Explainable Multiple Receptive Attention Network for Expert Cardiologist Compatible Incomplete Kawasaki Disease Diagnosis on Echocardiography

Abstract

Identifying incomplete Kawasaki disease (KD) has been challenging due to its atypical clinical manifestations. Differentiating it from other febrile illnesses, including COVID-19, is crucial for pediatric patients. Early detection of coronary artery abnormalities through echocardiographic examination is vital for accurate diagnosis and favorable outcomes. With the increased prevalence of KD among pediatric populations, there is a need for continued research and innovative diagnostic tools to improve early detection and management. To address this, we introduce a Multiple Receptive Attention Network (MRANet) incorporating a multi-receptive attention layer, designed to enhance the discrimination of incomplete KD from echocardiographic images, achieving better sensitivity and specificity. A total of 147 echocardiographic imaging datasets were utilized for training MRANet and other state-of-the-art deep learning models. The performance of MRANet was compared with compatible deep learning networks for evaluation. The results demonstrate that MRANet outperforms other advanced deep learning methodologies. MRANet’s performance is comparable to that of an experienced pediatric cardiologist in detecting coronary artery abnormalities for accurate KD diagnosis. This study highlights the potential of MRANet as a valuable tool for aiding early detection and management of complex conditions in medical imaging and computer vision. Further research and validation are warranted to establish MRANet as a reliable tool in pediatric cardiology practice. ©2024 IEEE.