Scholarlink

<p>Coronary arteries are the vessels which supply oxygen rich blood to the heart. Due to the build up of plaques along the walls of vessels that can obstruct the flow of blood and result in a cardiac event. Segmentation of these coronary arteries is critical for plaque analysis. The non-invasive imaging method such as Coronary Computed Tomography Angiography (CCTA) is frequently used for visual examination and diagnosis of coronary artery diseases. In this study, we propose an automated coronary artery segmentation approach based on deep learning. The method leverages the UNetr model, a Transformer-based encoder–decoder architecture. Our experiments are conducted on a subset of 460 3D CCTA volumes selected from the ImageCAS dataset. To address the significant class imbalance between the coronary arteries and background tissue and to enhance the segmentation of smaller vascular structures, we utilize the Tversky loss function during training. The Experiments are conducted using patch sizes of 16 and 32. Our proposed model has achieved a Dice Similarity Coefficient (DSC) of 0.82 for patch size 16 and 0.76 for patch size 32 demonstrating its efficacy in coronary artery segmentation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transformer Based Coronary Artery Segmentation Using Tversky Loss Function on 3D CCTA Images

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Kavita Chachadi,
  • S. R. Nirmala

摘要

Coronary arteries are the vessels which supply oxygen rich blood to the heart. Due to the build up of plaques along the walls of vessels that can obstruct the flow of blood and result in a cardiac event. Segmentation of these coronary arteries is critical for plaque analysis. The non-invasive imaging method such as Coronary Computed Tomography Angiography (CCTA) is frequently used for visual examination and diagnosis of coronary artery diseases. In this study, we propose an automated coronary artery segmentation approach based on deep learning. The method leverages the UNetr model, a Transformer-based encoder–decoder architecture. Our experiments are conducted on a subset of 460 3D CCTA volumes selected from the ImageCAS dataset. To address the significant class imbalance between the coronary arteries and background tissue and to enhance the segmentation of smaller vascular structures, we utilize the Tversky loss function during training. The Experiments are conducted using patch sizes of 16 and 32. Our proposed model has achieved a Dice Similarity Coefficient (DSC) of 0.82 for patch size 16 and 0.76 for patch size 32 demonstrating its efficacy in coronary artery segmentation.