Published before the number
An open-source breast ultrasound pipeline. Not a medical device. No accuracy claim in this release
Most medical-AI announcements arrive with a score. The split is hard to reconstruct. The excluded files are not listed. The code, if it exists, comes later.
This one is published in the other order.
Open Breast Ultrasound is an open-source research prototype for whole-image classification of breast ultrasound into three classes: benign, malignant, and normal. The public repository is the audit, the split, and the training code. It is Apache-2.0 licensed. Anyone can read it, rerun it, and say what is wrong with it.
The code is here:
https://github.com/milanstanisic-opensource/open-breast-ultrasound
What is in the repository
The release inventories a local BUSI-style folder without deleting images or masks. It skips segmentation files, records file hashes, and marks duplicate candidates. It writes a seeded group split that keeps known links together. It can train an EfficientNetB0 baseline only after an explicit inventory review, fit an optional temperature on validation data, evaluate a locked test split, and run a local demo.
It does not ship the dataset, ImageNet weights, or a trained checkpoint. A person who wants to rerun it downloads BUSI separately and accepts that dataset’s terms.
Open source here means the method is inspectable before anyone is asked to trust a number.
What this release is not
This software is not a medical device. It is not a diagnosis. No BUSI model has been trained or measured for this release. A held-out split from a public dataset is not external validation, and it is not patient-level separation. There is no clinical partner in this release, and there is no authorization.
I am stating that first because the usual order is the other way around. A score is published. The caveats come later, if they come at all.
Get Milan Stanisic’s stories in your inbox
Join Medium for free to get updates from this writer.
Why publish it empty of results
A reproducible pipeline is useful before it has a number. Another group can see which files were counted, how the split was made, and what the training config will be. Disputed public-data problems are documented. They are not silently removed.
Closed medical AI can report a strong score and still be impossible to rerun. An open pipeline can report a weak score and still be useful, because the weak score can be checked.
The next honest step is a documented BUSI baseline: run the audit, freeze the split, train, and publish the number that comes out, including a weak one.
Money
There is a public campaign:
https://whydonate.com/fundraising/open-breast-ultrasound-research
The displayed goal is 120,000 EUR. That figure is provisional. It is not a quotation. No supplier has priced this plan, and no institution has agreed to host it.
Money received will be used only for the next step that the amount can actually cover: a documented baseline, public code, and published results. The code stays open source either way. The money will not be described as payment for a clinical product or a regulatory submission.
If less is received, the work will be narrowed to what that amount can finish, and that will be said publicly. If more is received, it stays with this project.
A wider multi-year programme has been discussed as a provisional ambition of 300,000–500,000 EUR. That range is not this campaign’s budget, and it is not a claim that the current prototype already justifies it.
What I am asking for
Two things.
Readers who work with medical images, evaluation, or open scientific software can look at the repository and say what is missing.
Readers who want to fund the next step can use the campaign link. The code remains public either way.
Contact: aksentijeikechukwu@gmail.com
Milan Stanisic
