基于改进TransUnet网络的血管内超声图像边界提取方法研究

血管内超声(IVUS)图像是观测血管内部结构的首选影像学手段，基于IVUS图像的内膜和中外膜边界的提取是实现冠脉粥样硬化精准诊断的前提和关键。针对IVUS图像结构复杂、对比度低、边界提取困难等问题，本研究提出一种基于改进TransUnet网络的分割方法。首先，针对IVUS图像边界提取难点，对边血管、血管分叉、导丝伪像、阴影等4种图像结构进行建模，并基于建模结果予以定向数据增强；而后，结合IVUS图像的环状结构分布特点，在TransUnet网络中提出了Polar-bias归纳偏置的策略，并对IVUS图像进行像素级的分类；最后，基于分类结果优化GVF snake模型的外力场，进而提取IVUS图像的内膜和中外膜边界。采用国际标准IVUS图像数据集(两组不同中心频率，共512幅图像)对算法进行评测，引入JMard距离(JM),Hausdorff距离(HD)和面积差异百分比(PAD)等3个评测指标，在数据集A中JM为0.87, HD为0.87, PAD为0.18,数据集B中JM为0.91, HD为0.25, PAD为0.08。实验结果表明，所提出的算法在两组数据集的内膜及中外膜提取问题中的表现均...

Research on the IVUS Border Detection Method Based on Improved TransUnet

Intravascular ultrasound (IVUS) is a popular imaging technique that can observe the internal structure of blood vessels. The extraction of lumen and media-adventitia border in the IVUS image is vitally important for the accurate diagnosis of coronary atherosclerosis. To solve the problems of complex structure, low contrast and difficult boundary detection in the IVUS images, a modified TransUnet network was proposed in this paper to achieve pixel-wise classification results. Firstly, to overcome the difficulty of IVUS image border detection, data augmentation strategy is performed based on four kinds of image structure models, including edge vessel, vessel bifurcation, guide wire artifact and shadow. Then, a position named Polar-bias was proposed in the TransUnet network. The Polar-bias combines the ring structure distribution characteristics of IVUS images. The modified TransUnet network was applied to classify the IVUS images in pixel-wise level. Finally, the classified results were employed to optimize the external force field of GVF Snake model, which was used to detect the lumen and media-adventitia border in IVUS images. Two public datasets with different ultrasonic imaging center frequencies (a total of 512 images) were used for testing and verification, and three evaluation indexes of JMard measure (JM), Hausdorff distance (HD) and percentage of area difference (PAD) were introduced. The proposed method has achieved JM (0.87), HD (0.87) and PAD (0.18) in dataset A, and JM(0.91), HD(0.25) and PAD(0.08) in dataset B. The experimental results demonstrated that the proposed method outperformed other state-of-the-art ones in the border detection of two datasets.