Computer-assisted and unassisted optical diagnosis of colorectal polyps: results from a prospective pragmatic...

Computer-aided diagnosis (CADx) may support optical diagnosis of diminutive (≤5 mm) colorectal polyps, replacing histopathology. We described CADx usage and diagnostic performance under real-life, discretionary adoption.

This prospective implementation study enrolled patients undergoing colonoscopy at our tertiary center (2021-2025). Patients aged <40 years, with inflammatory bowel disease, polyposis syndromes, or no diminutive polyp were excluded. CADx use was at the endoscopists' discretion. Endoscopists' optical diagnoses with and without CADx assistance, and the CADx output were documented. Main outcomes were CADx usage and sensitivity for adenomas of CADx-assisted and CADx-unassisted optical diagnosis, using histopathology as reference.

868 patients (age 66.1 years, 47.0% female) with 1660 diminutive polyps (62.0% adenomas) were included. Endoscopists recorded CADx-assisted and CADx-unassisted optical diagnoses for 657 and 864 polyps, respectively, showing sensitivity of 85.1% (95%CI 78.3-90.1) and 85.1% (81.2-88.4), specificity 55.2% (45.8-64.3) and 61.1% (54.5-67.3), positive predictive value 76.2% (69.1-82.1) and 79.3% (75.5-82.9), negative predictive value (NPV) 60.8% (52.6-68.5) and 67.1% (60.7-73.0), and accuracy 62.7% (56.5-68.4) and 66.4% (63.0-69.7), respectively. In a simulated resect-and-discard strategy, CADx assistance marginally increased the proportion of polyps that could avoid histopathology (84.9% vs. 81.1%). For high-confidence rectosigmoid diagnoses, NPV for adenomas was 95.5% (86.9-98.5) for CADx-assisted and 93.8% (85.9-97.4) for CADx-unassisted predictions.

Both CADx-assisted and CADx-unassisted optical diagnoses showed good diagnostic performance in real-world practice. Assessment of the relationship between CADx use and optical diagnosis was limited by unavoidable selection bias relating to discretionary use and variable endoscopist behavior.