A retired breast-cancer screening model, frozen in PyTorch

What it does

Implements the view-wise and image-wise models from the 2019 paper. It takes four standard mammography views (two per breast), runs them through a cropping and heatmap-generation pipeline, and outputs benign/malignant probability scores for each breast. The repo includes both an image-only classifier and an image-and-heatmaps variant, plus a less accurate single-image model intended for transfer learning. It ships with four sample exams and pretrained PyTorch weights so you can run the full pipeline end-to-end.

The interesting bit

The authors are refreshingly blunt: the model is not clinically deployed and has been surpassed by their later work. That honesty makes the repo a useful historical artifact rather than a product pitch. The multi-stage pipeline—cropping, center extraction, patch-classifier heatmaps, then final classification—shows how medical-imaging models often rely on heavy preprocessing, not just a single forward pass.

Key highlights

• Two exam-level classifiers: image-only and image-and-heatmaps, plus a single-image model for transfer learning.
• Full preprocessing pipeline: breast cropping, optimal center calculation, and heatmap generation from a patch classifier.
• Input requires four standard views (L-CC, R-CC, L-MLO, R-MLO) as 16-bit PNGs.
• Includes four sample mammography exams with labels and pretrained PyTorch weights.
• Also offers a TensorFlow implementation of the image-wise model.

Caveats

• The authors explicitly state this 2019 model is not used clinically and lags far behind their newer models.
• Dependencies are pinned to old versions (Python 3.6, PyTorch 0.4.1), so expect environment friction.
• The image-wise model underperforms the view-wise model and is provided mainly for transfer-learning convenience.

Verdict

Worth a look if you are reproducing 2019 medical-imaging literature or studying how heatmap-guided architectures work. Skip it if you need a modern, clinical-grade screening tool—the authors themselves would tell you to email them for that.