An Efficient Segmentation-Driven Measurement Framework for Rapid Diastolic Dysfunction Detection

Abstract

Diastolic dysfunction (DD) is closely associated with the development of heart failure, yet its echocardiographic assessment remains labor-intensive and highly dependent on operator experience. The temporal features of the left ventricle (LV) and left atrium (LA) provide critical information for characterizing diastolic function, and semantic segmentation offers a fast, noninvasive means to automatically extract such functional information for DD evaluation and screening. In this study, we propose an automated framework for DD detection based on semantic segmentation-derived temporal features. Compared with conventional assessment pipelines, the proposed framework reduces operator dependency and enables rapid, objective decision-making. To ensure reliable feature extraction throughout the cardiac cycle, we develop LSNet, a lightweight segmentation network for accurate delineation of the LV and LA. LSNet incorporates a split-attention mechanism to reduce computational complexity while preserving contextual interactions. Furthermore, an edge-weighted cross-entropy (EWCE) loss is introduced to improve boundary delineation performance of both atrial and ventricular structures. By jointly leveraging spatiotemporal features derived from both the LV and LA, the proposed framework achieved a classification accuracy of 86.80 % $\pm ~5.89$ % for DD identification in a cohort of 126 subjects. Overall, this method enables fast and effective DD identification from echocardiography, demonstrating its potential for DD screening.