Skip to content
Computational Pathology Reading Roadmap

Cell, gland, and tissue modeling

Dense prediction turns histology into explicit objects: tissue compartments, tumour regions, glands, cells, nuclei, boundaries, and phenotypes. These tasks are valuable both as clinical endpoints and as intermediate representations for interpretable pathomics, graph models, spatial biology, and quality control.

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

Selected ICLR oral addition checked 19 September 2026; the broader literature-review boundary is unchanged.

This route is specific to histology and WSI-derived regions. General segmentation architecture prerequisites are in the computer-vision route; dataset access belongs in the WSI Data Atlas.

1. Define the object before choosing a model

Section titled “1. Define the object before choosing a model”

“Segmentation” can refer to different prediction contracts.

Task Output Typical reference standard Useful metrics
Tissue/semantic segmentation one class per pixel polygon or pixel mask class Dice/IoU, boundary score
Detection point or box per object centres or boxes precision/recall/F1 under a distance or IoU rule, FROC
Instance segmentation separate mask per object instance masks panoptic quality (PQ), aggregated Jaccard, object Dice
Instance classification type per detected instance typed masks/points per-class F1, detection-quality × classification-quality
Panoptic segmentation semantic “stuff” plus object “things” joint labels PQ and class-stratified errors
Counting/density count or spatial density objects or stereology MAE, correlation, calibration by region

These outputs are not interchangeable. A high semantic Dice can merge touching nuclei; a good detector can localize objects without recovering morphology; and a count can be accurate while cell identities are wrong.

Annotation is a measurement process

Section titled “Annotation is a measurement process”

Before training, document:

  • tissue, stain, magnification/MPP, and field-of-view selection;
  • whether labels came from H&E alone, IHC/IF restaining, consensus, or one annotator;
  • annotation instructions for overlap, truncation, mitosis, necrosis, and uncertain type;
  • instance inclusion at image borders;
  • inter- and intra-observer agreement;
  • correction of automatically proposed masks; and
  • patient/site separation across all patches.

Use reference standard, not “ground truth,” when labels contain expert judgment or measurement error. Preserve annotator IDs and uncertainty where governance permits it.

2. Tissue and gland segmentation

Section titled “2. Tissue and gland segmentation”

Tissue-scale models identify tumour, stroma, epithelium, necrosis, vessels, glands, or other compartments. They supply interpretable areas, define regions for cell analysis, and can constrain slide sampling.

The architectural baseline is U-Net (2015): an encoder–decoder with skip connections that joins context and localization. For pathology, the deeper lesson is to separate three scales:

  1. local boundary evidence at cell or gland resolution;
  2. regional architecture that distinguishes similar textures; and
  3. slide context that determines what a compartment means.

Gland masks are instances with topology, not merely foreground pixels. Read the GlaS challenge report to see how object-level and boundary errors complicate evaluation. For cell–tissue context, OCELOT provides overlapping local cell and larger-field tissue annotations.

3. Nuclei and cell instances

Section titled “3. Nuclei and cell instances”

Crowded nuclei create a recurring problem: foreground pixels are easier to predict than the boundary between touching objects. Major solution families include:

  • semantic masks followed by watershed or contour post-processing;
  • centre/offset, boundary, or horizontal–vertical distance fields;
  • star-convex or other parametric object shapes;
  • box/mask detection frameworks;
  • promptable interactive segmentation; and
  • foundation encoders with lightweight task-specific decoders.

Model lineage

Section titled “Model lineage”
Primary paper Year / status Contribution to understand Main caution
U-Net 2015 · MICCAI dense encoder–decoder and skip connections semantic foreground does not separate instances by itself
Cell Detection with Star-Convex Polygons (StarDist) 2018 · MICCAI object-centred star-convex shape representation violated for strongly non-star-convex objects
HoVer-Net 2020 · Medical Image Analysis joint nuclear pixels, horizontal/vertical maps, and type prediction post-processing and class taxonomy remain dataset-specific
Cellpose 2021 · Nature Methods generalist flow-field segmentation across microscopy broad microscopy generality does not guarantee H&E phenotyping
CellViT 2024 · Medical Image Analysis ViT/foundation encoder in a HoVer-style instance framework encoder advantage must be separated from training and decoder choices
Segment Anything for Microscopy (μSAM) 2025 · Nature Methods promptable and automatic segmentation across microscopy domains broad microscopy performance is not proof of pathology phenotype validity
PathoSAM 2025 · MIDL oral automatic and interactive SAM adaptation for histopathology nuclei prompt protocol and domain transfer are part of the comparison
CellSAM 2025 · Nature Methods generalist cell segmentation and detection across microscopy detection, mask quality, and phenotype transfer require separate pathology validation
CellViT++ 2026 · CMPB decoupled segmentation and data-efficient adaptation to new cell taxonomies zero-shot masks do not eliminate local phenotype validation
Revisiting Foundation Models for Cell Instance Segmentation 2026 · MIDL, peer reviewed controlled comparison of Cellpose-SAM, CellSAM, μSAM, and SAM-family variants a benchmark ranking does not establish one universal winner across pathology sites and phenotypes
Pathology foundation models as detection encoders for mitotic figures 2026 · arXiv preprint tests frozen PFM features inside a dense mitosis detector, including domain/species transfer; author code detector, feature level, object density, annotation/matching, and encoder cannot be collapsed into one FM score
CellDETR 2026 · arXiv preprint uses detection-guided pretraining to learn scalable cell representations from histopathology; author code the claimed transfer depends on detector targets, tissue distribution, annotation availability, and the currently non-peer-reviewed evidence

Do not rank these methods using one aggregate number copied from different papers. Compare them with the same resolution, annotation policy, class mapping, post-processing, test patients, and inference budget.

4. Dataset and benchmark lineage

Section titled “4. Dataset and benchmark lineage”

Benchmark papers teach as much about label construction as about algorithms.

Reading Year / status Scale and scientific lesson
GlaS challenge 2017 · Medical Image Analysis colon glands; object shape and malignant architecture make pixel-only evaluation insufficient
MoNuSeg challenge 2020 · IEEE TMI multi-organ nuclei; tests cross-organ generalization and instance separation
PanNuke extension, insights, and baselines 2020 · dataset paper/preprint about 200,000 typed nuclei across 19 tissues; broad taxonomy with semi-automatic annotation
MoNuSAC2020 2021 · IEEE TMI multi-organ segmentation plus four cell types and challenge-based comparison
Lizard 2021 · ICCV workshop nearly half a million colon nuclei assembled with pathologist-in-the-loop refinement
SegPath 2023 · Cell Reports Methods H&E labels derived through destain/restain and IF for eight cell types; supervision can come from another stain
OCELOT 2023 · CVPR overlapping cell and tissue fields across organs; formalizes context-aware cell detection
CoNIC challenge 2024 · Medical Image Analysis colon nuclei segmentation, classification, and counting with robustness analysis
PathoCellBench 2025 · MICCAI, peer reviewed 14 phenotypes and 88M cells; audit registered-multiplex-derived labels, automated processing, and single-center colon domain shift

Before pooling datasets, reconcile cell taxonomies. “Epithelial,” “neoplastic,” “connective,” and “stromal” are not universal synonyms. Create a mapping with exact, broader, ambiguous, and not represented states instead of silently forcing labels together.

5. Annotation-efficient and interactive learning

Section titled “5. Annotation-efficient and interactive learning”

Dense labels are expensive, so annotation tooling is part of the methodology.

  • NuClick (2020) studies one-click or squiggle-guided object annotation.
  • PathoSAM supports automatic and interactive use, but a fair study must count clicks, corrections, latency, and failed objects—not only final mask quality.
  • SegPath uses registered stain information to create labels that are difficult to infer consistently from H&E alone.
  • Pseudo-labeling and active learning can reduce expert effort, but training and test reference standards must remain independent.

Measure expert time per accepted object, disagreement after correction, and the phenotypes underrepresented among rejected suggestions. Automation can bias the final dataset toward objects the proposal model already handles well.

6. Evaluation that survives crowded tissue

Section titled “6. Evaluation that survives crowded tissue”

Match predictions to references once

Section titled “Match predictions to references once”

For instance tasks, define:

  1. the spatial match rule (IoU or centroid distance);
  2. one-to-one assignment procedure;
  3. treatment of split and merge errors;
  4. border-instance policy; and
  5. whether classification is evaluated only on matched detections.

Report detection quality separately from segmentation quality and phenotype accuracy. PQ is helpful because it combines recognition and mask overlap, but include per-class components so rare cells do not disappear inside a mean.

Evaluate at three levels

Section titled “Evaluate at three levels”
Level Question
Object was each nucleus/cell/gland found, separated, shaped, and typed correctly?
Region are densities, ratios, compartment areas, and spatial arrangements reliable?
Patient/slide does the derived biomarker reproduce across blocks, slides, sites, and scanners?

Downstream correlation does not validate individual masks, and excellent patch masks do not guarantee an unbiased patient-level biomarker.

Stress tests

Section titled “Stress tests”

Stratify errors by:

  • organ and histologic subtype;
  • stain, laboratory, scanner, and MPP;
  • cell density, size, overlap, and necrosis;
  • tumour, invasive margin, stroma, and normal compartment;
  • blur, folds, compression, and stain outliers; and
  • rare phenotype and ambiguous annotation.

7. From objects to pathomics

Section titled “7. From objects to pathomics”

Once instances are validated, build interpretable features in registered coordinates:

  • morphology: area, eccentricity, contour, texture, chromatin;
  • composition: type counts, proportions, and uncertainty;
  • neighbourhood: nearest-neighbour distances, mixing, clustering, exclusion;
  • architecture: gland shape, compartment adjacency, tumour–stroma interface; and
  • multiscale summaries: region, slide, and patient distributions.

Avoid thousands of correlated features with a small patient cohort. Define features from a biological question, stabilize them across segmentation perturbations, fit selection inside the training fold, and retain uncertainty from detection and typing. For learned spatial graphs and molecular outcomes, continue to spatial, molecular, and outcome modeling.

Key readings for translating validated objects into biological or patient-level measurements:

Primary paper Year / status Pathomics lesson
Next-Generation Morphometry 2023 · Nature Communications, peer reviewed bridge validated segmented structures to interpretable morphology features and patient-level analysis
HistoCell and sCellST 2025 / 2026 · Nature Communications, peer reviewed connect nuclei, cell states, topology, and predicted single-cell expression while carrying segmentation/deconvolution uncertainty forward
Colon histomorphology transferred to AVANT material 2025 · Nature Communications, peer reviewed 47 learned morphology clusters and external trial-material analysis; retrospective trial material is not an AI-guided treatment trial
Section titled “8. Recommended research reading sequence”

L1 — dense prediction foundations

Section titled “L1 — dense prediction foundations”
  1. U-Net (2015)
  2. GlaS (2017)
  3. StarDist (2018)

Checkpoint: explain semantic, instance, and panoptic segmentation and draw the failure produced by merging two touching nuclei.

L2 — pathology-specific instances and labels

Section titled “L2 — pathology-specific instances and labels”
  1. HoVer-Net (2020)
  2. MoNuSeg (2020)
  3. PanNuke (2020)
  4. NuClick (2020)
  5. Lizard (2021)
  6. MoNuSAC2020 (2021)

Checkpoint: reproduce one matching metric and manually categorize 100 errors as miss, duplicate, split, merge, boundary, or type error.

L3 — context, foundation models, and transfer

Section titled “L3 — context, foundation models, and transfer”
  1. OCELOT (2023)
  2. SegPath (2023)
  3. CoNIC (2024)
  4. CellViT (2024)
  5. PathoSAM (2025)
  6. CellViT++ (2026)
  7. PFMs as mitotic-figure detection encoders (2026)
  8. MSGR: Gene Ontology-guided spatial-expression prediction (2026)

Checkpoint: test one frozen/adapted model on an unseen site, then determine whether the failure is localization, mask shape, phenotype mapping, or acquisition shift.

9. Minimum experiment record

Section titled “9. Minimum experiment record”
unit: pixel | object | region | slide | patient
specimen_and_stain: exact definition
input_mpp: value
field_selection: sampling policy
reference_standard: annotators, protocol, source stain, uncertainty
split_unit: patient
label_taxonomy_version: immutable mapping
model_and_weights: exact release
postprocessing: thresholds and parameters
matching_rule: IoU/distance and assignment
metrics: object + class + region/patient
external_sites: list
failure_review: sampling and categories

The strongest dense-prediction paper is not the one with the longest model name. It is the one whose object definition, labels, matching rule, domain shift, and downstream measurement remain coherent.

Diffusion-boundary self-distillation and negative transfer

TRACE turns trajectory-derived attention boundaries into one-pass predictions through LoRA and an edge decoder. Read its natural-image gains alongside the reported MoNuSeg/TNBC degradation.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
TRACE: Your Diffusion Model Is Secretly an Instance Edge DetectorSanghyun Jo, Ziseok Lee, Wooyeol Lee, Jonghyun Choi, Jaesik Park, Kyungsu Kim2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. Diffusion-boundary self-distillation improves selected natural-image segmentation but degrades MoNuSeg/TNBC cell segmentation. One-pass speed is edge extraction only, with the large backbone retained. No WSI or cross-magnification validation; official code/checkpoints remain promised rather than released at the checked revision.TRACE extracts instance-boundary cues from self-attention inside a pretrained diffusion model, then self-distills a slower trajectory-based edge procedure into one backbone pass. Natural-image mask refinement improves, but the paper also reports worse cell segmentation on MoNuSeg and TNBC.OpenReview · Official project · Official repository — release pending

Reading question: Does the teacher prior delineate the target objects before it is distilled, and what remains in the deployed model?

Evidence boundary: A tiny decoder does not remove the large diffusion backbone. Edge-extraction speed is not end-to-end speed, and negative cell-level transfer is not WSI validation.

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
PAGET: Hierarchical Multi-Teacher Knowledge Distillation for Comprehensive Tumor Microenvironment SegmentationDaisuke Komura, Maki Takao, Mieko Ochi, Takumi Onoyama, Hiroto Katoh, Hiroyuki Abe, Hiroyuki Sano, Teppei Konishi, Toshio Kumasaka, Tomoyuki Yokose, Yohei Miyagi, Tetsuo Ushiku, and Shumpei Ishikawa2026 · MIDL · peer-reviewedPeer-reviewed MIDL 2026 proceedings, PMLR 315:256–279; official record verified 2026-10-02.PAGET compresses specialized pathology teachers into a SegFormer student that segments 13 tumor-microenvironment components plus background. Its main intervention is taxonomy-aware pseudo-label construction: resolve tissue context, major cell categories, and fine immune subtypes in an explicit order before teaching the student. This is heterogeneous-teacher hard-label distillation, not averaging interchangeable classifiers or matching foundation-model feature vectors.Official proceedings PDF · Official implementation
ConStruct: Structural Distillation of Foundation Models for Prototype-Based Weakly Supervised Histopathology SegmentationKhang Le, Ha Thach, Anh M. Vu, Trang T. K. Vo, Han H. Huynh, David Yang, Minh H. N. Le, Thanh-Huy Nguyen, Akash Awasthi, Chandra Mohan, Zhu Han, and Hien Van Nguyen2026 · MIDL · peer-reviewedPeer-reviewed MIDL 2026 proceedings, PMLR 315:2893–2905; official record verified 2026-10-02.ConStruct adds spatial organization to a pathology vision–language model for weakly supervised tissue segmentation. A frozen SegFormer teacher supplies token-to-token relations; trainable adapters refine frozen CONCH image features, and text-initialized prototypes turn the refined features into class activation maps. Training needs image-level tissue labels rather than dense pixel masks. The reported experiment is on BCSS-WSSS patches, not a whole-slide classification or cross-magnification benchmark.Official proceedings PDF · Official implementation

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.

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.

Complete maintained paper route

Section titled “Complete maintained paper route”

47 canonical readings · 2015–2026· 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 47 readings in maintained order

  20. #202023IEEE Journal of Biomedical and Health InformaticsPeer reviewedL3

    Deep Multi-Magnification Similarity Learning for Histopathological Image Classification

    Songhui Diao, Weiren Luo, Jiaxin Hou, Ricardo Lambo, Hamas A. AL-kuhali, Hanqing Zhao, Yinli Tian, Yaoqin Xie, Nazar Zaki, and Wenjian Qin

    Peer-reviewed journal article 27(3):1535–1545. DMSL learns paired same-field-of-view patch features at two magnifications and retains both magnification branches at inference; its reported comparison metrics are patch-level, not a WSI endpoint.

    Official artifacts: PubMed record

    Metadata checked

  26. #262025Nature MethodsPeer reviewedL2

    Segment Anything for Microscopy

    Anwai Archit et al.

    Published online in 2024 and assigned to the 2025 volume.

    Metadata checked

  33. #332025Scientific ReportsPeer reviewedL3

    Human Visual Attention-Inspired Knowledge Distillation Underlying Interpretable Computational Pathology

    Muzhou Yu, Zihan Zhong, Xingang Zhou, Yuekun Wang, Tingyu Liang, Jiamin Chen, Hongmin Huang, Junhan Zhou, Dachun Zhao, Bo Lei, Yu Wang, Wenbin Ma, and Kaisheng Ma

    Peer-reviewed open-access journal article 15:42124. Low- and high-magnification language motivates the design, but HVisKD transfers sample- and region-relation features between teacher and student given the same pathology patches; it is not cross-magnification distillation.

    Official artifacts: Open-access article · Official code

    Metadata checked

  41. #412026Scientific ReportsPeer reviewedL3

    KDTViT Knowledge Distillation, Transfer Learning and Transformer Based Deep Learning Framework for Efficient Histopathology Image Classification

    Aarv Mankodi, Praveen Kumar Shukla, and Hitesh Tekchandani

    Peer-reviewed open-access journal article 16:20721. A MobileViT-XS is sequentially adapted from 40× through 400× using the previous checkpoint as teacher; BreakHis consists of biopsy images rather than WSIs, and the same-capacity teacher/student design targets retention rather than model compression.

    Official artifacts: Open-access article

    Metadata checked

  42. #422026arXivPreprintL3

    One Model to Magnify Them All: Efficient Scale-Invariant Histopathology via Conditional Normalization and Continuous Magnification Training

    Agnieszka Florkowska, Henning Müller, and Marek Wodzinski

    arXiv v1 submitted 10 August 2026; no peer-reviewed venue was verified on 2026-09-02. A single CNN is conditioned on physical pixel size and trained over a continuous magnification range; each prediction still consumes one patch scale and the method is neither MIL nor knowledge distillation.

    Official artifacts: Official code

    Metadata checked

  43. #432026MIDLPeer reviewedL3

    PAGET: Hierarchical Multi-Teacher Knowledge Distillation for Comprehensive Tumor Microenvironment Segmentation

    Daisuke Komura, Maki Takao, Mieko Ochi, Takumi Onoyama, Hiroto Katoh, Hiroyuki Abe, Hiroyuki Sano, Teppei Konishi, Toshio Kumasaka, Tomoyuki Yokose, Yohei Miyagi, Tetsuo Ushiku, and Shumpei Ishikawa

    Peer-reviewed MIDL 2026 proceedings, PMLR 315:256–279; official record verified 2026-10-02.

    Official artifacts: Official proceedings PDF · Official implementation

    Metadata checked

  44. #442026MIDLPeer reviewedL3

    ConStruct: Structural Distillation of Foundation Models for Prototype-Based Weakly Supervised Histopathology Segmentation

    Khang Le, Ha Thach, Anh M. Vu, Trang T. K. Vo, Han H. Huynh, David Yang, Minh H. N. Le, Thanh-Huy Nguyen, Akash Awasthi, Chandra Mohan, Zhu Han, and Hien Van Nguyen

    Peer-reviewed MIDL 2026 proceedings, PMLR 315:2893–2905; official record verified 2026-10-02.

    Official artifacts: Official proceedings PDF · Official implementation

    Metadata checked

  45. #452026arXivPreprintL3

    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

  46. #462026MICCAIPeer 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

  47. #472026ICLRPeer reviewedL3

    TRACE: Your Diffusion Model Is Secretly an Instance Edge Detector

    Sanghyun Jo, Ziseok Lee, Wooyeol Lee, Jonghyun Choi, Jaesik Park, Kyungsu Kim

    Official ICLR 2026 oral and full proceedings manuscript checked. Diffusion-boundary self-distillation improves selected natural-image segmentation but degrades MoNuSeg/TNBC cell segmentation. One-pass speed is edge extraction only, with the large backbone retained. No WSI or cross-magnification validation; official code/checkpoints remain promised rather than released at the checked revision.

    Official artifacts: OpenReview · Official project · Official repository — release pending

    Metadata checked