基于人工智能的非放大内镜结直肠病变自动NICE分型:从构建到部署  

AI-Based non-magnifying endoscopic automatic NICE classification of colorectal lesions:From development to deployment

作　　者：陈健[1] 夏开建 卢勇达 丁雨 刘罗杰 徐晓丹 Chen Jian;Xia Kaijian;Lu Yongda;Ding Yu;Liu Luojie;Xu Xiaodan(Department of Gastroenterology,Changshu No.1 People's Hospital,Suzhou 215500,Jiangsu,China;Changshu Key Laboratory of Medical Artificial Intelligence and Big Data,Suzhou 215500,Jiangsu,China;Department of Gastroenterology,The First Affiliated Hospital of Suzhou University,Suzhou 215500,Jiangsu,China)

机构地区：[1]常熟市第一人民医院消化内科,江苏苏州215500 [2]常熟市医学人工智能与大数据重点实验室,江苏苏州215500 [3]苏州大学第一附属医院消化内科,江苏苏州215500

出　　处：《兰州大学学报(医学版)》2024年第12期18-31,共14页Journal of Lanzhou University(Medical Sciences)

基　　金：苏州市第二十三批科技发展计划项目(SLT2023006);常熟市医学人工智能与大数据重点实验室能力提升项目(CYZ202301);常熟市科技计划项目(CS202019,CS202452);苏州卫生信息与健康医疗大数据学会项目(SZMIA202402)。

摘　　要：目的近年,欧洲与中国的年轻人群中结肠癌发病率明显增加。尽管窄带成像(NBI)国际结直肠息肉内镜(NICE)分型为此提供了策略,但年轻医师在应用中仍面临挑战。本研究涵盖了从构建到部署的全流程,致力于构建基于NICE分型的人工智能深度学习模型,旨在提升结直肠病变鉴别的准确度并协助临床决策。方法基于3个数据集:数据集1(苏州大学附属常熟医院,n=2050)、数据集2(常熟市中医院,n=456)用于模型训练与测试,数据集3(苏州大学附属第一医院,n=99)作为外部测试集。纳入研究的结肠镜图像包括正常肠道及NICE 3个分型。图像经过预处理和增强后,采用基于卷积神经网络和Transformer的预训练模型进行迁移学习。模型训练采用交叉熵损失函数,使用Adam优化器,并实施学习率调度。模型的评估指标包括准确性、精确度、召回率、F1分数等,并进行模型和内镜医师在不同阈值下的预测准确性对比。为提高模型透明度,进行深入地可解释性分析,包括梯度加权类激活映射、引导式梯度加权类激活映射和沙普利加和解释等技术。最后,模型被转换为ONNX格式并部署到多种设备终端,以实现结直肠病变的实时NICE分型。结果在2605张结肠内窥镜图像中,EfficientNet模型的表现超越了其他5种卷积神经网络和Transformer模型,测试准确率达91.03%。该模型精确率、召回率和F1得分达91.82%、91.40%和91.60%。在外部的99张测试图像上,其准确率、曲线下面积、精确度和召回率分别为91.92%、99.43%、92.96%和91.92%。虽然模型整体表现卓越,但仍存在误分型。通过可解释性分析可识别模型决策的关键区域以及导致误分型的原因。此外,模型已成功转换为ONNX格式,并在多个平台上实现了超过50帧/s的实时NICE分型。结论本研究成功构建EfficientNet模型,针对结直肠病变的NICE分型,准确率高达91.92%,在关键性能指标上优�Objective In recent years,there has been a noticeable increase in colorectal cancer incidence among younger populations in Europe and China.Although the narrow band imaging(NBI)international colorectal eadoscopic(NICE)classification provides strategies,junior physicians still face challenges in getting themselves familliar with its application.This study encompassed the entire process from development to deployment,dedicated to constructing an AI deep learning model based on the NICE classification to enhance the accuracy of colorectal lesion differentiation and assist clinical decision-making.Methods Three datasets were utilized in this research:Dataset 1(from the Affiliated Changshu Hospital of Suzhou University,n=2050)and Dataset 2(from Changshu Traditional Chinese Medicine Hospital,n=456)for model training and testing,and Dataset 3(from the First Affiliated Hospital of Suzhou University,n=99)serving as an external validation set.The endoscopic images under investigation encompassed normal intestines and the three categories of NICE classification.After preprocessing and enhancement,images underwent transfer learning using pretrained models based on convolutional neural network(CNN)and Transformer architectures.Model training employed the cross-entropy loss function,utilized the Adam optimizer,and implemented learning rate scheduling.Model evaluation metrics included accuracy,precision,recall,F1-score,and a comparison of predictive accuracy between the model and endoscopists at various thresholds.For enhanced model transparency,extensive interpretability analysis was performed using techniques like gradient-weighted class activation mapping,guided gradient-weighted class activation mapping and Shapley additive explanations.Ultimately,the model was converted to the ONNX format and deployed across multiple device endpoints to facilitate real-time NICE classification of colorectal lesions.Results In the analysis of 2605 colonoscopy images,the Efficient-Net model outperformed 5 other CNN and Transformer models,ach

关 键 词：深度学习 窄带成像国际结直肠息肉内镜 结肠镜 卷积神经网络模型 Transformer模型 模型部署 

分 类 号：R574.6[医药卫生—消化系统]