Digital Pathology Software
Cytario
Open-core, cloud-native Image Management System (IMS) and web-based viewer for digital pathology and spatial biology. Built on React 19, Viv (Harvard Medical School HIDIVE Lab), deck.gl, and DuckDB-WASM/Apache Arrow to stream petabyte-scale OME-TIFF, OME-Zarr, and Parquet data directly from S3-compatible storage with Keycloak multi-tenancy — see dedicated tool note: Cytario.
GitHub: cytario/cytario-web — AGPL-3.0
Platform Portal: cytario.com
Cytomine
Open-source, web-based platform for multi-gigapixel whole-slide image management, collaborative multi-observer annotation (PostGIS-backed spatial geometries), and distributed algorithmic analysis. See dedicated tool note: Cytomine.
GitHub: cytomine/cytomine
Academic Portal: uliege.cytomine.org
Documentation: doc.uliege.cytomine.org
TIA Centre Cytomine Apps: TissueImageAnalytics/cytomine-app — Wraps TIAToolbox models (
cytomine-hovernet,cytomine-kongnet,cytomine-interactive-segmentation-nuclick) into Cytomine Docker apps.
NuClick
Interactive deep-learning framework for prompted segmentation of microscopic objects (nuclei, cells, and glands) from user point clicks (Jahanifar, Koohbanani, Tajeddin & Rajpoot, Medical Image Analysis 2020). Utilizes a 5-channel tensor (RGB + target inclusion map + neighbour exclusion map) with dedicated NuClick CNN and U-Net models (Dice 0.874+) to eliminate manual boundary drawing. Serves as the annotation and consensus foundation for Owkin's [[HistoPLUS: Towards Comprehensive Cellular Characterisation of H&E Slides]] — see dedicated tool note: NuClick.
PyTorch GitHub: mostafajahanifar/nuclick_torch — CC BY-NC-SA 4.0
Cytomine App: Included in TissueImageAnalytics/cytomine-app (
cytomine-interactive-segmentation-nuclick)
CellQuant-Net
End-to-end WSI nuclei quantification pipeline integrating deep-learning artifact quality assessment (WSI-QA), prior-guided nuclei detection and 3-class classification (CellPrior-Net / CP-Net), and spatial cell neighborhood network analysis (Jabar et al., Journal of Pathology Informatics 2026). Combines a UniRepLKNet-N large-kernel CNN with a 4-channel input (RGB + hematoxylin Difference of Gaussians prior) and GPU watershed post-processing, matching transformer accuracy while running 2x to 3x faster than CellViT. Exports directly to QuPath GeoJSON and computes tumor-infiltrating lymphocyte (TIL) prognostic density metrics — see dedicated tool note: CellQuant-Net and literature review [[CellPrior-Net: Prior-Guided Nuclei Detection and Classification for H&E Whole-Slide Images]].
GitHub: Falah-Jabar-Rahim/CellQuant-Net
Video Tutorial: YouTube
RepLKNet
Foundational large-kernel convolutional architecture and PyTorch framework scaling depthwise convolutions up to 31×31 via structural re-parameterization and custom CUDA Implicit GEMM acceleration (Ding et al., CVPR 2022). By establishing that large kernels deliver an ultra-wide Effective Receptive Field (ERF) and high shape bias matching Vision Transformers, RepLKNet laid the theoretical and architectural foundation for next-generation, high-speed computational pathology encoders—including UniRepLKNet and the UniRepLKNet-N backbone powering [[CellQuant-Net]] / [[CellPrior-Net: Prior-Guided Nuclei Detection and Classification for H&E Whole-Slide Images]]—enabling whole-slide image analysis at 2x to 3x the speed of CellViT. See dedicated tool note: RepLKNet and literature review [[Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs]].
PyTorch GitHub: DingXiaoH/RepLKNet-pytorch — Apache-2.0
ePMA.start – universal whole slide image viewer for digital pathology An end-user viewer and tile server in one convenient package****
Micro-Manager
Open-source microscope control and acquisition automation, integrated with ImageJ — see Micro-Manager.
OpenFlexure Microscope
Open-source, 3D-printable robotic digital microscope featuring sub-100 nm positioning precision via monolithic compliant flexure mechanics, RMS DIN finite/infinity optics, Sangaboard motor control, Python REST API, automated whole-slide scanning, and validated applications in parasitology and telepathology — see dedicated tool note: OpenFlexure Microscope and literature review [[Robotic microscopy for everyone: the OpenFlexure Microscope]].
Website: openflexure.org/projects/microscope
GitHub: rwb27/openflexure_microscope | GitLab: openflexure/openflexure-microscope
Paper: Collins et al., Biomedical Optics Express 2020 (DOI: 10.1364/BOE.385729)
M-Optimus
Multimodal biological world model and digital pathology foundation platform developed by Bioptimus (Paris, France). Connects routine Hematoxylin and Eosin (H&E) whole-slide histology with molecular genetics (bulk RNA-seq and spatial transcriptomics) to perform in silico spatial transcriptomics—reconstructing spatial expression maps for 6,000+ genes directly from H&E whole-slide images alone or paired with bulk RNA-seq. Validated against 10x Genomics Xenium Prime ($r = 0.81$) and deployed on archival clinical cohorts (TCIA) to identify spatial microenvironmental biomarkers of targeted therapy resistance — see dedicated tool note: M-Optimus and literature review [[From Model to Patient: M-Optimus-1 for Spatial Biomarker Discovery in Ovarian Cancer]].
Website: bioptimus.com/m-optimus
Technical Launch: bioptimus.com/introducing-m-optimus
STELA Data Consortium: bioptimus.com/stela
PathoActivationAtlas
Open-source PyTorch and Captum-based interpretability framework developed by the Kather Lab (Else Kröner Fresenius Center for Digital Health, TU Dresden) to audit what morphological concepts are organized inside pathology foundation models (Gustav et al., Cell Reports Medicine 2026). Generates Fourier-parameterized class visualizations (CVs) and layer-wise activation atlases (AAs) with t-SNE grid aggregation, and includes an interactive desktop GUI viewer and blinded pathologist annotation software — see dedicated tool note: PathoActivationAtlas and literature review [[Class visualizations and activation atlases for computational pathology]].
GitHub: KatherLab/PathoActivationAtlas — MIT License (Captum components under BSD-3-Clause)
Paper: Gustav et al., Cell Reports Medicine 2026 (DOI: 10.1016/j.xcrm.2026.103054)
TRIDENT
Industrial-grade, open-source Python toolkit developed by the Mahmood Lab (Harvard Medical School / Brigham and Women's Hospital) for large-scale whole-slide image processing, tissue segmentation, and feature extraction across 33+ patch encoders (UNI, CONCH, Virchow, GigaPath, H-Optimus) and 8+ slide encoders (TITAN, PRISM, GigaPath, CHIEF). Features deep neural segmentation (HEST, GrandQC), multi-GPU sharding, QuPath GeoJSON contour exports, asynchronous SSD caching, crash-resilient deadlock recovery, and integration with the 42-task Patho-Bench suite — see dedicated tool note: TRIDENT and literature review [[Accelerating Data Processing and Benchmarking of AI Models for Pathology]].
GitHub: mahmoodlab/TRIDENT
Documentation: trident-docs.readthedocs.io
Paper: Zhang et al., arXiv 2025 (DOI: 10.48550/arXiv.2502.06750)
STAMP
Solid Tumor Associative Modeling in Pathology, an end-to-end weakly supervised deep learning pipeline developed by the Kather Lab (Else Kröner Fresenius Center for Digital Health, TU Dresden; RWTH Aachen; University Medical Center Mainz) for clinical biomarker prediction, multi-target classification, continuous regression, and survival analysis directly from whole-slide images. Features standardized preprocessing across 18+ foundation model extractors (UNI2, Virchow2, CONCH1.5, H-Optimus, CTransPath), patient-level virtual slide concatenation, slide contextualizers (TITAN, PRISM, GigaPath, COBRA2, EAGLE, MADELEINE, CHIEF), multiple MIL architectures (ViT, TransMIL, Barspoon), attention heatmaps, class maps, top/bottom predictive tile extraction, and native Model Context Protocol (FastMCP) server orchestration (mcp/server.py) paired with the STAMP-Workbench web UI — see dedicated tool note: STAMP and published protocol review [[From whole-slide image to biomarker prediction: end-to-end weakly supervised deep learning in computational pathology]].
GitHub: KatherLab/STAMP — MIT License
Workbench UI: KatherLab/STAMP-Workbench
Paper: El Nahhas et al., Nature Protocols 20, 293–316 (2025) (DOI: 10.1038/s41596-024-01047-2, PMID: 39285224)
Celldega
Open-source Python + JavaScript library from the Broad Institute (Platform Innovation Lab / Spatial Technology Platform) for scalable, interactive visualization and analysis of spatial-omics and single-cell data.
Paper: Fernandez N, Ishar J, Wang H, Ben Saad A, Lipinski M, Farhi SL. Celldega: Integrated Toolkit for Visualization and Analysis of Spatial Data. bioRxiv preprint, posted 2026-08-21 (DOI: 10.64898/2026.08.13.744672).
The Challenge in Spatial Pathology
Spatial-transcriptomics and high-plex spatial proteomics pair high-dimensional molecular measurements with high-resolution microscopy to characterize cellular phenotypes, cell–cell communication, and microenvironmental tissue architecture. However, as spatial technologies have matured, two major computational bottlenecks have emerged:
Visualization & Scale Limits: Open-source visualization tools fail to scale past ~1 billion transcripts and millions of segmented cells, suffering severe memory bloat or interface lag.
Ecosystem & Usability Barriers: Commercial viewers are costly, closed-source, vendor-locked, and lack integration with custom data pipelines (such as Scanpy, Squidpy, and AnnData/SpatialData).
Technical Architecture & Engineering Innovations
Celldega introduces an end-to-end framework combining Python preprocessing, optimized columnar file formats, and high-performance WebGL/WebAssembly frontends:
DegaFiles & Row Group Storage Mode:
Traditional tile-based architectures generate 50,000+ individual files (transcripts, cell polygons, gene matrices, and WebP Deep Zoom image tiles), exceeding cloud host limits.
Celldega implements a consolidated Apache Parquet / GeoParquet format with an analytical formula-based index (
row_group_index = tile_x * num_tiles_y + tile_y).Uses ParquetWASM (Rust compiled to WebAssembly) and Apache Arrow in-memory structures to execute HTTP Range Requests in the browser, streaming only the exact byte ranges required for the visible viewport without downloading entire datasets. Reduces file counts from >50,000 to ~10 consolidated Parquet files.
Hardware-Accelerated Rendering: Leverages deck.gl for GPU-accelerated rendering of multi-channel microscopy image pyramids, polygon segmentation masks, and single-molecule transcript locations.
Specialized Visualization Modes
Celldega provides an integrated suite of views designed for different spatial analysis workflows:
Landscape: 2D interactive multi-layer tissue canvas displaying pyramidal microscopy (H&E, DAPI, multichannel fluorescence), cell segmentation boundaries, individual transcripts, and spatial neighborhood overlays, complete with real-time gene search and categorical summary histograms.Yearbook: Synchronized grid of cell "portraits" (configurable rows/cols) centered on individual cells for morphological comparison and rapid phenotyping. Supports stateless browser-side querying (front_end_query) or programmatic selection from AnnData.CellCloud&NeighborhoodCloud: 3D orbit-camera views rendering millions of cell centroids and alpha-shape neighborhood polygons across serial sections and thick tissues (with on-demand centroid streaming to maintain 60 FPS).Clustergram&Composition: Hierarchically clustered heatmaps with interactive dendrogram cutting (built on Clustergrammer), bi-directionally linked to spatial views (spatial_clustergram) so selecting clusters or genes instantly highlights corresponding cells on the tissue section.Enrich: Integrated real-time gene-set enrichment widget connecting directly to the Enrichr API (e.g.CellMarker_2024, GO) to functionally annotate marker gene lists from selected tissue regions.
Validated Platforms & Scale
Supported Technologies: 10x Genomics Xenium & Xenium Prime, NanoString CosMx SMI, 10x Visium HD, Vizgen MERSCOPE, and single-cell Chromium.
Scale Benchmarks: Demonstrated on datasets exceeding 1 billion transcripts, multi-slice serial alignment datasets, and a full 3D reconstruction of a developing whole mouse head comprising over four million cells.
Cloud & Pipeline Integration: Integrates with WDL workflows on Terra.bio (
stp_segmentation_wdl) for reproducible automated preprocessing with Cellpose, InstanSeg, and Starfish.
Relevance to Digital & Computational Pathology
Celldega bridges the gap between raw spatial-omics data processing and diagnostic/translational slide evaluation:
Enables lightweight, zero-install, browser-based exploration and public web galleries without demanding dedicated GPU server infrastructure.
Empowers pathologists to inspect single-cell morphology, verify segmentation boundaries against histological ground truth, and interrogate local microenvironmental gradients (e.g. tumor-immune interfaces, islet neighborhoods, tertiary lymphoid structures) in a single unified workspace.
HistoMetPath
Not software to adopt — a single-author MIL research framework whose models are 30-slide development artefacts (best AUROC 0.68, no external validation). Read instead for its evaluation-governance code: a held-out test set that can be opened exactly once, enforced by checksums, a config-load refusal of the test split, and an execution counter spent before the first pixel is read — see HistoMetPath.
WSInfer
Applies an already-trained patch-classification model across a whole slide and returns a per-tile probability map — as a QuPath overlay, or as CSV and GeoJSON from the command line. Eight models (TCGA tumour classifiers, pan-cancer lymphocytes, PatchCamelyon lymph-node metastasis, NCT-CRC-HE-100K colorectal tissue) come from a Hugging Face zoo; your own TorchScript model plus a config works too. Use the QuPath extension, not the Python package: the runtime has had no commit since July 2024 and its last two bug fixes are unreleased, while the extension is maintained by the QuPath group and requires QuPath 0.6 — see WSInfer.
nanopath
A lean harness from MedARC for pretraining tile-level pathology foundation models from scratch, modelled on Karpathy's nanochat: one GPU, 1,000,000 TCGA tile presentations, a 1e18-FLOP cap, then a fixed 12-dataset probe suite (drawn from THUNDER, PathoBench and LEOPARD) covering tile classification, nuclear segmentation, slide-level progression, mutation, survival and robustness. The point is to test a training idea cheaply before spending real compute on it. Apache-2.0; 139 commits since April 2026, essentially one author (Paul Scotti). The 120 GB pre-tiled TCGA parquet set is on Hugging Face as medarc/nanopath, so no WSI handling is needed to start — but reproducing the tiling from raw SVS needs the full ~13 TB open-access TCGA slide set.
Its real value here is the frozen-baseline table: UNI-2-h, H-optimus-0, Virchow, Prov-GigaPath, Midnight-12K, OpenMidnight, EXAONE-Path-2.5, GenBio-PathFM and three DINOv2 sizes all run through one identical probe suite. On that scale the best community model (0.6676, single GPU) lands above Virchow (0.6591) and GigaPath (0.6456) but below GenBio-PathFM (0.6917), UNI-2-h (0.6782) and H-optimus-0 (0.6763) — with the caveat the maintainers state themselves, that each backbone's native feature width (384-d to 4608-d) feeds the probe heads uncontrolled. The methodological detail worth copying is in benchmarking/: every probe carries a randomised-weight null audit — 20 seeds of an untrained DINOv2-small pushed through the same path — and the per-dataset notes say plainly which probes fail it. UCLA-Lung progression has a null of 0.692 ± 0.004, above every natural-image DINOv2 checkpoint and level with the current leaderboard leader, and the note calls it "a caution flag" and "not a clean representation-quality readout in isolation". LEOPARD BCR is similar (null 0.633 ± 0.012). Neither top-level README carries those caveats into the headline mean_probe_score, which averages all eight probe families equally — so the ranking absorbs columns the per-dataset notes flag as noise. There is a noise guard, though not that one: a new leader must beat the old by 0.006 and survive a maintainer rerun at a different rng seed.
HoVer-NeXt
Modernized, high-throughput nuclear instance segmentation and classification pipeline from the University of Bern (Digital Pathology group / Prof. Inti Zlobec; Baumann et al., MIDL 2024). Re-engineers the classic HoVer-Net architecture using ConvNeXt-V2 backbones (Tiny/Base/Large), decoupled multi-head U-Net decoders (5-channel distance maps and multi-class cell typing), native multi-threaded whole-slide inference (OpenSlide, CZI, Zarr), and direct export to QuPath .tsv measurement tables. Pretrained checkpoints available on Zenodo for Lizard-Mitosis (7 classes) and PanNuke (5 classes across 19 tissues) — see HoVer-NeXt.
Training Code: digitalpathologybern/hover_next_train
Inference Pipeline: digitalpathologybern/hover_next_inference
Paper: MIDL 2024 (OpenReview)
From Samples to Knowledge 2025: QuPath Training Course
Comprehensive hands-on curriculum from the La Jolla Institute for Immunology (Dr. Zbigniew Mikulski & Dr. Sara McArdle) targeting modern QuPath v0.6.0+ bioimage workflows. Covers 18-plex multiplex immunofluorescence (RareCyte Orion), native deep-learning cell segmentation via Deep Java Library (DJL / InstanSeg), Groovy scripting pipelines for batch cohort processing, complex multi-marker composite phenotyping, multimodal image registration (Warpy / qupath-extension-warpy), spatial proximity/density metrics, and interactive Python/Jupyter bridging (QuBylab / Paquo) — see dedicated course note: From Samples to Knowledge 2025: QuPath Training Course.
Course Book: saramcardle.github.io/FS2K
Repository: saramcardle/FS2K
Video Playlist: YouTube Playlist (14 videos)
Registration Tool: BIOP/qupath-extension-warpy
Celldega
High-throughput client-side spatial omics and multiplexed digital pathology visualization platform developed by the Broad Institute (Fernandez et al., bioRxiv 2026). Streams billion-transcript datasets directly in the browser via parquet-wasm and Apache Arrow using DegaFiles (GeoParquet indexed by spatial row groups) with GPU-accelerated deck.gl multi-layer rendering — see dedicated tool note: Celldega.
Documentation & Viewer: broadinstitute.github.io/celldega
GitHub: broadinstitute/celldega — BSD-3-Clause
Interactive Notebook Demo: marimo interactive notebook on molab
HistoCAM
Ambient, real-time digitization and datafication platform for glass-slide microscopy developed by researchers across Tulane University and Kitware, Inc. (Maira et al., Nature Communications 2026). Couples a 31-megapixel high space-bandwidth camera with a C++ streaming engine (pathcam) to passively composite multi-resolution WSI pyramids (2X to 40X) during standard optical examination, record pathologist search trajectories and dwell times, and execute real-time TensorRT AI segmentation directly at the microscope — see dedicated tool note: HistoCAM and literature review Ambient, real-time digitization and datafication of glass slide microscopy towards AI-at-the-microscope.
Zenodo Benchmark Data: Zenodo Record 21970445
In this section
Last updated