Iterative Development of a Pneumothorax Detection System Using Transfer Learning on Chest Radiographs

Pneumothorax, the abnormal collection of air in the pleural cavity that causes lung collapse, is a potentially life-threatening condition that requires rapid radiological diagnosis. This paper presents the iterative development of an automated pneumothorax detection system trained on chest radiographs from the NIH ChestX-ray14 dataset. Three iterations of increasing complexity are documented: (1) a Support Vector Machine (SVM) classifier trained on two hand-crafted geometric features extracted from 200 images, achieving 62.5% accuracy; (2) a ResNet18 convolutional neural network using transfer learning on the same 200 images, achieving 80% accuracy but exhibiting severe overfitting; and (3) a scaled ResNet18 system trained on 4,694 images with class-balanced loss and data augmentation, achieving 83% test accuracy on 694 unseen radiographs with a macro-averaged F1-score of 0.83. Results demonstrate that while transfer learning provides substantial improvements over hand-crafted features, dataset scale and balancing strategies are critical for producing reliable and generalizable medical image classifiers. The final model is deployed publicly as a web-based diagnostic aid at https://pneumothorax.streamlit.app.