Machine learning for image analysis in the cervical spine: systematic review of the available models and methods

Abstract

Background: Automating parts of the radiological image analysis process using Machine Learning could provide more accurate and consistent assessment, eliminating interobserver variability with an increased time efficiency. The aim of this systematic review is to create the first overview of publications focusing on automated segmentation and studies using Machine Learning to study characteristics of the cervical spine on radiological imaging.

Methods: A systematic literature review was conducted in medical and technical libraries. Studies were selected using predefined selection criteria and risk of bias was assessed. Data on study characteristics and model performance was collected. Included publications were divided into either the ‘segmentation’ category (further specified into ‘Conventional Machine’ and ‘Deep Learning’ subgroups) or the ‘cervical spine analysis’ category.

Results: In segmentation, Deep Learning models show promising results with the application of fully automatic convolutional neural networks (CNN) for MR and CT imaging. In cervical spine analysis, the biomechanical features are most often studied using finite element models. The application of artificial neural networks and support vector machine models looks promising for other classification purposes. The reported outcome measures vary widely, leaving room for standardization in this field. Conclusions: This systematic review may function as a reference for all authors conducting research on computer aided diagnostics of the cervical spine. In future research, individual studies should aim to increase external validity by reporting similar outcome metrics, include uncertainty metrics in their reporting, as well as share code to increase reproducibility.