Physiological Model Based Deep Learning Framework for Cardiac TMP Recovery

• Published in: Medical Image Computing and Computer Assisted Intervention - MICCAI 2022 Vol. 13432; pp. 433 - 443
• Main Authors: Huang, Xufeng, Yu, Chengjin, Liu, Huafeng
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Deep learning; Kalman filtering; Physiological model; TMP
• ISBN: 3031164334; 9783031164330
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-16434-7_42

Recovering cardiac transmembrane potential (TMP) from body surface potential (BSP) plays an important role in the noninvasive diagnosis of heart diseases. However, most current solutions for TMP recovery are typically proposed and designed to follow a static mapping paradigm between TMP and BSP, which ignores the inherent dynamic activation process of cardiomyocytes during the cardiac cycle. In this paper, we propose to introduce the physiological information of this dynamic activation process in the objective functions. Based on this, we further establish a physiological model based deep learning framework for cardiac TMP recovery. First, the objective functions of our physiological model are deduced via a two-variable diffusion-reaction system, where the static mapping and the dynamic activation process of cardiomyocytes are jointly modeled. Then, a data-driven Kalman Filtering network (KFNet) is adopted to solve the proposed objective functions. Specifically, the KFNet consists of two components: a state transfer network (SSNet) is employed for directly predicting the prior estimation; furthermore, a Kalman gain network (KGNet) is employed for adaptively learning the gain coefficients. In our experiments, the proposed physiological model is verified on the 1200 simulated subjects. The quantified analysis shows the proposed method can accurately recover the TMP, with the low LE values 10.5 for the ectopic pacing location task and the high SSIM values 0.75 for the myocardial infarction detection task. These powerful performances completely verify the effectiveness of our model.