Cholangioscopy-based convoluted neuronal network vs. confocal laser endomicroscopy in identification of neoplastic biliary strictures.

Background and study aims Artificial intelligence (AI) models have demonstrated high diagnostic performance identifying neoplasia during digital single-operator cholangioscopy (DSOC). To date, there are no studies directly comparing AI vs. DSOC-guided probe-base confocal laser endomicroscopy (DSOC-pCLE). Thus, we aimed to compare the diagnostic accuracy of a DSOC-based AI model with DSOC-pCLE for identifying neoplasia in patients with indeterminate biliary strictures. Patients and methods This retrospective cohort-based diagnostic accuracy study included patients ≥ 18 years old who underwent DSOC and DSOC-pCLE (June 2014 to May 2022). Four methods were used to diagnose each patient's biliary structure, including DSOC direct visualization, DSOC-pCLE, an offline DSOC-based AI model analysis performed in DSOC recordings, and DSOC/pCLE-guided biopsies. The reference standard for neoplasia was a diagnosis based on further clinical evolution, imaging, or surgical specimen findings during a 12-month follow-up period. Results A total of 90 patients were included in the study. Eighty-six of 90 (95.5%) had neoplastic lesions including cholangiocarcinoma (98.8%) and tubulopapillary adenoma (1.2%). Four cases were inflammatory including two cases with chronic inflammation and two cases of primary sclerosing cholangitis. Compared with DSOC-AI, which obtained an area under the receiver operator curve (AUC) of 0.79, DSOC direct visualization had an AUC of 0.74 ( P = 0.763), DSOC-pCLE had an AUC of 0.72 ( P = 0.634), and DSOC- and pCLE-guided biopsy had an AUC of 0.83 ( P = 0.809). Conclusions The DSOC-AI model demonstrated an offline diagnostic performance similar to that of DSOC-pCLE, DSOC alone, and DSOC/pCLE-guided biopsies. Larger multicenter, prospective, head-to-head trials with a proportional sample among neoplastic and nonneoplastic cases are advisable to confirm the obtained results.