Computer Aided Diagnosis (CADx) for Colorectal Polyps Resect-and-Discard Strategy (CADx)

Computer Aided Diagnosis (CADx) for Colorectal Polyps Resect-and-Discard Strategy: a Multi-centre Randomized Controlled Trial

Colonoscopic removal of adenomatous polyps reduce both the incidence and mortality of colorectal cancer (CRC). The common clinical management of colorectal polyp detected during colonoscopy is to remove them and send for histopathology to determine the subsequent surveillance interval. More than 80% of polyps detected during screening or surveillance colonoscopy are diminutive (≤5mm). As the chance of diminutive polyps to harbor cancer or advanced neoplasia is low, leave-in-situ and resect-and-discard strategies using optical diagnosis are recommended for non-neoplastic polyps by the American Society for Gastrointestinal Endoscopy (ASGE) and the European Society for Gastrointestinal Endoscopy (ESGE) so as to reduce the financial burden of polypectomy and histopathology. The societies proposed leave-in-situ strategy if optical diagnosis can achieve a negative predictive value (NPV) of >90% for rectosigmoid polyp and resect-and-discard if an agreement of more than 90% concordance with histopathology-based post-polypectomy surveillance interval can be achieved. However, optical diagnosis is operator dependent and most endoscopists are reluctant to adopt this strategy in routine practice because of the need of strict training and auditing and fear of incorrect diagnosis.

In the past decade, with the exponential increase in computational power, reduced cost of data storage, improved algorithmic sophistication, and increased availability of electronic health data, artificial intelligence (AI) assisted technologies were widely adopted in various healthcare settings to improve clinical outcomes, especially the quality of colonoscopy in the area of gastroenterology. Real time use of computer-aided diagnosis (CADx) for adenoma using AI systems were developed and proven to be useful to help endoscopists to distinguish neoplastic polyps from non-adenomatous polyps. However, these studies only examined diminutive polyp but not polyp of larger size (>5mm). They were conducted with small sample size of less than few hundred subjects and the study settings were open-label and non-randomized.

The investigators aim to conduct a large scale randomized controlled trial to evaluate the performance of colorectal polyp characterization of all size polyps by real-time CADx using AI system against conventional colonoscopy with optical diagnosis.

Study Overview

• Status: Recruiting
• Conditions: Colonic Polyp
• Intervention / Treatment: Procedure: AI-powered computer-aided diagnosis (CADx)

Detailed Description

Study setting

This is an international, prospective, multi-centre, single-blind, non-inferiority, randomized controlled trial conducted in 6 university-affiliated endoscopy centres China (4 centres), Hong Kong and Singapore. This study will be conducted according to the CONSORT-AI and SPIRIT-AI guideline and complied with ICH-GCP and the declaration of Helsinki.

Artificial intelligence polyp characterization system

The investigators refer to the NICE (Narrow Band Imaging International Colorectal Endoscopic) classification as the standard to establish a deep neural network model for polyp type differentiation. To build the polyp characterization model, the investigators collected 3762 images of polyps under NBI for model training and testing, including 1483 cases of hyperplastic polyps, 1993 cases of adenomas and 286 cases of advanced tumors.

The difference of image features among the three types of NICE classification is obvious, and it is easy to distinguish them by computer under endoscope. Considering the processing capacity of the computer hardware equipped with the model, in order to achieve real-time analysis under limited computing resources, the investigators chose a lightweight network architecture called Mobile-Net to build the model. In a Mobile-Net framework, the first is a 3x3 standard convolution layer, followed by a heap of depth-wise separable convolution layers. Some of the depth-wise convolution layers will be down sampled through streets set as 2. The following average pooling layer changes the features to 1x1. According to the predicted category size, a full connection layer is then added, and finally a soft-max layer is added. If the depth-wise convolution layers and point-wise convolution layers are calculated separately, the entire network has only 28 layers ( Avg Pool and Softmax are not included). At present, the model has not been clinically validated. The investigators used a five-fold cross validation to evaluate the accuracy of the model. The results showed that the classification accuracy of the model in the current dataset exceeds 99%

Standardized training of optical diagnosis

Before the initiation of the study, participating endoscopists will be given a standardized online training workshop on the operation of the AI polyp characterization system, principles of optical diagnosis and an image-based quiz.

Colonoscopy procedures

Study colonoscopies will be performed by 12 non-expert endoscopists (colonoscopy experience <2000) and 12 expert endoscopists (colonoscopy experience ≥2000), with 2 non-expert and 2 expert endoscopists from each of the 6 centres. All colonoscopies will be performed by using high-definition endoscopy system and colonoscope (EVIS Lucera Elite, 290 series, Olympus, Co Ltd, Tokyo, Japan). All enrolled patients received low-fiber diets 3 days prior to colonoscopy and underwent bowel cleansing with polyethylene glycol solution in split dose based on institutional protocol. Colonoscopies will be performed under conscious sedation. Bowel preparation quality will be rated by the Boston Bowel Preparation Scale (BBPS) with adequate bowel preparation being defined as BBPS score ≥6 and any segmental BBPS score ≥2.

For patients randomized to the conventional colonoscopy (CC) with optical diagnosis group, the endoscopists will turn on narrow band imaging (NBI) mode once a polyp is detected. An optical diagnosis will be provided without magnification. Hyperplastic and sessile serrated polyps are categorized as non-neoplastic, adenoma and malignancy as neoplastic. The diagnosis will be recorded with the level of confidence of the assessment (high or low). For patients randomized to the AI-powered CADx colonoscopy (AI) group, endoscopists will turn on CADx upon detection of colorectal polyp. The polyp will then be marked by a blue tracking box and shown on the same high-definition monitor of the endoscopy system. The AI will provide a diagnosis (non-neoplastic or neoplastic) next to the tracking box. For the purpose of this study, all polyps irrespective of optical or AI diagnosis will be removed by endoscopic polypectomy/endoscopic mucosal resection (EMR)/endoscopic submucosal dissection (ESD) during withdrawal phrase and send for histopathology. Endoscopists are not allowed to leave any identified polyp in-situ at their discretion.

Location, size and morphology of all removed polyps will be documented. Proximal colon is defined as segment from cecum to transverse colon, while distal colon is defined as segment from splenic flexure to rectum. Polyps will be classified into pedunculated or non-pedunculated morphology and non-pedunculated lesions will be further characterized as sessile, flat or depressed lesions. Polyps removed will be placed in separate specimen bottles. Colonic polyp specimens will be evaluated by pathologists of individual study centre who are blinded to study group assignment. Colonic polyp specimens will be evaluated according to the 2019 WHO classification of tumours of the digestive system. Advanced adenomas are defined as adenomas with size ≥10mm, villous component, or high-grade dysplasia. Diagnostic yield of any existing colonic pathology will not be comprised if the subject is allocated to the AI group. Patients will be monitored for any immediate adverse event post colonoscopy and will be discharged after recovery from sedation.

• Study Type: Interventional
• Enrollment (Estimated): 1764
• Phase: Not Applicable

Contacts and Locations

• Study Contact: Name: Felix Sia; Phone Number: 26370428; Email: felixsia@cuhk.edu.hk

Study Locations

• Hong Kong
  • Hong Kong, Hong Kong
    • Recruiting
    • Combined Endoscopy Unit, ALice Ho Miu Ling Nethersole Hospital
    • Contact: Felix Sia; Phone Number: 26370428

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: No

Description

Inclusion Criteria:

• undergoing elective colonoscopy with any indication (screening, surveillance or diagnostic) and complete colonoscopy (caecal intubation) with at least one colorectal polyp detected will be recruited

Exclusion Criteria:

• personal history of CRC or inflammatory bowel disease, prior colorectal surgery, receiving anticoagulant therapy
• lack of informed consent

Study Plan

How is the study designed?

Design Details

• Primary Purpose: Diagnostic
• Allocation: Randomized
• Interventional Model: Parallel Assignment
• Masking: Single

Number of Arms

2

Arms and Interventions

Participant Group / Arm	Intervention / Treatment
Experimental: AI arm; AI will be used to diagnose polyps	Procedure: AI-powered computer-aided diagnosis (CADx); Real time AI will be used to diagnosis polyps found during colonoscopy.
No Intervention: Control arm; Conventional colonoscopy without AI will be perform to diagnose polyps

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
diagnostic accuracy of polyp histology in %	the NICE classification (NICE I: non-neoplastic polyp: NICE II: adenoma; NICE III: invasive tumor) given by the AI CADx and conventional optical diagnosis will be compared against the polyp histopathology (reference standard)	24 months

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
sensitivity (%)	sensitivity to correctly predict polyp histopathology polyps.	24 months
specificity (%)	specificity to correctly predict polyp histopathology polyps	24 months
positive predictive value (%)	positive predictive value to correctly predict polyp histopathology polyps	24 months
negative predictive value (%)	negative predictive value to correctly predict polyp histopathology polyps	24 months
agreement in assigning post-polypectomy surveillance intervals with pathology-based diagnoses (%)	agreement in assigning post-polypectomy surveillance intervals with pathology-based diagnoses according to the latest USMSTF, ESGE and Asia-Pacific guidelines	24 months

Collaborators and Investigators

• Sponsor: Chinese University of Hong Kong

Study record dates

Study Major Dates

• Study Start (Actual): September 29, 2023
• Primary Completion (Estimated): September 28, 2025
• Study Completion (Estimated): December 31, 2025

Study Registration Dates

• First Submitted: September 6, 2023
• First Submitted That Met QC Criteria: September 28, 2023
• First Posted (Actual): October 2, 2023

Study Record Updates

• Last Update Posted (Actual): October 11, 2023
• Last Update Submitted That Met QC Criteria: October 9, 2023
• Last Verified: October 1, 2023

More Information

Terms related to this study

• Additional Relevant MeSH Terms: Pathological Conditions, Anatomical; Intestinal Polyps; Polyps; Colonic Polyps
• Other Study ID Numbers: 2022.525

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No