目的 构建基于机器学习的胃肠道疾病舌诊模型,以寻求更加方便、经济的方式实现对常见胃肠道疾病的非侵入性诊断.方法 前瞻性收集接受电子内镜检查的 948 名受试者的舌象图片,经过质量筛选,最终获得符合应用标准的 3140 张图片构成本研究使用的舌象数据集.对原始舌象数据进行预处理、特征提取与模式识别,在传统机器学习方法的基础之上,提出一种从特征融合和决策融合两个方面实现信息融合的方法,以此构建以舌象特征为输入的胃肠道疾病舌诊模型.结果 本研究构建的基于舌象的信息融合诊断模型的曲线下面积(area under the curve,AUC)为0.808,高于单一手工特征(AUC=0.769)和深度特征(AUC=0.779)模型;使用BSFCM混合采样方法进行样本增强提高了该模型对幽门螺杆菌(Helicobacter pylori,H.pylori)感染(AUC=0.816)、胆汁反流(AUC=0.829)、反流性食管炎(AUC=0.800)、胃糜烂(AUC=0.833)和十二指肠糜烂(AUC=0.818)的分类性能.结论 本研究构建的基于机器学习的智能舌诊模型对多种胃肠道疾病具有较高的区分度,或为胃肠道疾病的诊断与筛查提供一种新的、有价值的思路与方法.
Construction of a machine learning-based tongue diagnosis model for gastrointestinal diseases
Objective To construct a machine learning(ML)-based tongue diagnostic model for the diagnosis of gas-trointestinal diseases so as to realize the non-invasive auxiliary diagnosis of common gastrointestinal diseases in a more convenient and faster way.Methods Tongue images of 948 subjects who underwent electronic endoscopy were prospectively collected.After quality screening,3,140 images that met the application criteria were finally obtained to constitute the tongue image data set,which underwent preprocessing,feature extraction and pattern recognition.On the basis of traditional machine learning methods,a method to realize information fusion in terms of feature fusion and deci-sion fusion was proposed,and a tongue diagnosis model of gastrointestinal diseases was constructed.Results The area under the curve(AUC)of the model was 0.808,which was higher than that of the single handcrafted feature(AUC= 0.769)and deep feature(AUC=0.779)models.Sample enhancement using the BSFCM hybrid sampling method im-proved the model's performance for Helicobacter pylori(H.pylori)infection(AUC= 0.816),bile reflux(AUC= 0.829),reflux esophagitis(AUC= 0.800),gastric erosion(AUC= 0.833)and duodenal erosion(AUC= 0.818).Conclusion The intelligent tongue diagnostic model based on ML constructed in this study shows a high degree of dif-ferentiation for a variety of gastrointestinal diseases,and may provide a new and valuable idea and method for the diag-nosis and screening of gastrointestinal diseases.
Artificial intelligenceTongue imageGastrointestinal diseasesMachine learningTongue diagnosis model
NETL
NSTL
万方数据
维普
