Prospective Multicenter Validation of an Artificial Intelligence-Assisted Optical Diagnosis Strategy (CADx) for the Real-Time Characterization of Diminutive Colorectal Polyps: A STARD-Compliant Diagnostic Accuracy Study

Colonoscopy with polypectomy is the cornerstone of colorectal cancer (CRC) screening and prevention. The majority of polyps detected during colonoscopy are diminutive (≤5 mm) and carry a very low risk of harbouring advanced pathology. In Canada, over 1.9 million colonoscopies are performed annually, and sending all diminutive polyps for histopathological examination (at a cost of approximately $150-$300 CAD per specimen) generates a substantial economic burden with limited benefit for cancer prevention.

Two international guideline frameworks define the performance thresholds required for clinical adoption of optical diagnosis strategies for diminutive colorectal polyps. The ASGE PIVI initiative (2011) established (1) ≥90% negative predictive value (NPV) for adenomatous histology among diminutive rectosigmoid polyps diagnosed with high confidence ("diagnose-and-leave" threshold), and (2) ≥90% agreement between optical-diagnosis-based and pathology-based post-polypectomy surveillance intervals ("resect-and-discard" threshold). The more recent ESGE SODA position statement (2022) adopted sensitivity and specificity as the foundational, prevalence-independent performance measures for high-confidence real-time characterization of colorectal neoplasia, applicable across all colonic locations: sensitivity ≥90% and specificity ≥80% for the rectosigmoid subset, and sensitivity ≥80% and specificity ≥80% across all diminutive colorectal polyps regardless of location.

The investigators' research group has previously conducted several single-centre studies evaluating early-generation, Health Canada-approved computer-aided diagnosis (CADx) systems in routine clinical practice, generating one of the largest real-world CADx implementation datasets in the literature (868 patients, 1,660 diminutive polyps). These studies consistently demonstrated suboptimal overall diagnostic accuracy (62.7%-66.4%), attributable to restricted training datasets and limited polyp-classification capabilities in early-generation devices. GI Genius (Medtronic), the newer-generation, Health Canada-approved CADx module evaluated in this study, was trained on a larger and more diverse dataset of annotated endoscopic images and videos. Whether this translates into measurable gains in diagnostic performance sufficient to meet international competence thresholds has not been established in an independent, prospective, multicentre study, and is the question this study addresses.

This is a prospective, three-centre diagnostic accuracy study conducted in accordance with the Standards for Reporting Diagnostic Accuracy Studies (STARD 2015). It is a single-arm study validating a CADx-assisted optical-diagnosis strategy - rather than autonomous or standalone CADx performance - in which the endoscopist's final CADx-assisted optical diagnosis constitutes the primary index test. For each diminutive polyp (≤5 mm) identified during an already-scheduled elective colonoscopy, the following sequential steps are performed: (1) the endoscopist records a CADx-unassisted optical diagnosis (adenoma, hyperplastic polyp, or other) and confidence level (high or low), which is electronically locked before any device output is viewed; (2) the CADx module is activated and displays its own characterization of the polyp; (3) the endoscopist records a final CADx-assisted optical diagnosis and confidence level, remaining free to agree or disagree with the device; (4) a research assistant documents polyp size, morphology (Paris classification), and colonic location; and (5) the polyp is resected per standard technique and sent for histopathological examination, which serves as the reference standard. This sequential-locking design - recording and locking the unassisted diagnosis before CADx activation - is the study's key strategy for minimizing contamination bias between the two diagnostic strategies, and allows each polyp to generate two paired diagnostic data points (CADx-unassisted and CADx-assisted) compared against a single histopathological reference standard.

The study will enroll consecutive patients aged 45-80 years undergoing elective colonoscopy (screening, surveillance, or diagnostic indication) at three Canadian academic centres: Santé Québec - CHUM (Montreal), the McGill University Health Centre (MUHC/CUSM, Montreal), and St. Paul's Hospital (Vancouver), who provide written informed consent before the colonoscopy and before sedation.