A novel enhanced normalization technique for a mandible bones segmentation using deep learning: Batch normalization with the dropout

Several cases of oral and maxillofacial surgery require 3D virtual surgical planning, which is essential for craniofacial tumor resection and flap reconstruction of the mandible. This could only be achieved if the mandible bone was segmented accurately using computed Tomography (CT) images. The convolutional Neural Network (CNN) has achieved high accuracy and more robust segmentation within less processing time in segmentation. In this research, we propose a CNN-based system to improve the accuracy and performance of the segmentation. The proposed system consists of U-Net-based on CNN for the segmentation of mandible bone using the dropout technique and batch normalization in fully connected layers of a convolutional neural network to avoid over-fitting and instability of the process. This method provides 3D segmentation of mandible bones from 2D segmented regions from three different orthogonal planes. Four different types of planar data were used to achieve better accuracy and processing time of the segmentation of mandible bones. Dataset was taken from Public Domain Database for Computational Anatomy (PDDCA). Greyscale computed tomography (CT) images were used. 310 CT scan images were used. A confusion matrix has been used to measure the accuracy, i.e., true positive, false positive, and false negative. In contrast to the state-of-art solutions, Results of the proposed solution show that the accuracy of mandible bones’ segmentation has been improved by 21%, on average, and the processing time has been reduced by 30% second. Our proposed enhanced system is based on the accurate segmentation of mandible bones in datasets from two different kinds of planes, i.e., single-planar and multi-planar. And single planar data has further been divided into three types i.e., axial, sagittal, and coronal planes.