Automatic Segmentation of the Left Ventricle Through the Cardiac Cycle in Pediatric Echocardiography Videos Using SegFormer Architecture

Echocardiography generates real-time results, aiding in examination and diagnosis. It is widely used for detecting congenital heart disease (CHD), evaluating risk, and guiding treatment strategies in pediatric cardiology. However, the complexity of these images makes their interpretation and analysis challenging, often leading to inter-observer variability. This research aims to develop an automated left ventricle (LV) segmentation method throughout the full cardiac cycle for pediatric echocardiography videos using a semantic Transformer model known as SegFormer. The goal is to support the analysis of clinical imaging techniques. In recent years, semantic Transformers have demonstrated significant effectiveness in segmentation tasks, making them highly suitable choice for this application. To achieve accurate LV segmentation through the cardiac cycle, the SegFormer model is trained using the EchoNet-Peds dataset, which consists of annotated pediatric echocardiography videos. The initial training phase includes segmenting the left ventricle images at the end of systole and the end of diastole, with performance evaluated based on accuracy, mean absolute error (MAE), recall and Dice score metrics to compare with other pediatric segmentation methods. As a final result, this research produces segmented left ventricle videos throughout the cardiac cycle for different pediatric echocardiography videos.Clinical RelevanceBy applying a semantic Transformer to pediatric echocardiography for automated LV segmentation, the quantification of key cardiac parameters such as ejection fraction, end of diastole, and end of systole is improved, leading to greater accuracy and providing more valuable information for medical staff. Consequently, a tool capable of efficiently processing large volumes of data can significantly facilitate and support the diagnosis process for pediatric patients.