Whole Heart Segmentation with Deep Neural Networks

<p dir="ltr">This thesis presents an approach for whole heart segmentation in computed tomography (CT) scans using convolutional neural networks (CNNs), to produce a segmentation map for cardiac substructures from a given CT scan volume. The objective of this research is to develop techniques that enhance the performance of a fully automatic segmentation model, called <i>U-Net</i>, both in terms of the accuracy of the predicted segmentations and the generalisability to unseen data. The study was conducted on the training set of the Multi-modality Whole Heart Segmentation (MM-WHS) dataset, which has been used to evaluate the performance of whole heart segmentation models.</p><p dir="ltr">The first contribution is the application of rotation augmentation to the cardiac volumes in the dataset. Specifically, slice-wise rotations are applied within a ±10° range along the axial axis, with a probability of <i>p </i>= 0.33. This rotation setting is shown to improve segmentation accuracy, particularly in boundary areas, and enhances the generalisability across different cross-validation folds.</p><p dir="ltr">The second contribution is the introduction of the adaptive composite loss, which combines binary cross-entropy loss (BCE) and Boundary DoU loss. The weight of each constituent loss function is dynamically adjusted based on the performance of the current weight composition. When used in conjunction with the proposed rotation augmentation, the adaptive BCE-BoundaryDoU composite loss applied to the U-Net model yields strong performance, setting a robust benchmark.</p><p dir="ltr">In addition, the proposed integration of the Swin Transformer with convolution based residual blocks demonstrates performance comparable to the U-Net model with rotation augmentation and adaptive composite loss, but does not surpass it. This is particularly evident in small regions of interests (ROIs) at the apex and the basal slices of the cardiac volumes. Given the strong performance of convolutional neural networks, we propose future work on recurrent convolutional neural networks (RCNNs) to better utilise spatial information across slices and further improve segmentation accuracy.</p>