U-Net based automatic carotid plaque segmentation from 3D ultrasound images

Abstract

Ultrasound image assessment plays an important role in the diagnosis of carotid artery atherosclerosis. The segmentation of plaques from carotid artery ultrasound images is critical for the atherosclerotic diagnosis. In this paper, a novel automatic plaque segmentation method is presented based on U-Net deep learning network which allows to train the network end-to-end for pixel-wise classification. A large number of labeled examples are required for traditional supervised learning techniques as to obtain the global optimization. However, in this task, it is unavailable to obtain so many labeled examples since manually segmentation of plaques is a time-consuming task and its reliability relies to the experience of experts. In order to solve the problem of lack of labeled samples, an unsupervised learning technique, the deep convolutional encoder-decoder architecture, was proposed to pre-train the parameters of U-Net by amount of unlabeled data. Then the parameters learned from the deep convolutional encoder-decoder network were applied to initialize a U-Net from the labeled images for fine-tuning. Algorithm accuracy was examined on the common carotid artery part of 26 3D carotid ultrasound images (34 plaques) by comparing the results of our algorithm with manual segmentations and the Dice similarity coefficient (DSC) is 90.72±6.2% which was better than the previous level set method with the DSC of 88.2±8.3%. The automatic method provides a more convenient way to segment carotid plaques in 3D ultrasound images.