ADAPTING FACE DETECTION AND RECOGNITION MODELS FOR ECHOCARDIOGRAPHIC DEFECT DIAGNOSIS

Abstract

Accurate detection of atrial and ventricular septal defects (ASD, VSD) in pediatric echocardiography remains a clinical challenge, especially in resource-constrained settings. In this work, we propose a novel deep learning framework that adapts state-of-the-art face detection and recognition pipelines—RetinaFace and ArcFace—for medical image analysis. Our method comprises a two-stage architecture: (1) a RetinaFace-inspired module that detects and crops the cardiac septal region from transthoracic echocardiogram frames, followed by (2) a ResNet-50-based classification network trained with an ArcFace loss function to learn discriminative embeddings of defect types. We evaluate the system on a curated pediatric echocardiography dataset labeled as Normal, ASD, or VSD. The model achieves 94.4% classification accuracy and an AUC of 0.976 in a pairwise verification setting, outperforming conventional CNN classifiers. These findings suggest that embedding-based metric learning, as pioneered in face recognition, can be successfully translated to medical imaging tasks requiring fine-grained classification. These findings illustrate that face recognition pipelines can be effectively repurposed for medical image analysis, offering a promising new direction for computer-aided echocardiographic diagnosis of congenital defects.