AI-Assisted Colonoscopy: How Computer-Aided Detection Improves Polyp and Adenoma Detection

What Is AI-Assisted Colonoscopy and How Does It Work?

Computer-aided detection (CADe) in colonoscopy represents one of the most successful clinical applications of medical artificial intelligence. These systems use deep learning algorithms, primarily convolutional neural networks (CNNs), trained on hundreds of thousands of annotated colonoscopy images and videos. During a procedure, the AI processes each video frame in real time (typically within tens of milliseconds), identifying regions suspicious for polyps and displaying visual alerts (boxes, highlights, or markers) on the endoscopy monitor to direct the operator's attention.

The need for AI assistance stems from the known miss rate in conventional colonoscopy. Tandem colonoscopy studies have shown that approximately 20–26% of adenomas are missed during standard colonoscopy, including clinically significant lesions. The adenoma detection rate (ADR) varies significantly between endoscopists, with rates ranging from below 15% to above 50%. Since ADR is inversely correlated with interval colorectal cancer risk — a landmark study by Corley et al. (New England Journal of Medicine, 2014) demonstrated that each 1% increase in ADR is associated with a 3% decrease in cancer risk — improving detection has direct clinical impact.

Several CADe systems have received regulatory clearance. GI Genius (Medtronic, originally developed by Cosmo Pharmaceuticals) received FDA De Novo authorization in April 2021, becoming the first AI-based tool for colonoscopy cleared in the United States. EndoScreener (Wuhan EndoAngel Medical Technology) has CE marking in Europe. Other systems include CAD EYE (Fujifilm) and ENDO-AID (Olympus), with additional products in development from other endoscopy manufacturers. These systems are designed as assistive tools that supplement, not replace, the endoscopist's clinical judgment.

What Does the Evidence Show About AI Improving Adenoma Detection?

The evidence base for AI-assisted colonoscopy has grown rapidly, with multiple randomized controlled trials (RCTs) now published. A comprehensive meta-analysis by Hassan et al., published in Gastrointestinal Endoscopy (2023), pooled data from multiple RCTs encompassing thousands of patients. The analysis demonstrated that AI-assisted colonoscopy significantly increased the adenoma detection rate (ADR), with a pooled relative risk in the range of 1.30–1.50, representing a roughly 30–50% improvement depending on the analysis. The mean number of adenomas detected per colonoscopy also increased significantly.

Individual landmark trials have confirmed these benefits. The first major RCT by Wang et al., published in Gut in 2019, randomized over 1,000 patients and found that CADe increased ADR from approximately 20% to 29% (p<0.001). A multicenter RCT by Repici et al. in Gastroenterology (2020) showed a significant ADR improvement with GI Genius, with AI-assisted ADR exceeding 50% compared to approximately 40% in the control arm. Importantly, most studies show that AI primarily improves detection of diminutive (less than 5 mm) and small (6–9 mm) polyps, with less impact on larger or advanced lesions that are already well-detected by experienced endoscopists.

One concern has been whether increased detection of small polyps translates into meaningful clinical benefit or merely leads to more polypectomies of clinically insignificant lesions. However, modeling studies suggest that even the detection of additional diminutive adenomas reduces future advanced neoplasia risk. Furthermore, AI has shown particular value in detecting flat (nonpolypoid) lesions, which are a minority of all polyps but are associated with higher malignant potential and are more commonly missed during standard colonoscopy.

Can AI Characterize Polyps in Real Time to Guide Treatment?

Beyond detection, AI is being developed for real-time polyp characterization, known as computer-aided diagnosis (CADx). These systems analyze the surface pattern, vascular morphology, and color features of detected polyps to predict histological type (adenomatous, hyperplastic, sessile serrated) during the procedure. The goal is to enable "optical diagnosis" or "resect and discard" strategies, where diminutive polyps can be characterized and managed in real time without formal pathological examination, potentially reducing healthcare costs and patient burden.

The American Society for Gastrointestinal Endoscopy (ASGE) has established the Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) thresholds for optical diagnosis: at least 90% negative predictive value for adenomatous histology of diminutive rectosigmoid polyps, and at least 90% agreement with histopathology-based surveillance intervals. Several AI-CADx systems have approached or met these thresholds in validation studies. Fujifilm's CAD EYE, which integrates both CADe and CADx functions, has achieved accuracy exceeding 90% in distinguishing neoplastic from non-neoplastic polyps in clinical evaluations.

Despite promising results, widespread adoption of AI-based optical diagnosis faces several challenges. Medicolegal concerns about mischaracterizing a malignant polyp as benign, the need for high-quality imaging (narrow-band imaging or similar enhanced modalities), and variable performance across different polyp subtypes remain barriers. Current guidelines recommend that optical diagnosis strategies be limited to expert endoscopists using validated technology. The integration of CADe and CADx into a single workflow, where AI first detects a polyp and then characterizes it in real time, represents the next frontier in AI-assisted gastrointestinal endoscopy.