Introduction

Colonoscopy as gold standard for colorectal cancer screening remains subject to subjective interpretation. Computer-aided detection (CADe) improves colorectal polyp detection and needs to be clinically tested as an objective measure.

Aims & Methods

This trial aimed to validate a novel CADe system for colorectal polyp detection versus trained endoscopists in an innovative endoscopist blinded set-up with real-time unblinding by a second observer. This study design allows to assess the performances of both CADe and endoscopists.
The CAD-ARTIPOD trial is an investigator-initiated prospective multicenter trial conducted in 9 European centers. The CADe system is unique in design combining a convolutional neural network with a recurrent neural network for detection of colorectal polyps trained in all imaging modalities. Sample size analysis for superiority in polyp detection required 2048 polyps assuming a difference in sensitivity of 2% with endoscopic diagnosis as gold standard. All participating endoscopists (adenoma detection rate (ADR) 20-50%) were blinded to the CADe output. A trained second observer assessed and classified in real time the CADe output in
(1) true positive, either detected or not detected by the endoscopist,
(2) False negative or
(3) clinically relevant or irrelevant false positive detection.
The endoscopist was the gold standard for the primary endpoint (endoscopic lesion detection) and if a CADe detected polyp was not picked-up by the endoscopist, the endoscopist was unblinded for confirmation of the finding. All polyps detected and confirmed endoscopically were resected. Histological assessment was performed by two independent histopathologists.

Results

A total number of 2141 polyps were collected in 946 patients of which 1958 (91.5%) polyps were resected and 1608 (82.1%) were histologically confirmed, with 475 hyperplastic polyps , 989 adenomas, and 119 sessile serrated lesions, 15 inflammatory pseudopolyps and 10 carcinomas.
For the primary endpoint superiority was not reached between CADe and endoscopists with a sensitivity of 0.946 and 0.961, respectively (non-inferiority margin of 5%, p = 0.0216). The CADe system detected 3.9% extra polyps, missed by the endoscopist. False positive rate was 1.8/min clean withdrawal time and 36.3% of those were deemed clinically relevant.
With histology as gold standard, CADe outperformed the endoscopists with a sensitivity of 0.9602 versus 0.9496 (p=0.03) and showed an extra detection rate of 5.0%. Remarkably, baseline polyp detection rate (PDR) increased from 0.47 to 0.70 during the trial (D0.48, OR 2.76 (95% CI: 2.26 – 3.37)). Similarly, ADR increased from 0.38 at baseline to 0.45 (D0.21, OR 1.52 (95% CI: 1.26 ; 1.85)).

Conclusion

This multiplatform CADe system, used in an innovative and competitive study design, demonstrated non-inferiority in comparison to endoscopists’ polyp detection. However, with histology as gold standard, CADe outperformed endoscopists. Our study also shows that “watching over a shoulder” in itself leads to higher PDR and ADR.