An optimized EfficientNetB0 framework with CLAHE-based preprocessing for accurate multi-class chest X-ray classification

Chest radiography remains an essential diagnostic tool for thoracic diseases, yet interpreting overlapping anatomical structures is particularly challenging when multiple pathologies co-occur a common clinical scenario often oversimplified in deep learning approaches. This study presents an optimized EfficientNetB0 framework designed explicitly for multi-label classification of chest X-rays using the NIH dataset, integrating CLAHE-based contrast enhancement, strategic class balancing, and a comparative transfer learning strategy that preserves the dataset's inherent multi-label complexity. The proposed model achieved superior diagnostic performance with a macro-average AUC of 0.906 and recall of 0.824, outperforming DenseNet121 and MobileNetV2, and demonstrated strong per-class discrimination, especially for Pneumonia (AUC = 0.950) and Cardiomegaly (AUC = 0.946). These results confirm that the framework effectively balances learning capacity and generalization in a realistic multi-label clinical setting, offering a robust, interpretable solution suitable for computer-aided diagnosis where accurate detection of co-occurring thoracic pathologies is critical.