Artificial Intelligence for Diminutive Polyp Characterization

Efficacy and Cost-effectiveness of an Artificial Intelligence System (GI-Genius) on the Characterization of Diminutive Colorectal Polyps Within a Colorectal Cancer Screening Program: a Multicenter Randomized Controlled Trial (ODDITY Trial)

Artificial intelligence is a promising tool that may have a role in characterizing colon epithelial lesions (CADx), helping to get a reliable optical diagnosis regardless of the endoscopist experience. Performances of the different CADx systems are variable but it seems that, in most cases, high accuracy and sensitivities are achieved. However, these CADx systems have been developed and validated using still pictures or videos, and a real-world accurate test is lacking. No clinical trials have tested this technology in clinical practice and, therefore, performance in real colonoscopies, practical problems, applicability, and cost are unknown.

The resect-and-discard (R&D) and diagnose-and-leave (D&L) strategies have been proposed as a means to reduce costs in the evaluation of colorectal polyps avoiding a substantial number of pathology evaluations. A pre-requisite for this paradigm shift is an accurate optical diagnosis (HOD). However, performance results for HOD have been highly variable among endoscopists representing a barrier for the adoption of the R&D and the D&L strategies. Artificial intelligence is a promising tool that may have a role in characterizing colon epithelial lesions (CADx), helping to get a reliable optical diagnosis regardless of the endoscopist experience. Performances of the different CADx systems are variable but it seems that, in most cases, high accuracy and sensitivities are achieved. However, these CADx systems have been developed and validated using still pictures or videos, and a real-world accurate test is lacking. No clinical trials have tested this technology in clinical practice and, therefore, performance in real colonoscopies, practical problems, applicability, and cost are unknown. Methods and analysis: The ODDITY trial is a European multicenter randomized, parallel-group superiority trial comparing GI-Genius artificial intelligence optical diagnosis (AIOD) to human optical diagnosis (HOD) of colon lesions ≤ 5 mm performed by endoscopists, using histopathology as the gold standard. A total of 643 patients attending a colonoscopy within a CRC screening program (either FIT- or colonoscopy-based) or because of post-polypectomy surveillance will be randomized to the ADI group or the HOD (control) group. A computer-generated 1:1 blocking randomization scheme stratified for center and endoscopist will be used.

The ODDITY trial is a European multicenter randomized, parallel-group superiority trial comparing GI-Genius artificial intelligence optical diagnosis (AIOD) to human optical diagnosis (HOD) of colon lesions ≤ 5 mm performed by endoscopists, using histopathology as the gold standard. A total of 643 patients attending a colonoscopy within a CRC screening program (either FIT- or colonoscopy-based) or because of post-polypectomy surveillance will be randomized to the ADI group (group 1) or the HOD (control, group 2) group. A computer-generated 1:1 blocking randomization scheme stratified for center and endoscopist will be used.

Patients will be blinded to group allocation. The endoscopist in group 2 will be blinded to the AIOD. However, the endoscopist in group 1 will not be blinded to the CADx diagnosis because the output helps the endoscopist to focus the lesion properly for a AIOD diagnosis. In group 2, the person in charge of handling the GI-Genius output will communicate with the endoscopist when the AIOD of a particular lesion has been obtained. There is no need to mask personnel who enters HOD or AIOD data on the CRD because at that point results of pathology are not available.

The examinator will provide a HOD for every lesion (regardless of their size) found during the examination (adenoma vs non-adenoma) following one of the available validated classifications (NICE, JNET, BASIC). He/she will also give a level of confidence in his/her diagnosis (high/low confidence). However, only diminutive lesions will be considered when analyzing the main outcome. The time to get a HOD will be recorded. An in situ surveillance interval will be provided if possible.

GI-Genius will provide an artificial intelligence diagnosis (AIOD) for every lesion detected (adenoma vs non-adenoma). Only diminutive lesions will be considered for the analysis of the main outcome. However, data on larger lesions will be recorded to describe GI-Genius´ performance in detail (secondary outcome). The time to get an AIOD will be recorded. An in situ surveillance interval will be provided if possible

The software allows for the real-time characterization of framed polyps during a colonoscopy classifying them on adenoma or non-adenoma.

Comparison of the AIOD and HOD accuracy of the post-polypectomy surveillance interval assignment with respect to the surveillance interval assigned by pathology

A surveillance interval will be assigned using optical diagnosis of ≤ 5 mm polyps (Arm 1: AIOD; Arm 2: HOD of polyps diagnosed with high confidence) plus histopathology of > 5 mm polyps and polyps ≤ 5 mm diagnosed with low confidence. For each patient included, the optical-diagnosis surveillance assignment will be matched with the histology-directed one, and a concordance rate will be calculated. The post-polypectomy surveillance interval will be calculated using the ESGE 2020 and the USMSTF 2020 guidelines. Per-patient analysis.

Comparison of the AIOD and HOD negative predictive value (NPV) for adenoma in rectosigmoid polyps ≤ 5 mm with respect to histology

The optical diagnosis of ≤ 5 mm rectosigmoid polyps (Arm 1: AIOD; Arm 2: HOD, only high-confidence diagnosis) reliability on ruling out the presence of an adenoma will be calculated using histopathology as the gold standard. Per-lesion analysis. NPV = number of confirmed hyperplastic polyps/number of hyperplastic optical diagnosis

Comparison of the AIOD and HOD diagnostic accuracy parameters of polyps ≤ 5 mm (Arm 1: AIOD; Arm 2: HOD) with respect to histology

Operative characteristics (sensitivity, specificity, positive and negative predictive value and positive likely hood ratio) using histopathology as the gold standard. Per-lesion analysis

The economic burden of applying the AIOD and HOD to assign the post-polypectomy surveillance intervals compared to the histology-driven strategy. A direct cost evaluation will be performed including medical and non-medical costs. Per-patient analysis.

The occurrence and severity of adverse events in colonoscopies with and without the AIOD device will be monitored during the 30-days period after the procedure. Adverse events are defined as: abdominal pain or discomfort, post-polypectomy bleeding, perforation, post-polypectomy syndrome and infection. Per-patient analysis

Proportion of patients accepting to have their polyps diagnosed by the AI system or human optical diagnosis (designed questionnaire)

The proportion of patients willing to have their polyps diagnosed by an AI system or HOD will be assessed using a structured questionnaire. Per-patient analysis.

Inclusion Criteria: Patients attending a colonoscopy within a population-based CRC screening program (FIT- or colonoscopy-based) or because of post-polypectomy surveillance, Written informed consent before the colonoscopy, Exclusion Criteria: None, patient included Previous history of inflammatory bowel disease. Previous history of CRC Previous CR resection Polyposis or hereditary CRC syndrome Coagulopathy/Anticoagulants Unwillingness to participate