Lung Disease Classification via U-Net-Guided Lung Segmentation and Enhanced Chest X-Ray Cropping

Chest radiography (CXR) remains pivotal for the diagnosis of pulmonary diseases; however, the presence of non-relevant anatomical structures and background noise in full CXR images often limits the effectiveness of deep learning models. In this study, we propose a preprocessing strategy that leverages U-Net-based lung segmentation to extract segmentation-guided cropped lung regions, thereby enhancing the focus of classification models. We systematically evaluated the impact of this approach across multiple architectures, including ResNet50, MobileNetV2, InceptionV2, DenseNet121, ViT, Swin Transformer, MobileViT, and PulmoNetX. Comparative experiments demonstrated that models trained on segmentation-guided cropped lung regions consistently outperformed their counterparts trained on original full images. Among all models, PulmoNetX achieved the highest performance on the processed dataset, attaining an accuracy of 90.93%, precision of 90.80%, recall of 89.98%, and F1-score of 90.75%. These findings underscore the significance of anatomical-focused preprocessing in boosting classification performance and provide a scalable pathway for improving AI-driven diagnostics in chest radiology.