Radiomics-Based Differentiation of Enchondroma from Atypical Cartilaginous Tumour and Chondrosarcoma: A Systematic Review and Appraisal.
Authors
Affiliations (10)
Affiliations (10)
- Department of Upper Gastrointestinal Surgery, St John of God Murdoch Hospital, Perth, WA, 6150, Australia. [email protected].
- Department of Colorectal Surgery, St John of God Murdoch Hospital, Perth, WA, 6150, Australia.
- Department of Radiology, St James's Hospital, Dublin, D08 NYH1, Ireland.
- Trinity St James's Cancer Institute, Dublin, D08 NHY1, Ireland.
- Department of Plastic Surgery, Fiona Stanley Hospital, Perth, WA, 6150, Australia.
- Department of Health Sciences, National University of Ireland, Galway, Galway, H91 TK33, Ireland.
- Department of General Surgery, Royal Perth Hospital, Perth, WA, 6847, Australia.
- School of Surgery, University of Western Australia, Perth, WA, 6009, Australia.
- Department of Trauma and Orthopaedics, University Hospital Waterford, Waterford, X91 ER8E, Ireland.
- Department of Trauma and Orthopaedics, University College Cork (UCC), Cork, T12 AK54, Ireland.
Abstract
The objective was to systematically review radiomics models that differentiate enchondroma from atypical cartilaginous tumour (ACT) or chondrosarcoma and to appraise their methodological quality and readiness for clinical use. PubMed, Embase, the Cochrane Central Register of Controlled Trials and Scopus were searched from inception to 24 October 2026 for radiomic, machine learning or deep learning models applied to CT, MRI, radiography or bone SPECT/CT (PROSPERO registration number withheld for blinded review). Performance was synthesised descriptively, not pooled. Quality was appraised with PROBAST + AI, TRIPOD + AI, the Radiomics Quality Score and METRICS. Fourteen reports representing 13 retrospective cohorts (1447 participants, including 79 control scans) were included; two probably overlapping reports were counted once. Eight cohorts used MRI, three CT, one both and one bone SPECT/CT; three reports evaluated deep learning. Eight reports supplied an exact or derivable 2 × 2 table for one of two clinical targets. Sensitivity ranged from 0.800 to 0.935 and specificity from 0.727 to 1.000; preferred areas under the curve (AUC) ranged from 0.801 to 1.000. Development quality concern was high in all reports, and all 22 diagnostic evaluation datasets had high risk of bias. One study used geographically external validation; none was prospective or reported decision curve analysis. Median complete TRIPOD + AI reporting was 48.3%. Reported discrimination was promising but methodologically preliminary and does not establish clinical usefulness; prospective assessment and independent multicentre external validation are required.