Medical Shape Pattern Analysis with Meshcnn

Accurate recognition and knowledge of anatomical shapes are important in both medical diagnosis and treatment planning. Improving techniques for 3D shape analysis would help gain clinical knowledge, be useful for efficient treatment delivery, and minimise side effects. In this paper, a new approach was proposed for shape variation analysis using MeshCNN. The classification network of the MeshCNN framework was used along with the density visualisation approach for feature extraction. The approach was applied on a hippocampus dataset with both normal control and Alzheimer's disease (AD) groups to identify the localised shape variations between the two groups. The result highlighted the subiculum subfield as the region with highest local shape variation which was consistent with previous study on AD's effect on hippocampus shape. The proposed approach showed ability to identify regions of clinical interest and to extract non-linear features of 3D shapes.