Carotid atherosclerosis is a major cause of cerebrovascular events such as is-chemic strokes. B-mode ultrasound (US) is a safe, widely available, and cost-effective imaging tool. However, inter-operator variability in the interpretation of ultrasound images may lead to misdiagnosis. In this work, we developed a fully automatic convolutional neural network (CNN)-based method to 1) detect and segment plaques from US (longitudinal/transverse), 2) quantify plaque parameters: thickness (T), total plaque area (TPA), and total plaque volume (TPV), and 3) classify symptomatic plaque. Patients (n=141) with severe carotid atherosclerotic plaques underwent US examination prior to carotid endarterectomy. A radiologist annotated US images (467), longitudinal (326) and transverse (141). First, a U-Net-based-semantic segmentation model was trained to classify plaque pixels. Second, an automatic PCA-based method was used for plaque quantification. Third, six different CNN-based classifiers were trained to classify plaque instability and symptomatic into four classes. The accuracy of the plaque segmentation method is evaluated in comparison to manual segmentation performed by a radiologist. The Intersection Over Union between manual and algorithm-based segmentation was 0.80±0.18 (mean ± standard deviation) on 46 unseen US images. The mean absolute error of (∆T), (∆TPA), and (∆TPV) were 0.22 mm, 0.73 mm2, and 0.89 mm3. The total time (detection/quantifications) was <1 second per im-age. The best performing model for plaque instability and symptomatic classification achieved accuracy of 77(%). |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.