Awesome-AI4DigitalPathology
A community-curated index of 453 works in computational pathology, organised into fifteen sections that map almost exactly onto this wiki’s own structure — worth adopting as the standing entry point for a topic, but its bundled “literature skill” ships an index whose descriptive fields are empty in every record.
Purpose
Verdict: keep the list as a standing reading index; the bundled skill is not usable as shipped.
(adopted: false in the frontmatter refers to software, as it does on the other external pages — nothing here is installed or depended on. The list is worth reading regularly; the skill is not worth running.)
This is an awesome-list, so the thing being evaluated is a bibliography rather than software. On that basis it is good: 453 entries with a consistent badge format carrying paper, code, dataset and model links; 236 GitHub links and 57 Hugging Face links; real venue badges (Lancet Oncol 2019, Mod Pathol 2022, Nat Rev Clin Oncol 2019); 262 commits and 16 contributors, with recent additions arriving as pull requests from people outside the core group. It is maintained, not dumped.
The reason to care here specifically is that its taxonomy is close to this wiki’s, which makes it usable as a gap-finder rather than just a reading list. Where its sections have no counterpart here, that is a real absence rather than a difference of vocabulary.
Data used
Not a data project. What it indexes, by section, with entry counts read 2026-07-28:
| Section | Entries | Nearest page here |
|---|---|---|
| Vision-Language Models and Pathology Agents | 76 | Agentic Slide Navigation, LLM Adaptation Techniques |
| Datasets and Benchmarks | 56 | wiki/datasets/ |
| Multiple Instance Learning | 46 | Weakly Supervised Learning |
| Computational Pathology with Multi-Omics | 44 | Multimodal Fusion |
| Cytology and Cervical Cytology | 35 | none — see Cervical Cytology Screening |
| Patch-Level Foundation Models | 34 | Foundation Models in Pathology |
| Generative Models for Computational Pathology | 31 | partial — Feature-Space Distance covers evaluation only |
| Clinical Tasks and Applications | 25 | scattered |
| Slide-Level Foundation Models and Slide Encoders | 19 | Foundation Models in Pathology, Foundation Model Cache |
| Dense Prediction | 19 | Cell Detection and Counting |
| Resources, Toolkits, Open-Source Projects | 17 | QuPath Tooling |
| Surveys, Reviews, and Perspectives | 16 | — |
| Pathology Image Registration and Spatial Alignment | 16 | Cross-Stain Registration |
| Federated Learning in Computational Pathology | 15 | none — see Federated Learning in Pathology |
| Digital Slide Scanners and File Formats | 4 | Whole Slide Imaging, Slide Metadata Extraction |
Two sections had no counterpart in this wiki at all, and both are now written: federated learning and cervical cytology. That is the concrete payoff of reading the list — not the individual papers, but noticing which whole areas were missing.
One taxonomy caveat worth recording: the sections bleed into each other. The cytology section contains most of the nucleus-segmentation canon (HoVer-Net, Cellpose, PanNuke, MoNuSeg, CoNIC, CellViT, Mesmer, Lizard, NuCLS), which belongs under dense prediction and maps here to Cell Detection and Counting. Anyone searching by section will miss things.
Methods
Read at source level, which for this repo means the README and the bundled skill directory.
The bundled literature skill does not deliver what it describes
ai4pathology_literature_skill/ is an agent skill with a SKILL.md, a JSON schema, a JSONL index of 500 records, query and update scripts, and a GitHub Actions workflow. SKILL.md defines a rich canonical schema and query_prompt.md promises retrieval by category, task, method family, year, venue, disease or organ, modality, and code/dataset/model availability.
In the shipped data/papers.jsonl, eleven schema fields carry no information in any record:
short_name,task,method_family,disease_or_organ,modality,code_url,dataset_url,model_url— empty in all 500bibtexandlimitation— absent entirelytype—"paper"for all 500, despite a nine-value enum
arxiv_id is filled in 156/500. venue holds only two placeholder values — README (404) and arXiv (96).
This is by construction, not backlog. scripts/update_papers.py hard-codes those fields empty in both code paths (lines 296–317 and 398–419) and no path ever populates them, so re-running the updater cannot fix it. It also discards data it already holds: 233 README entries carry a GitHub link and 57 a Hugging Face link, but parse_readme_papers keeps only the first external link (line 277); 400 entries carry a real venue in their badge, and line 302 hard-codes "venue": "README" after the parser has already matched the year out of that same badge.
The consequence lands on the only retrieval tool shipped. scripts/query_papers.py scores candidates against short_name, task and method_family — 5.5 of 12.5 total field weight against fields that are always empty — and has no README fallback, contrary to SKILL.md, which says to fall back only “if no index exists”. The index exists, so it is used, and it is hollow.
The refresh has never run and cannot. The workflow sits at ai4pathology_literature_skill/.github/workflows/update_papers.yml; GitHub Actions only executes workflows from the repository root .github/workflows/, which does not exist here. Its relative paths also assume the skill directory as working directory while actions/checkout lands at the root. Git history agrees: the entire skill including papers.jsonl arrived in a single commit (529cec6, 2026-04-28, one author) and has not been touched in the 235 commits since.
What still works. tags is populated on all 500 records with eight inferred topical values (wsi 132, foundation-model 80, segmentation 60, vlm 58, mil 56, diagnosis 39, survival 36, federated-learning 16), and 315/500 carry at least one. Keyword search over title, tags and contribution therefore functions as a lossy proxy. Filtering by modality, disease, real venue or code availability does not.
And the curated half is fine. An earlier draft of this page described contribution as uniformly broken because 483/500 lack terminal punctuation. Split by source that is misleading: the 404 README-derived records are genuine human-written one-liners with a median length of 44 characters, missing their full stop only because the parser calls .strip("."). Only the 96 arXiv-derived records are 300-character truncated abstracts. The correction came from an independent reviewer asked to refute the finding, and it is the difference between “the index is broken” and “the index throws away the structure, keeps the prose”.
Fixing it would be cheap, and is worth saying because the list deserves it
Moving the workflow to the repository root and correcting its paths would make the index refresh — but would populate none of the eleven dead fields; that needs generator changes. Keeping all external links rather than the first, and reading the venue out of the badge the parser already matches, would restore code_url, model_url and venue from data already in hand. That is a small patch against a genuinely useful list, and worth offering upstream rather than only recording.
Current state / open questions
- Should the fix be offered upstream? The parser changes are small and the maintainers are demonstrably responsive to pull requests. Recording the defect here helps nobody but us.
- Which of the 453 entries are worth actually reading? This evaluation deliberately did not appraise individual papers — everything catalogued below is a pointer, not an endorsement, and nothing on the two new wiki pages claims a finding from a paper that has not been read here.
- The list is a snapshot dependency: it is maintained now, and a wiki page that leans on it inherits whatever happens to it later. The two new concept pages cite specific works by name and URL rather than citing “the list”, so they survive it going stale.
- Does anything in Datasets and Benchmarks (56 entries) overlap what is already recorded in
wiki/datasets/? Not checked.[unverified]
How it connects
Federated Learning in Pathology — the first of two whole areas this wiki was missing, and the one that speaks directly to the standing external-validation gap.
Cervical Cytology Screening — the second, and the largest deployed application of pathology AI anywhere.
Cell Detection and Counting — where the list’s nucleus-segmentation canon belongs, filed under its cytology section rather than dense prediction.
Foundation Models in Pathology — 53 entries across the two foundation-model sections, and the closest thing to a complete published inventory of that landscape.
QuPath Tooling — the hub this list’s Resources section is the external counterpart of.
Derived from: repository cloned 2026-07-28 at 8047cde (full clone, 262 commits); README.md read in full (766 lines, 453 entries, link and section counts computed programmatically); ai4pathology_literature_skill/SKILL.md, papers.schema.json, query_prompt.md, scripts/update_papers.py, scripts/query_papers.py, .github/workflows/update_papers.yml and all 500 records of data/papers.jsonl read or analysed directly; LICENSE and git log read. The index finding was independently checked by a reviewer instructed only to refute it, which corrected the field count from ten to eleven and corrected the characterisation of contribution.