Machine Intelligence Laboratory, College of Computer Science, Sichuan University, Chengdu, Sichuan, PR China.
Department of Cardiology, West China Hospital, Sichuan University, Chengdu, Sichuan, PR China.
Comput Biol Med. 2024 Oct;181:109045. doi: 10.1016/j.compbiomed.2024.109045. Epub 2024 Aug 24.
Coronary artery segmentation is crucial for physicians to identify and locate plaques and stenosis using coronary computed tomography angiography (CCTA). However, the low contrast of CCTA images and the intricate structures of coronary arteries make this task challenging. To address these difficulties, we propose a novel model, the DFS-PDS network. This network comprises two subnetworks: a discriminative frequency segment subnetwork (DFS) and a position domain scales subnetwork (PDS). DFS introduced a gated mechanism within the feed-forward network, leveraging the Joint Photographic Experts Group (JPEG) compression algorithm, to discriminatively determine which low- and high-frequency information of the features should be preserved for latent image segmentation. The PDS aims to learn the shape prototype by predicting the radius. Additionally, our model has the consistent ability to guarantee region and boundary features through boundary consistency loss. During training, both subnetworks are optimized jointly, and in the testing stage, the coarse segmentation and radius prediction are produced. A coronary-geometric refinement method refines the segmentation masks by leveraging the shape prior to being reconstructed from the radius map, reducing the difficulty of segmenting coronary artery structures from complex surrounding structures. The DFS-PDS network is compared with state-of-the-art (SOTA) methods on two coronary artery datasets to evaluate its performance. The experimental results demonstrate that the DFS-PDS network performs better than the SOTA models, including Vnet, nnUnet, DDT, CS-Net, Unetr, and CAS-Net, in terms of Dice or connectivity evaluation metrics.
冠状动脉分割对于医生使用冠状动脉计算机断层血管造影术 (CCTA) 识别和定位斑块和狭窄至关重要。然而，CCTA 图像对比度低，冠状动脉结构复杂，使得这项任务具有挑战性。为了解决这些困难，我们提出了一种新的模型，即 DFS-PDS 网络。该网络由两个子网组成：一个是鉴别频率分割子网 (DFS)，另一个是位置域尺度子网 (PDS)。DFS 在前馈网络中引入了门控机制，利用联合图像专家组 (JPEG) 压缩算法，有鉴别地确定特征的低频和高频信息中哪些应该保留用于潜在图像分割。PDS 旨在通过预测半径来学习形状原型。此外，我们的模型通过边界一致性损失具有保证区域和边界特征的一致能力。在训练过程中，两个子网是联合优化的，在测试阶段，生成粗分割和半径预测。冠状动脉几何细化方法通过利用形状先验，从半径图中重建，细化分割掩模，从而降低从复杂周围结构中分割冠状动脉结构的难度。将 DFS-PDS 网络与两个冠状动脉数据集上的最先进 (SOTA) 方法进行比较，以评估其性能。实验结果表明，DFS-PDS 网络在 Dice 或连通性评估指标方面优于 SOTA 模型，包括 Vnet、nnUnet、DDT、CS-Net、Unetr 和 CAS-Net。
