HRU-Net: A Transfer Learning Method for Carotid Artery Plaque Segmentation in Ultrasound Images

Abstract

Carotid artery stenotic plaque segmentation in ultrasound images is a crucial means for the analysis of plaque components and vulnerability. However, segmentation of severe stenotic plaques 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 plaques, using the limited images. The HRU-Net is based on the U-Net encoder–decoder paradigm, and cross-domain knowledge is transferred for plaque segmentation by fine-tuning the pretrained ResNet-50. Moreover, a cropped-blood-vessel image augmentation is customized for the plaque position constraint during training only. Moreover, hybrid atrous convolutions (HACs) are designed to derive diverse long-range dependences for refined plaque segmentation that are used on high-level semantic layers to exploit the implicit discrimination features. The experiments are performed on 115 images; Firstly, the 10-fold cross-validation, using 40 images with severe stenosis plaques, shows that the proposed method outperforms some of the state-of-the-art CNN-based methods on Dice, IoU, Acc, and modified Hausdorff distance (MHD) metrics; the improvements on metrics of Dice and MHD are statistically significant (p < 0.05). Furthermore, our HRU-Net transfer learning method shows fine generalization performance on 75 new images with varying degrees of plaque stenosis, and it may be used as an alternative for automatic noisy plaque segmentation in carotid ultrasound images clinically.

UltraAIGenomics: Artificial Intelligence-Based Cardiovascular Disease Risk Assessment by Fusion of Ultrasound-Based Radiomics and Genomics Features for Preventive, Personalized and Precision Medicine: A Narrative Review

“…Table 2 (Ref. [ 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 , 87 , 88 ]) shows the studies that used DL-based radiomics (covariates) to segment carotid B-mode ultrasound (cBUS). Most of the studies used UNet [ 79 , 80 , 81 , 82 , 89 , 90 ], UNet++ [ 83 ], and convolution neural network (CNN) [ 84 ] as classifiers and segmentation for the cIMT region in carotid scans.…”

“…The challenge with DL solutions is that they need optimization during training using hyperparameters [ 38 , 80 ]. DL-based training requires several epochs, the best learning rate, batch size, batch normalization, and dropout layers to avoid overfitting or generalization without memorization [ 198 , 199 ].…”

Advances in Intelligent Segmentation and 3D/4D Reconstruction of Carotid Ultrasound Imaging

“…Since dataset is difficult to obtain and making label needs expensive resources, Transfer learning and unsupervised learning have been applied to plaque segmentation. Yuan et al [36] used cross-domain knowledge for plaque segmentation by fine-tuning the pre-trained ResNet-50. In addition, hybrid atrous convolutions (HACs) were deisgned to derive different long-range dependencies for fine plaque segmentation, which can obtain more receptive fields to distinguish similar textures between plaques and speckle noise, and the model has good generalization.…”