Ilmu Komputer & AI editorial
Open AccessOA2026
Gaze responses to false-positive computer-aided detection prompts during colonoscopy: a paired-video and real-time eye-tracking study
Event-locked eye tracking reveals that false-positive CADe prompts capture endoscopists' gaze far beyond their brief on-screen visibility, with attention occupation amplified up to 17.55-fold.
Te Luo; Yan Zhu; Peiyao Fu; Ruijie Yang; Xian Yang; Quanlin Li; Pinghong Zhou; Shuo Wang· 2026· DOI 10.48550/arXiv.2609.25581
The core problem
Computer-aided detection (CADe) systems are increasingly deployed during colonoscopy to highlight suspected lesions in real time. While their sensitivity for polyp detection has been a central focus, the attentional cost of false-positive prompts remains poorly quantified. A false-positive prompt—an on-screen alert that does not correspond to an expert-annotated lesion—may transiently divert the endoscopist's gaze away from the mucosal surface being inspected. The authors note that the attentional impact of individual prompts is unclear, motivating an event-locked eye-tracking approach that can measure both whether a prompt attracts gaze and how long attention remains occupied after the prompt disappears. The study combines a retrospective paired-video experiment with a prospective real-time clinical study to capture gaze behavior in controlled and ecologically valid settings. The central hypothesis is that false-positive CADe prompts frequently capture gaze and that attention persists beyond prompt visibility, potentially imposing a hidden burden on endoscopic attention.
Innovation
False-positive prompts attracted gaze in **48.6% (68/140)** of retrospective observations and **65.2% (533/817)** of prospective events. Among attraction events with complete recovery, median attention occupation was **1000 ms** in the retrospective study and **1100 ms** in the prospective study. Corresponding median prompt durations were only **33 ms** and **267 ms**, yielding median time amplifications of **17.55-fold** and **5.15-fold**, respectively. In paired retrospective comparisons, visible artifact prompts drew gaze closer to the prompted region than did the same-coordinate unassisted reference. Secondary retrospective analyses showed high lesion gaze recognition both without and with CADe (**98.0% versus 99.0%**). First gaze entry into lesion regions occurred **147.8 ms earlier** with CADe. These results indicate that false-positive prompts not only attract gaze but also hold attention well beyond their brief display, with the retrospective setting showing a particularly large amplification because prompts were very short-lived.
Computer-aided detection (CADe) systems are increasingly deployed during colonoscopy to highlight suspected lesions in real time. While their sensitivity for polyp detection has been a central focus, the attentional cost of false-positive prompts remains poorly quantified. A false-positive prompt—an on-screen alert that does not correspond to an expert-annotated lesion—may transiently divert the endoscopist's gaze away from the mucosal surface being inspected. The authors note that the attentional impact of individual prompts is unclear, motivating an event-locked eye-tracking approach that can measure both whether a prompt attracts gaze and how long attention remains occupied after the prompt disappears. The study combines a retrospective paired-video experiment with a prospective real-time clinical study to capture gaze behavior in controlled and ecologically valid settings. The central hypothesis is that false-positive CADe prompts frequently capture gaze and that attention persists beyond prompt visibility, potentially imposing a hidden burden on endoscopic attention.
The study used complementary retrospective and prospective designs with event-locked eye tracking.
Why it matters
The findings demonstrate that false-positive CADe prompts impose a measurable attentional burden during colonoscopy. The high attraction rates—nearly half of retrospective observations and about two-thirds of prospective events—suggest that endoscopists are drawn to artifact prompts even when no lesion is present. The median attention occupation of roughly one second is clinically meaningful: at typical colonoscopy withdrawal speeds, a one-second diversion can correspond to a substantial segment of mucosa being inspected less carefully. The time amplification factors quantify this mismatch; in the retrospective setting, attention persisted about 17.55 times longer than the prompt itself, while in real-time use the amplification was 5.15-fold. The paired-video comparison strengthens causal interpretation by showing that the same coordinates attracted more gaze when an artifact prompt was visible. Importantly, CADe did not impair lesion recognition—recognition remained high (98.0% without vs. 99.0% with CADe) and first gaze entry into lesion regions was 147.8 ms earlier with CADe—indicating that the attentional cost of false positives coexists with a detection benefit. The authors conclude that prompt-related attentional burden should be considered in CADe evaluation and design, for example by reducing false-positive alert frequency, shortening alert salience, or adapting prompt timing to the endoscopic task. Limitations include the modest number of endoscopists, the artificial nature of the retrospective video task, and the focus on senior operators in the prospective arm. Future work could link gaze disruption to polyp miss rates and integrate attention-aware metrics into CADe benchmarking.
Who should read this
CS practitioners and researchers
Opening member content…