Automated segmentation of the fibula from CT imaging using two-stepped deep learning in 3D U-Net architectures

Abstract

This study proposes a fully automatic segmentation of the fibula bone from CT images for application in pre-operative planning of reconstructive surgery. The objective is to make use of new developments in the image segmentation field to optimize and reduce the costs of patient-specific surgery planning. Two different approaches are proposed to perform the fibula bone segmentation, both based on a two-step segmentation method using a 3D-UNet architecture. To account for the symmetry of the left and right fibula bones, input images of the right fibula are mirrored to the left side. The accuracy of the trained models is measured using common evaluation metrics, together with specific metrics focused on facial reconstructive surgery. Both of the described approaches achieve high-accuracy results. For the best-trained model, an average Dice score of 0.95 and Average Surface Distances below 0.31 mm is measured on the test set in the region of interest for the surgery. Both approaches are robust segmentation techniques and permit data pre-processing for further application in the context of preoperative surgical planning of procedures for facial reconstruction with bony transplants.