Mammography Segmentation

Delineate lesions on a mammogram at pixel level (segmentation mask), with an open-source model evaluated on Dice/IoU.

2 participants · own branch each

Why this challenge exists

Classification tells you that something is suspicious - segmentation tells you where. Pixel-level lesion masks are what make a model's output actually usable by a radiologist: they can be overlaid on the image, checked against what the reader sees, and reused downstream for measurement or follow-up. Public datasets like CBIS-DDSM ship ROI annotations, which makes open, reproducible segmentation work possible. This challenge is the localization counterpart of the classification challenge, feeding the same clinical validation phase.

The brief

What has to be shipped:

Fine-tune an open-source segmentation model to delineate lesions on mammograms at pixel level, producing a binary or per-lesion mask, trained exclusively on public datasets
Reach the best Dice / IoU you can on the evaluation setup, starting from CBIS-DDSM and its ROI annotations as the baseline dataset
Keep the full pipeline reproducible: anyone should be able to retrain your model from your code and your dataset alone

What a reviewer should receive at the end, submitted step by step from the ML workspace:

A Kaggle dataset: the curated training data with masks, sources and preparation documented
A Kaggle model with its reported Dice / IoU
A GitHub repo with the full training code (data loading, training, evaluation)
A GitHub repo packaging the model as a callable API - image in, mask out