Feasibility of a Deep-Learning Model for Objective Visualization-Quality Assessment During Cholangioscopy-Guided...

During peroral cholangioscopy-guided electrohydraulic lithotripsy (EHL), stone fragmentation obscures the visual field (the "snowstorm effect"), and visualization quality has been assessed only subjectively. This proof-of-concept study explored the feasibility of a deep-learning method for objectively quantifying visualization quality during EHL and, as an exploratory application, compared viscous gel immersion with saline irrigation. A ResNet-18 model was trained on expert-annotated frames (5-point scale) from three development cases. For independent validation, 120 frames from 12 previously unseen comparison cases were scored by two blinded endoscopists (inter-reader weighted κ = 0.98). The model showed only moderate agreement with their consensus (weighted κ = 0.60; accuracy 70.8%); because the validation sample was enriched for predicted-usable frames, metrics re-weighted to the natural frame distribution gave sensitivity 29.8% and specificity 93.5%, indicating systematic under-detection of usable views. Agreement did not differ systematically between gel and saline frames. The High-Quality View Rate (HQVR) was the proportion of EHL-phase frames scored Good/Excellent. In a patient-level comparison, median HQVR was 22.7% for gel versus 8.4% for saline (exact Mann-Whitney U, p = 0.15, not significant); in post hoc exploratory analyses, a threshold-independent mean predicted score differed (1.65 vs. 1.35, p = 0.005) and the longest continuous usable interval favored gel (median 24.4 vs. 5.2 s). Mild cholangitis occurred in 1/5 gel and 4/7 saline patients. Training such a model is feasible, but it is not yet a validated measurement tool, and the gel-versus-saline signal is hypothesis-generating only. Larger multi-case training and prospective validation are required. Trial Registration: N/A.