Mammography View Classifier

Problem

Mammography studies hang correctly only if laterality and view position are right — but the DICOM tags that carry them are frequently missing, wrong, or vendor-inconsistent. Getting it wrong mis-hangs the study on the radiologist’s screen.

Architecture

A stacked ensemble that plays each signal to its strength.

• Model A — CatBoost on metadata: 8 categorical + 14 numeric DICOM features with missingness indicators and TF-IDF over SeriesDescription, temperature-calibrated, class-weighted for rare views.
• Model B — ConvNeXt-Tiny on pixels: grayscale-adapted backbone, multi-task heads for laterality and view, focal loss for the hard 8-way view problem, 1024² input with Otsu breast masking. Rotation and flip augmentation are deliberately disabled — orientation is the label.
• Model C — CatBoost meta-learner: combines A and B probabilities with agreement flags and a trust gate — when structured tags are present and Model A is >0.95 confident, its probabilities are upweighted 2×; otherwise the pixels win.

Confidence is the top-1 minus top-2 probability gap, a better-calibrated signal than raw max.

Results

99.83% F1 on laterality and 95.75% F1 on view position across 8 classes, on real clinical data — with the ensemble correctly trusting metadata when it’s there and falling back to vision when it isn’t.

Impact

Studies hang right, automatically, even when the upstream DICOM tags can’t be trusted.