Abstract 4146002: Advancing Cardiac Function Assessment with a Real-Time, Lightweight Deep Learning Network for Precise Left Ventricular Internal Diameter Measurement

The purpose of this study is to develop a real-time, lightweight deep learning network to accurately detect the Left Ventricular Internal Diameter (LVID) for enhanced precision and efficiency in cardiac function assessment. Our approach redefines the task of locating LVID endpoints by framing it as a classification problem along two axes, thereby significantly reducing the computational burden. This novel approach ensures high accuracy and enables real-time deployment in clinical settings. We conducted extensive training using EchoNet-LVH, a primary dataset comprising 12,000 meticulously labeled echocardiogram videos sourced from diverse patients exhibiting various cardiac conditions. Furthermore, we evaluated our model on two external datasets and collected specialized datasets to assess its generalizability: a pediatric dataset consisting of 892 key frames from 138 children to evaluate accuracy in young patients with distinct cardiac dynamics, and an adult dataset encompassing 298 key frames from 39 adults representing a wide range of cardiac health statuses for testing effectiveness across diverse demographics. On the EchoNet-LVH dataset, the model exhibited exceptional accuracy by achieving a mean absolute error (MAE) of 2.39mm and a mean absolute percentage error (MAPE) of 6.5%, thereby establishing a new benchmark for precision in left ventricular internal diameter (LVID) measurement. Notably, the model also demonstrated excellent generalizability across diverse patient populations in external datasets, yielding MAPEs of 6.4% in the pediatric dataset and 5.6% in the adult dataset. Moreover, its efficiency was underscored by an impressive forward inference speed of up to 401.82 FPS on advanced GPU, representing a substantial improvement over previous approaches. This accelerated processing capability enables real-time analysis, which is crucial for clinical applications as it facilitates immediate diagnostic feedback essential for effective patient management and care. The proposed model significantly enhances the efficiency of LVID measurements while maintaining a high level of precision, enabling real-time and automated assessments of cardiac function that are highly suitable for clinical application. In addition, its robust performance suggests a strong potential for widespread adoption in routine cardiac evaluations. Our approach provides rapid, accurate assessments of LVID makes it an invaluable tool for cardiac function evaluation.