Segmentation-Guided Dual-Channel Ensemble Learning for Automated Multi-Label Thoracic Disease Detection in Chest Radiographs

Abstract

Automated detection of multi-label thoracic disease in chest X-rays is essential for improving the efficiency and reducing clinical burden, although it remains challenge because of co-occurring pathologies and visual feature overlap. Proposed work presents a novel dual-channel ensemble architecture that combines VGG-16 and ResNet-50 through independent parallel processing pathways, where each fed with segmentation-guided lung regions of interest (ROI), which facilitating complementary feature extraction and fusion-based classification of 14 thoracic pathologies. The dual-channel design leverages architectural complementarity, VGG-16 excels at fine-grained texture and boundary characterization while ResNet-50 captures hierarchical, multi-scale semantic features through parallel independent processing with learned weighted averaging fusion to produce robust multi-label predictions. Comprehensive evaluation on the NIH-14 dataset of 112,120 Chest X-ray images demonstrates superior ensemble performance, achieve mean AUC 0.895. it is 11.5 percentage point improvement over single VGG-16 and 7.5 percentage point improvement over ResNet-50.