Problem
Pneumothorax is a sparse, class-imbalanced segmentation problem on chest X-rays. A network that reports a global accuracy number without looking at collapsed lung boundaries is not a clinical model.
Approach
I trained a multi-task U-Net with residual blocks on the SIIM-ACR pneumothorax dataset.
- Loss: Dice + BCE with positive-class up-weighting, because empty masks dominate the set.
- Augmentation: elastic deform and random affine, tuned for chest X-ray morphology rather than natural-image defaults.
- Metrics: Dice and IoU, plus a failure-mode pass across subgroups so confidence is something I measured, not assumed.
Result: IoU 0.65 and 80% accuracy on the held evaluation. The useful artifact is the imbalance handling and the subgroup analysis, not the headline number.