Development and validation of a three-dimensional deep learning-based system for assessing bowel preparation on colonoscopy video

• DOI: 10.3389/fmed.2023.1296249
• Publication Date: 2023-12-18T06:41:32Z
• Authors: Lina Feng; Jiaxin Xu; Xuantao Ji; Liping Chen; Shuai Xing; Bo Liu; Jian Han; Kai Zhao; Junqi Li; Suhong Xia; Jialun Guan; Chenyu Yan; Qiaoyun Tong; Hui Long; Juanli Zhang; Ruihong Chen; Dean Tian; Xiaoping Luo; Fang Xiao; Jiazhi Liao
• Keywords: Virtual colonoscopy

Background The performance of existing image-based training models in evaluating bowel preparation on colonoscopy videos was relatively low, and only a few used external data to prove their generalization. Therefore, this study attempted develop more precise stable AI system for assessing video. Methods We proposed named ViENDO assess the quality, including two CNNs. First, Information-Net identify filter out video frames unsuitable Boston scale (BBPS) scoring. Second, BBPS-Net trained tested with 5,566 suitable short clips through three-dimensional (3D) convolutional neural network (CNN) technology detect BBPS-based insufficient preparation. Then, applied complete withdrawal from multiple centers predict BBPS segment scores clinical settings. also conducted human-machine contest compare its endoscopists. Results In clips, determining inadequate generated an area under curve up 0.98 accuracy 95.2%. When full-length videos, assessed cleanliness 93.8% internal test set 91.7% dataset. demonstrated that slightly superior compared most endoscopists, though no statistical significance found. Conclusion 3D-CNN-based model showed good It has potential as substitute endoscopists provide assessments during daily practice.