ADWECA-Net: An Asymmetric Dual-Weight External with Coordinate Attention for Multi-Class Thoracic Disease Classification and Lesion Localization

Abstract

Automated multi-class disease classification and lesion localization remain challenging because of overlapping radiographic patterns and large anatomical variability. In this study, we propose ADWECA-Net, an asymmetric dual-weight external framework that integrates convolutional neural networks with transformer-based global attention mechanisms for simultaneous multi-class thoracic disease classification and localization. The proposed architecture consists of a local branch based on EfficientNet-B0 enhanced with Dual-Weighted External Coordinate Attention (DWECA) to preserve spatial coordinate information, and a global branch employing window-based multi-head self-attention to capture long-range anatomical dependencies. Grad-CAM visualization is used to confirms anatomically consistent localization behaviour, enhancing interpretability. The framework is evaluated on the NIH ChestX-ray14 dataset and further VinDr-CXR dataset used for localization. Experimental results demonstrate with precision ranging from 0.836 to 0.962 and recall between 0.850 and 0.940. The majority of disease categories achieve F1-scores above 0. 90. The proposed ADWECA-Net demonstrates robust multi-class performance.