Abstract:

Application of artificial intelligence based on deep learning in colonoscopy        Ayimukedisi·yalikong*, ZHUANG Hui-jun, CAI Shi-lun, et al. *Department of Endoscopy Center, Zhongshan Hospital, Fudan Universiry, Shanghai 200032, China
Corresponding author:ZHONG Yun-shi，E-mail：13564623481@126.com
Ayimukedisi·yalikong and ZHUANG Hui-jun are the first authors who contributed equally to the article.
Abstract    Objective    To achieve the standardization of colonoscopy by using the artificial intelligence-assisted polyp classification algorithm and colonoscopy quality evaluation algorithm. Methods    A total of 18 962 images obtained from January 2018 to August 2018 were collected from endoscopic database in Zhongshan Hospital, Fudan University. Among them, 7140 images were used for the establishment of colonoscopy quality control evaluation algorithm. A total of 11 822 images were used for the establishment of polyp classification algorithm. The images were used as the input of the convolutional neural network （CNN）to train the end-to-end CNN. Results    The quality control evaluation model showed the accuracy rate of 95.27% of ileocecal valve recognition. The AUC was up to 0.9769. The performance of the quality control evaluation model was satisfactory. The total accuracy of the model in identifying the colonoscopy images based on the four categories of Boston scoring standard was 76.96%. For the polyp classification algorithm，the AUC was 0.8695. Conclusion The deep learning model established in the study has good sensitivity and AUC value. It can assists doctors to evaluate the quality of colonoscopy examination and conduct real-time assessment of intestinal polyps, so as to achieve standardization of colonoscopy examination and improve the level of colonoscopy diagnosis and treatment.

Key words: colonoscopy, deep learning, artificial intelligence, quality control, colorectal polyp

摘要： 目的    构建人工智能辅助的结肠镜质量评估算法及肠息肉形态分类算法，客观评估肠镜检查质量、息肉形态，实现结肠镜检查的规范化和统一化。方法    收集复旦大学附属中山医院2018年1月至8月，共18 962张肠镜图片。其中7140张用于肠镜质量评估算法建立，11 822张用于肠息肉形态分类算法建立。把肠镜图像作为卷积神经网络（CNN）的输入，端到端训练卷积神经网络，实现肠镜图像的分类任务，从而建立算法。其中包括3个模型：（1）肠道准备质量评分（四分类）。（2）回盲瓣的识别（二分类）。（3）无蒂和有蒂息肉的分类（二分类）。结果    肠镜质量评估模型对回盲瓣识别的准确率为95.27%，受试者工作特征（ROC）曲线下的面积（AUC）为0.9769，对基于波士顿评分标准四分类的图像的识别总精度为76.96%。肠息肉形态分类模型的AUC值为0.8695。结论    该深度学习模型用于肠镜检查质量的评估和肠息肉形态学的分类,具有良好的特异度、敏感度和AUC值，可辅助医师对肠镜检查质量进行评价，并对肠息肉进行分类，实现规范化和统一化。

关键词: 结肠镜检查, 深度学习, 人工智能, 质量控制, 结肠息肉