AI-assisted digital cholangioscopy for indeterminate and malignant biliary strictures: 5 studies, 675 lesions, pooled sensitivity 95% and specificity 88%
Synopsis
This systematic review and meta-analysis pooled 5 studies (675 lesions; 2,685,674 cholangioscopic images) using PRISMA, MOOSE, and Cochrane Diagnostic Test Accuracy methodology with a bivariate model, and found that AI-assisted cholangioscopy—mostly deep learning systems using convolutional neural networks at roughly 30 to 60 frames per second—achieved a pooled sensitivity of 95% (95% CI: 85-98), specificity of 88% (95% CI: 76-94), and diagnostic accuracy (SROC) of 97% (95% CI: 95-98) for indeterminate or malignant biliary strictures, with a CNN-only sensitivity analysis (4 studies, 538 patients) showing consistent results, leading the authors to call the approach promising.
Interpretation
AI-assisted cholangioscopy shows high pooled diagnostic performance for indeterminate and malignant biliary strictures: sensitivity 95%, specificity 88%, and SROC accuracy 97%. ERCP and cholangioscopic sampling for indeterminate biliary strictures were previously described as suboptimal; this work is the first systematic review and meta-analysis to pool the scattered AI cholangioscopy studies into a single diagnostic accuracy estimate. Pooled analysis of 5 studies, 675 lesions, and 2,685,674 images, using a bivariate model per Cochrane Diagnostic Test Accuracy working group methodology, with 95% confidence intervals reported.
The pooled result held in a sensitivity analysis restricted to CNN deep learning studies: 4 studies and 538 patients showed pooled sensitivity 95%, specificity 88%, and SROC accuracy 97%. Narrowing the conclusion from a mixed set of technologies to predominantly convolutional neural network deep learning systems shows the pooled performance is not driven by non-CNN approaches. The sensitivity analysis is limited to the CNN subset, a smaller sample than the main analysis, but point estimates match the main analysis with wider confidence intervals.
The included AI systems operated at near real-time image processing speeds, averaging about 30 to 60 frames per second. Presenting diagnostic accuracy alongside usable processing speed indicates the technology is positioned for deployment in real-time endoscopic workflows. Descriptive summary from the included studies, all but one of which analyzed CNN deep learning systems; no quantitative link between speed and accuracy was reported.
Perspective
This work is aimed at clinicians and endoscopy researchers and applies to diagnostic settings where digital cholangioscopy images are used to assess indeterminate and malignant biliary strictures; its pooled estimates can be used directly to gauge the potential value of AI-assisted interpretation in that setting and as a reference benchmark for prospective validation and clinical pathway design. For readers who want to understand how far AI has been translated into endoscopic imaging, or who need quantitative grounds for biliary stricture diagnostic decisions, this paper provides currently citable pooled sensitivity, specificity, and SROC accuracy values.
The pooled estimates come from 5 studies and 675 lesions, a limited evidence base, and the confidence intervals for sensitivity and specificity remain of moderate width; the included studies are predominantly CNN deep learning systems with image processing speeds of about 30 to 60 frames per second, but the relationship between speed and diagnostic accuracy is not quantified here. How differences in patient selection, reference standards, and interpretation workflows across studies affect the pooled results, and whether this high accuracy holds in real clinical workflows, remain open questions for further research.