Dual-Encoder framework for multi-abnormality classification of VinDr-CXR dataset

The basic examination technique used to analyze chest abnormalities is Chest X-Ray (CXR) imaging. Computer Aided Diagnosis (CAD) systems have revealed a valuable second opinion for radiologists but most of the studies have addressed multi-class detections from chest scans. However, expert radiologists often diagnose multiple abnormalities from scans depending on their expertise and experience. Moreover, there are challenges of inter and intra-variability among experts. This paper reports an automated model for multi-label finding from chest radiographs to address such a problem. For this, one of the recent publicly available datasets that include 14 position-specific abnormalities, named VinDr-CXR is applied. This paper introduces a dual encoder framework with a vision transformer and MobileNet networks for diagnosing multi-label abnormalities from CXR images. The dual encoder network extracts and fuses features of both global and local information with fine-grained spatial and contextual details. Our algorithm achieves the highest ROC AUC score of 0.92 in comparison to 0.89 and 0.90 scores of individual encoder networks of MobileNet and vision transformer respectively. These results validate that our network is efficient in improving the multi-abnormality classification performance and can support clinical decisions.