benchmark; match checkpoint, transform, parameter count, throughput, and robustness rather than assuming size preserves invariance
XMAG Medical Image Analysis (first online 2026), peer reviewed; earlier 2025 manuscript the inspected earlier manuscript uses a frozen 20× UNI2 teacher: sixteen registered tile CLS targets supervise the mean student CLS and sixteen pooled student regions at 5× official code and model; journal publication metadata verified 26 October 2026, but the final full text was not inspected in this pass. Earlier scientific notes and timing inconsistencies remain version-bound; journal abstract timing is encoder-level, not end-to-end WSI latency
MAD-NP arXiv (2025), preprint aligned low-magnification context and high-magnification detail supervise one self-supervised scale-aware embedding space official model; audit whether invariance transfers beyond the reported neuropathology tasks and whether detail-sensitive endpoints are lost
H0-mini versus XMAG mechanism control H0-mini changes capacity at a fixed model input contract; XMAG changes both capacity and magnification do not attribute XMAG’s WSI patch-count reduction to parameter compression or H0-mini’s robustness to low-magnification transfer
Pathryoshka arXiv manuscript and official ECCV record (2026), accepted several pathology foundation-model teachers supervise a compact student with nested embedding widths official Pathryoshka-B model; benchmark every deployed width, not only the largest nested vector
LaGuadia Official proceedings (2026), MICCAI, peer reviewed MedSigLIP-guided meta-teacher routes knowledge from GigaPath, UNI, and Virchow2 into an 87M-parameter student across MIL and multimodal tasks official code; official MICCAI proceedings verified 2 October 2026; text-guided routing needs text-only and single-teacher controls
MuCoDi / MuCoEdge Official EMA workshop proceedings (2026), peer reviewed workshop, not the main MICCAI track three frozen pathology teachers provide cached keys; two student augmentations and teacher-specific projections learn contrastive tile identity; both teacher and student use 5× author code; 23 teacher-eligible tasks, frozen-encoder MIL, and Raspberry Pi single-tile timing support an edge-compression trade-off, not 20×→5× transfer or complete WSI speed

For a compact-encoder comparison, publish a lineage table with teacher checkpoint, student architecture, training images and patients, native MPP/crop, matched field of view, transferred layers/tokens, loss weights, frozen components, output dimension, parameter and activation memory, tiles per WSI, throughput, license, and external robustness. Then place each model on a quality–robustness–systems-cost frontier.

Important research lines without a verified public checkpoint

Section titled “Important research lines without a verified public checkpoint”

Read these for hypotheses about curation and scaling, not as immediately reproducible baselines.

Research line Year and status Contribution to study Public-weight status on review date
RudolfV arXiv (2024), preprint pathologist-guided curation over 103,000 slides, 750M patches, more than 15 labs, 58 tissues, and 129 stains no public checkpoint located on the official Aignostics model page
Atlas / Atlas2 Atlas (2025) and Atlas2 (2026), preprints extends the RudolfV curation line to 1.2M and then 5.5M WSIs; useful for studying institutional and disease diversity no public checkpoint located
PLUTO arXiv (2024), preprint a lightweight FlexiViT trained on 195M multi-site, multiscale tiles with hybrid self-supervision no public checkpoint located; see the official PathAI announcement
PLUTO-4 arXiv (2025), preprint PLUTO-4S/4G trained from 551,164 WSIs, 137,144 patients, more than 50 institutions, 60 disease types, and 100 stains no verified public weights located

Part III — tile vision–language encoders

Section titled “Part III — tile vision–language encoders”

A vision–language encoder has at least two output spaces:

  • unprojected visual features for a linear probe or MIL;
  • projected, normalized image and text features for zero-shot scoring or retrieval.

They are not interchangeable. Record the prompt set, text tokenizer, projection, normalization, temperature/logit scale, and whether prompts were tuned on validation or test data.

Model Year and evidence Pretraining signal and scale Released capability and caveat
PLIP Nature Medicine (2023), peer reviewed CLIP adaptation on OpenPath: 208,414 pathology images paired with natural-language descriptions curated from public social-media and web sources Author code and model; text quality, consent/provenance, and temporal web-data leakage require scrutiny
CONCH Nature Medicine (2024), peer reviewed CoCa image–text contrastive plus captioning losses over 1.17M H&E, IHC, and special-stain image–caption pairs; ViT-B/16 vision and 12-layer text encoders Official code and gated model; public decoder weights were removed, but image/text encoders remain
KEP ECCV proceedings (2024), peer reviewed encodes a pathology knowledge tree and uses it to guide image–text alignment on OpenPath or Quilt-1M Author code and checkpoints; distinguish KEP from the later KEEP model
MUSK Nature (2025), peer reviewed unified masked modeling over 50M pathology images from 11,577 patients and one billion pathology-related text tokens, then alignment with one million image–text pairs Author code and gated model; 384-pixel native input and separate image/text feature modes
CONCH v1.5 Official model card (release only); used inside peer-reviewed TITAN (2025) ViT-L/16 initialized from UNI and CoCa-adapted; the card reports 1.17M pairs while the TITAN version of record reports 1.26M no standalone paper or code repository on the card; preserve this source discrepancy, exact revision, 512-pixel TITAN recipe, and gated non-commercial terms
MedSigLIP-448 Google technical report (2025), preprint and official card 400M-parameter vision and 400M-parameter text towers, 448-pixel images and up to 64 text tokens; trained across radiology, dermatology, ophthalmology, histopathology, and other medical data a useful general-medical comparator, not a pathology-specialist model; click-through Health AI terms
KEEP Cancer Cell article (2026), peer reviewed knowledge graph with 11,454 diseases and 139,143 attributes reorganizes noisy web pairs into 143,000 semantic groups Official model card and author repository; evaluate knowledge-source coverage and rare-class prompt sensitivity

Do not call zero-shot WSI classification “slide encoding” when it merely scores tiles and pools their logits. The slide representation and pooling rule still need to be named.

Part IV — slide and case encoders

Section titled “Part IV — slide and case encoders”

Dependency map

Section titled “Dependency map”
CTransPath tile features ─────────────► CHIEF slide model
Prov-GigaPath ViT-g tile features ────► GigaPath LongNet slide encoder
Virchow tile features ────────────────► PRISM Perceiver + BioGPT
CONCH v1.5 512-pixel tile features ───► TITAN slide encoder
Virchow2 class-token features ────────► PRISM2 slide encoder + decoder


tile model + generic ABMIL/CLAM/TransMIL ≠ a pretrained slide foundation model

The upstream tile checkpoint is part of a slide model’s identity. Swapping it without retraining is a new method, not an implementation detail.

Slide-encoder reading and artifact registry

Section titled “Slide-encoder reading and artifact registry”
Model Year and status Input dependency and pretraining Output / use Official artifact and caution
HIPT CVPR (2022), peer reviewed nested 256-pixel tile and 4,096-pixel region representations learned with hierarchical DINO region/slide hierarchy for weakly supervised WSI tasks Author release; reproduce coordinate and hierarchy construction, not only the lowest-level encoder
CHIEF Nature (2024), peer reviewed CTransPath tile features followed by unsupervised tile learning and weak slide supervision using anatomical-site labels; study reports 60,530 WSIs, 19 sites, and 44 TB slide representation for cancer detection, origin, molecular profile, and prognosis Author code and model; it is not a text-aligned zero-shot model
PANTHER CVPR (2024), peer reviewed fits task-agnostic morphological prototypes to tile embeddings compact, interpretable WSI representation Author code; prototype count, tile encoder, and sampling are part of the representation
TANGLE CVPR (2024), peer reviewed aligns slide representations with paired transcriptomic profiles slide vector enriched for morphology–expression relations Author code; paired RNA is training supervision, not a free test-time input
MADELEINE ECCV (2024), peer reviewed global and local alignment across matched H&E, IHC, and special-stain slides multistain slide representation Author code; matching/registration and stain availability define the valid population
Prov-GigaPath slide Nature (2024), peer reviewed official GigaPath tile embeddings plus coordinates; LongNet with masked autoencoding over whole-slide sequences contextualized tiles and/or slide representation Author code and weights; tile and slide checkpoints, coordinates, and sequence length must all be pinned
PRISM arXiv (2024), preprint Virchow tile embeddings, a Perceiver slide encoder, and BioGPT trained with CoCa-style objectives on 587,000 WSIs and 195,000 reports slide–text retrieval, zero-shot tasks, and report-related representations Official gated model card; still a preprint and not interchangeable with Virchow alone
mSTAR Nature Communications (2025), peer reviewed WSI–report and WSI–gene-expression contrastive learning, followed by slide-to-tile self-taught transfer both slide and knowledge-enriched tile representations Author code/model; record which stage/checkpoint supplies a reported feature
TITAN Nature Medicine (2025), peer reviewed 512-pixel CONCH v1.5 features on a spatial grid; visual SSL plus report alignment over 335,645 WSIs and 423,122 synthetic captions slide vector, cross-modal retrieval, zero-shot use, and report generation in the full architecture Author code and gated model; released decoder is omitted to reduce PHI risk, and CONCH v1.5 is a required dependency
PRISM2 arXiv (2025), preprint class-token-only Virchow2 features; 2.3M WSIs, 685,507 specimens, 200,692 patients, then 14M pathology QA pairs approximately 0.6B-parameter WSI encoder inside a 4.4B-parameter multimodal system Official gated model card; pin WSI encoder versus full decoder and retain preprint status
COBRA CVPR (2025), peer reviewed combines frozen features from several tile foundation models in one unsupervised slide-learning objective foundation-model-agnostic slide vector preserve the constituent tile checkpoints and test whether gains come from ensembling rather than the slide objective alone
GECKO ICCV (2025), peer reviewed contrasts gigapixel visual evidence with curated pathology concepts concept-aligned slide representation concept supervision is not label-free SSL; audit concept coverage and prompt sensitivity
TopoSlide CVPR (2026), peer reviewed predicts persistent-homology descriptors of the spatial arrangement of tile-embedding clusters topology-aware slide vector and conditional retrieval Author code and models; a focused inductive-bias study, not evidence that hundreds of slides replace scale in every task
CARE CVPR (2026), peer reviewed combines self-supervised morphology, RNA/protein guidance, and adaptive modeling of irregular coherent regions molecular-guided region and slide representation name the molecular supervision and compare adaptive regions against fixed tiles under equal evidence and compute
MOOZY arXiv (2026), accepted at ECCV 2026 according to the author project; proceedings forthcoming masked slide self-distillation over 77,134 public feature grids (about 1.67B patches), then a case transformer across all slides from a patient with supervision over 333 tasks from 56 datasets slide and patient/case representations Project · code · weights; record v3’s 16 held-out tasks, audit cohort/task overlap, and do not compare a case vector as though it were a single-slide embedding
GigaPath-Flash / GigaTIME-Flash arXiv (July 2026 preprint, v2) distils tile and slide components of the GigaPath family efficient tile/slide representations reproduce the author-reported efficiency/performance trade-off and retain preprint status until independently validated

Patient/case-level multi-slide representation

Section titled “Patient/case-level multi-slide representation”

MOOZY belongs here and in the MIL/WSI chapter—not in the pathology-VLM branch. Its defining unit is the patient case: Stage 1 learns a vision-only slide encoder, and Stage 2 aggregates one or more slide vectors with a case transformer. A fair evaluation must therefore report how slides are assigned to cases, how missing or extra slides are handled, whether any cohort/task families appear in pretraining, and whether the comparator receives the same multi-slide evidence.

What counts as a fair slide comparison?

Section titled “What counts as a fair slide comparison?”

There are three different scientific questions:

  1. Tile representation: hold the slide sampler and aggregator fixed while changing only the tile checkpoint.
  2. Slide representation: use frozen, released slide features with the same probe and patient-level folds.
  3. End-to-end system: compare each native tile-plus-slide pipeline, then report that the result measures the entire pipeline—not “encoder quality.”

A fourth experiment, swapping tile encoders under one pretrained slide transformer, is valid only when the slide model is retrained or the distribution shift is explicitly studied.

Part V — preprocessing is part of the checkpoint

Section titled “Part V — preprocessing is part of the checkpoint”

Physical field of view

Section titled “Physical field of view”

For an image of width P pixels sampled at m micrometres per pixel:

field of view in micrometres = P × m

Thus, 224 pixels at 0.5 µm/px covers 112 µm, while 512 pixels at 0.5 µm/px covers 256 µm. Resizing both arrays to 224 pixels does not make their biological context equal. Nominal magnification is also scanner-dependent; “20×” is often near 0.5 µm/px, but slide metadata or a calibrated estimate should be used.

Native-recipe ledger

Section titled “Native-recipe ledger”

This is an orientation table, not a substitute for the linked model card at the pinned revision.

Checkpoint Native image / scale convention Feature rule to preserve Frequent mistake
Phikon / Phikon-v2 official Hugging Face processor, 224-pixel model input; v2 training tiles sampled at 20× class token: 768 / 1,024 dimensions manually copying another ViT’s normalization
Path Foundation 224 pixels; training spans approximately 5×/10×/20× official 384-dimensional embedding API assuming its multiscale training makes test MPP irrelevant
UNI official timm transform resizes to 224 and uses ImageNet normalization; training sampled at 20× 1,024-dimensional class token treating the 256/512-pixel stored training crops as two supported inference inputs
UNI2-h official timm transform; 224-pixel model input 1,536-dimensional class token; eight registers exist internally pooling registers as tissue tokens
Virchow 224 pixels at 0.5 µm/px either 1,280-dimensional class token or official 2,560-dimensional class-plus-mean feature—state which comparing one extraction rule with the other under the same name
Virchow2 224 pixels; trained at 2.0/1.0/0.5/0.25 µm/px official 2,560-dimensional class-plus-mean feature excludes four registers claiming mixed-MPP pretraining removes the need to choose inference MPP
Prov-GigaPath tile extract 256-pixel tiles at 0.5 µm/px; official transform resizes to 256 then centre-crops 224 1,536-dimensional tile vector; retain (x,y) coordinates for slide encoding feeding unordered tile vectors to the LongNet slide checkpoint
H-optimus-0 / 1 224 pixels at 0.5 µm/px with the card’s custom mean and standard deviation 1,536-dimensional feature applying ImageNet normalization because the backbone is a ViT
Hibou-B / L use the author transform and scale guidance for the exact B or L checkpoint pin class/patch/register-token handling writing “Hibou” without version or feature rule
Kaiko sample at an explicit 5×/10×/20×/40× policy; official pipeline resizes/crops to 224 checkpoint-specific class token collapsing a multiscale checkpoint experiment into “20×”
MUSK resize and centre-crop to 384 with the official multiscale flag recorded visual class token before or after the retrieval head, depending on task mixing with_head, normalization, or multiscale modes
CONCH v1.5 → TITAN 512-pixel tiles at 20× arranged as a spatial grid 768-dimensional tile features expected by TITAN extracting ordinary 224-pixel UNI features because v1.5 was initialized from UNI
Virchow2 → PRISM2 224 pixels at 0.5 µm/px in the official PRISM2 preprocessing class-token-only N × 1280 feature sequence passing Virchow2’s 2,560-dimensional class-plus-mean WSI feature

Two defensible evaluation tracks

Section titled “Two defensible evaluation tracks”

Run both when resources permit:

  1. Native-recipe track: each checkpoint receives the exact input, MPP, transform, token rule, and normalization specified by its authors. This tests the released system.
  2. Matched-physical-field track: choose one biological field of view and sampling MPP, derive every model input from the same source pixels, and document any crop, interpolation, or padding required. This better isolates representation differences but may move a checkpoint off its native distribution.

Never silently use a “universal” extractor for both tracks.

Extraction manifest

Section titled “Extraction manifest”

For every feature store, preserve:

source:
  slide_id: ""
  patient_id: ""
  slide_checksum: ""
  scanner: ""
  native_mpp: null
tiling:
  requested_mpp: null
  tile_pixels_before_model_transform: null
  stride_pixels: null
  tissue_mask_version: ""
  qc_version: ""
  coordinates_reference_level: 0
encoder:
  repository: ""
  checkpoint_id: ""
  checkpoint_revision_or_hash: ""
  input_transform: ""
  model_input_pixels: null
  token_rule: cls | mean_patch | cls_plus_mean | projected | other
  output_dimension: null
  feature_normalization: none | l2 | other
runtime:
  library_versions: {}
  precision: fp32 | fp16 | bf16
  device: ""
  extraction_date: YYYY-MM-DD

Access and license are selection criteria

Section titled “Access and license are selection criteria”

Terms change, so the linked card at the pinned revision is authoritative. The categories below describe the state on the review date.

Access pattern Examples Operational consequence
Gated, academic/non-commercial UNI, UNI2-h, CONCH, CONCH v1.5, MUSK, Virchow2, H-optimus-1, PathOrchestra, PRISM/PRISM2, TITAN, mSTAR every collaborator may need individual approval; derived features, redistribution, hosted demos, and commercial collaboration may be restricted
Click-through with permissive code/weight license plus extra terms Virchow, Prov-GigaPath, H-optimus-0, Hibou “Apache-2.0” in metadata does not erase card-specific access or acceptable-use conditions
Downloadable but non-commercial model terms Phikon, Phikon-v2, Kaiko families public availability is not permission for every downstream use
Health-AI terms Path Foundation, MedSigLIP review the current Google terms; do not assume a standard open-source license
Paper/code but no verified public weights RudolfV, Atlas/Atlas2, PLUTO/PLUTO-4 suitable for reading and hypothesis generation, not a reproducible checkpoint comparison

Before adopting any model, save the license/terms revision, approval date, permitted use, redistribution rule, derivative-feature rule, clinical-use disclaimer, and the contact used for clarification. Nothing on this page is legal or regulatory advice.

Model selection without a global leaderboard

Section titled “Model selection without a global leaderboard”

Choose a panel that spans the scientific uncertainty:

Research question Minimum defensible panel
Is pathology pretraining useful? one controlled ImageNet vision baseline, one compact pathology encoder, and one recent pathology encoder under the same tiles/head
Does scale help this endpoint? two checkpoints from one lineage where objective and recipe are as similar as possible, such as Phikon → Phikon-v2 or Virchow → Virchow2
Is image–text alignment useful? one vision-only encoder and one VLM with comparable adaptation; evaluate unprojected visual features separately from projected zero-shot features
Does a pretrained slide model help? fixed tile encoder + simple mean/ABMIL control versus the native tile-plus-slide pipeline
Are non-H&E stains in scope? an H&E-heavy control plus a model with disclosed IHC/special-stain training, such as StainNet, CONCH, UNI2-h, or PathOrchestra
Is deployment constrained? at least one compact and one large encoder; report throughput, memory, feature size, gate latency, and license fit
Is external robustness central? models with different institutions/data sources, evaluated at a held-out health system and stratified by site/scanner/stain

Do not select a panel solely because the models are easy to download. Access convenience, scientific coverage, and deployment permission should be reported as separate criteria.

Primary benchmark route

Section titled “Primary benchmark route”

Read model papers first, then these comparisons:

Reading Year / status What it can establish What it cannot establish
Clinical benchmark of public self-supervised pathology foundation models 2025, Nature Communications standardized public tile encoders on real clinical cohorts from multiple health systems, including compute-aware comparison a permanent winner for unreleased checkpoints, other heads, or other populations
Benchmarking foundation models as feature extractors for weakly supervised computational pathology 2025, Nature Biomedical Engineering 19 models across 31 tasks, 6,818 patients, 9,528 slides, 13 cohorts, and external validation tile quality independent of the selected WSI aggregation protocol
PathoROB: Towards robust foundation models for digital pathology 2026, Nature Communications robustness/site-signature analysis over 20 models, 99,392 patches, 28 biological classes, and 34 centres clinical safety or generalization to every stain and workflow
Benchmark study of vision and pathology foundation models 2026, Nature Communications; publisher page currently labels the manuscript unedited 32 general/pathology vision and VLM encoders across patch, WSI, external, and out-of-domain tasks immunity to split choices, future benchmark contamination, or model-card recipe errors
Adaptive fusion of pathology foundation models 2026, arXiv preprint whether complementary encoder signals can be selected per sample instead of assuming one globally best checkpoint that a more complex fusion controller is clinically robust or compute-neutral; author code
What carries the signal in pathology-FM atlases? 2026, arXiv preprint patient-level controls for composition, segmentation, and slide/patient hierarchy in a breast atlas generalization beyond the audited atlas, cohort, and access-controlled segmentation pipeline
Foundation models as mitotic-figure detection encoders 2026, arXiv preprint dense detector transfer rather than another frozen classification ranking; includes cross-domain/cross-species stress a universal detector winner independent of feature level, detector, annotation, and object-density shift
Beyond Counts: a distributional robustness margin 2026, arXiv preprint distribution-level cell-composition shifts rather than only mean counts or average task accuracy a complete safety metric; representation, cell model, matching, and cohort definitions remain part of the estimand

Useful author-maintained evaluation infrastructure:

  • Patho-Bench for standardized pathology foundation-model evaluation;
  • TRIDENT for WSI preprocessing and supported feature extractors;
  • eva for reproducible representation evaluation;
  • the Campanella live benchmark for updated public-checkpoint comparisons.

Using a toolkit does not remove the need to inspect its pinned transform, checkpoint, split, and aggregation implementation.

Minimal fair-comparison protocol

Section titled “Minimal fair-comparison protocol”

1. Freeze the scientific question

Section titled “1. Freeze the scientific question”

Declare one of: tile information, slide information, adaptation capacity, zero-shot alignment, external robustness, dense prediction, or deployment fit. Do not mix frozen linear probing with full fine-tuning and call the difference “encoder quality.”

2. Freeze the cohort and sampling

Section titled “2. Freeze the cohort and sampling”
  • split by patient before tile extraction or normalization fitting;
  • use identical tissue masks, QC, coordinates, tile budget, labels, and folds;
  • stratify by institution, scanner, tissue preparation, stain, and class prevalence;
  • construct a pretraining-overlap matrix for TCGA, CPTAC, GTEx, CAMELYON, PANDA, TCIA collections, and local institutions.

3. Run the two preprocessing tracks

Section titled “3. Run the two preprocessing tracks”

Report native-recipe and matched-field-of-view results separately. Save image examples and coordinate overlays for each extractor.

4. Equalize adaptation

Section titled “4. Equalize adaptation”
  • frozen tile test: same linear head, k-NN, retrieval metric, or segmentation decoder;
  • WSI test: same tile bags and mean pooling/ABMIL/TransMIL head;
  • slide-model test: same frozen slide probe;
  • fine-tuning test: same labelled data, trainable-parameter class, optimizer search, early stopping, seeds, and accelerator budget.

5. Report uncertainty and failure strata

Section titled “5. Report uncertainty and failure strata”

Report confidence intervals or repeated folds, paired tests where justified, calibration, abstention, rare-class recall, and site/scanner/stain strata. A small mean improvement without uncertainty is not a model-selection argument.

6. Report systems cost

Section titled “6. Report systems cost”

Measure tiles per second, peak device memory, wall time, energy when available, feature bytes per tile, WSI feature-store size, slide-encoder sequence limit, checkpoint access time, and license fit.

Failure modes to look for

Section titled “Failure modes to look for”

Data and evaluation

Section titled “Data and evaluation”
  • public-cohort recycling: a public “test” cohort may appear in pretraining;
  • unit leakage: serial sections, multiple slides, or tiles from one patient cross folds;
  • site shortcut: stain/scanner/lab signatures dominate tissue morphology;
  • test-time prompt selection: prompts or class descriptions are tuned on test labels;
  • one-head conclusions: an encoder is declared best using one MIL head and one tile budget;
  • benchmark aging: later pretraining can absorb a once-independent public test set.

Representation and implementation

Section titled “Representation and implementation”
  • MPP mismatch: nominal 20× is treated as calibrated 0.5 µm/px;
  • token mismatch: class, patch mean, class-plus-mean, register, projected, and unprojected features are mixed;
  • dependency mismatch: PRISM, TITAN, GigaPath, or PRISM2 receives the wrong tile checkpoint or dimension;
  • coordinate loss: a spatial slide encoder receives shuffled or level-mismatched coordinates;
  • transform drift: ImageNet normalization replaces a checkpoint-specific transform;
  • feature-store ambiguity: old and new checkpoint outputs share one directory without hashes.

Interpretation and deployment

Section titled “Interpretation and deployment”
  • attention-as-explanation: attention or nearest neighbours are treated as causal pathology evidence;
  • scale mythology: slide count, tile count, patient count, and images seen are compared as one number;
  • open-weight mythology: gated or non-commercial weights are described as unrestricted open source;
  • clinical overreach: retrospective transfer is described as clinical readiness;
  • population erasure: rare tissues, non-neoplastic disease, paediatric disease, non-H&E stains, frozen tissue, and underrepresented regions are hidden inside a pan-cancer average.

Advanced research exercises

Section titled “Advanced research exercises”

Exercise A — checkpoint identity audit

Section titled “Exercise A — checkpoint identity audit”

Select UNI, Virchow2, and one smaller encoder. Reproduce one feature vector per checkpoint from the official example, then save transform, revision, token rule, shape, checksum, precision, and tolerance. Deliberately change one factor and quantify the drift.

Exercise B — MPP × encoder crossover

Section titled “Exercise B — MPP × encoder crossover”

On one external WSI cohort, extract the same tissue at 0.25, 0.5, 1.0, and 2.0 µm/px. Compare native and matched-field tracks with a fixed head. Ask whether the model ranking changes with biological scale and whether that interaction is pathology-plausible.

Exercise C — token ablation

Section titled “Exercise C — token ablation”

For a ViT with patch and register tokens, compare class token, mean patch token, class-plus-mean, and an intentionally incorrect mean including registers. Evaluate linear probing, retrieval, and WSI MIL. Explain the failure from architecture, not only the score.

Exercise D — tile versus slide factorization

Section titled “Exercise D — tile versus slide factorization”

Using identical patient folds and tiles, compare:

  1. mean pooling;
  2. one simple supervised MIL head;
  3. one pretrained slide encoder with its native tile dependency.

Then replace the tile encoder and retrain the MIL control. Attribute gains to tile representation, aggregation, pretraining, or interaction only when the design identifies them.

Exercise E — VLM prompt audit

Section titled “Exercise E — VLM prompt audit”

For PLIP, CONCH, MUSK, or KEEP, pre-register class names and prompt templates using a development cohort. Compare single prompts, an ensemble, synonyms, definitions, and negation. Report image-only linear probes beside zero-shot results and test prompt transfer to another institution.

Exercise F — site-signature stress test

Section titled “Exercise F — site-signature stress test”

Train a probe to predict institution/scanner from frozen features, then measure target performance before and after stain augmentation or site balancing. Inspect whether removing site signal also removes clinically useful morphology.

Exercise G — mechanism-level replication

Section titled “Exercise G — mechanism-level replication”

Choose one specialized release—PathOrchestra, StainNet, TopoSlide, KEEP, or ALICE—and reproduce only one mechanism-level claim. First verify publication status, exact checkpoint, access, cohort overlap, and whether independent code/weight reproduction is possible. A carefully documented non-reproduction is a valid result.

Versioned model record

Section titled “Versioned model record”

Use one record per checkpoint, not one per marketing name:

identity:
  display_name: ""
  checkpoint_id: ""
  checkpoint_revision_or_hash: ""
  release_date: YYYY-MM-DD
  representation_level: tile | region | slide | case | image-text | slide-text
evidence:
  primary_url: ""
  status: peer-reviewed | abstract-only | preprint | model-card-only
  venue: ""
  publication_year: null
  code_url: ""
  weights_url: ""
architecture:
  family: ""
  parameters: null
  input_pixels: null
  patch_pixels: null
  register_tokens: null
  output_features:
    name: cls | mean_patch | cls_plus_mean | projected | slide
    dimension: null
pretraining:
  objective_terms: []
  patients: null
  cases_or_specimens: null
  slides: null
  regions: null
  tiles_or_images: null
  image_text_pairs: null
  tissues: null
  stains: []
  sites: null
  public_collections: []
inputs:
  mpp: null
  magnification_policy: ""
  physical_field_um: null
  transform_source: ""
  normalization: ""
dependencies:
  tile_checkpoint: ""
  coordinate_convention: ""
  aggregator_or_slide_checkpoint: ""
access:
  gate: none | click-through | application | unavailable
  license_spdx_or_name: ""
  noncommercial: null
  redistribution: ""
  clinical_use: ""
evaluation:
  frozen_or_tuned: ""
  downstream_head: ""
  split_unit: patient
  external_sites: []
  pretraining_overlap: none | possible | known | unknown
verification:
  checked_on: YYYY-MM-DD
  checked_by: ""
  notes: ""

Maintenance policy

Section titled “Maintenance policy”

At each quarterly review:

  1. search the Paper Radar and primary venue indexes for pathology encoders, WSI pretraining, image–text alignment, multistain models, case-level models, and external benchmarks;
  2. verify peer-review status against the publisher—never promote a preprint from a search snippet;
  3. pin the model-card and repository revision, then rerun transform, output-shape, and checksum tests;
  4. update access and license fields separately from scientific evidence;
  5. preserve source units and source disagreements instead of silently reconciling them;
  6. add a model only when it changes the scientific coverage or has a usable checkpoint; otherwise leave it in the Paper Radar or an internal staging queue until it belongs in a scientific subsection;
  7. archive benchmark version, cohort-overlap audit, and evaluation date with every score;
  8. update the Last literature review date only after every changed record has passed these checks.

The practical output of this chapter is not “the best encoder.” It is a defensible shortlist of exact checkpoints whose data, objective, input physics, representation level, access terms, and evaluation protocol match a stated research question.

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 · evidenceCore idea and why to read itOfficial 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 · evidenceCore idea and why to read itOfficial 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.

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 · evidenceCore idea and why to read itOfficial 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 · evidenceCore idea and why to read itOfficial 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.

Tissue distillation: masks, features, hierarchy and relations

Compare staged pixel/feature KD from Virchow2, PAGET’s hierarchical multi-teacher pseudo-labels, and ConStruct’s transfer of SegFormer structural affinities into adapted CONCH features for prototype-guided segmentation. Their annotation contracts and deployed components differ.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Computationally Efficient Pathology Segmentation using Knowledge Distillation from Foundation ModelsJiaqi Lv, Yijie Zhu, Saki Okada, Maxwell Stanley Renna, Robert Goldin, Stefan Antonowicz, and Shan E Ahmed Raza2026 · arXiv · preprintarXiv v1 (2026-09-03), submitted to Medical Image Analysis according to the authors, not accepted. Verified methods and selected results on 2026-10-02. TIAToolbox is the stated release destination; availability of these exact weights was not independently confirmed. Patch-throughput measurements are not end-to-end WSI latency.Transfer Virchow2 segmentation knowledge into SegFormerB5 or EUNet-M using a staged combination of pixel predictions and global features. The deployed model is a compact segmentation student, not the foundation-model teacher. This changes model cost at fixed patch input, not optical magnification.Versioned manuscript · TIAToolbox — stated release destination

Reading question: Which supervision remains observable at student resolution, and which teacher or post-processing components still run at inference?

Evidence boundary: PAGET-H retains HoverNet. ConStruct’s residual adapters are not residual-error targets. Pixel segmentation, patch throughput and patient-level WSI outcomes are distinct evidence.

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 · evidenceCore idea and why to read itOfficial 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 · evidenceCore idea and why to read itOfficial 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.

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 · evidenceCore idea and why to read itOfficial 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.

Spatial evidence: separate coherence, molecular agreement, and anatomy

SpaPath-Bench compares frozen image features through a controlled spatial learner and clustering readout on paired histology–ST sections. Its three reference regimes expose different strengths: a smooth partition, a transcriptomic proxy, and an expert anatomical boundary are not interchangeable targets.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Benchmarking Pathology Foundation Models for Spatial Domain UnderstandingBokai Zhao, Yiyang Zhang, Yuanchi Zhu, Hanqing Chao, Long Bai, Tai Ma, Minfeng Xu, Ming Song, and Tianzi Jiang2026 · MICCAI · peer-reviewedMICCAI 2026 chapter, LNCS 16895, pages 35–45, first online 25 September 2026; the volume imprint says 2027. Review uses the inspected publisher PDF. The 42 slides comprise 12 expert-annotated DLPFC slides and 30 slides assessed using transcriptomic clustering proxies. The linked public repository is a prepared-results dashboard, not a verified end-to-end benchmark implementation.SpaPath-Bench asks whether frozen pathology features support meaningful tissue partitions after a controlled spatial readout. It compares 19 encoders on 42 paired H&E–spatial-transcriptomics slides, separating spatial coherence, molecular-partition agreement and expert-label agreement. The evaluated object is an encoder-plus-readout system, not innate global spatial reasoning by a local tile encoder.Publisher full chapter · Official MICCAI open-access record · Official accepted manuscript · Author results dashboard · Dashboard repository and documented limits

Reading question: Which spatial relationship is preserved, and how much is supplied by the readout rather than the encoder?

Evidence boundary: Only 12 DLPFC slides carry expert labels; the other 30 use molecular clustering proxies. Repeated configurations are not independent specimens, and single-slide visual examples are not cohort means.

Edge encoders: contrastive multi-teacher compression

MuCoDi caches fixed pathology-teacher vectors and trains compact students to match tile identity contrastively. MuCoEdge is evaluated with frozen-encoder MIL and single-tile CPU timing. Teacher and student both use 5×; capacity reduction and magnification reduction must remain separate hypotheses.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Multi-Teacher Contrastive Distillation for Edge-Efficient Pathology Foundation ModelsTim Lenz, Maurice Heide, Marco Gustav, Nic G. Reitsam, and Jakob Nikolas Kather2026 · MICCAI Efficient Medical AI workshop (EMA) · peer-reviewedOfficial MICCAI satellite proceedings confirm the 2nd Efficient Medical AI workshop (EMA), not the main MICCAI track. LNCS volume 17265 is listed; DOI and pagination remain pending on the checked record. Full 11-page accepted manuscript inspected. Teachers and students use 2.0 µm/px (5×); this is capacity compression, not a demonstrated 20×→5× transfer method. Raspberry Pi timings concern individual encoder forward passes, not complete WSI processing.MuCoDi learns compact pathology encoders by multi-teacher contrastive distillation from fixed Virchow2, UNI2 and H-Optimus-1 tile embeddings. Its MuCoEdge students keep much of the teachers' downstream performance with small MobileOne or RepViT backbones. Crucially, the paper uses the same 5× sampling for teacher and student: it compresses model capacity, not a 20× evidence stream into a 5× one.Official workshop accepted manuscript · Author implementation and documented scope

Reading question: Which performance survives after reducing capacity, at the same physical sampling and downstream readout?

Evidence boundary: The 23 tasks are teacher-eligible; encoder timing excludes WSI IO and aggregation. Close average AUROC is not formal non-inferiority or cross-magnification validation.

Complete maintained paper route

Section titled “Complete maintained paper route”

72 canonical readings · 2016–2026· 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 72 readings in maintained order

  31. #312026ECCVAccepted · 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

  45. #452026ECCVAccepted · 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

  47. #472026MICCAIPeer reviewedL3

    Harnessing Adversarial Distillation to Customise Debiased, Disease-Specific Pathology Foundation Models for Breast Cancer

    Zhiwei Chen, Yang Hu, Yuxiang Xiao, Yakun Ju, Tianyang Zhang, Yingxue Xu, Wei Li, Hao Chen, Jens Rittscher, and Kaixiang Yang

    Official MICCAI 2026 proceedings record and accepted-version open-access PDF verified 2026-10-02; Springer DOI and pagination remain pending.

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

    Metadata checked

  56. #562026Medical Image AnalysisPeer reviewedL3

    Streamline Pathology Foundation Model by Cross-Magnification Distillation

    Ziyu Su, Abdul Rehman Akbar, Usama Sajjad, Anil V. Parwani, and Muhammad Khalid Khan Niazi

    Journal record verified through publisher metadata, Crossref and PubMed PMID 42874158 on 26 October 2026: electronic publication 10 October 2026, assigned to volume 115 (January 2027), article 104352. The stable ID retains the 2025 preprint year. Journal full text was unavailable; detailed scientific claims and visuals remain explicitly tied to the previously inspected arXiv v3, not asserted to match the journal version. The journal abstract describes seven tasks and encoder-level throughput, not an end-to-end speedup.

    Official artifacts: Inspected author manuscript (v3) · PubMed journal metadata and abstract · Official code · Official model

    Metadata checked

  57. #572025arXivPreprintL3

    Magnification-Aware Distillation (MAD): A Self-Supervised Framework for Unified Representation Learning in Gigapixel Whole-Slide Images

    Mahmut S. Gokmen, Mitchell A. Klusty, Peter T. Nelson, Allison M. Neltner, Sen-Ching Samson Cheung, Thomas M. Pearce, David A. Gutman, Brittany N. Dugger, Devavrat S. Bisht, Margaret E. Flanagan, and V. K. Cody Bumgardner

    Preprint; arXiv comments report submission to the AMIA 2026 Informatics Summit, not acceptance.

    Official artifacts: Official model

    Metadata checked

  58. #582026ECCVAccepted · proceedings pendingL3

    Pathryoshka: Compressing Pathology Foundation Models via Multi-Teacher Knowledge Distillation With Nested Embeddings

    Christian Grashei, Christian Brechenmacher, Rao Muhammad Umer, Jingsong Liu, Carsten Marr, Peter J. Schüffler, and Ewa Szczurek

    ECCV 2026 poster status reverified against the official complete program (poster 4454). No oral or spotlight designation is recorded. A Springer version of record was not verified on 2026-09-12; the existing review retains its declared manuscript revision.

    Official artifacts: Official ECCV record · Official model · Conference manuscript

    Metadata checked

  60. #602026ECCVAccepted · 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

  61. #612026ECCVAccepted · 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

  63. #632026arXivPreprintL3

    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

  64. #642026arXivPreprintL3

    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

  65. #652026arXivPreprintL3

    Computationally Efficient Pathology Segmentation using Knowledge Distillation from Foundation Models

    Jiaqi Lv, Yijie Zhu, Saki Okada, Maxwell Stanley Renna, Robert Goldin, Stefan Antonowicz, and Shan E Ahmed Raza

    arXiv v1 (2026-09-03), submitted to Medical Image Analysis according to the authors, not accepted. Verified methods and selected results on 2026-10-02. TIAToolbox is the stated release destination; availability of these exact weights was not independently confirmed. Patch-throughput measurements are not end-to-end WSI latency.

    Official artifacts: Versioned manuscript · TIAToolbox — stated release destination

    Metadata checked

  66. #662026arXivPreprintL3

    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

  67. #672026MICCAIPeer 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

  68. #682026MICCAIPeer reviewedL3

    Benchmarking Pathology Foundation Models for Spatial Domain Understanding

    Bokai Zhao, Yiyang Zhang, Yuanchi Zhu, Hanqing Chao, Long Bai, Tai Ma, Minfeng Xu, Ming Song, and Tianzi Jiang

    MICCAI 2026 chapter, LNCS 16895, pages 35–45, first online 25 September 2026; the volume imprint says 2027. Review uses the inspected publisher PDF. The 42 slides comprise 12 expert-annotated DLPFC slides and 30 slides assessed using transcriptomic clustering proxies. The linked public repository is a prepared-results dashboard, not a verified end-to-end benchmark implementation.

    Official artifacts: Publisher full chapter · Official MICCAI open-access record · Official accepted manuscript · Author results dashboard · Dashboard repository and documented limits

    Metadata checked

  69. #692026MICCAI Efficient Medical AI workshop (EMA)Peer reviewedL3

    Multi-Teacher Contrastive Distillation for Edge-Efficient Pathology Foundation Models

    Tim Lenz, Maurice Heide, Marco Gustav, Nic G. Reitsam, and Jakob Nikolas Kather

    Official MICCAI satellite proceedings confirm the 2nd Efficient Medical AI workshop (EMA), not the main MICCAI track. LNCS volume 17265 is listed; DOI and pagination remain pending on the checked record. Full 11-page accepted manuscript inspected. Teachers and students use 2.0 µm/px (5×); this is capacity compression, not a demonstrated 20×→5× transfer method. Raspberry Pi timings concern individual encoder forward passes, not complete WSI processing.

    Official artifacts: Official workshop accepted manuscript · Author implementation and documented scope

    Metadata checked