Pneumonia Detection with Data Augmentation Strategies
A systematic study of data augmentation pipelines for pneumonia classification from chest X-rays, targeting performance improvement under limited data conditions.
Problem Statement
Medical image datasets are often small and class-imbalanced, leading to models that overfit or fail to generalize. Choosing the right augmentation strategy is critical but underexplored.
Method & Approach
We evaluate 12 augmentation strategies (geometric, photometric, mixup-based) across DenseNet-121, EfficientNet-B3, and ResNet-50. Performance measured with AUC, sensitivity, specificity, and F1-score.
Dataset / Data Source
NIH Chest X-ray14 (pneumonia subset), Kaggle Chest X-ray Images (Pneumonia).
Results
CutMix combined with geometric augmentation yielded the best generalization, improving AUC by ~3% over no-augmentation baseline on the held-out test set.
My Contribution
Augmentation pipeline design, model training, statistical evaluation, and ablation study.