Utilizing Weighted Spatio-Temporal Information for Automated Transthoracic Echocardiography View Classification

Transthoracic echocardiography (TTE) is a critical imaging modality for assessing cardiac function and diagnosing cardiomyopathies. These videos, captured from various heart angles, must be accurately categorized for comprehensive diagnosis, making automated view classification essential. This study develops and evaluates multiple model architectures and feature fusion techniques to classify eleven standard TTE views from video input. We examine both spatial and temporal feature extraction approaches alongside a Dynamic Feature Fusion technique (DFF) that adaptively balances the importance of either feature in analyzing TTE studies. Our results highlight the superior performance of a model architecture utilizing an EfficientNet and Dilated Convolutional Networks, achieving a micro F1 Score of 0.951. Additionally, we leverage a custom cycle detector algorithm to partition a TTE sequence into individual cardiac cycles, i.e., heartbeats, providing consistent input for model training and enabling an ensemble technique for a more reliable prediction. Our findings present valuable insights into effective TTE analysis, identifying the most effective model architecture for echocardiogram interpretation and advancing clinical workflows.Clinical RelevanceThis approach improves echocardiogram view classification, aiding clinicians by streamlining work-flows and reducing errors. The method can also be applied to broader TTE data analysis tasks, such as disease classification, while offering insights into interpretable spatial and temporal patterns.