Automatic 3D Segmentation in CT images of Congenital Heart Defects using Deep Learning

Abstract: 3D segmentations of hearts with congenital heart defects are routinely used today. They are used to study hearts and to prepare before surgery which makes them an important part of patient care. The 3D segmentations are usually created manually, which is a time-consuming process. By automating this process, less time can be spent creating these models, and more time can be spend on patient care. The purpose of this project was to create a fully automatic method that created 3D segmentation of CT images of hearts with congenital heart defect. The aim was to create this method using deep learning, more specifically the U-net structure. This was done by manually segment images, which were used as ground truth segmentations during network training. The classes used for training were the Blood Pool, Bone and Pericardium, which were segmented separately. The ground truth segmentations were done by the author of this thesis, and evaluated by others. The hyper parameters of the network were adjusted to fit this project and data augmentations was applied to the training data to extend the training data set. Training was done on 21 patients, and testing was done on seven. The class used for performance evaluation was the Blood Pool and the final output from the network was the Cardiac Model, which included segmentation of the Blood Pool and Pericardium. The final network in this project segmented the Blood Pool, that scored an average Dice score of 93,5% (±2,9) and a Jaccard Index of 88,0% (±5,0), when evaluated on seven test images. The results of this thesis shows that creating automatic 3D segmentations of congenital heart defects is possible with the use of deep learning. It also shows that even if the automatically created segmentations are imperfect, they can be used clinically to improve the patient care.