Class-wise image distribution.

The research creates a Vision Transformer (ViT) diagnostic system which identifies pneumonia and pneumothorax from chest radiographs through analysis of the NIH ChestX-ray14 dataset. The research methodology solves medical imaging problems through three essential components which include (i) radiograph-specific augmentation for simulating authentic imaging conditions and (ii) multi-label imbalance handling through WeightedRandomSampler with class-specific weight application to stop all-normal predictions and (iii) optimization improvements that include CosineAnnealingWarmRestarts scheduling and sigmoid-based classification head optimization and disease-specific threshold optimization. The evaluation of model performance uses AUC and sensitivity and specificity and precision and F1-score because accuracy proves ineffective when dealing with severe class imbalances. The ViT models achieve 70−75% accuracy and 0.63–0.71 AUC values for both target conditions during non-leaking and noise-aware experiments because of the weak labels and restricted supervision in the ChestX-ray14 dataset. The system enhances its ability to detect rare conditions while providing better interpretability through Vision Transformer attention-based visualization of important radiological areas. The research demonstrates that ViT performance improves significantly through medical-focused data preparation methods and training approaches which demonstrate potential for radiology assistance in high-volume and resource-constrained environments.