AI-assisted MRI for diagnosis and characterization of pituitary adenomas and sellar lesions: A systematic review of diagnostic performance

Abstract

Background: Pituitary adenomas and sellar lesions account for 10-15 of intracranial tumors. MRI is the primary diagnostic modality, but interpretive variability and lesion heterogeneity remain challenges. This systematic review evaluated the diagnostic performance and evidence certainty of AI-assisted MRI for pituitary adenomas and sellar lesions. Methods: This review followed PRISMA 2020 guidelines. PubMed, Embase, Scopus, Web of Science, Cochrane Library, and IEEE Xplore were searched from inception to February 15, 2026. Studies evaluating machinelearning or deep-learning models applied to MRI for detection, segmentation, classification, invasiveness prediction, or prognostic assessment were included. Two reviewers independently screened studies, extracted data, and assessed risk of bias using QUADAS-2. Owing to heterogeneity, qualitative synthesis was performed. Certainty of evidence was assessed using GRADE. Results: Of 1381 records identified, 13 studies were included. The median patient sample size was 204 across 12 studies reporting patient-level samples; one study reported only dynamic contrast-enhanced phase-level samples. Most studies were retrospective and single-center. Preoperative applications showed encouraging reported performance, with AUC values ranging from 0.76 to 0.95. Hybrid and multiparametric models often reported stronger performance, particularly for microadenoma detection and subtype classification. Postoperative segmentation was poor in the single study evaluating this task. Risk of bias was high in the index test domain in 8/13 studies and in patient selection in 5/13 studies. Certainty of evidence was low for preoperative diagnostic, characterization, surgical-planning, and prognostic applications, and very low for postoperative segmentation. Conclusions: AI-assisted MRI shows encouraging reported performance for selected preoperative applications in retrospective studies. However, current evidence remains low certainty because of high risk of bias, limited external validation, and substantial methodological heterogeneity. Prospective multicenter validation and studies evaluating workflow impact and patient outcomes are required before routine implementation.