A Novel Convolutional Neural Network Model as an Alternative Approach to Bowel Preparation Evaluation Before Colonoscopy in the COVID-19 Era: A Multicenter, Single-Blinded, Randomized Study

Published: 26 October 2021| Version 1 | DOI: 10.17632/ydsg7y56gr.1
Contributors:
YANGBOR LU,
,
,
,
,
,
,
,
,
,

Description

An adequate bowel preparation is the key to a successful colonoscopy for detecting adenomas and preventing colorectal cancer. We developed an artificial intelligence (AI) platform using a convolutional neural network (CNN) model to evaluate the quality of bowel preparation before colonoscopy. This was a colonoscopist-blinded, randomized study. A total of 1,434 patients were enrolled and randomized into groups that used AI-CNN model to evaluate the quality of bowel preparation (AI-CNN, n = 730) or performed self-evaluation per routine practice (Control, n = 704). The primary outcome was the consistency between two methods. Secondary outcomes included quality of bowel preparation according to Boston Bowel Preparation Scale (BBPS), polyp detection rate (PDR), and adenoma detection rate (ADR). There was no significant difference in evaluation results (pass or not pass) in respect of the adequacy of bowel preparation per BBPS score between groups, suggesting that the AI-CNN model and routine practice were generally consistent in the evaluation of bowel preparation quality. The mean BBPS score, PDR, and ADR were also similar in both groups. Additionally, it's worth noting that the mean BBPS score of patients with pass results was significantly higher for the AI-CNN group than for the Control group, indicating that the AI-CNN model may improve the quality of bowel preparation further in patients who showed adequate bowel preparation. The novel AI-CNN model, demonstrating comparable outcomes to the routine practice, may potentially be an alternative approach for evaluating the bowel preparation quality before colonoscopy.

Files

Steps to reproduce

Outpatients, inpatients, or people having health check aged from 18 to 60 years old, with scheduled colonoscopy, using the smartphone, and being able to read and understand webpages were eligible to participate. Exclusion criteria were: (1) contradictions for colonoscopy; (2) familial adenomatous polyposis; (3) known colorectal polyps; (4) received gastrointestinal surgery; (5) unable to give informed consent; and (6) refused to receive colonoscopy. Demographics and clinical characteristics were recorded at the time of appointment for colonoscopy. Colonoscopies were performed by experienced colonoscopists who were unaware of the bowel preparation method that patients used. Nurses working in endoscopic centers were also blinded to patients’ preparation methods before, during, and after the procedures. Fujinun and Olympus colonoscope were used to conduct colonoscopy. The blinded colonoscopist recorded the quality of bowel preparation, colonoscopic findings, cecal intubation time, and withdrawal time for patients. Bowel preparation quality was assessed by BBPS, which is a validated tool for the assessment of colon cleansing quality.Each segment of the colon (right, transverse, and left colon) was scored from 0 to 3, where 0 represents a poorly prepared colon segment and 3 represents successful preparation with visible mucosa of the colon. Scores of three segments were added up as the total BBPS scores, varying from 0 to 9. Adequate bowel preparation was defined as a total BBPS score ≥ 6. A CNN model was employed to evaluate the bowel preparation images. To train the model, 4,302 images were collected from Zhongshan Hospital Xiamen University. All the images were labeled with “pass” or “not pass” by an experienced nursing staff from the endoscopy center of Chang Gung Memorial Hospital, Taiwan. The parameters of the CNN model were set randomly in the beginning. In the training process, the labeled images were put into the CNN model to adjust the parameters of the model iteratively. When the training process was finished, the CNN model can give either a “pass” sign or a “not pass” sign to a new uploaded bowel preparation image. The bowel-preparation-image-analysis CNN model was developed as a web service for convenient use by using an architecture called “MobileNet” because we aimed to serve many patients from different hospitals simultaneously and give feedback immediately. Compared with other CNN architectures, MobileNet can solve the classification problem with a smaller-scale network that has fewer parameters. The MobileNet architecture was therefore employed for its less computational resources requirement and fast inference speed.

Institutions

Xiamen Chang Gung Hospital, Shenzhen Hospital of Southern Medical University

Categories

Colonoscopy

License