Real-time computer-aided detection (CADe) flags suspicious lesions during live colonoscopy
Colonoscopy is the reference-standard procedure for colorectal cancer (CRC) screening and prevention, allowing direct visualization and removal of precancerous polyps. But it depends entirely on one variable that is notoriously inconsistent: the attentiveness and visual acuity of the endoscopist scanning a fast-moving, cluttered video feed for subtle lesions.
Colonoscopy reduces colorectal cancer incidence and mortality by enabling detection and immediate removal of adenomatous polyps before they progress to cancer — a process (the adenoma-carcinoma sequence) that typically takes 7-15 years. Landmark studies (the National Polyp Study, and subsequent large cohort data) show that colonoscopic polypectomy reduces CRC incidence by roughly 76-90% and CRC mortality by approximately 53%.
But the protective effect is not uniform. It depends heavily on the quality of the exam — specifically on the adenoma detection rate (ADR), defined as the proportion of screening colonoscopies in which at least one histologically confirmed adenoma is found. ADR varies enormously between endoscopists performing the identical procedure on similar patient populations — published ranges span roughly 7% to 53% among individual physicians at the same institution.
A landmark 2014 NEJM cohort study (Corley et al., >314,000 colonoscopies) found that every 1-percentage-point increase in an endoscopist's ADR was associated with a 3% decrease in the risk of interval colorectal cancer, and a 5% decrease in fatal interval cancer. ADR is the single strongest quality metric in colonoscopy.
Even skilled, attentive endoscopists miss polyps. Tandem colonoscopy studies — where two exams are performed back-to-back on the same patient — consistently show a miss rate for adenomas of roughly 20-26% overall, rising to 26-30% for small (<5 mm) and flat/sessile lesions that blend into surrounding mucosa.
Several factors drive this: withdrawal speed (endoscopists rushing under time pressure), incomplete mucosal inspection (folds and flexures hide lesions from direct view), fatigue over a long procedure list, and simple lapses in sustained visual attention — the video stream delivers roughly 1,800-3,600 individual frames over a 6-10 minute withdrawal, and a human cannot devote full foveal attention to every one.
Flat and sessile serrated lesions (SSLs) are disproportionately missed because they lack the polypoid, mushroom-like silhouette that draws the eye; instead they present as subtle color and texture changes, mucus caps, or indistinct borders — exactly the kind of low-contrast pattern that convolutional neural networks excel at flagging.
Unlike radiology, where a computer-aided detection system can flag a static image for a radiologist to review at leisure, colonoscopy is a live, dynamic procedure: the endoscopist must decide whether to biopsy or resect a lesion in the moment it appears on screen, before the scope moves past it. This defines the engineering requirement for polyp-detection AI: it must run inference on every frame at video rate (processing at or above 25-30 fps) with end-to-end latency low enough (typically well under 100-200 ms) that the bounding box stays visually locked onto the correct location on screen as the endoscopist steers.
This is a fundamentally different computational problem from analyzing a single high-resolution photograph — the system must be simultaneously fast, robust to motion blur, reflections, bubbles, stool debris and specular highlights, and sensitive enough to catch a 3 mm flat lesion without drowning the operator in false alarms.
Modern real-time polyp detectors are deep convolutional neural networks (CNNs) — often built on object-detection architectures like YOLO, Faster R-CNN, or SSD variants — trained on millions of labeled endoscopic frames to recognize the visual signature of adenomatous and serrated tissue amid normal mucosa, blood vessels, folds and instrumentation artifacts.
A typical real-time CADe pipeline (exemplified by commercial systems like Medtronic's GI Genius, Fujifilm's CAD EYE, or Olympus's EndoBRAIN-EYE) passes each incoming video frame through a stack of convolutional layers that progressively extract low-level features (edges, color gradients, texture) and combine them into higher-level representations (vascular pattern irregularity, mucosal surface disruption, pit-pattern distortion).
A detection head — commonly a single-shot architecture such as YOLO (You Only Look Once) or a region-proposal network — then scans these feature maps and proposes candidate bounding boxes, each carrying a confidence score representing the network's estimated probability that the enclosed region contains a polyp.
Because colonoscopy is video rather than isolated stills, most systems also incorporate temporal information — tracking a candidate detection across consecutive frames and requiring persistence over multiple frames before triggering an alert, which sharply reduces spurious single-frame false positives caused by bubbles, light reflections, or fecal debris.
GI Genius (Medtronic) became the first FDA-authorized real-time AI colonoscopy device in April 2021, trained on more than 13 million polyp images drawn from over 2,000 procedures, and cleared based on pivotal RCT data showing a substantial ADR increase over unassisted colonoscopy.
Training a robust CADe model requires enormous, carefully curated datasets: video frames from thousands of procedures, each with pixel-level or box-level polyp annotations drawn by expert endoscopists and cross-referenced against histopathology reports confirming true adenoma status. Data augmentation (rotation, brightness/contrast jitter, simulated motion blur, synthetic bubble/debris overlay) helps the network generalize to the huge variability of real-world endoscopic conditions — different scope manufacturers, bowel-prep quality, lighting modes (white light vs. narrow-band imaging), and patient anatomy.
A persistent challenge is domain shift: a model trained predominantly on one endoscope manufacturer's image characteristics, one population's polyp morphology (e.g., Western vs. East Asian cohorts, which show different serrated-lesion prevalence), or one bowel-prep standard can lose accuracy when deployed in a different clinical setting. Multi-center, multi-device training data and prospective external validation are now expected components of regulatory submission for these devices.
The network outputs a continuous confidence score (0-100%) for each candidate detection; a threshold decides which candidates actually trigger an on-screen alert. This threshold is the central engineering and clinical tuning knob:
• Lower threshold → higher sensitivity: catches more true polyps, including subtle flat lesions, but also raises more false alarms on folds, vessels, bubbles and light reflections • Higher threshold → higher specificity: fewer nuisance alerts, smoother workflow, but risks silently missing genuine small or low-contrast lesions
Most commercial systems ship with a manufacturer-preset default threshold calibrated against validation datasets to balance these forces, though some systems allow institutional or even per-procedure adjustment. This is precisely the tradeoff modeled by the sensitivity slider in this simulation — sliding it up recruits more true-positive detections but also more false-positive boxes flashing on frames without any true lesion.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| GI Genius (Medtronic) | CADe — polyp localization | Deep CNN, marker overlay + audio cue, FDA-cleared 2021 | Largest pivotal RCT evidence base; ADR +14.4 pts in US trial |
| CAD EYE (Fujifilm) | CADe + CADx | Integrated with 700-series scope processor, real-time overlay | Combined detection + optical characterization pathway |
| EndoBRAIN / EndoBRAIN-EYE (Cybernet/Olympus) | CADx — histology prediction | Trained on endocytoscopy + NBI ultra-magnified images | High reported accuracy for optical diagnosis of diminutive polyps |
| SKOUT (Iterative Health / Wision AI) | CADe — polyp localization | Deep learning bounding-box overlay, multi-center RCT validated | FDA-cleared; validated in US community-practice settings |
Detecting a polyp computationally is only useful if the signal reaches the endoscopist's attention fast enough, clearly enough, and with enough trustworthiness to change behavior — without becoming a distracting nuisance. The human-computer interaction design of the alert is as clinically important as the underlying neural network.
When the model's confidence score for a tracked region crosses the alert threshold, the system typically overlays a colored bounding box or corner brackets directly on the live video around the suspicious area, often accompanied by a green frame border and a short audio chime — designed to be perceptible in the operator's peripheral vision without forcing them to look away from the screen. The box must track the lesion smoothly as the scope and target both move, updating at full video frame rate so it does not lag or "swim" off the target, which would erode operator trust.
Most systems display an unobtrusive marker rather than a large opaque overlay, explicitly to avoid obscuring the lesion itself or nearby mucosa the endoscopist needs to inspect for biopsy planning, vascular pattern, or margin assessment.
Some systems display a numeric or graphical confidence indicator alongside the box, giving the endoscopist a sense of how strongly the network believes a true lesion is present. Well-calibrated confidence scores allow experienced operators to mentally down-weight low-confidence alerts on frames with obvious artifacts (bubbles, stool, specular glare) while treating high-confidence, persistent alerts as high-priority.
However, over-reliance on displayed confidence carries risk: studies of automation bias in radiology and endoscopy alike show that operators can anchor too strongly on an AI system's stated certainty, either dismissing true lesions flagged with modest confidence or over-trusting confidently-flagged false positives. Device labeling and training programs increasingly emphasize that CADe output is a decision-support signal, not a diagnostic verdict — the endoscopist retains full responsibility for the final call to biopsy, resect, or pass.
In practice, an alert changes the physical exam in real time: the endoscopist typically slows or stops withdrawal, re-centers the scope on the flagged region, may apply water-jet washing or air insufflation to improve the view, and then makes the diagnostic/therapeutic decision (biopsy forceps, cold snare polypectomy, or observation). Because the scope can be repositioned within seconds, the alert-to-action loop closes quickly when the system is well-tuned.
A well-designed system also needs a “dismiss” behavior: once the endoscopist has visually confirmed and addressed a flagged lesion (or judged it a clear artifact), the alert should not continue firing repeatedly on the same already-assessed region, which is one of several strategies used to control alert fatigue over a multi-minute withdrawal.
Randomized trials measuring withdrawal time with vs. without AI assistance generally find only a modest increase (commonly under one to a few minutes), because most flagged regions are dismissed quickly by the endoscopist as soon as they are visually re-examined.
Once a lesion is localized, a second and distinct AI task begins: predicting its likely histology — adenoma, hyperplastic polyp, or sessile serrated lesion — directly from its optical appearance, without waiting for formal pathology. This CADx capability underpins "resect-and-discard" and "diagnose-and-leave" strategies that could meaningfully reduce cost and specimen-processing burden.
CADe answers "is there a lesion here?" CADx answers a harder question: "what kind of lesion is it?" This second model typically analyzes fine-grained optical features — surface pit pattern, microvascular architecture, color, and texture — often using enhanced-imaging modes like narrow-band imaging (NBI), blue-light imaging (BLI), or magnifying endocytoscopy, which accentuate the vascular and glandular features pathologists rely on under the microscope, in vivo and in real time.
The clinical target is the diminutive polyp (≤5 mm), which makes up roughly 80% of all colorectal polyps detected at screening colonoscopy. Because the cancer risk in diminutive lesions is extremely low, an accurate in-vivo prediction of "adenoma" vs. "hyperplastic" could allow either resection-without-pathology-submission (resect-and-discard, for adenomas) or complete avoidance of resection (diagnose-and-leave, for small rectosigmoid hyperplastic polyps), consistent with the ASGE's PIVI (Preservation and Incorporation of Valuable endoscopic Innovations) performance thresholds.
The ASGE PIVI statement set explicit numeric bars that any real-time optical diagnosis technology — AI-based or human expert-based — must clear before being used to guide management without pathology: for resect-and-discard, ≥90% agreement between optical and pathological assessment of post-polypectomy surveillance intervals; for diagnose-and-leave of diminutive rectosigmoid hyperplastic polyps, ≥90% negative predictive value (NPV) for adenomatous histology, with high confidence predictions.
Several CADx systems and expert-augmented AI models have reported optical accuracy in the 85-95%+ range in prospective and retrospective validation studies, with some meeting PIVI thresholds in controlled settings — but real-world, multi-center generalization performance has been more variable, and no CADx system has yet achieved the same breadth of regulatory clearance and routine clinical adoption as first-generation CADe (pure detection) systems.
Because the clinical and medicolegal cost of a false "leave it" call on a lesion that later proves to harbor advanced neoplasia is high, most gastroenterology societies still recommend that CADx output be used as an adjunct to, not a replacement for, expert visual assessment and standard pathology workflows outside of specific validated protocols.
Emerging CADx applications extend beyond simple adenoma/non-adenoma classification: estimating lesion size from endoscopic images (historically prone to endoscopist over- or under-estimation), assessing margins and submucosal invasion depth for lesions being considered for endoscopic mucosal resection (EMR) or endoscopic submucosal dissection (ESD), and flagging features suspicious for early invasive cancer that would change the resection strategy entirely (e.g., non-lifting sign, depressed morphology, or disrupted pit pattern under magnification).
These higher-order tasks require substantially more sophisticated training data — expert-annotated correlation between optical appearance and confirmed depth of invasion on resected specimens — and remain an active area of research rather than routine, widely-deployed clinical practice as of the mid-2020s.
The ultimate test of AI-assisted colonoscopy is not laboratory sensitivity/specificity numbers but whether it changes real clinical outcomes across large populations: does it find more adenomas per procedure, does it reduce the interval-cancer rate that follows missed lesions, and does it do so without imposing an unacceptable burden of false alarms, added procedure time, or unnecessary resections.
Multiple randomized controlled trials — conducted in Italy, China, Japan, the US and elsewhere, and subsequently pooled in several meta-analyses — have compared AI-assisted (CADe-on) colonoscopy against standard unassisted colonoscopy in screening and surveillance populations. The consistent finding: AI assistance increases the adenoma detection rate by roughly 8 to 14 percentage points in absolute terms — for example, from a baseline ADR around 25-40% up to roughly 35-50% — corresponding to a relative increase of about 20-30%.
The additional adenomas found by AI are disproportionately small (<5 mm), flat, and non-polypoid lesions and sessile serrated lesions — precisely the morphologies human tandem-colonoscopy studies show are missed most often — while detection of larger, more obviously polypoid adenomas is essentially unchanged between AI-assisted and standard exams, since experienced endoscopists rarely miss those regardless.
A widely cited pooled meta-analysis of RCTs (Barua et al. and related pooled analyses, 2021-2023) found AI-assisted colonoscopy increased adenomas detected per colonoscopy without meaningfully lengthening total procedure time, and several individual RCTs (including a US multi-center GI Genius trial) reported ADR increases exceeding 14 percentage points versus the control arm.
ADR is a validated surrogate for outcomes that matter most — post-colonoscopy (interval) colorectal cancer incidence and CRC-specific mortality — based on the strong epidemiological relationship established in large cohort studies (each 1-point ADR increase associated with roughly a 3% reduction in interval cancer risk). However, because interval cancers develop over years, no completed RCT has yet directly demonstrated that AI-assisted colonoscopy reduces interval CRC incidence or mortality; that evidence would require multi-year follow-up of large screened populations and is still accumulating.
Most of the clinical case for AI-CADe adoption therefore currently rests on the strength of the ADR surrogate relationship plus consistent RCT evidence on detection metrics, rather than on completed hard-outcome trials — an important nuance in how the technology is discussed in professional guidelines.
The tradeoff for improved sensitivity is an increase in false-positive alerts — the system flashing a box around a fold, bubble, reflection, or bit of debris that is not a true lesion. Reported false-positive rates vary substantially by system and threshold setting, from roughly one false alert every few minutes to several per minute in more sensitive configurations.
Frequent false alarms create two distinct problems: alert fatigue, where endoscopists become desensitized and may begin ignoring or dismissing alerts (including eventually true ones) without full visual verification; and unnecessary biopsies or resections of lesions that a careful human observer would have recognized as benign artifacts, adding cost, tissue-pathology burden, and small but real procedural risk (bleeding, perforation) without benefit. Ongoing engineering focus is on suppressing artifact-driven false positives (temporal consistency filtering, artifact-specific rejection classes) while preserving sensitivity to genuine flat, low-contrast lesions.
A notable controversy has emerged around whether continuous AI assistance, if uniformly adopted, could erode endoscopists' own detection skills over time (a "deskilling" or automation-complacency concern) — some observational studies have reported modestly lower ADR when the same endoscopists later perform unassisted exams after a period of AI-assisted practice, though this remains an active area of investigation rather than settled fact.
Professional society guidance (e.g., from the ASGE and European Society of Gastrointestinal Endoscopy) has generally moved toward conditional or qualified recommendations supporting AI-CADe use in average-risk screening colonoscopy, given the consistent ADR benefit, while emphasizing that it is an adjunct to — not a substitute for — meticulous technique: adequate bowel preparation, adequate withdrawal time (≥6 minutes), and thorough fold examination, all of which remain the foundation of exam quality regardless of AI assistance.