Differential diagnosis outlines organised by organ system and bench.
- Free tier
- Runs on any laptop
- Stays on your machine
The library
Filter by what you actually have to work with. Every combination is a shareable link β send a colleague the exact view you are looking at.
125 entries
Differential diagnosis outlines organised by organ system and bench.
Pathology image-caption dataset drawn from textbooks and articles.
Prognostic AI test using digital pathology images for breast cancer.
Answers a different question from a detection tool: not 'is there cancer' but 'how is this likely to behave'.
A no-frills slide viewer. Opens big scanner files quickly and lets you draw regions.
Research-focused. Useful if you want the academic papers rather than usable tools.
Lymph node metastasis detection challenge dataset.
The reason synoptic reporting exists: every element you fill in becomes a field a registry or a trial can actually use.
If your laboratory is going digital, this is the document your validation study will be judged against. Read it before you buy a scanner, not after.
Usable from QuPath via an extension if you would rather not touch Python.
Vision-transformer nucleus segmentation and classification.
Checklist for artificial intelligence in medical imaging.
Attention-based multiple-instance learning for weakly supervised whole-slide classification.
Worth knowing about because it takes seriously a problem most benchmarks ignore: whether a model can score well without actually looking at the slide.
Vision-language foundation model for pathology image-text tasks.
Trial reporting and protocol guidelines for AI interventions.
Unified multimodal model spanning patch and whole-slide analysis.
Agent that navigates a slide the way a pathologist moves a microscope.
Runs on a server your institution controls, so several people can annotate the same slides in a browser. Needs IT support to set up.
A useful corrective: impressive correlation numbers in this field are sensitive to how the test set was built.
The stage most pathology AI has never been through. Worth asking a vendor whether their product has.
Uses spatial context from surrounding tissue to improve expression prediction.
The reason a slide from one scanner can be opened by another vendor's software. Ask about DICOM support before buying anything.
Ask an AI vendor how their results are stored. If the answer is a bespoke database, your results are locked in.
A good starting point if your department is considering going digital and wants peer experience rather than vendor material.
Server platform for managing, annotating and analysing large slide collections.
Instruction tuning grounded in symbolic clinical reasoning.
Preference alignment to reduce hallucination in pathology VLMs.
'CE marked' under the old directive and under IVDR are not the same claim. Check which one a vendor means.
How a cleared AI device may be updated without a new submission.
Single-cell resolution spatial gene expression from histology.
Short, plain and non-technical. The quickest way to see what regulators expect of a medical AI product.
Run this before trusting any result computed on a batch of scanned slides. Most 'model failure' turns out to be slide quality.
A realistic entry point for a department that cannot justify a bulk scanner. Small enough to sit on a desk.
Practical rather than clever β the sort of thing that saves ten minutes a day.
Paired histology and spatial transcriptomics dataset and benchmark.
Transformer and graph approaches to expression prediction from histology.
Python toolkit for histology image analysis and feature extraction.
Checks a batch of scanned slides for problems like blur or pen marks before anyone spends time on them.
If a vendor cannot tell you how their output reaches your LIS, this is the vocabulary in which to ask the question.
Simultaneous nucleus segmentation and classification in histology.
Faster successor to HoVer-Net for nucleus segmentation and classification.
Runs after you have signed out, and raises a flag if it disagrees. Designed as a safety net, not a first opinion.
Topography and morphology coding for cancer registries.
International standard for medical device software lifecycle processes.
Anatomic pathology workflow and image analysis platform.
Embedding-based cell segmentation that runs acceptably on a CPU.
Risk management standard for medical devices.
The idea worth knowing: morphology carries enough signal to interpolate molecular data between measured points β but it is inference, not measurement.
Read this before trusting any performance figure. Most published accuracy is measured under conditions that do not resemble a working laboratory.
The scanner itself, not an AI tool β it produces the digital slides everything else works on.
Biomedical vision-language assistant trained rapidly from PubMed figures.
Universal codes for laboratory observations and report sections.
Mitotic count drives grading in several tumours, and it is one of the least reproducible things pathologists do β which is exactly why this benchmark measures generalisation rather than headline accuracy.
Vision-language foundation model for precision oncology.
A good starting dataset if you are learning β small enough to work with on a laptop.
Matters if slides are going to the cloud β proprietary scanner formats were not designed for that.
Image data management server for microscopy, including whole-slide images.
The genuinely low-cost option. You print the body, add optics and a Raspberry Pi, and get automated scanning for telepathology or teaching.
C library with Python bindings for reading proprietary whole-slide formats.
Low-cost high-throughput whole slide imaging from open-source hardware.
Notable as an early example of pathology analysis reachable from a general assistant rather than a dedicated application.
Pathology language-vision assistant built on PLIP with two-stage training.
Whole-slide image viewer for primary diagnosis.
An assistive second look at prostate biopsies. It flags suspicious areas β it does not make the diagnosis, and you remain responsible for the report.
Prostate biopsy Gleason grading challenge dataset.
Pan-cancer nucleus instance segmentation and classification dataset.
Training-free LLM agent that explores whole slides and reasons over findings.
Digital pathology image management and viewing platform.
Generative pathology assistant with a domain-adapted CLIP encoder.
Pathology image-caption corpus released with PathAsst.
Multimodal generative copilot for human pathology.
Practice cases and self-testing for pathologists and trainees.
Fine-grained language-image pretraining grounding captions to regions.
The most useful part is the structure, not the individual prompts: it shows how to build a prompt that can be evaluated rather than just admired.
Pathology instruction-tuning dataset released with PathAsst.
Python library for computational pathology preprocessing and modelling.
Expert-validated multimodal pathology understanding benchmark.
The interesting idea here is grounding: answers point back to a textbook page instead of being asserted from nowhere.
Standardised evaluation suite for pathology slide-level foundation models.
Reinforcement-learning pathology reasoner trained on chain-of-thought data.
Not an AI tool, but the reference most pathologists already use β included because AI outputs should always be checked against a trusted source.
Many pathologists already know the teaching platform; the clinical viewer is a separate, cleared product.
Knowledge-guided structured reasoning for whole-slide pathology.
GRPO-trained reasoning model for pathology visual questions.
Pathology visual question answering benchmark.
Notable because it is genuinely open β no access request, no non-commercial restriction.
Whole-slide imaging system for primary diagnosis.
Pathology CLIP model trained on image-text pairs mined from social media.
Independent confirmation that acquisition site leaks into the pixels.
Slide-level multimodal model generating reports from tile embeddings.
Digital pathology platform for diagnostic viewing and workflow.
Slide-level foundation model with a tile encoder and long-context aggregator.
Histopathology image-text pairs mined from educational videos.
Instruction-tuned pathology assistant trained from educational video narration.
Free, installs on an ordinary laptop, and does not upload your slides anywhere. The usual first stop if you have never used digital pathology software.
Another automatic cell-detection option inside QuPath. Needs a one-time Python setup.
Newer segmentation option that runs without a graphics card.
Adds automatic nucleus detection to QuPath. Runs on a normal laptop, no coding needed.
Whole-slide scanner and viewing software for diagnostic use.
Worth knowing about if your hospital already uses Sectra for radiology β the pathology module reuses that infrastructure.
Its practical value in pathology today is speeding up manual annotation, not replacing it.
The most important cautionary result in computational pathology. It explains why a model reporting excellent accuracy on TCGA may fail completely on your slides.
Vision-language assistant operating on whole slides rather than patches.
End-to-end deep learning pipeline for whole-slide images.
Slide-level prompt learning for few-shot MIL.
Mixture-of-experts reasoner unifying ROI and whole-slide tasks.
The difference between a report a computer can count and one it can only store.
Early demonstration that gene expression can be predicted from H&E alone.
Why a model trained in one laboratory degrades in another.
Reporting guideline for AI diagnostic accuracy studies.
Star-convex polygon nucleus detection.
Pan-cancer whole-slide and molecular data archive.
Python toolbox for computational pathology pipelines.
Multimodal whole-slide foundation model producing slide-level embeddings.
Stain normalisation as a differentiable, GPU-capable operation.
Toolkit for whole-slide preprocessing and foundation-model feature extraction.
A quick way to judge a paper: if it does not report these items, you cannot tell whether the model would work in your laboratory.
Self-supervised vision foundation model for pathology tile encoding.
Dual-scale vision-language MIL for whole-slide classification.
Large pathology vision foundation model trained on a very large slide corpus.
Not AI, but included deliberately β AI output should always be checked against a trusted image reference.
Whole-slide VQA benchmark organised around morphological description.
Multimodal model for whole-slide images with a morphology-aware benchmark.
Runs pretrained patch classification models across whole slides, with a QuPath front end.
Nothing matches that combination. Loosen a filter β the counts beside each option show what is still available.