Abstract
Background
Artificial intelligence-based optical diagnosis systems (CADx) have been developped to assist in eliminating the need for histologic diagnosis of diminutive colorectal polyps (resect-and-discard and diagnose-and-leave strategies). However, these systems have not yet been implemented in routine clinical practice.
Aims
We were interested in evaluating the performance of CADx without human input to diagnoses performed by endoscopists assisted by CADx.
Methods
We performed a randomized clinical trial of patients undergoing elective colonoscopies at the CHUM. Patients were randomized into two arms: 1) optical diagnosis of colorectal polyps using CADx without human input; 2) endoscopists performed optical diagnosis after consulting a real-time CADx diagnosis (Human in the Loop [HiL]). Primary outcome was accuracy in optical diagnosis for both arms.
Results
467 patients were randomized (229 in the CADx group, 238 in the HiL group). Overall accuracy for optical diagnosis was 76.3% in the CADx group and 70.5% in the HiL group (p=0.19). Sensitivity, specificity, PPV and NPV for adenoma diagnosis were 89.7%, 58.4%, 82.2%, and 72.5% respectively in the CADx group vs 85.3%, 69.6%, 81.8%, and 74.8% in the HiL group. Sensitivity, specificity, PPV and NPV did not differ significantly between the two groups.
Conclusions
Optical diagnosis of colorectal polyps had similar accuracy when using CADx without human input compared to endoscopist-based diagnosis assisted by CADx. Resect and discard and diagnose and leave strategies can therefore potentially be implemented without need for endoscopist optical diagnosis.
Funding Agencies
CAGFujifilm