Medical image registration (MIR) is the process of aligning images for comparison and analysis. Traditional image registration methods, which rely on optimisation over voxel intensities, are computationally expensive and often struggle in settings with large deformations, such in abdominal scans. In recent years, deep learning (DL) methods have demonstrated improved computational efficiency while offering comparable performance in MIR.

The broader clinical application of the project, of which this thesis forms a part, is planning and evaluation of minimal invasive liver surgery. This thesis investigates the integration of hepatic vessel landmarks within a deep learning-based pipeline for intra-patient liver registration on Computed Tomography (CT) scans. The objective is to incorporate anatomical knowledge into the registration process by generating graph-based representations of the hepatic vessels. Experiments were conducted with three deep learning models: a small Graph Convolutional Network (GCN), Dynamic Graph CNNs (DGCNNs), and PointNet. The models were trained on synthetic datasets.

The resulting registration had a Normalised Cross-Correlation (NCC) of 0.775, which is compared with a traditional registration method called elatsix that had a NCC for the same images of 0.993, where NCC=1.0 corresponds to theoretically optimal registration. Qualitative analysis through visual inspection of the registered images suggests that the inclusion of anatomical information may offer advantages however, the simple pipeline had many shortcomings that needs to be addressed to improve its results. Future use of clinical data would be essential to validate the model's robustness in practical scenarios.