Deep learning for the detection and characterization of the carotid artery in ultrasound imaging

Abstract

[en] Atherosclerosis is the main process causing most Cardio Vascular (CV) diseases. The measurement of Intima Media Thickness (IMT) in artery ultrasound images can be used to detect the presence of atherosclerotic plaques, which may appear in several territories of the artery. Moreover, it is well known that disruption of atherosclerotic plaque plays a crucial role in the pathogenesis of CV events. Several works have tried to automatize the detection of the IMT and the classification of the plaque by its composition. Traditionally, the methods used in the literature are semi-automatic. Furthermore, very little work has been done using Deep Learning approaches in order to solve this problems. In this thesis, we explore the effectiveness of Deep Learning techniques in attempting to automatize and improve the diagnosis of atheroma plaques. To achieve so we tackle the following problems: ultrasound image segmentation and plaque tissue classification. The techniques applied in this work are the following. For the segmentation of the common carotid artery IMT we replicate a state of the art Fully Convolutional Network approach and explore the implementation of a trained network to another dataset. Regarding the plaque classification problem, we explore the performance of Convolutional Neural Networks as well with two baseline methods. These techniques are applied on two datasets: REGICOR and NEFRONA. These datasets are provided by two research groups of IMIM and IRBLleida in collaboration in a larger project with the UB. A data exploration analysis is also presented on the patient’s data of NEFRONA to justify the importance of detecting the atherosclerotic plaques and thus the techniques we explore.