Regression Approach for Cranioplasty Modeling

Abstract

Patient-specific implants provide important advantages for patients and medical professionals. The state of the art of cranioplasty implant production is based on the bone structure reconstruction and use of patient’s own anatomical information for filling the bone defect. The present work proposes a two-dimensional investigation of which dataset results in the closest polynomial regression to a gold standard structure combining points of the bone defect region and points of the healthy contralateral skull hemisphere. The similarity measures used to compare datasets are the root mean square error (RMSE) and the Hausdorff distance. The objective is to use the most successful dataset in future development and testing of a semi-automatic methodology for cranial prosthesis modeling. The present methodology was implemented in Python scripts and uses five series of skull computed tomography images to generate phantoms with small, medium and large bone defects. Results from statistical tests and observations made from the mean RMSE and mean Hausdorff distance allow to determine that the dataset formed by the phantom contour points twice and the mirrored contour points is the one that significantly increases the similarity measures.