Abstract

Info

Title:: A dual-encoder U-Net based algorithm for right ventricle MRI segmentation

Author(s):: DING Weibin1; JIANG Shaohua1; XU Ting1; HUANG Lijuan2; 1. College of Information Science and Engineering, Hunan Normal University, Changsha 410081, China 2. Hunan Intelligent Rehabilitation Robot and Auxiliary Equipment Engineering Technology Research Center, Changsha 410004, China

Keywords:: right ventricle segmentation cardiac magnetic resonance imaging feature repurposing multi-scale feature hybrid loss function

Abstract:: The accurate segmentation of the right ventricle is crucial for cardiac disease research, but its low contrast with surrounding tissues and complex structure make segmentation challenging. To address these issues, a dual-encoder segmentation model combining nested multi-scale feature fusion and feature repurposing modules is proposed. Specifically, the nested multi-scale feature fusion module captures boundary detail features through multi-scale dilated convolutions and reduces the semantic gap between the encoder and decoder using short skip connections, while the feature repurposing module enhances feature extraction ability by leveraging fine-grained features from shallow layers. Ablation experiments show that the inclusion of these two modules improves the Dice similarity coefficient of U-Net by 3.14%. On the ACDC dataset, the proposed model achieves a Dice similarity coefficient of 90.31% and a mean Hausdorff distance of 5.21 mm, outperforming other comparative models. Additionally, its generalization ability is validated on the M&Ms dataset. Experimental results demonstrate the excellent performance and robustness of the proposed model in right ventricle segmentation.