Deep Learning for Carotid Plaque Segmentation using a Dilated U-Net Architecture

Abstract

Carotid plaque segmentation in ultrasound longitudinal B-mode images using deep learning is presented in this work. We report on 101 severely stenotic carotid plaque patients. A standard U-Net is compared with a dilated U-Net architecture in which the dilated convolution layers were used in the bottleneck. Both a fully automatic and a semi-automatic approach with a bounding box was implemented. The performance degradation in plaque segmentation due to errors in the bounding box is quantified. We found that the bounding box significantly improved the performance of the networks with U-Net Dice coefficients of 0.48 for automatic and 0.83 for semi-automatic segmentation of plaque. Similar results were also obtained for the dilated U-Net with Dice coefficients of 0.55 for automatic and 0.84 for semi-automatic when compared to manual segmentations of the same plaque by an experienced sonographer. A 5% error in the bounding box in both dimensions reduced the Dice coefficient to 0.79 and 0.80 for U-Net and dilated U-Net respectively.

Automatic localization of Common Carotid Artery in ultrasound images using Deep Learning

“…Unlike the other state of the arts that select the initial bounding box manually as Nirvedh H. Meshram et al [24] proposed a semi-automatic segmentation approach that involves the input of a sonographer to provide limited inputs, such as bounding boxes or seed points, to achieve reasonable segmentation output on these carotid B-mode images in patients with a substantial plaque and related shadowing artifacts. Destrempes et al [25] provided a system by which the user got a manual segmentation of the carotid plaque according to which motion prediction and the Bayesian model were used to approximate plaque boundary.…”

Multimodality carotid plaque tissue characterization and classification in the artificial intelligence paradigm: a narrative review for stroke application

“…Some of these are the U-Net, Autoencoders, recurrent neural networks (RNNs), long short-term memory (LSTM), Regionbased Convolutional Neural Networks (R-CNN), and Mask R-CNN. U-Net (a combination of encoders and decoders) plays an important role in the segmentation and classification processes (55,(161)(162)(163). Figure 10B shows the architecture of U-Net.…”

Segmentation of common and internal carotid arteries from 3D ultrasound images based on adaptive triple loss

“…As a 3D measurement, VWV has a comparable dynamic range as TPV, 16 and therefore, serves as a strong alternative for carotid disease quantification in placebo‐controlled clinical trials investigating the effects of medical and dietary interventions 13,17 ; (3) Plaque segmentation is more difficult to automate. Existing plaque segmentation algorithms require users to identify a region of interest (ROI) for segmenting each plaque (i.e., carotid arteries with multiple plaques would require ROI identification multiple times) 18,19 . The initialization process would be even more time‐consuming for 3D analysis, in which a user would be required to crop an ROI on multiple image slices encompassing each plaque.…”

“…Existing plaque segmentation algorithms require users to identify a region of interest (ROI) for segmenting each plaque (i.e., carotid arteries with multiple plaques would require ROI identification multiple times). 18,19 The initialization process would be even more time-consuming for 3D analysis, in which a user would be required to crop an ROI on multiple image slices encompassing each plaque.…”