This study aims to develop, validate and evaluate an artificial intelligence-based system for real-time automatic localization of colorectal anatomical sites during colonoscopy and detection of colonoscopic inspection extent, so as to improve examination quality and the precision of lesion management. The hypothesis of this study is that a colonoscopic image recognition model built on deep convolutional neural networks can accurately extract intestinal anatomical features from endoscopic video frames. In prospective and retrospective independent validation cohorts, the artificial intelligence model is hypothesized to achieve an overall accuracy of no less than 85% for the multi-class anatomical site localization task across colonic segments, with an expected sensitivity of 90% and specificity of 85%. The model predictions are expected to be in high agreement with assessments by senior endoscopists, with an anticipated Kappa coefficient of 0.80. This will verify the technical feasibility and accuracy of deep learning-driven real-time anatomical localization during colonoscopy.
Aged 18-75 years, male or female;
Undergoing colonoscopy;
Successful cecal intubation with the endoscope reaching the ileocecal valve;
Adequate colonic cleansing, with a Boston Bowel Preparation Scale (BBPS) score ≥2 points in each of the three colonic segments.
Reduced life expectancy;
Active gastrointestinal bleeding;
History of lower gastrointestinal tract surgery;
Previous history of colorectal cancer or adenomas;
Participants aged 18-75 years scheduled for colonoscopy, with successful cecal intubation and adequate bowel preparation. Colonoscopic videos are collected to validate an AI system for real-time anatomical site recognition and localization.
Other
Other
Artificial intelligence-based colonoscopic anatomical site recognition system
Other interventionThis artificial intelligence system performs real-time automatic recognition and localization of colorectal anatomical sites and detects the colonoscopic inspection extent during colonoscopy. The model is built with deep convolutional neural networks and validated using endoscopic video frames.
Accuracy
The percentage of samples in which the artificial intelligence model correctly identifies the corresponding intestinal anatomical sites across all endoscopic video frames, relative to the total sample size enrolled in the study. This metric reflects the overall classification accuracy of the model.
Sensitivity
The proportion of all true positive samples annotated as a specific intestinal segment by endoscopy experts that are correctly identified as this segment by the artificial intelligence model. This metric primarily evaluates the model's ability to avoid missed detection of the target anatomical site.
F1-score
To address the problem of "class imbalance" caused by variable colonoscopy examination time across different intestinal segments, this study adopts the F1-score for comprehensive evaluation. The calculation is performed as follows: first, the precision for a specific intestinal segment is obtained; then the harmonic mean of this precision and sensitivity is calculated. To assess overall performance, the arithmetic mean of the F1-scores for all intestinal segment classifications is finally computed (i.e., macro-averaged F1-score).
Macro-AUC
The receiver operating characteristic (ROC) curve reflects the model's overall discriminative performance under different decision thresholds. For this multi-class classification task, the calculation proceeds as follows: each specific intestinal segment is treated as the positive class one by one; the corresponding ROC curve is plotted and the area under the curve (AUC) is calculated for each segment. Thereafter, the arithmetic mean of AUC values across all anatomical categories is computed to obtain the macro-averaged AUC, which serves as the composite metric for evaluating the overall multi-class discriminative ability of the model.
Click on the lead sponsor to view all of their trials.
Xiuli Zuo
Lead sponsor
Shandong University
Sponsor institution
This trial is not recruiting at the moment. You can still explore other options:
