Articles on computational, digital, and mathematical pathology
DCTD Workshop on Foundation Models for Cancer: Advancing Diagnosis, Prognosis, and Treatment Response
Reviewed in DCTD Workshop on Foundation Models for Cancer: Advancing Diagnosis, Prognosis, and Treatment Response — 3-day NCI Division of Cancer Treatment and Diagnosis (DCTD) workshop exploring the clinical, computational, and regulatory landscape of self-supervised foundation models in oncology. Evaluates multimodal data integration (unifying gigapixel whole-slide pathology, volumetric radiology, and molecular omics into joint latent spaces with cross-attention), digital patient trajectory simulation (predicting therapeutic response and resistance), federated learning for privacy-preserving multi-institutional consortia, diagnostic quality enhancement, and the validation rigor (suppressing hospital shortcut learning and batch confounding) needed for clinical trial and clinical decision support translation (March 24–26, 2026; program lead: Asif Rizwan, Ph.D.). Proceedings archived across six sessions on NCI Vbrick Rev.
From whole-slide image to biomarker prediction: end-to-end weakly supervised deep learning in computational pathology
Reviewed in From whole-slide image to biomarker prediction: end-to-end weakly supervised deep learning in computational pathology — presents STAMP (Solid Tumor Associative Modeling in Pathology), a standardized, modular 5-stage protocol and open-source software pipeline developed by the Kather Lab (TU Dresden / RWTH Aachen) for end-to-end weakly supervised deep learning. Connects routine H&E whole-slide images to clinical endpoints (binary mutation status, multi-class subtyping, continuous regression, survival analysis, and joint multi-target prediction via the Barspoon transformer). Features automated luminance-based tissue thresholding, pyramidal tiling, reusable caching, unified feature extraction across 18+ foundation models (UNI2, Virchow2, CONCH1.5, H-Optimus, CTransPath), patient-level virtual slide concatenation, spatial attention heatmaps, class maps, top/bottom predictive patch harvesting, and native FastMCP server orchestration (mcp/server.py) for AI agents (El Nahhas et al., Nature Protocols 20:293–316, 2025; DOI: 10.1038/s41596-024-01047-2; PMID: 39285224). Code at KatherLab/STAMP; see dedicated tool note [[STAMP]].
Accelerating Data Processing and Benchmarking of AI Models for Pathology
Reviewed in Accelerating Data Processing and Benchmarking of AI Models for Pathology — presents TRIDENT (Toolkit for Large-Scale Whole-Slide Image Processing) and Patho-Bench, an open-source software ecosystem from the Mahmood Lab (Harvard Medical School / Brigham and Women's Hospital). TRIDENT standardizes high-throughput, fault-tolerant WSI processing, neural tissue segmentation (HEST, GrandQC, Otsu) with artifact/penmark cleanup, interactive QuPath GeoJSON contour exports, and embedding extraction across 33+ patch encoders (UNI, CONCH, Virchow, GigaPath, H-Optimus) and 8+ slide encoders (TITAN, PRISM, CHIEF) with asynchronous SSD caching and deadlock recovery. Patho-Bench establishes a standardized, reproducible benchmark suite across 42 clinically relevant pathology tasks spanning morphological subtyping, grading, biomarker expression, driver mutation prediction, survival, and treatment response with generalized attention pooling for multi-slide case fusion (Zhang et al., arXiv:2502.06750, 2025). Code at mahmoodlab/trident and mahmoodlab/patho-bench; see dedicated tool note [[TRIDENT]].
HERO: Histology Encoder for Robust Representation in Oncology
Reviewed in HERO: Histology Encoder for Robust Representation in Oncology — 1.1-billion-parameter ViT-G/14 foundation model from Caris Life Sciences pre-trained on a morphology-balanced corpus of 500 million tiles from approximately 575,000 WSIs across 21 organ groups and 3 scanner platforms. Solves the acquisition confounder trap through cluster-quota curation (capping frequent stroma and common carcinomas while preserving rare phenotypes), DINO/iBOT pretraining with a Kernel Density Estimator (KDE) regularizer, and high-resolution Gram-anchored dense refinement (Stage 2). Establishes state-of-the-art robustness on PathoROB (Robustness Index 0.892 vs Virchow2 0.861) and PLISM (median matched consistency 0.933, cross-stain retrieval 0.424), ranks first on average across 39 slide-level clinical tasks in Patho-Bench (0.673, leading mutation prediction, treatment response, and survival), and achieves the lowest average rank (2.61) across six public benchmark suites (Li et al., Caris Life Sciences / arXiv:2609.35943, 2026).
Standardization in digital pathology: Supplement 145 of the DICOM standards
Reviewed in Standardization in digital pathology: Supplement 145 of the DICOM standards — foundational review detailing DICOM Standards Committee Working Group 26 (WG-26) Supplement 145, establishing the vendor-neutral Whole Slide Microscopic Image Information Object Definition (IOD) and SOP Classes. Outlines the multi-resolution pyramidal tiled image storage architecture, standardized (X, Y, Z) physical slide frame of reference, Z-plane focal stacks, decoupled annotation series, and Modality Worklist (MWL) / MPPS integration with hospital PACS and VNAs (Singh, Chubb, Pantanowitz & Parwani, Journal of Pathology Informatics 2:23, 2011, DOI: 10.4103/2153-3539.80719, PMID: 21633489).
Class visualizations and activation atlases for computational pathology
Reviewed in Class visualizations and activation atlases for computational pathology — concept-level interpretability framework that adapts class visualizations and activation atlases to a frozen UNI pathology foundation model. Four pathologists showed that synthetic concepts remain recognizable for distinct colorectal tissue classes but become ambiguous as cancer labels grow morphologically overlapping: NCT Fleiss' κ fell from 0.755 on real patches to 0.313 on class visualizations, while activation atlases ranged from κ = 0.82 for coarse cancer groupings to κ = 0.11 for subclasses. The method is positioned as an audit of representation structure and taxonomy—not a diagnostic or causal explanation tool (Gustav et al., Cell Reports Medicine 7:103054, 2026, DOI: 10.1016/j.xcrm.2026.103054). Code at KatherLab/PathoActivationAtlas.
Towards robust foundation models for digital pathology
Reviewed in Towards robust foundation models for digital pathology — PathoROB, a public benchmark showing all 20 pathology foundation models tested encode the contributing hospital strongly enough to cause diagnostic failures (Kömen et al., Nature Communications 17, 5218, 2026), with code at bifold-pathomics/PathoROB.
Confirmation bias and time pressure during human–AI collaboration in computational pathology
Reviewed in When Two Wrongs Don't Make a Right: Examining Confirmation Bias and the Role of Time Pressure During Human-AI Collaboration in Computational Pathology — 28 pathologists estimated tumour cell percentage twice; when AI agreed with an initially wrong estimate, reliance on prior judgement collapsed (p=0.09) and AI advice dominated, while time pressure attenuated confirmation bias only because automation bias eclipsed it (Rosbach et al., CHI '25 / arXiv:2411.01007).
Cognitive biases in AI–assisted medical decision making: a primer for pathology
Reviewed in Cognitive biases in AI-assisted medical decision making: A structured review as a primer for veterinary and human pathology — structured review across ACM, IEEE, and PubMed identifying 12 cognitive biases in AI-assisted medicine; revealed that across all medical specialties, only one primary study came from pathology, prompting author-developed hypothetical pathology vignettes as a practical primer (Rosbach et al., Veterinary Pathology 2026, PMID 42557856).
Screening efficiency over experience: Rapid target detection in digital cytology
Reviewed in Screening efficiency over experience: Rapid target detection in low-power field as a modifiable cognitive biomarker for diagnostic accuracy in digital cytology — eye-tracking study of 100 cytotechnologists and 28 students demonstrating the "experience paradox": years of experience showed no significant correlation with diagnostic accuracy ($r = 0.189$), while shorter fixation on low-power targets ("pop-out" detection) was the sole independent predictor ($p = .045$), with trainees acquiring expert selective noise neglect within 3 months (Abe et al., Cancer Cytopathology 2026, DOI: 10.1002/cncy.70132).
A distributional robustness margin for pathology foundation models
Reviewed in A distributional robustness margin for pathology foundation models — argues PathoROB's Robustness Index is structurally unfit for cross-model comparison and replaces it with CRoMa, a per-sample signed margin (Grisi, van der Laak & Litjens, arXiv:2607.25497), with the library evaluated in CRoMa.
Weakly supervised multiple instance learning histopathological tumor segmentation
Reviewed in Weakly supervised MIL histopathological tumor segmentation — MIL tumor segmentation from slide-level labels only (Lerousseau et al., MICCAI 2020), with code and 6,481 released TCGA tumor maps at MarvinLer/tcga_segmentation.
Tumor budding T-cell graphs: assessing the need for resection in pT1 colorectal cancer patients
Reviewed in Tumor budding T-cell graphs for pT1 colorectal cancer — GNNs over tumor-bud/T-cell hotspot graphs raise the specificity of lymph-node-metastasis prediction by ~20 points over guideline stratification at equal sensitivity (Studer et al., MIDL 2023), with the pT1-HBTG dataset on Zenodo and code at digitalpathologybern/pT1-HBTG-MIDL2023.
Reporting tumor budding in colorectal cancer: ITBCC 2016 consensus recommendations
Reviewed in Recommendations for reporting tumor budding in colorectal cancer based on the International Tumor Budding Consensus Conference (ITBCC) 2016 — landmark international consensus defining tumor budding as detached single cells or clusters $\le 4$ cells, standardizing the 0.785 mm² hotspot on H&E and introducing the 3-tier grading system (Bd1–Bd3) to guide surgical escalation in pT1 and adjuvant chemotherapy in stage II CRC (Lugli et al., Modern Pathology 2017).
What AI Can and Cannot Do in Pathology
Reviewed in What AI Can and Cannot Do in Pathology — pathCast lecture by Dr. Rajendra Singh (UPenn; founder of PathPresenter), 2026-08-18. Argues the real risk is not replacement but bypass — slides shipped to commercial vendors whose predictions reach the oncologist directly — and that the answer is for departments to re-validate vendor models on their own data and own the governance layer — AI may be invisible in the workflow, but never unaudited in the record. Summarised from auto-generated captions; slides and demos not captured.
TRICARE: Deep-learning triage of 3D pathology datasets
Reviewed in TRICARE: Deep-learning triage of 3D pathology datasets — 2.5D context-aware deep learning triage for 3D open-top light-sheet microscopy, prioritizing high-risk 2D slices in prostate and Barrett's esophagus biopsies (Gao et al., Nature Biomedical Engineering 2026), with dataset at TCIA, code at alecgao066/TRICARE, and Zenodo record at zenodo.20052262.
RepLKNet: Revisiting Large Kernel Design in CNNs (Scaling Up Kernels to 31x31)
Reviewed in Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs and RepLKNet — landmark architectural framework demonstrating that scaling depthwise convolutional kernels up to $31 \times 31$ coupled with structural re-parameterization allows pure CNNs to match or surpass Vision Transformers on ImageNet, ADE20K semantic segmentation, and COCO detection. Establishes the 5 design guidelines for large-kernel CNNs, shows that large kernels provide an ultra-wide Effective Receptive Field (ERF) and high shape bias comparable to human perception, and introduces custom CUTLASS Implicit GEMM CUDA kernels. Forms the direct architectural ancestor of modern fast digital pathology backbones including UniRepLKNet and the UniRepLKNet-N visual encoder in CellQuant-Net / [[CellPrior-Net: Prior-Guided Nuclei Detection and Classification for H&E Whole-Slide Images]] (Ding et al., CVPR 2022, DOI: 10.1109/CVPR52688.2022.01167, arXiv:2203.06717). Code at DingXiaoH/RepLKNet-pytorch.
CellPrior-Net / CellQuant-Net: Prior-guided nuclei detection and classification for H&E whole-slide images
Reviewed in CellPrior-Net: Prior-Guided Nuclei Detection and Classification for H&E Whole-Slide Images and CellQuant-Net — prior-guided nuclei detection and 3-class classification (tumor, immune, other) pipeline coupling a lightweight UniRepLKNet-N large-kernel CNN with a 4-channel input tensor (RGB + hematoxylin Difference of Gaussians prior via Macenko stain deconvolution) and GPU-accelerated watershed post-processing. Benchmarked against 10 state-of-the-art pipelines across 8 multi-center datasets (~10.4M nuclei), CP-Net defines the optimal Pareto efficiency frontier, matching CellViT panoptic quality while cutting whole-slide inference time by 2x to 3x (2.16 min on 20× and 2.77 min on 40×). Integrated into CellQuant-Net with deep-learning WSI Quality Assessment (artifact filter excluding blur, folds, pen ink), spatial TIL neighborhood graph analysis, and native QuPath GeoJSON export, demonstrating significant survival stratification in hepatocellular carcinoma (Jabar et al., Journal of Pathology Informatics 2026, DOI: 10.1016/j.jpi.2026.100716, PII S2153-3539(26)00178-1; arXiv:2607.00802). Code at Falah-Jabar-Rahim/CellQuant-Net.
HistoPLUS: Comprehensive cellular characterisation of H&E slides
Reviewed in HistoPLUS: Towards Comprehensive Cellular Characterisation of H&E Slides — cell detection, boundary segmentation, and 13-class classification framework integrating the compact distilled Bioptimus H0-mini pathology foundation model (86M params) within a 3-branch CellViT architecture. Trained via active learning on HistoTRAIN (108,722 nuclei, 6 indications) with NuClick boundary propagation and validated across 6 MOSAIC cohorts (HistoVAL: 69,108 consensus nuclei from multi-pathologist review). Outperforms current state-of-the-art models by +5.2% in detection quality and +23.7% in classification F1 while matching ViT-Huge models with 5× fewer parameters, unlocking 7 understudied cell populations, predicting FGFR3 mutation status in bladder cancer, and generalizing zero-shot to unseen breast and ovarian cancers (Adjadj et al., Journal of Pathology Informatics 2026, DOI: 10.1016/j.jpi.2026.100696, PII S2153-3539(26)00156-2; arXiv:2508.09926). Code at owkin/histoplus, weights on Hugging Face.
CytoFormer: Molecularly supervised cell foundation model for histopathology cell classification
Reviewed in CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification — replaces manual pathologist cell annotation with molecular ground truth derived from paired in situ spatial transcriptomics (81 Xenium sections, 15.4M cells, 16 organs, 23 cell types). Leverages a ViT-giant backbone with 16 per-organ linear routing heads (84.6% accuracy, 0.78 macro-F1), outperforming 6 pathology foundation models on public benchmarks (PanNuke, CoNSeP, MoNuSAC, PUMA), transferring zero-shot to unseen organs, and achieving label-efficient active learning (F1 0.82) on TissueLab (Yao, Li, Yu & Huang, Precision Pathology 2026, DOI: 10.1016/j.prpath.2026.100006, PII S3117-678X(26)00006-5). Code at zhihuanglab/CytoFormer, weights on Hugging Face, and interactive WSI viewer at zhihuanglab.github.io/CytoFormer.
ConvMixerSSM: Hybrid convolution and state-space model for WSI cancer subtyping
Reviewed in A Hybrid MIL Approach Leveraging Convolution and State-Space Model for Whole-Slide Image Cancer Subtyping — hybrid weakly supervised MIL architecture combining depthwise separable convolutions (ConvMixer) for local tissue textures with linear state-space sequence modeling (Mamba/SSM) and ReLU-gated attention for sparse instance selection; sets top performance on TCGA-NSCLC (AUC 97.83%, ACC 91.82%, F1 91.18%) and CAMELYON16 (AUC 98.95%) with 1.65 ms inference latency per gigapixel slide (Bi & Zhang, Mathematics 2025, DOI: 10.3390/math13132178).
The pathology report as a boundary object: From clinical communication to computational representation
Reviewed in The pathology report as a boundary object: From clinical communication to computational representation — foundational theoretical framework examining why pathology reports are contextually complete for expert clinical readers yet semantically underdetermined for secondary computational reuse (AI, registries, biobanks, foundation models). Analyzes the pathology report through Star & Griesemer's boundary object theory, articulates the mechanisms of selective semantic compression, differentiates latent from absent information, illustrates why naive NLP co-occurrence fails, and introduces Ontology-Separated Pathology Representation (OSPR) to decouple biological state, procedural context, and observational findings (Wang, Precision Pathology 2026, DOI: 10.1016/j.prpath.2026.100002).
NPIC Quality Coordination Centre: Digital Pathology Quality Assurance and Metrology
Reviewed in NPIC Quality Coordination Centre: Digital Pathology Quality Assurance and Metrology — full-lifecycle quality assurance framework and physical metrology standards developed by the UK National Pathology Imaging Co-operative (NPIC) and Leeds Teaching Hospitals NHS Trust. Covers objective chemical stain quantification via biopolymer Tango slides (National Staining Survey, Dunn et al., Diagnostic Pathology 2024), WSI scanner variation benchmarking (Pye et al., 2022), display and viewing luminance science (arXiv:2312.00475), the web-based Point-of-Use Quality Assurance (POUQA) tool, and the foundational Leeds Guides to Digital Pathology.
OpenFlexure: Robotic microscopy for everyone
Reviewed in [[Robotic microscopy for everyone: the OpenFlexure Microscope]] and OpenFlexure Microscope — foundational open-hardware and robotic digital microscopy framework developed by Dr. Richard Bowman, Julian Stirling, and colleagues across the University of Bath and Cambridge. Replaces precision-machined sliding translation stages with a monolithic 3D-printed flexure stage capable of sub-100 nm positioning resolution. Integrates modular RMS DIN finite/infinity optics, Sangaboard motor control, a Python REST API / WebSocket daemon (openflexure-microscope-server), automated 2D tiled whole-slide scanning, and live streaming for telepathology. Validated in field trials for automated malaria screening in Tanzania, decentralized whole-slide imaging (WSI), and remote telepathology consultations (Collins et al., Biomedical Optics Express 2020, DOI: 10.1364/BOE.385729; Sharkey et al., Review of Scientific Instruments 2016, DOI: 10.1063/1.4941068). Repositories at openflexure/openflexure-microscope and rwb27/openflexure_microscope.
M-Optimus-1: In silico spatial biomarker discovery in ovarian cancer
Reviewed in [[From Model to Patient: M-Optimus-1 for Spatial Biomarker Discovery in Ovarian Cancer]] and M-Optimus — multimodal foundation biological world model developed by Bioptimus (Paris, France) that infers spatially resolved gene expression profiles across 6,000+ genes directly from routine H&E whole-slide images (with or without paired bulk RNA-seq). Evaluated against physical 10x Genomics Xenium Prime ground truth ($r = 0.81$) on unseen stage III-B ovarian papillary serous carcinoma, and deployed across an archival retrospective cohort of 288 WSIs (TCIA) to demonstrate that spatial colocalization between VEGFA+ tumor cells and cancer-associated fibroblasts (CAFs) serves as a statistically significant spatial biomarker of bevacizumab anti-angiogenic resistance ($p = 0.003$). Unsupervised spatial niche clustering correlated with measured Xenium cellular compositions at $r = 0.899$, demonstrating how foundation models transform archival digital pathology slides into high-dimensional discovery search spaces (Bioptimus, 2026). Platform at bioptimus.com/m-optimus.
From Samples to Knowledge 2025: QuPath Training Course
Reviewed in From Samples to Knowledge 2025: QuPath Training Course — 14-session practical curriculum from the La Jolla Institute for Immunology (Mikulski & McArdle) establishing QuPath v0.6.0+ best practices for 18-plex RareCyte Orion immunofluorescence, deep learning boundary segmentation (InstanSeg / DJL), headless Groovy batch scripting, composite object phenotyping, multimodal H&E registration (Warpy), spatial proximity transforms, and Paquo/QuBylab Python connectivity.
The benefits of building and working with interactive simulations Interactive simulations for better model intuition
Survival Prediction in Pancreatic Ductal Adenocarcinoma by Quantitative Computed Tomography Image Analysis
Computational, Digital & Mathematical Pathology
CompPath Dr. Shikhar Uttam, Ph.D. Lecture (April 14th 2017)
https://www.youtube.com/watch?v=6NFyB19D_fU&list=PLwdWByS9hJ3IpNyN1Ge4dbfSV43296jLn&index=2
Introduction to Computational and Systems Biology
Local Alignment (BLAST) and Statistics
https://www.youtube.com/watch?time_continue=6&v=6Udqou3vmng
https://www.youtube.com/results?search_query=Philips+healthcare+pathology
Biologists would love to program cells as if they were computer chips
Using the principles of evolution to treat and prevent cancer
https://www.statnews.com/2018/06/27/cancer-treatment-prevention-evolution/
Last updated