Development of a Convolutional Neural Network Based Skull Segmentation in MRI Using Standard Tesselation Language Models

Rodrigo Dalvit Carvalho da Silva; Thomas Richard Jenkyn; Victor Alexander Carranza

doi:10.3390/jpm11040310

Development of a Convolutional Neural Network Based Skull Segmentation in MRI Using Standard Tesselation Language Models

J Pers Med. 2021 Apr 16;11(4):310. doi: 10.3390/jpm11040310.

Authors

Rodrigo Dalvit Carvalho da Silva^{1

2}, Thomas Richard Jenkyn^{1

2

3

4

5}, Victor Alexander Carranza^{1

6}

Affiliations

¹ Craniofacial Injury and Concussion Research Laboratory, Western University, London, ON N6A 3K7, Canada.
² Faculty of Engineering, School of Biomedical Engineering, Western University, London, ON N6A 3K7, Canada.
³ Department of Mechanical and Materials Engineering, Western University, London, ON N6A 3K7, Canada.
⁴ Faculty of Health Sciences, School of Kinesiology, Western University, London, ON N6A 3K7, Canada.
⁵ Wolf Orthopaedic Biomechanics Laboratory, Fowler Kennedy Sport Medicine Clinic, London, ON N6A 3K7, Canada.
⁶ Faculty of Engineering, School of Biomedical Engineering, Collaborative Specialization in Musculoskeletal Health Research, and Bone and Joint Institute, Western University, London, ON N6A 3K7, Canada.

Abstract

Segmentation is crucial in medical imaging analysis to help extract regions of interest (ROI) from different imaging modalities. The aim of this study is to develop and train a 3D convolutional neural network (CNN) for skull segmentation in magnetic resonance imaging (MRI). 58 gold standard volumetric labels were created from computed tomography (CT) scans in standard tessellation language (STL) models. These STL models were converted into matrices and overlapped on the 58 corresponding MR images to create the MRI gold standards labels. The CNN was trained with these 58 MR images and a mean ± standard deviation (SD) Dice similarity coefficient (DSC) of 0.7300 ± 0.04 was achieved. A further investigation was carried out where the brain region was removed from the image with the help of a 3D CNN and manual corrections by using only MR images. This new dataset, without the brain, was presented to the previous CNN which reached a new mean ± SD DSC of 0.7826 ± 0.03. This paper aims to provide a framework for segmenting the skull using CNN and STL models, as the 3D CNN was able to segment the skull with a certain precision.

Keywords: CT; MRI; convolutional neural network; image segmentation; standard tessellation language.