Diagnosis of Coronary Heart Disease Through Deep Learning-Based Segmentation and Localization in Computed Tomography Angiography

Abstract

Coronary heart disease (CHD), a leading cause of global mortality, requires precise and early diagnosis for effective intervention. Coronary computed tomography angiography (CCTA) has emerged as a non-invasive modality for detailed coronary artery visualization; however, automatic and accurate segmentation of coronary structures from CCTA images remains challenging. Conventional convolutional neural networks (CNNs), despite their success in medical imaging, face limitations in capturing the complex, long-range dependencies in coronary artery images due to their localized receptive fields. Vision transformers, with their self-attention mechanisms, offer a global perspective, yet demand extensive data and computational resources, making them less adaptable for the often limited medical imaging datasets. This research addresses these challenges by proposing TransCHD, a hybrid CNN-Transformer architecture developed for coronary artery segmentation in CCTA. TransCHD incorporates a Contextual Representation Learning (CRL) module and a Spatially-Aware Feature (SAF) module, enabling both local feature extraction and global contextual awareness within a unified architecture. The CRL module mitigates spatial continuity disruptions caused by standard patch-based transformers, while the SAF module enhances spatial locality and preserves fine-grained anatomical details essential for accurate segmentation. The segmentation outcomes are clinically significant as they provide quantitative assessments of arterial stenosis, plaque characterization, and ischemia-prone regions, supporting risk assessment and treatment planning. Trained and evaluated on the CorArtTS2020 dataset, TransCHD achieved superior performance compared to state-of-the-art CNN- and transformer-based models, with a Dice score of 0.81 and an Intersection over Union (IoU) of 0.65. Results show that our proposed TransCHD is effective in CCTA segmentation.

Keywords

• Convolutional neural networks
• vision transformer
• cardiovascular segmentation
• deep learning