Computer-Aided Diagnosis System Bombs in a Real-World Colon Polyp Histology Prediction Trial

In a prospective study from Singapore, the computer-aided diagnosis (CADx) program CAD EYE (Fujifilm, Tokyo, Japan) was assessed in 661 consecutive polyps of all sizes in 320 patients. The findings revealed that 61.7% of polyps were neoplastic. The overall accuracy of CADx prediction was 71.6%, which was lower than that for the endoscopists (75.2%). CADx sensitivity for neoplastic polyps was 61.8% versus 70.3% for the endoscopists. For polyps ≤5 mm, CADx sensitivity was 53.6% versus 62.7% for endoscopists. CADx performed worse when polyps were in difficult locations. It also performed worse for inexperienced endoscopists compared with experienced endoscopists.

Both CADx (80.6%) and endoscopists (83.3%) had low negative predictive value for diminutive rectosigmoid polyps. With regard to supporting the resect-and-discard paradigm, CADx for diminutive lesions plus histology for lesions >5 mm produced agreement in surveillance intervals (using the U.S. Multisociety Task Force on Colorectal Cancer recommendations) in 70.8% of cases. Endoscopist prediction resulted in agreement in 71.9%.