Bi-directional semi-3D network for accurate epicardial fat segmentation and quantification using reflection equivariant quantum neural networks.
Authors
Affiliations (2)
Affiliations (2)
- Department of Networking and Communications, SRM Institute of Science & Technology, Kattankulathur, Tamil Nadu, India. Electronic address: [email protected].
- Department of Electronics and Communication Engineering, P.S.R. Engineering College, Sivakasi, Tamil Nadu, India. Electronic address: [email protected].
Abstract
The process of detecting and measuring the fat layer surrounding the heart from medical images is referred to as epicardial fat segmentation. Accurate segmentation is essential for assessing heart health and associated risk factors. It plays a critical role in evaluating cardiovascular disease, requiring advanced techniques to enhance precision and effectiveness. However, there is currently a shortage of resources made for fat mass measurement. The Visual Lab's cardiac fat database addresses this limitation by providing a comprehensive set of high-resolution images crucial for reliable fat analysis. This study proposes a novel method for epicardial fat segmentation, involving a multi-stage framework. In the preprocessing phase, window-aware guided bilateral filtering (WGBR) is applied to reduce noise while preserving structural features. For region-of-interest (ROI) selection, the White Shark Optimizer (WSO) is employed to improve exploration and exploitation accuracy. The segmentation task is handled using a bidirectional guided semi-3D network (BGSNet), which enhances robustness by extracting features in both forward and backward directions. Following segmentation, quantification is performed to estimate the epicardial fat volume. This is achieved using reflection-equivariant quantum neural networks (REQNN), which are well-suited for modelling complex visual patterns. The Parrot optimizer is further utilized to fine-tune hyperparameters, ensuring optimal performance. The experimental results confirm the effectiveness of the suggested BGSNet with REQNN approach, achieving a Dice score of 99.50 %, an accuracy of 99.50 %, and an execution time of 1.022 s per slice. Furthermore, the Spearman correlation coefficient for fat quantification yielded an R<sup>2</sup> value of 0.9867, indicating a strong agreement with the reference measurements. This integrated approach offers a reliable solution for epicardial fat segmentation and quantification, thereby supporting improved cardiovascular risk assessment and monitoring.