MICCAI OpenSourcePapers: eight years of medical imaging papers with code, indexed in Markdown

This is a hand-built index, not a piece of software

The repository description is the whole pitch: MICCAI 2019-2026 Open Source Papers. It has 1,294 stars, 224 forks, zero open issues, an Apache-2.0 licence and two topics, deep-learning and medical-imaging. The last push to the master branch was 2026-09-23, which is recent enough that the current year is being kept up with.

What it is not is worth stating early, because the name invites a wrong assumption. There is no package to install, no build step, no command line tool and no API. The files list is empty, there are no releases, and there is no example directory. The entire repository is documentation.

That is a legitimate and useful artefact. The MICCAI conference accepts a large volume of papers each year, publishes them through an open-access proceedings site, and asks authors whether they will release code. That last question is asked once, at submission time, and the answer lives in a submission form rather than anywhere a reader can search. This repository is the collected answer, transcribed by a person, which is why it has stars and forks despite containing nothing executable. Forking it is the expected way to build on it.

One Markdown table per year, with five columns that matter

The file tree is the design. There is MICCAI-2019.md, MICCAI-2020.md, MICCAI-2021.md, MICCAI-2022.md, MICCAI-2023.md, MICCAI-2024.md and MICCAI-2025.md, alongside a README.md, a LICENSE file and a WordCloud directory holding WordCloud-MICCAI2026.png. The README itself opens by naming its source, the MICCAI 2026 open-access proceedings, and then goes straight into the table.

The table has five columns and each one answers a different question. Title links to the paper PDF on the proceedings site. First author is the name as it appears on the submission, which matters more than it might seem because it is often the fastest route to a code repository. Code links to the author's repository. Dataset information is free text rather than a link. Paper links to the proceedings review page.

Every cell is a hyperlink or a name, so the whole thing renders on GitHub with no tooling and reads fine in a terminal. There is also no front end to keep alive and no dependency to upgrade, which is a real operational advantage over a scraper that would break whenever the proceedings site changes its markup.

What the 2026 entries show about how this research publishes

The 2026 table is worth reading as a sample of the field rather than as a list. One row is a paper on three-dimensional cerebrovascular shape completion from biplane angiography with a CTA prior, by Janik Jehkul, with code at github.com/janik-j/cta-dsa-fusion. Another classifies paramagnetic rim lesions in multiple sclerosis through asymmetric quantitative susceptibility mapping and FLAIR modelling, by Veronica Pignedoli, with code at github.com/veronicapignedoli/FRODO and a dataset column reading N/A. A third covers a clinical guideline-grounded hybrid agentic framework for holistic epilepsy management, by Duy Khoa Pham, with code at github.com/khoapham154/epi_guide.

Three things fall out of that sample. The first is that the dataset column is honest. The cerebrovascular paper lists the TopBrain CTA Dataset with its own link, and then states that the clinical CTA and DSA dataset from Brigham and Women's Hospital is not publicly available because of PHI constraints. Protected health information is the reason, and it is a constraint that no amount of open-source intent can remove. The second is that N/A appears as a real value, so an empty cell is a claim about the paper rather than an oversight. The third is that the volume of agentic and clinical framing language has risen sharply. A 2019 table would not contain that third row.

The dataset column is the reason to keep this open

In much of machine learning the binding constraint is not code, it is data. Code for a medical imaging method is a few thousand lines and a PyTorch dependency. The data behind it is a hospital's protected cohort, an institutional review board, a data use agreement and often a de-identification pipeline that takes longer than the paper.

Because the dataset column is free text, the table can record all four of those outcomes and not just the happy one. A row can point at a public challenge dataset, note that a companion clinical cohort was withheld, and link the challenge directly. That is more useful than a yes-or-no code flag, and it is the kind of detail that a reader would otherwise only learn by emailing the corresponding author.

It also has a planning use. Before committing to a baseline, reading a few hundred rows of this column tells you which public datasets the community is actually reusing, which methods are being applied to which anatomy, and where the gaps are. The WordCloud directory serves the same summarising purpose more loosely, rendering the 2026 title corpus as an image so the vocabulary drift across years is visible at a glance.

Coverage, provenance and the limits of a curated list

The structure has one asymmetry worth knowing about. Years 2019 through 2025 each have their own named file, but there is no MICCAI-2026.md in the tree. The 2026 table lives in README.md itself, which is consistent with the README being titled as the 2026 papers with code. That is a reasonable choice while a year is still in progress, since a file rewritten on every batch of submissions produces noisy diffs, but it does mean a script that walks the tree looking for a 2026 file will find nothing.

Provenance is the second limit. The README names its source, which is the part that matters most, but there is no per-row timestamp and no record of when a given year was last checked against the proceedings. A transcription can drift as papers are added, corrections are made, or a code link rots. The repository's own activity is a reasonable proxy for how current the whole thing is, given the September 2026 push.

Finally, a blank code cell is not proof of no code. It means the author did not supply a link on the submission form, or the link was not caught. Some papers release code after acceptance, on a personal page or under an institutional group, and some never release it at all. Treat the table as a strong starting set rather than a census.

How to actually use the index

Clone it and grep it. The tables are plain text with consistent column headers, so searching for an anatomical region, an imaging modality, a method name or a repeated code host across all eight years takes one command. If you are surveying a subfield, open the year file closest to your area and read it end to end, which is faster than any search interface over the proceedings.

When you find a row, follow the links in the order the table suggests. Read the paper PDF for the method, open the code repository to judge whether it is a complete pipeline or a training script, and then check the dataset information before planning an experiment. That last step will save you the most time, because it is where the practical obstacles are documented.

Apache-2.0 is about as permissive as a licence gets, so if you want to build a paper-counting script, a benchmark shortlist, or a leaderboard of which methods are most reproducible, you can do so without asking. The fork count suggests people have done exactly that.

For the wider picture, the related-search signal around this repository is dominated by MICCAI 2026 and 2027 logistics, workshop listings and submission systems. If you are here for those administrative details, the proceedings site it links from is the authoritative answer. The repository does one job and does it in a form that will still be readable in ten years.

Editorial conclusion

The value of MICCAI-OpenSourcePapers is that it exists at all. Nobody has to run a scraper, wait for a crawl to finish, or sort a search result to find out that a MICCAI paper published its code and what dataset it used. The work of reading hundreds of proceedings pages and transcribing five fields per paper is already done, one file per year, in a format that greps, diffs and renders without a build step. Three caveats travel with it. It is a curated index and not an authoritative registry, so a blank code cell means the author did not make a link available rather than proof that none exists. The dataset column is often more valuable than the code column, and it frequently records that the data cannot be released at all. And the coverage is uneven across years, with the most recent year living in the README rather than in a file named after it. Use it as a map into the proceedings, then read the paper.

Frequently asked questions

Official sources

1. Issues
2. JunMa11/MICCAI-OpenSourcePapers on GitHub
3. License: Apache-2.0
4. README