A Deep Learning Framework Integrating Multi-View Morphologic and Hemodynamic Features for Pericardial Disease Classification

Sihyeon Jeong, Jina Lee, Yeonggul Jang, Jaeik Jeon, Dawun Jeong, In Tae Moon, Seung-Ah Lee, Yeonyee E. Yoon, Hyuk-Jae Chang

Keywords: echocardiography, deep learning, multi-view, fusion

TL;DR: We propose a lightweight, multi-view deep learning framework that fuses echocardiographic video features with IVC-derived indicators to enable more accurate classification of pericardial diseases.

Abstract: Pericardial diseases require accurate and timely diagnosis, yet echocardiography analysis typically depends on expert interpretation. In this study, we introduce a novel two-stage deep learning framework designed to improve diagnostic accuracy by integrating multi-view echocardiographic video data— PLAX (parasternal long-axis), A4C (apical 4-chamber), modified A4C (modified apical 4-chamber), S4C (subcostal 4-chamber)— with hemodynamic indicators derived from IVC (inferior vena cava). In Stage 1, a tailored spatiotemporal convolutional neural network (CNN) effectively captures dynamic cardiac patterns, enabling precise classification of pericardial effusion severity and pericardial thickening (accuracy 0.921). In stage 2, embedding and integration of IVC-derived hemodynamic features substantially enhance sensitivity for detecting clinically significant cases (positive 0.969, negative 0.618). Our findings highlinght the clinical benefit of combining spatiotemporal echocardiographic features with functional indicators, potentially reducing reliance on subjective interpretation while ensuring compatibility with existing clinical workflows.

Submission Number: 45