The automatic design of cranial implants is an important and challenging task. The implants must be designed according to the individual characterization of the patient’s defect. This makes the process tedious and time consuming. However, if possible, the personalized implants should be designed and fabricated during the surgical procedure that requires the implant modeling to be as efficient as possible.

The design of the cranial implants may be improved and accelerated by deep learning-based segmentation networks. This approach transfers the computational burden to the training phase, allowing a real-time inference. Moreover, the practical method should be fully automatic, without the need for manual parameter tuning related to the defect characterization. Therefore, a single, universal model is desirable during practical usage. Nevertheless, deep learning-based solutions require large amount of training data that is difficult to acquire and annotate.

To address this problem, we propose a method to connect the two training sets from the AutoImplant challenge, together with a dedicated U-Net based segmentation network. The datasets are combined by the affine and non-rigid registration, and then are further augmented by random affine transformations. The segmentation method consists of two sequential networks responsible for general structure modelling and the preservation of fine details respectively.

We evaluate the proposed results using test sets for all tasks from the AutoImplant 2021 challenge. Three evaluation metrics are used: Dice Score, Boundary Dice Score, and the 95th percentile of the Hausdorff distance. The method achieves mean Dice Score close to and above 0.9 for Task 1 and 3 respectively. The mean Hausdorff distance is close to 1.5 mm. This shows the method good accuracy and robustness. The qualitative results for Task 2 are unavailable at the moment of writing the manuscript.