Segmentation-Aware Multi-Label Chest X-Ray Classification and Visual Explainability Analysis

Abstract

Chest X-ray (CXR) interpretation remains one of the most challenging tasks in medical imaging due to overlapping anatomical structures and the frequent coexistence of multiple thoracic pathologies. This study introduces a dual-scenario deep learning framework that integrates lung segmentation and multi-label classification with Grad-CAM–based interpretability. Two complementary preprocessing approaches are evaluated: (1) Scenario A, where the original image is concatenated with its corresponding lung mask (two-channel input), and (2) Scenario B, where only the segmented lung region is retained (masked input). Both configurations are tested using ResNet50 and ConvNeXt-Tiny backbones on the NIH ChestX-ray14 dataset comprising 112,120 radiographs labelled with 14 diseases. Experimental results demonstrate that integrating anatomical priors improves model stability, interpretability, and classification accuracy. The best-performing setup—ConvNeXt-Tiny under the masked-lung scenario—achieves an average ROC-AUC exceeding 0.85, outperforming or matching recent state-of-the-art methods. Furthermore, Grad-CAM visualizations confirm that segmentation-driven masking enhances spatial coherence between model activations and true pathological regions. Overall, the proposed framework offers a reproducible, anatomy-aware, and interpretable approach for automated thoracic disease detection, bridging the gap between deep learning and clinically meaningful radiological reasoning.

DOI: 10.61416/ceai.v28i2.9953