X-ray Classification

Medical imaging data is messy, imbalanced, and unforgiving — which is exactly why this project was worth doing carefully.

The Brief

A funded medical-imaging research effort needed a multi-label classifier for chest X-rays, trained and evaluated against the well-known NIH Chest X-ray dataset, narrowed to the six most frequently occurring diagnostic labels.

The Challenge

Severe class imbalance in the data, combined with how genuinely hard chest radiographs are to interpret, made reliable multi-label prediction a real challenge rather than a routine training run.

The Outcome

The resulting classifier reached roughly 75% accuracy — enough to satisfy the client's benchmark and close out a tight, week-long engagement successfully.