Class Imbalance Corrected Multi Label Classification of Thoracic Pathologies in Resource Constrained Chest Radiograph Screening

Chest radiograph screening in resource constrained settings demands models robust to severe imbalance under limited compute. We train dense and convolutional neural networks on ChestMNIST, covering 14 thoracic pathologies across patient disjoint train, validation and test splits, and correct a 96.6 times imbalance ratio using a class weighted focal loss with square root dampened inverse frequency weights. Replacing the fixed 0.5 decision threshold with per label, validation set F1 optimal thresholds raises mean test F1 from 0.0057 to 0.1786 and mean recall from 0.0031 to 0.3600. A controlled ablation against plain binary cross entropy at matched optimized thresholds shows the focal loss model achieves higher mean recall and mean F1 at comparable mean area under the curve, isolating loss from thresholding. Three architectures are compared under an identical training protocol, with the convolutional network outperforming both dense variants using 5.2 times fewer parameters.