Shape Completion by U-Net: An Approach to the AutoImplant MICCAI Cranial Implant Design Challenge

Abstract

Reconstruction of the craniomaxillofacial (CMF) skeleton requires patient specific implants that restore cosmesis and protect the neural structures. Designing 3D patient specific geometries is challenging and labor intensive because of the lack of pre-injury information. We present an automated shape completion framework for the MICCAI AutoImplant Challenge 2020. The automated workflow selected standardized segmented skull volumes from the skull base to the apex. A U-Net style encoder/decoder framework was used to create the predictive model. The training data consisted of defective skulls with matched intact skulls. The challenge training set (100 cases) was augmented by randomly placed cubic and spherical defects on the same 100 cases for a total of 300 samples split 75/25% by case into a training and validation set. Probability volumes of the predicted skulls were generated by the U-Net and segmented to create an intact skull. Subtraction with defect skulls was used to isolate the implant geometry and were denoised with a connected region extraction of the single largest object, followed by a spherical topological filter. Dice Score (DSC) was 0.86 and Hausdorff distance (HD) was 14.2 mm for the validation set of 25 skulls (×3 defect types). Filtering improved the predicted implants with DSC of 0.87 and HD of 6.72. The automated pipeline for generating implants, produced geometries suitable for integration into a clinical pipeline that could dramatically decrease design time, cost, and increase reconstruction accuracy.