Artificial Intelligence and Endoscopist Diagnostic Agreement as a Framework for Colorectal Polyp Optical Diagnosis Implementation

Aims Artificial intelligence (AI) has enabled the development of computer-aided diagnosis (CADx) systems which offer real-time endoscopic pathology prediction of colorectal polyps [1]. However, the clinical benefit of CADx assisted optical diagnosis remains questionable due to limited diagnostic improvement compared to CADx unassisted optical diagnosis [2]. This study aimed to assess diagnostic performance of a novel CADx implementation framework in which the high/low endoscopist confidence-based approach was replaced by a CADx and endoscopist agreement-based approach.

Methods We performed a secondary analysis of a large prospective cohort undergoing optical polyp diagnosis at our center. Polyps measuring≤5mm (diminutive) with available CADx unassisted and CADx assisted optical polyp diagnosis documentation were included in our analysis. For CADx assisted cases, first the CADx diagnostic output was documented, followed by the endoscopist’s final diagnosis after seeing the CADx output. Only cases where endoscopist and CADx agreed on a diagnosis were retained. Primary outcome was sensitivity for adenoma diagnosis of CADx assisted optical diagnosis based on the agreement framework versus CADx unassisted optical diagnosis, using pathology results as a reference standard. Secondary outcomes included accuracy, diagnostic characteristics, and optical diagnosis surveillance interval agreement with pathology-based United States Multi-Society Task Force (USMSTF) guidelines.

Results A total of 810 polyps, of which 444 and 366 underwent CADx assisted and CADx unassisted optical diagnosis respectively, were included in our analysis. In CADx assisted cases, endoscopists and CADx agreed in their diagnosis in 72.3% of cases, meaning 321/444 diminutive polyps could undergo optical diagnosis according to an agreement-based approach. Sensitivity for adenoma diagnosis using agreement-based CADx assisted optical diagnosis was 93.4% (95% CI 89.9-96.8) versus 83.8% (95% CI 77.9-89.7) for the 366 polyps undergoing CADx unassisted optical diagnosis (p=0.008). Diagnostic accuracy of agreement-based CADx assisted and CADx unassisted optical diagnosis was 82.2% (95% CI 78.0-86.3) and 74.8% (95% CI 69.5-80.0) respectively (p=0.026). Excluding cases with CADx and endoscopist disagreement from CADx assisted optical diagnosis filtered out cases with low diagnostic accuracy (123 polyps; accuracy 35.0% (95% CI 26.2-43.8)).

Conclusions The CADx and endoscopist agreement framework effectively filters out cases with low diagnostic accuracy from undergoing CADx assisted optical diagnosis. Applying this framework allows endoscopists and CADx to work in synergy resulting in diagnostic performance superior to CADx unassisted OD performance.

Publikationsverlauf

Artikel online veröffentlicht:
27. März 2025

• References

• 1 Hassan C, Misawa M, Rizkala T. et al. Computer-Aided Diagnosis for Leaving Colorectal Polyps In Situ: A Systematic Review and Meta-analysis. Ann Intern Med 2024; 177 (07): 919-928
• 2 Byrne MF, Chapados N, Soudan F. et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model. Gut 2019; 68: 94-100