Automation and deep (machine) learning in temporomandibular joint disorder radiomics: A systematic review

Abstract

AbstractObjectiveThis review aimed to systematically analyse the influence of clinical variables, diagnostic parameters and the overall image acquisition process on automation and deep learning in TMJ disorders.MethodsArticles were screened in late 2022 according to a predefined eligibility criteria adhering to the PRISMA protocol. Eligible studies were extracted from databases hosted by MEDLINE, EBSCOHost, Scopus, PubMed and Web of Science. Critical appraisals were performed on individual studies following Nature Medicine's MI‐CLAIM checklist while a combined appraisal of the image acquisition procedures was conducted using Cochrane's GRADE approach.ResultsTwenty articles were included for full review following eligibility screening. The average experience possessed by the clinical operators within the eligible studies was 13.7 years. Bone volume, trabecular number and separation, and bone surface‐to‐volume ratio were clinical radiographic parameters while disc shape, signal intensity, fluid collection, joint space narrowing and arthritic changes were successful parameters used in MRI‐based deep machine learning. Entropy was correlated to sclerosis in CBCT and was the most stable radiomic parameter in MRI while contrast was the least stable across thermography and MRI. Adjunct serum and salivary biomarkers, or clinical questionnaires only marginally improved diagnostic outcomes through deep learning. Substantial data was classified as unusable and subsequently discarded owing to a combination of suboptimal image acquisition and data augmentation procedures. Inadequate identification of the participant characteristics and multiple studies utilising the same dataset and data acquisition procedures accounted for serious risks of bias.ConclusionDeep‐learned models diagnosed osteoarthritis as accurately as clinicians from 2D and 3D radiographs but, in comparison, performed poorly when detecting disc disorders from MRI datasets. Complexities in clinical classification criteria; non‐standardised diagnostic parameters; errors in image acquisition; cognitive, contextual or implicit biases were influential variables that generally affected analyses of inflammatory joint changes and disc disorders.