Efficient Thoracic Abnormalities Detection Using Mobile Deep Learning Models

Indonesia faces a critical shortage of radiologists, with only 1.2 radiologists per 100,000 individuals, leading to delays in diagnosing thoracic abnormalities such as pneumothorax, cardiomegaly, nodule/mass, consolidation and infiltration. Chest X-ray (CXR) interpretation remains challenging due to overlapping radiological features, necessitating AI-assisted solutions. This study evaluates three lightweight deep learning models—MobileNetV2, ShuffleNetV2, and EfficientNetB0—for automated thoracic abnormality detection using the ChestX-ray8 dataset. We assessed model performance using accuracy, precision, recall, F1-score, and AUC-ROC, selecting the best model based on the highest per-fold F1-score. EfficientNetB0 emerged as the top-performing model, achieving a macro-average F1-score of 0.556 and AUC-ROC of 0.765, outperforming MobileNetV2 (0.494, 0.719) and ShuffleNetV2 (0.481, 0.713). Grad-CAM analysis revealed strong localization for pneumothorax and consolidation but misclassifications in cardiomegaly and nodule/mass detection due to poor feature differentiation. The findings highlight EfficientNetB0’s potential as an AI-assisted diagnostic tool for low-resource settings, while also underscoring the need for segmentation-based pretraining and multi-scale feature extraction to enhance detection accuracy. Future work should focus on optimizing sensitivity to subtle abnormalities and ensuring clinical trust through improved interpretability techniques.