The following is a summary of “Real-World Validation of a Computer-Aided Diagnosis System for Prediction of Polyp Histology in Colonoscopy: A Prospective Multicenter Study,” published in the August 2023 issue of Gastroenterology and Nutrition by Li, et al.

Computer-aided diagnosis (CADx) has the potential to assist endoscopists in making clinical decisions regarding polyp histology. However, its validation in a real-world setting was lacking.

For a prospective multicenter study, researchers compared real-time polyp histology predictions made by CADx and endoscopists during colonoscopy. Experienced endoscopists provided optical diagnoses based on visual examination of polyps, followed by recording the automated output from the CADx tool. All resected polyps underwent histological assessment. The primary outcome measured the difference in diagnostic performance between CADx and endoscopist predictions of polyp histology. Subgroup analysis considered polyp size, bowel preparation, difficulty of polyp location, and endoscopist experience.

Between March 2021 and July 2022, 661 eligible polyps were resected from 320 patients aged ≥40 years. CADx exhibited an overall accuracy of 71.6% (95% CI 68.0–75.0), whereas endoscopists achieved 75.2% accuracy (95% CI 71.7–78.4) (P = 0.023). The sensitivity of CADx for neoplastic polyps was 61.8% (95% CI 56.9–66.5), whereas endoscopists achieved 70.3% sensitivity (95% CI 65.7–74.7) (P < 0.001). Moderate interobserver agreement was observed between CADx and endoscopist predictions of polyp histology (83.1% agreement, κ 0.661). Concordance between CADx and endoscopist predictions led to an increased accuracy of 78.1%.

Experienced endoscopists demonstrated higher overall diagnostic accuracy and sensitivity for neoplastic polyps compared to CADx predictions, with moderate interobserver agreement. Achieving concordance between CADx and endoscopist predictions improved diagnostic accuracy. Further research was needed to enhance CADx performance and establish its role in clinical practice.

Source: journals.lww.com/ajg/Abstract/2023/08000/Real_World_Validation_of_a_Computer_Aided.16.aspx