Computer vision scoring of endoscopically traced line figures in an inanimate scope tip coordination training model.

introductionFine control of the moving scope tip with an instrument extended from the scope tip is a critical skill in performing Endoscopic Submucosal Dissection (ESD) and hybrid ESD/Endoscopic Mucosal Resection (EMR) polypectomy. An inanimate training model has been utilized for several years where trainees trace line figures (e.g., grid, circle, S-shape, and figure of eight) using a sclerotherapy catheter "pen" within a plastic colon tube containing a paper insert. This exercise requires fine coordination similar to that needed for needle knife use. Traced figures are traditionally scored based on completion time and the number of deviations greater than 2 mm from the intended line. This study assesses the efficacy of a computer vision (CV) algorithm versus manual scoring in evaluating these tracings.

methodsA training session involved 8 tracings (4 figures, traced twice) resulting in 352 tracings manually scored for deviations over 2 mm. A CV model was employed to score the tracings by aligning each output with a reference template using perspective transformation. Color space segmentation isolated the ink, and deviations were highlighted by subtracting the tracing from the template. A pixel-based deviation analysis converted pixel distance to millimeters, classifying deviations greater than 2 mm as large deviations for final analysis. Additional morphological processing ensured accurate detection of significant deviations.

resultsOn average, the CV algorithm detected 1.28 more deviations than manual scoring. Detection rates by figure were as follows: grid pattern (1.23), circle (1.27), figure of eight (1.61), and S-shape (1.02). A Bland-Altman plot demonstrated agreement between the CV algorithm and manual scoring, with most deviations falling within the 96% confidence interval (± 1.96 SD). A noted confounder was the variability in pen stroke thickness, which occasionally resulted in thick lines being misconstrued as deviations. The introduction of orienting marks (circular or diamond-shaped) to each tracing facilitated more accurate evaluations.

conclusionThe CV algorithm effectively assessed the accuracy of tracings in the inanimate model by accurately detecting deviations compared to human scoring. Future developments include a CV application that provides real-time feedback for trainees, potentially integrating AI evaluation into advanced endoscopy educational programs.