Learning Volumetric Shape Super-Resolution for Cranial Implant Design

Abstract

Cranioplasty is the process of repairing cranial defects or deformations, which may be the result of injuries or necessary medical treatments such as brain tumor surgery. For this procedure, it is necessary to generate a high-quality cranial implant, which needs to be shaped individually for each skull and each defect. This tends to be a very time consuming task and requires also in-depth knowledge of various CAM/CAD programs. In this work, we present a novel automatic three-stage implant generation pipeline. First, skull completion is conducted in low resolution using a trained artificial neural network (ANN). Second, the completed low-resolution skull is sent to a super-resolution network, which up-samples the low-resolution skull to higher resolution while, at the same time, filling the skull surface with geometric details. Finally, by simple subtraction and blob filtering, the desired high-resolution implant is generated.