Pneumonia Detection from Chest X-Rays

The CV documents 988 chest X-rays split 620/368 train/validation. I built train.py around EfficientNetV2B0A efficient convolutional architecture from Google—strong transfer-learning backbone for medical imaging with moderate compute, not a from-scratch CNN—data volume does not support training millions of weights cold.

Focal lossA loss function that down-weights easy examples and focuses learning on hard misclassified cases—useful for class imbalance plus class balancingOversampling or reweighting so the minority class influences gradients proportionally addressed pneumonia vs. normal skew. Two-phase fine-tuningFreeze backbone layers first, then unfreeze top blocks for a lower learning rate pass stabilized transfer from ImageNet-like features to radiograph texture.

Medical screening cares about errors, not a single accuracy headline. I exported confusion matricesTabular breakdown of true/false positives and negatives, ROC/AUCReceiver operating characteristic and area under curve—threshold-independent separability metric, and precision–recall curvesEspecially informative when positives are rare to outputs/ alongside checkpoints.

Test-time augmentationAverage predictions over flipped/rotated variants at inference to reduce variance and explicit threshold tuningSweeping decision boundaries to trade sensitivity vs. false alarms prevented overfitting to a default 0.5 cutoff.

Held-out validation landed at 83% accuracy per the CV—reported honestly with the full metric bundle, not a cherry-picked epoch. Structured outputs/ holds models and plots so runs are comparable week to week.

The project informed my later research interests in medical image analysis: preprocessing discipline and error-type analysis matter more than leaderboard placement.