TMN-Net: A Hybrid 2.5D Multi-Branch Transformer Network for Coronary Artery Segmentation in Cardiac Diagnosis
Year: 2025
Type: article
Abstract: Coronary artery disease (CAD) continues to be a leading cause of global mortality, necessitating accurate and efficient diagnostic tools for early detection and treatment planning. Coronary computed tomography angiography (CCTA) serves as a non-invasive imaging modality critical for visualizing coronary anatomy. However, manual coronary artery segmentation (CAS) is labor-intensive, prone to inter-observer variability, and limited in scalability due to the anatomical complexity of coronary arteries, particularly in small and distal branches. Current deep learning approaches, including 2D and 3D segmentation frameworks, exhibit trade-offs between computational efficiency and spatial context awareness. While traditional methods such as U-shaped architectures provide significant advancements, they fall short in capturing multi-scale contextual information and handling imaging artifacts. This study proposes the Transformer Multi-branch Network (TMN-Net), an innovative hybrid framework that integrates augmented 2.5D and transformer-based 3D processing pathways. TMN-Net leverages dual transformer blocks to capture global context and long-range dependencies, while its multi-scale feature fusion and attention mechanisms enhance the extraction of fine-grained details, enabling robust segmentation of intricate coronary structures. Novel post-processing fusion strategies, including slice fusion and maximum fusion methods, further refine boundary alignment and mitigate false negatives. TMN-Net outperforms existing models, achieving Dice Similarity Coefficient (DSC) improvements of 1.63% and 3.22% over CAS-Net and MC-DFN, respectively, on the CAT08 dataset, with similar gains on the ASOCA dataset. Qualitative evaluations and attention heatmaps highlight the model’s ability to preserve vessel continuity, capture distal branches, and ensure accurate boundary delineation. These advancements confirm TMN-Net’s efficacy in handling the anatomical complexity and imaging variability inherent in coronary artery segmentation.
Source: IEEE Access
Authors: Bo Zhao, Jianjun Peng, Kai Zhang, Yongyan Fan, Ce Chen, Yang Zhang
Institutions: Capital Medical University, Beijing Shijitan Hospital
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W4411949189