Introduction

Histopathology remains the gold standard for the assessment of diminutive (≤5 mm) colorectal polyps, but it requires considerable cost and resources allocation. Optical diagnosis supported by artificial intelligence (AI) has shown promise, though its integration into clinical decision-making is limited. Our prior work showed the feasibility of replacing histopathology with AI-assisted human optical diagnosis in real-time practice.¹

Aims & Methods

This study aimed to assess patient acceptance, and diagnostic performance of autonomous computer-aided optical diagnosis (CADx) when used as the sole determinant in resect-and-discard (RD) and diagnose-and-leave (DL) strategies.
This prospective study was conducted at the University of Montreal Hospital Center between August and November 2024 (ID: CER23.095, NCT06059378). Consecutive patients aged 45–80 years undergoing elective colonoscopy were approached and provided detailed information about the use of CADx, and the principles of the RD and DL strategies. For consented patients, procedures were performed using CADEYE system, which autonomously provided an optical diagnosis for detected diminutive polyps. Lesions predicted to be neoplastic were removed and discarded without histopathologic analysis (RD), while rectosigmoid lesions predicted to be hyperplastic were left in place (DL). All colonoscopies were video-recorded. To establish the reference standard, the videos of all polyps undergoing optical diagnosis were independently reviewed by two expert endoscopists. In cases of disagreement, a third reviewer arbitrated the final decision. The primary outcome was the proportion of patients consenting to undergo autonomous CADx. Secondary outcomes included the accuracy of the RD strategy, the negative predictive value (NPV) of the DL approach, and the agreement between CADx- and expert-driven surveillance interval recommendations.

Results

95/102 patients approached (93.1%) consented to undergo CADx-based optical diagnosis (mean age 65.7 years). Among 149 diminutive polyps detected (51 (34.2%) in the rectosigmoid and 98 (65.8%) in the proximal colon), CADx provided optical diagnosis in 98.7% (n=147) of cases. The RD and DL strategies were applied to 113 (76.9%) and 30 (20.4%) of polyps, respectively. After excluding indeterminate cases (n=5) and sessile serrated lesions (SSLs) (n=14), the RD strategy achieved an accuracy of 83.5% and the DL strategy demonstrated a NPV of 92.3%. Surveillance interval agreement between CADx and expert recommendations was 100% (95%CI 96.2-100.0). Sensitivity analyses showed reduced performance when SSLs were classified as neoplastic, but Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) benchmarks were met when SSLs were considered hyperplastic.

Table 1. Diagnostic performance of computer- assisted optical diagnosis (CADx), using expert optical diagnosis as the reference.

Conclusion

Autonomous CADx achieved high patient acceptance and met key performance benchmarks for managing diminutive polyps using RD and DL strategies. These findings support its potential for safe, real-world implementation, though further development is needed to address current limitations in SSL detection and to enable quality optical diagnosis practice.

References

1. Taghiakbari M, Rex DK, Pohl H, et al. Pragmatic Resect and Discard Implementation Using Computer-Assisted Optical Polyp Diagnosis. Gastroenterology 2025;168:154-156.e2.

Disclosure

Daniel von Renteln has received research funding from ERBE Elektromedizin GmbH, Ventage, Pendopharm, Fujifilm and Pentax, and has received consultant or speaker fees from Boston Scientific Inc., ERBE Elektromedizin GmbH, and Pendopharm. Roupen Djinbachian has received speaker fees from Fujifilm. The remaining authors declare that they have no conflict of interest.