Automated assessment of 3D facial asymmetry: a systematic review.

内容摘要

This systematic review aims to evaluate the clinical applicability of automated 3D facial asymmetry assessment methods based on 3D facial scans. A comprehensive search of electronic databases (PubMed, Web of Science, EMBASE, Medline, and Scopus) and manual literature searches were conducted in April 2025. Studies that were published in English and evaluated the validity or reliability of automated asymmetry assessment methods in 3D facial scans for medical or biological settings were included. Two reviewers independently screened the articles for eligibility. Risk of bias was assessed using QUADAS-2, and the certainty of evidence was graded using the Grading of Recommendations Assessment, Development, and Evaluation framework. Fourteen studies met the inclusion criteria and were analyzed for methodology, validity, and reliability. Methodologies for assessing facial asymmetry were categorized into four approaches: landmark-based, depth-stratified, original-mirror alignment, and template-based approaches. While landmark and depth-stratified methods rely on sparse data, original-mirror and template-based methods enable comprehensive surface analysis. Six studies evaluating validity against alternative methods, synthetic ground truth, or human ratings consistently demonstrated moderate-to-strong correlation coefficients and classification accuracy. Reliability was examined across nine studies using repeated measurements and multi-observer designs, generally showing minimal measurement variation. Notably, methodological analysis revealed that original-mirror alignment, typically implemented using unconstrained iterative closest point (ICP)-based best-fit registration, is susceptible to registration errors (the "Pinocchio effect") in cases of severe asymmetry, whereas template-based methods mitigate this through correspondence transfer and weighted registration strategies that stabilize anatomical alignment. The review is limited by a high risk of bias in primary studies and significant methodological heterogeneity. Despite "very low" certainty evidence, template-based approaches that transfer anatomical correspondence and apply weighted registration appear preferable due to their robustness against localized deformities. Conversely, original-mirror alignment, typically implemented via unconstrained ICP, remains a practical alternative for mild asymmetry. Future research should prioritize end-to-end deep learning automation, dynamic analysis, and the development of accessible, open-source tools to bridge the gap between technical innovation and routine clinical practice. PROSPERO (CRD420251025105).