A real-time comparative study of CADx and sizing devices for colorectal polyps during colonoscopy: A total solution to implement resect and discard strategy?

Aims Despite resect and discard strategy being recommended by multiple GI societies, it has not been widely implemented. One of the major barriers has been the lack of available technology for accurately predicting polyp histology and size. The aim of this study was to compare two commercially available AI systems for prediction of polyp histology (CADx) and size.

Methods Patients having screening/surveillance colonoscopies were recruited into the study. The two AI platforms used in the study were WISEVISION (CADx and computer-aided sizing) from NEC, Japan and CAD-EYE (CADx)+SCALE-EYE (laser-guided sizing) from FUJIFILM. Two expert endoscopists were present in the room and both AI platforms were used simultaneously with outcomes projected on two separate screens. Ground truth for CADx was histology. Sizing ground truth was established by two experts using measuring devices like biopsy forceps or snares.

Results 122 polyps from 36 consecutive colonoscopies were included. Of them, 60 (49.2%) were diminutive and 62 non-diminutive. 42 (34.4%) were rectosigmoid and 80 non-rectosigmoid. 95 (77.9%) were flat while 27 were polypoidal.

CADe and CADx performance: All polyps were detected on both platforms. For WISEVISION, the sensitivity, specificity and accuracy of CADx to diagnose adenomatous polyps were 91.43%, 80.77% and 86.89%. Those for CAD-EYE were 89.55%, 75.51% and 83.62%. Those of expert’s optical diagnosis were 94.92%, 90.38% and 92.62%. There is no statistical difference in any performance parameter between the two systems and also in sensitivity between experts and any of the two AI systems.

Sizing performance: For computer-guided sizing with WISEVISION, the sensitivity, specificity and accuracy to size as diminutive polyps were 85%, 88.71% and 86.89%. When using laser-guided sizing with SCALE-EYE, there was a failure rate of 7.02% and the performance is 88.33%, 82.26% and 85.25% respectively. There is no statistical difference between the two systems.

Conclusions This is the first ever real-time comparative study between two different AI systems on same polyps for CADx and sizing. Our study demonstrates that the sensitivity of both CADx systems is similar and as good as experts. However, specificity of both AI systems is lower than that of experts suggesting that if these were used for resect & discard and diagnose & leave strategy, slightly more hyperplastic polyps might get resected as compared to experts.

Conflicts of interest

Professor Bhandari has received research grants or is the advisory board for Fujifilm, Boston, Olympus, Pentax, 3-D matrix, NEC (Japan), Medtronic.

Publication History

Article published online:
15 April 2024