Objective: To compare the accuracy of magnetic resonance imaging (MRI) staging for nasopharyngeal carcinoma (NPC) between central and provincial hospitals in Vietnam; to quantify agreement between institutional reports and expert consensus; and to identify institutional, technical, and workflow factors associated with staging discrepancy.

Methods: This multicentre retrospective comparative imaging study included 536 MRI examinations from 536 patients with histologically confirmed NPC treated or evaluated at five Vietnamese hospitals (three central and two provincial). Institutional MRI reports were compared with a blinded consensus re-reading by two senior head-and-neck radiologists. Six staging criteria were assessed: parapharyngeal extension, skull-base or bony invasion, retropharyngeal lymph-node involvement, suspicious subcentimetric cervical lymph nodes, intracranial extension, and cavernous sinus involvement. Agreement was quantified with Cohen’s kappa. Case-level logistic regression was used to identify predictors of any staging discrepancy. Criterion-level associations were explored separately to avoid unit-of-analysis error.

Results: Central hospitals showed higher point estimates of sensitivity, specificity, and positive predictive value across all six MRI criteria. The largest inter-tier gaps were observed for suspicious subcentimetric cervical lymph nodes (sensitivity 79.6% vs. 51.3%) and retropharyngeal lymph-node detection (83.4% vs. 57.9%). Agreement between institutional reports and expert consensus was excellent or substantial at central hospitals (overall κ=0.83) but only moderate at provincial hospitals (overall κ=0.52). Overall staging discrepancy affected 23.7% of central cases and 75.2% of provincial cases (p<0.001). A treatment-plan change was recommended after expert review in 5.6% of central cases and 29.8% of provincial cases. In the case-level model, provincial hospital tier (adjusted odds ratio [aOR] 3.14), absence of double reading (aOR 2.47), use of 1.5T MRI (aOR 2.03), absence of regular multidisciplinary team meetings (aOR 1.94), and absence of PACS/RIS (aOR 1.72) were the strongest predictors of discrepancy.

Conclusions: MRI staging discrepancy in NPC was substantially more frequent at provincial hospitals, particularly for nodal criteria that require high spatial resolution and subspecialty experience. Double reading, structured reporting, MDT review, PACS/RIS connectivity, and staged MRI infrastructure upgrades are the most actionable priorities. These findings identify the staging criteria most suitable for AI-assisted second-reader tools designed to reduce diagnostic inequality across hospital tiers.