Introduction

Computer-aided diagnosis (CADx) systems for automated prediction of polyp histology in colonoscopy are commercially available. These systems provide real-time binary outputs of “neoplastic” and “hyperplastic” in detected polyps to guide clinical decision-making. While false positive polyp detections with computer-aided detection models has been reported, these do not taken into account the predicted histology of the polyps as they are unable to generate such an output. The magnitude and clinical relevance of false positive predictions of polyp histology in colonoscopy in a clinical setting has not been studied. The clinical characteristics of false positive CADx predictions and how it affects polyp surveillance intervals are also unknown.

Aims & Methods

We aim to determine the frequency and characteristics of false positive CADx predictions and evaluate its effect on recommendations for polyp surveillance.
Patients more than 40 years old with indications for colonoscopy were recruited to undergo colonoscopy using a CADx system (Fujifilm Corp., Tokyo) in four large tertiary referral centres. Exclusion criteria included patients with known unresected colorectal cancer, incomplete colonoscopy and inflammatory bowel disease. False positive CADx predictions were defined as non-adenomatous and non-hyperplastic polypoid lesions for which the CADx system generated a definite output for predicted histology during colonoscopy. All polyps which had a CADx prediction generated were resected for final histology, which formed the reference standard for polyp surveillance intervals according to US Multi-society Task Force (USMTF) and European Society of Gastrointestinal Endoscopy (ESGE) guidelines. Polyp characteristics and patient demographics were recorded in the study. The frequency and characteristics of false positive CADx predictions were analysed on a per-polyp level, while its effect on polyp surveillance intervals were interpreted on a per-patient level.

Results

775 polyps had an automated CADx prediction in 326 patients between February 2021 and October 2022. 8.9% (69/775) were false positive CADx predictions in 51 patients (15.6% of patients; mean age 61.8 ± 9.0 years; 43.1% female). 20.3% (14/69) and 79.7% (55/69) of the false positive CADx predictions were “neoplastic” and “hyperplastic”, respectively. The mean polyp size for false positive predictions was 3.3 ± 1.7mm. Final histology for false positive predictions were polypoid mucosal folds (84.1%; 58/69), inflammatory polyps (8.7%; 6/69), juvenile polyps (2.9%; 2/69), lipomas (2.9%; 2/69) and leiomyoma (1.4%; 1/69). False positive CADx predictions overestimated the recommended polyp surveillance interval for patients based on final histology in 9.8% (5/51) according to USMTF guidelines and 2.0% (1/51) according to ESGE guidelines.

Conclusion

False positive CADx predictions were seen in 8.9% of polyps and 15.6% of patients in our study. The false positive detections occurred mainly in diminutive polyps. While an output of “hyperplastic” accounted for most of these false positive predictions, there were also “neoplastic” outputs generated which resulted in an overestimation of the recommended surveillance intervals in 9.8% and 2.0% of affected patients according to USMTF and ESGE guidelines, respectively.

References

1. Tang CP, Lin TL, Hsieh YH, Hsieh CH, Tseng CW, Leung FW. Polyp detection and false-positive rates by computer-aided analysis of withdrawal-phase videos of colonoscopy of the right-sided colon segment in a randomized controlled trial comparing water exchange and air insufflation. Gastrointest Endosc. 2022 Jun;95(6):1198-1206.e6
2. Holzwanger EA, Bilal M, Glissen Brown JR, Singh S, Becq A, Ernest-Suarez K, Berzin TM. Benchmarking definitions of false-positive alerts during computer-aided polyp detection in colonoscopy. Endoscopy. 2021 Sep;53(9):937-940