Comparison of graphic user interfaces for computer-aided detection in Barrett's neoplasia

Human-computer interaction, particularly the graphic user interface (GUI) for displaying detection results, is an important but underexplored aspect of computer-aided detection (CADe) systems in endoscopy. We explored whether the use of a bounding box GUI or a heatmap GUI resulted in different performance of endoscopists when using CADe for Barrett's neoplasia.

Thirty-seven endoscopists from 6 countries with varying expertise assessed 70 Barrett's esophagus videos. All videos were evaluated by our previously developed CADe system and comprised, at some point, a CADe detection, whether the video contained neoplasia or not. The study had 2 phases. Initially, videos were shown with either a bounding box or heatmap. Then, after a 2-week washout, the same videos were reordered and displayed with the alternate GUI. Endoscopists marked perceived neoplastic lesions and biopsy sampling sites, also noting their subjective GUI preference. Primary endpoints were objective classification and localization performance, whereas the secondary endpoint was the subjective preference.

No statistically significant difference in classification performance was found when endoscopists were provided with the bounding box or the heatmap visualization (sensitivity, 83% vs 83% [P = .29]; specificity, 86% vs 86% [P = .09]). Also, the comparison of localization accuracy between the bounding box and heatmap methods showed no significant differences, with both methods yielding a median score of 97%. Subjectively, 23 endoscopists preferred the heatmap and 14 the bounding box (P = .04).

Although endoscopists expressed a preference for the heatmap GUI, this was not associated with a statistical difference in performance outcomes.