Using convolutional neural networks to classify ultrasound images of the carotid artery wall

Abstract: Cardiovascular diseases are very prominent in western countries. This thesis examines three different convolutional neural networks (CNN) and their abilities to classify ultrasound images of the carotid artery wall to a risk level of atherosclerosis. The models are LeNet, VGG16, and VGG19. They are evaluated in terms of accuracy, precision, recall and f-score. Two datasetsare used. One with three classes and one with two. This means every metric except accuracy is calculated for each individual class. The results from the experiments are also put into context to other studies within the area of medical image analysis. LeNet is shown to be the only model that actuallymanages to find features that distinguish each class from one another. Using a percentage split of 66/33 it reaches a 44.57% accuracy for the three class dataset, and 69.41% accuracy for the two class. Cross validation gives 42.08% and 69.13% accuracy respectively. The results show LeNet to be the better model. But a further examination deduces that this does not have to be the case. When making a comparison the other studies, it becomes obvious that the simple nature of the experiments are not enough to draw a decisive conclusion.