Validation of a Novel AI System (CADEYE) for in VIVO Characterization of Colorectal POLYPS

Aims The aim of this study is to evaluate the diagnostic performance of a novel artificial intelligence system for characterization of colorectal polyps and compare its performance to endoscopists.

Methods We validated a recently developed AI system for polyp detection and characterization (EW10-EC02 CAD-EYE system from Fujifilm Japan). This is the only CE marked and commercially available system for real time characterization of polyps. The AI model was developed using convolutional neural networks CNN. It was pre trained on more than 50000 frames for detection and characterization of colorectal polyps. We assessed this system on real time unedited videos in two phases. In phase 1, we internally validated the system on unedited recorded colonoscopy videos from our endoscopy library. In phase 2, we prospectively evaluated the CADEYE system on real time colonoscopy videos and compared it to endoscopists’ performance as part of an ongoing optical diagnosis study. We collected data on sensitivity, specificity, NPV, accuracy and concordance between histology and AI based surveillance intervals.

Results

Tab. 1

Number of polyps =90 (60 adenomatous)	CADEYE	Endoscopists
Sensitivity	95.00 %	81.67 %
Specificity	96.67 %	80.00 %
NPV	90.62 %	68.57 %
Accuracy	95.56 %	81.11 %

The system was assessed on a total of 150 polyps (59.33 % neoplastic).Overall, sensitivity, specificity, NPV and accuracy of the AI system were 94.38 %, 95.08 %, 92.06 % and 94.67 % respectively.In phase (2) we prospectively included 90 polyps including 60 (66.6 %) neoplastic,sensitivity, specificity, NPV and accuracy of the AI system were 95.0 %, 96.67 %, 90.62 % and 95.56 % respectively, compared to 81.67 %, 80.0 %, 68.57 % and 81.11 % in the endoscopists group, respectively. [Table (1)] summarizes phase(2) results.Agreement between histology and AI-based surveillance decisions was 94.44 % based on BSG guidelines, and 97.22 % based on ESGE and ASGE guidelines.

Conclusions This AI system diagnosed colorectal polyps on prospectively recorded unaltered endoscopy videos with high degree of accuracy, regardless of polyp size, morphology or location. If proven in real time studies, this could support the implementation of resect and discard strategy, with significant clinical and cost implications.

Citation: Abdelrahim M, Hossain E, Subramaniam S et al. OP120 VALIDATION OF A NOVEL AI SYSTEM (CADEYE) FOR IN VIVO CHARACTERIZATION OF COLORECTAL POLYPS. Endoscopy 2021; 53: S50.

Publication History

Article published online:
19 March 2021