Design of 3D-printed cranioplasty moulds using Neural Network

Abstract: Cranioplasty is surgical repair of a skull bone defect due to a previous surgery or injury. Cranioplasty is most often performed with autologous bone flap, i.e the patient's own saved bone from previous surgery if this is available. If autologous bone is not available then custom procedure is to manually mould an implant using bone cement from the plastic polymethyl methacrylate (PMMA). Manually moulding implants during surgery has, albeit being a clinical routine, disadvantages and therefore Skåne University Hospital have developed a technique to 3D print patient-specific cranioplasty moulds based on a computed tomography (CT) scan of the skull bone. The shape of the mould is created by a combination of manual design process, mirroring and interpolation. This partly limits the technique to unilateral defects and the method can be tricky and time consuming for complicated cases. The purpose of the thesis was to develop a method based on a neural network to reconstruct missing parts of the skull bone and overcome the limitations with the current mirroring method when designing implants or moulds for cranioplasty. The process included developing a method to extract data from CT images for training of the neural networks. During the process numerous neural network structures and models were developed and evaluated with the best performing network being a convolutional autoencoder with skip connections. The network was trained with data from a total of 240 patients with simulated defects. The results of the network shows that it is able to handle both unilateral and bilateral defects with a mean error of 1.07mm. In comparison to the currently used method it performed as well or even better in some cases. Overall the developed method showed good enough results for it to be implemented as clinical routine.