PFHFF: Parameter-Friendly Hierarchical Feature Fusion Module for Enhanced Epicardium Segmentation in CT Images

View at publisher Summarise with VIRAT
Authors: Ibuki Naka, Jain Rahul Kumar, Yin-Hao Li, Xianhua Han, Atsuyuki Wada, Atsunori Kashiwagi, Yutaro Iwamoto, Kiyosumi Maeda, Yen-Wei Chen, Yuji Tezuka
Cited by: 0
References: 15

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Epicardial adipose tissue (EAT) is strongly correlated with diabetes risk, but its diagnosis is easily affected by external factors such as diet. CT-based quantification of EAT requires accurate epicardium segmentation, which remains challenging due to low contrast and thin boundaries. Conventional CNN-based models like DenseUNet effectively capture local details but fail to model global context, leading to over-extraction and irregular contours. To address this, we propose a dual-encoder network combining MedViT and CNN, integrated through a novel Parameter-Friendly Hierarchical Feature Fusion (PFHFF) module. PFHFF aligns multi-scale features from both encoders and fuses them using element-wise multiplication, functioning as a lightweight gating mechanism to emphasize regions jointly activated by global and local features. This design enhances boundary delineation while maintaining computational efficiency. Experiments show that our method achieves smoother contours closer to the ground truth and improves F1 score by over 3.5% and IoU by 2.4% compared to conventional approaches.

Fields of Study & Subjects

Medical Image Segmentation Techniques Cardiovascular Disease and Adiposity Cardiac Imaging and Diagnostics

Publication details

Journal / source
International Conference on Consumer Electronics
Pages
1–4
Publication date
3 February 2026
Language
EN