DEEP MAT DEFORMATION: FROM VOLUMETRIC MEDICAL IMAGING TO ORGAN SURFACE RECONSTRUCTION

Shape reconstruction from volumetric medical images is a crucial task in biomedical research and for multiple clinical applications. To date, common workflows rely on voxel-wise segmentation of the volume and a number of post-processing techniques to eventually generate 3D surface representations. However, these approaches tend to suffer from issues like disconnected regions, incorrect surface topology, and stair-case artifacts. To overcome these limitations, new deep learning methods have been developed to directly generate high-resolution 3D surfaces from discrete 3D imaging volumes. Nevertheless, the high resolution comes at the cost of a computationally heavier model with more parameters. In this thesis addressed this problem by suggesting Deep MAT Deformation (DMD). DMD is a novel approach that leverages a deep neural network to predict a shape by deforming the medial axis transform (MAT) of a template surface with similar topology. By employing the MAT, it is shown that DMD can effectively capture intricate surface details while maintaining a lighter model compared to existing high-resolution methods. This not only enhances the accuracy of shape reconstruction from volumetric medical images but also significantly reduces the required computational resources.