Deep Learning for Virtual Patient-Specific Skull Modelling and Reconstruction
Skull segmentation from 3D patient data and virtual reconstruction of the defective skull shape are the most challenging steps required for creation of patient-specific models of skull.These models are used in cranioplasty practice for surgery planning, patient education and patient-specific implant design, but their utility is currently limited by the amount of manual processing time required to reach sufficient virtual model quality.This thesis aims to streamline this virtual workflow by utilizing deep learning methods.
The thesis proposes a novel solution that consists of an automatic skull segmentation method based on a combination of convolutional neural networks and graph cut algorithm and an automatic virtual skull reconstruction method based on convolutional network cascade.Both of these components are demonstrated to achieve state-of-the-art accuracy. This work also aims to improve reproducibility of the skull reconstruction research by providing a structured synthetic dataset for development and benchmarking of automatic methods.
The main focus of this work is on applicability in clinical practice.While the proposed skull segmentation method is already successfully deployed to clinical workflow, the integration of automatic virtual skull reconstruction presents some additional challenges, such as low tolerance towards shape imperfections around the defect border.This work therefore also proposes an extension of the skull reconstruction method that allows its adaptation to target population and the desired type of cranial implant shape, which can vary between different clinical sites.The results of expert's evaluation show that the shape outputs of this method reach enough quality to be deployed into clinical practice along with the segmentation method.
Deep learning; Medical imaging; Cranioplasty; Surgery planning; Skull segmentation, Virtual skull reconstruction.