Artificial Intelligence in Gallbladder Imaging: A Rapid Evidence Review and Exploratory Meta-Analysis of Diagnostic...

Artificial intelligence (AI) is increasingly being investigated for gallbladder imaging, but its clinical diagnostic performance, added value over conventional interpretation, and generalizability remain uncertain. We conducted a rapid evidence review with exploratory meta-analysis using a previously published scoping review of 16 ultrasound studies as the initial evidence anchor; 12 reports met the present eligibility criteria. A targeted PubMed search through June 2026 was supplemented by publisher-website screening, reference-list review, and backward and forward citation tracking. Because Embase, Scopus, Web of Science, and the Cochrane Library were not independently searched, the evidence-identification strategy was not intended to be comprehensive. The included studies were classified into two evidence strata: Stratum A, clinically referenced diagnostic-accuracy studies evaluated against histopathology or another clearly defined patient- or lesion-level clinical reference standard, and Stratum B, technical/non-clinical image-classification studies based primarily on image or dataset labels. Seventeen peer-reviewed reports were included: 15 clinically referenced studies and two technical/non-clinical studies. Across gallbladder-polyp studies, reported performance varied according to target definition and validation design. Independent or temporal validation generally produced more conservative estimates than internal validation, supporting the need for external testing before clinical implementation. Only two independent or temporal cohorts provided reconstructable 2×2 data for neoplastic-polyp classification. Individual-study sensitivity ranged from 74.3% to 78.6%, while specificity ranged from 81.8% to 92.1%. Because only two studies were available, formal bivariate random-effects or hierarchical summary receiver operating characteristic synthesis was not considered reliable. Separate random-effects summaries of sensitivity (76.1%; 95% confidence interval (CI), 64.1%-85.1%) and specificity (87.8%; 95% CI, 73.9%-94.8%) were therefore presented strictly as descriptive exploratory estimates rather than conventional joint diagnostic test-accuracy meta-analysis estimates. The exploratory diagnostic odds ratio was 24.0 (95% CI, 10.1-57.1). Three reader-assistance studies reported absolute improvements in accuracy ranging from 6.6 to 15.1 percentage points with AI support. Because paired case-level and reader-level variance data were unavailable, these findings were summarized descriptively and were not meta-analyzed. Gallbladder-cancer studies were too heterogeneous for quantitative pooling, while the two technical studies demonstrated image-classification feasibility but did not provide equivalent evidence of patient-level clinical diagnostic accuracy. AI shows potential for gallbladder-polyp characterization and reader support, particularly for less-experienced clinicians. However, confidence remains limited by retrospective and surgery-enriched populations, small independent validation cohorts, heterogeneous validation methods, and the targeted rather than comprehensive literature search. Broader evidence synthesis and prospective multicenter clinical validation are required before routine or autonomous clinical implementation.