A Transfer Learning and Image Augmentation Method for Carotid Artery Vulnerable Plaque Segmentation in Ultrasound Images

Abstract

Evaluating carotid artery plaque by ultrasound technique is a crucial factor in the screening of atherosclerosis. However, vulnerable plaque segmentation remains a challenging task because of the heterogeneities of inter-plaques and intra-plaques, and obscure boundaries of plaques. In this paper, we propose an automated HRU-Net transfer learning method for segmenting carotid vulnerable plaques. Based on the U-Net encoder-decoder paradigm, cross-domain knowledge from natural images is transferred for plaque segmentation using pre-trained ResNet-50. Besides, a cropped blood vessel image augmentation is tailored for the limited images during only training. Moreover, to exploit the implicit discrimination feature of high-level plaque semantic information, the hybrid atrous convolutions are applied to obtain various scale long-range dependence of plaques for refining segmentation. 10-fold cross-validation using 40 carotid ultrasound images with severe stenosis shows that the proposed method yields a Dice value of 0.821, IoU of 0.701, Acc of 0.977, and modified Hausdorff distance (MHD) of 1.69 for the segmentation results, it outperforms some of the state-of-the-art CNN-based methods, and the improvements on metrics of Dice and MHD are statistically significant (p < 0.05). The proposed method can be used as an alternative for automatic vulnerable plaque segmentation in carotid ultrasound images clinically.