Automatic diagnosis of common carotid artery disease using different machine learning techniques

Common carotid artery (CCA) diagnosis is very important for carrying out an assessment of the severity of vascular disease and being able to suggest treatment solutions, whether with careful surgical planning or even an interventional radiological surgery. Early diagnosis of carotid atherosclerosis is an essential step in preventing stroke from occurring. This is the motivation for us to develop a novel Computer-Aided Diagnosis (CAD) system for CCA disease diagnosis. Our novel CAD system contains four phases named: segmentation, localization, intima-media thickness (IMT) measurement, and classification of the CCA as normal and abnormal. Each phase in our integrated system has its role and novelty contribution that distinguishes it from any previous studies and researches. These roles and contributions of all phases will be discussed later in this paper. These phases have been applied for the CCA in transverse and longitudinal sections to help in the early diagnosis of atherosclerosis providing a complete diagnosis approach. The CCA has been localized in the transverse section images based on a deep learning technique called faster regional proposal convolutional neural network (Faster R-CNN). The IMT measurement of the CCA has been accomplished in a longitudinal section based on edge detection techniques. The CCA-lumen segmentation has been made in a longitudinal section using active contour criteria. The CCA longitudinal section has been classified as normal and abnormal using the transfer learning of the pre-trained convolutional neural network (CNN) called AlexNet. Experiments have been performed on three different ultrasound image datasets that were manually collected. The comparison between our suggested localization phase circles and the clinician’s delineations shows an average Jaccard similarity of 90.86% with an accuracy of 97.5%. The mean ± standard deviation (SD) of our method and the experts for IMT measurements are 0.7573 ± 0.52 mm and 0.7604 ± 0.52, respectively. The obtained classification results show 100% for specificity, sensitivity, and accuracy. These results, show the superiority of the proposed system over other systems in the literature.