Elvin Alcevska presenting her master thesis on 3 November

Welcome to Elvin Alcevska´s master thesis presentation with the title “Segmentation of the left ventricle of the heart in 2D ultrasound images using convolutional neural networks” on Thursday, November 3, 2016 at 13.15 pm. in the EDIT conference room (room 3364), Hörsalsvägen 11, Chalmers. Examiner: Fredrik Kahl

When an experienced cardiologist studies a 2D ultrasound image of the heart he or she can manually segment (i.e. outline) the correct border of the left ventricle and thereafter for example calculate the ventricle volume. The information provided from the segmentation (i.e. the delineation) is used in modern cardiovascular medicine for diagnosis, disease progression, schedule and choice of treatment et cetera. New guidelines on how to segment the left ventricle were published 2015, but the new guidelines are still not general knowledge among cardiologists. Therefore, an automatic segmentation method based on the new guidelines is needed. In this thesis, an automatic segmentation method following the new guidelines is implemented. The method includes pixel classification using a multilayer convolutional neural network, where supervised learning is used as learning method. The network output is a probability map indicating the probability of each image pixel belonging to the left ventricle. Post-processing methods such as multi-atlas segmentation and graph cuts are used to obtain the final segmentation. The data consists of 2D ultrasound images with a 2-chamber view of the heart plus a corresponding manual delineation of the left ventricle. 30 images were used to train and validate the network, and 6 test images where used to evaluate the final segmentation framework. The segmentation results were evaluated by calculating the Dice coefficient for the test images, i.e. measuring the similarity between the automatically segmented area and the corresponding manual delineation. The average Dice coefficient for the test images was 0.82 when thresholding was used to obtain the final segmentation. The Dice coefficient increased to 0.92 when the network output was restricted to a region of interest defined by a multi-atlas approach and simple thresholding was replaced by graph cuts.