Aortic Segmentation in Real-Time Flow Exercise Cardiac Magnetic Resonance Images using Convolutional Neural Networks

Abstract: Background: Cardiovascular disease is common and it is therefore important to have tools to investigate pathophysiology and normal physiology of the cardiovascular system. One such tool is using cardiovascular magnetic resonance in order to measure blood flow in vessels, particularly while the participant imaged is exercising. This enables studying the blood flow during exercise which is especially important for people with diseases which only manifest during physical stress. In order to measure the blood flow of the aorta during multiple heart cycles it is necessary to delineate the images depicting the aorta. Typically this is done manually, but since this is a time-consuming process it could benefit from a segmentation algorithm. Aims: This thesis’ aims to develop a semi-automatic segmentation model using a convolutional neural network. The training data consisted of real-time flow and gated images acquired during rest, and real-time flow images acquired during exercise were used for validation and testing. Method: The model studied in this thesis is a kind of convolutional neural network called U-Net. Multiple variations of the U-Net architecture were explored by varying the number of input channels, the encoder depth of the network and using different loss functions. Additionally, the effects of using L2 regularization and adaptive learning rates during training were investigated. The performance of the models are measured using the accuracy, intersection over union and mean BF score of the segmented aorta class. Additionally, calculations of the sequence volumes based on the segmentations of the flow images were compared to the ground truth sequence volumes. Results: In this project, the best results were produced using a U-Net model with encoder depth 4 and using both a velocity-encoded and magnitude image as input. Binary cross-entropy was found to be the best choice for loss function, but further testing could be done using Dice. Using L2 regularization and adaptive learning rates did not improve the result. Conclusion: The networks developed in this thesis project prove that it is possible to train U-Net models on real-time flow and gated images acquired during rest in order to segment real-time flow images acquired during exercise. However, more studies into inter- and intra-observer variability of the data as well as improving the sequence volume calculations are needed before the algorithm could be of clinical use.