Vision Transformer-Based View Classification for Pediatric Cardiac Septal Defect

Abstract

Congenital Heart Disease (CHDs), particularly Cardiac Septal Defects (CSDs), remains a major cause of morbidity in children. This study proposes a deep learning–based approach using Vision Transformer (ViT) to automatically classify view echocardiographic images as a fundamental step toward improving diagnostic accuracy in pediatric CSDs. A total of 158 videos for the four-chamber view (4CV), five-chamber view (5CV), and subcostal (SC) views; 161 videos for the parasternal long-axis (PLAX) view; and 165 videos for the parasternal short-axis (PSAX) view were used in this study, and converted into 48,000 image frames. Four ViT variants (ViT-B/16, ViT-B/32, ViT-L/16, and ViT-L/32) were evaluated under identical experimental settings using 100 training epochs, the AdamW optimizer, and consistent input configurations. Model performance was assessed using accuracy, precision, and recall metrics. Experimental results showed that ViT-L/32 achieved the best performance with an accuracy of 92.20%, accompanied by superior precision and recall compared to other models. The proposed ViT-based framework provided an effective and reliable solution for automatic echocardiographic view classification, supporting early and accurate diagnosis of CSD