Oral: A Sequential Transfer Learning Approach for Congenital Heart Defect Diagnosis from Ultrasound Videos

This study presents a methodology for analyzing fetal cardiac defects using ultrasound videos. The dataset consists of both normal and abnormal cases, focusing on four cardiac views: four-chamber (4C), three-vessel (3V), left ventricular outflow tract (LVOT), and right ventricular outflow tract (RVOT). Fetal heart images were extracted, preprocessed, and annotated in four views and further labelled as' Normal' or 'Abnormal'. The classification framework employs a sequential approach. 4C images are classified first, followed by LVOT and pairwise discrimination between RVOT and 3V. These cardiac views are further classified as 'Normal' or 'Abnormal'. The heterogeneity of Congenital Heart Defects (CHD) makes detection and classification challenging. A convolutional neural network (CNN) was employed for training using the Mask R-CNN model. The dataset includes major CHDs such as ventricular septal defect (VSD), atrioventricular septal defect (AVSD), tetralogy of Fallot (TOF), total anomalous pulmonary venous connection (TAPVC), Ebstein's anomaly, hypoplastic left heart syndrome (HLHS) and transposition of great arteries (TGA). The models achieved a diagnostic accuracy for unseen data, with detection rates of 4C-99%, LVOT-98%, RVOT-98%, and 3V-98%. Further classification into 'Normal' or 'Abnormal' yielded accuracies of 95%, 92%, 98%, and 98%, respectively. These results highlight the potential of the approach to assist in the prenatal diagnosis of CHD during examination.