基于Hill分级建立胃食管阀瓣功能自动评估的人工智能辅助系统  

Establishment of Artificial Intelligence-Assisted System for Automatic Evaluation of Gastroesophageal Flap Valve Function Based on Hill Classification

作　　者：王甘红 夏开建 周燕婷 陈健 徐璐 WANG Ganhong;XIA Kaijian;ZHOU Yanting;CHEN Jian;XU Lu(Department of Gastroenterology,Changshu Hospital of Traditional Chinese Medicine(Changshu Hospital Affiliated to Nanjing University of Chinese Medicine),Suzhou Jiangsu 215500,China;Changshu Key Laboratory of Medical Artificial Intelligence and Big Data,Suzhou Jiangsu 215500,China;Department of Gastroenterology,Changshu Hospital Affiliated to Soochow University,Suzhou Jiangsu 215500,China)

机构地区：[1]常熟市中医院(南京中医药大学常熟附属医院)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]苏州大学附属常熟医院消化内科,江苏苏州215500

出　　处：《中国医疗设备》2024年第11期17-26,共10页China Medical Devices

基　　金：常熟市医药卫生科技计划项目(CSWS202316);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)。

摘　　要：目的利用深度学习技术开发基于Hill分级的人工智能模型,提升内镜医师对食管胃连接部形态的分类效率和准确性。方法本研究采集了4个数据集,包括总计1143张胃食管瓣膜(Gastroesophageal Flap Valve,GEFV)图像和17个胃镜视频,涵盖HillⅠ、Ⅱ、Ⅲ、Ⅳ级的GEFV图像。图像经过预处理和增强,利用基于卷积神经网络(Convolutional Neural Network,CNN)和Transformer架构的预训练模型进行迁移学习。评估模型性能时,考量准确度、精准度、召回率、F1得分等指标,并将其与内镜医师在不同阈值下的预测准确性进行对比。为增强模型透明度,采用多种可解释性分析技术,包括t-SNE、梯度加权分类激活映射(Grad-CAM)和SHAP技术。最终,模型被转化为开放神经网络交换(ONNX)格式,并部署到多种设备终端上,以实现GEFV形态的实时Hill分级。结果EfficientNet-Hill模型的性能表现超越了其他6种CNN和Transformer模型,在外部测试集上的准确度达83.32%,略低于高级内镜医师(86.51%),但高于初级内镜医师(75.82%)。另外,该模型在精准度、召回率和F1得分方面分别可达84.81%、83.32%和83.95%。模型在多终端设备部署后实现了超过50fps的实时自动Hill分级。结论利用深度学习构建EfficientNet-Hill人工智能模型,实现了针对GEFV形态的自动Hill分级,能够辅助内镜医师提升内镜下分级的诊断效率和准确性,促进Hill分级纳入常规内镜报告和全球胃食管反流病评估中。Objective To develop an artificial intelligence model based on Hill classification using deep learning technology,to enhance the efficiency and accuracy of endoscopists in classifying the morphology of the esophagogastric junction.Methods Four datasets were collected,comprising a total of 1143 gastroesophagealflap valve(GEFV)images and 17 gastroscopic videos,covering GEFV images of Hill gradesⅠ,Ⅱ,ⅢandⅣ.The images were preprocessed and enhanced,then transfer learning was performed using pre-trained models based on convolutional neural network(CNN)and Transformer architectures.When evaluating model performance,metrics such as accuracy,precision,recall,and F1 score were considered,and the predictive accuracy of endoscopists at different thresholds was compared.To enhance model transparency,various interpretable analysis techniques were employed,including t-SNE,Grad-CAM,and SHAP.Finally,the model was converted to ONNX format and deployed to multiple device endpoints to achieve real-time Hill grading of GEFV morphology.Results The EfficientNet-Hill model outperformed the other six CNN and Transformer models,achieving an accuracy of 83.32%on the external test set,slightly lower than that of senior endoscopists(86.51%),but higher than that of junior endoscopists(75.82%).Moreover,the model achieved precision,recall,and F1 scores of 84.81%,83.32%,and 83.95%,respectively.Following deployment on multi-terminal devices,the model achieved real-time automatic Hill grading at over 50 fps.Conclusion The development of the EfficientNet-Hill artificial intelligence model using deep learning,has realized automatic Hill grading of GEFV morphology.It can assist endoscopists in enhancing the diagnostic efficiency and accuracy of endoscopic grading,promoting the inclusion of Hill grading in routine endoscopic reports and gastroesophageal reflux disease assessment.

关 键 词：胃食管阀瓣 深度学习 卷积神经网络 Hill分级 可解释性分析技术 人工智能辅助系统 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] R197.3[自动化与计算机技术—计算机科学与技术]