End-to-end Pipeline to Learning-based Approaches for Automatic Cranial Implant Design

Cranioplasty is a complex procedure that aims to repair a defect in the skull. Generally, the surgeon can replace part of the skull, repair damage to the skull, or reshape the skull. In any case, it requires an implant that perfectly fits the missing region in the skull of the patient. The key is to have an accurate cranial implant design in a short time from the CT image of the patient. Recently, numerous successful attempts have been made to generate the implant design from the skull image where the most notable results come from deep learning-based methods. The generating models were trained from a well-prepared, normalized, and standardized dataset where each skull image was ideally made by well-trained radiologists. In this paper, we fulfill the gap by proposing an end-to-end pipeline to generate implant design from the actual patient CT image that is severely degraded by noise and non-ideal acquisition conditions. The pipeline consists of several image processing steps to prepare the skull image of the patient before passing it through the implant-generating module. It also consists of a pre-processing step to stabilize the result and a post-processing step to make the implant design practically usable according to the criterion suggested by the surgeon. The experimental results with several patient data confirm the ability of the proposed method to accurately and robustly generate the implant design from the CT images, thereby enabling the use of existing pre-trained models in practice.