Systematic Review of AI-Driven Applications for Screening Nasopharyngeal Carcinoma Using MRI.
Authors
Affiliations (4)
Affiliations (4)
- Department of Imaging and Interventional Radiology, Faculty of Medicine, The Chinese University of Hong Kong, Prince of Wales Hospital, Hong Kong, China.
- The CU Lab of AI in Radiology, Department of Imaging and Interventional Radiology, Faculty of Medicine, The Chinese University of Hong Kong, Prince of Wales Hospital, Hong Kong, China.
- Department of Imaging and Interventional Radiology, Prince of Wales Hospital, Hospital Authority, Hong Kong, China.
- JC School of Public Health and Primary Care, The Chinese University of Hong Kong, Hong Kong, China.
Abstract
Nasopharyngeal carcinoma (NPC) can be detected early on MRI, but adoption for screening is limited by a shortage of experienced specialists. MRI artificial intelligence (AI) algorithms for diagnosing NPC in the literature may address this, but most studies focus on routine clinical care, and their applicability in screening requires close examination. To (i) evaluate the diagnostic performance of AI for NPC detection, (ii) analyze the impact of study protocols, and (iii) assess the applicability of existing studies to NPC screening. Systematic review. Thirty-eight studies were included, including a total of 23,398 patients (20,693 NPC and 2705 non-NPC). 1.5 T to 3.0 T. Following PRISMA guidelines, PubMed, Scopus, and Embase records from January 1, 2009 to August 16, 2025 were screened by two independent reviewers for studies on NPC detection, localization, and/or diagnosis using MRI. Included studies were reviewed for applicability and risk of bias. Lesion-localization performance (Dice similarity score [DSC]) and NPC vs. non-NPC discrimination (sensitivity/specificity) were extracted and summarized. Subgroup analyses assessed the impact of intravenous contrast on AI performance. Meta-analysis used standard random-effects univariate restricted maximum likelihood model and hierarchical summary receiver operator characteristics for performance pooling and subgroup analysis, taking p value < 0.05 as significance. Pooled performance from 30 localization and 6 discrimination studies was DSC = 0.80 (CI: 0.78-0.82) and sensitivity/specificity = 97.1% (CI: 88.0%-99.3%)/87.8% (CI: 78.9%-93.2%), respectively. Intravenous contrast did not significantly affect AI performance for either task (p > 0.05). Only seven studies were fully applicable to screening. AI-driven NPC screening using MRI is feasible at high sensitivity. Using non-contrast MRI did not significantly impact AIs' performance. However, applicability and the suboptimal ratio of NPC: non-NPC patients remain the weaknesses of existing studies, warranting further investigation with cohorts that better reflect the screening scenario. The systematic review protocol of this system is registered with the Inplasy registry (Ref. INPLASY202570076).