基于眼底照片影像组学特征的糖尿病足分类模型构建

A diabetic foot classification model based on radiomics features of fundus photographs

目的开发和验证基于眼底照片影像组学特征的糖尿病足分类预测模型。方法回顾性收集2011年12月至2018年12月在南方医院诊断为2型糖尿病患者眼底照片共2035张[糖尿病足(DF)患者282张,糖尿病(DM)患者1753张],采用计算机随机数按7∶3将所有眼底照片随机分成训练集(1424张)与测试集(611张)。眼底照片进行图像预处理后,通过Radiomic工具包提取基于灰度矩阵的4128个纹理特征,并用ToolboxDESC工具包提取了11339个其他特征。采用LASSO算法选取出与DF最相关的30个特征,再使用Bootstrap+0.632自助采样法进一步筛选出7个最佳组合,进行logistic回归分析求得回归系数,建立最终的糖尿病足分类模型。绘制ROC曲线,计算训练集和测试集的AUC、敏感度、特异度及准确率以验证其预测效能。结果本研究筛选出7个糖尿病足患者眼底照片影像组学标记物,且构建的DF/DM分类预测模型在训练集中的AUC、敏感度、特异度、准确率分别为:0.9586、0.9840、0.9200、0.9280;在测试集中得到的AUC、敏感度、特异度、准确率分别为0.9271、0.9889、0.8810、0.8969。结论本研究采用了影像组学技术筛选出7个糖尿病足眼底影像组学标记物,基于此构建了高精度的可简便应用的DF/DM分类模型。该技术具有提高糖尿病足筛查效率的潜力。

Objective To construct a diabetic foot classification prediction model based on radiomics features of fundus photographs.Methods A total of 2035 fundus photographs of patients with type 2 diabetes diagnosed at Nanfang Hospital between December 2011 and December 2018 were retrospectively collected[282 photographs from patients with diabetic foot(DF),and 1753 from patients with diabetes mellitus(DM)].All fundus photographs were randomly divided into a training set(1424 photos)and a test set(611 photos)using a computer generated random number at 7∶3.After pre-processing the fundus photographs,a total of 4128 texture features based on the gray matrix were extracted by the Radiomic toolkit,and 11339 other features were extracted using the ToolboxDESC toolkit.The LASSO algorithm was used to select the 30 features most relevant to DF,and then the Bootstrap+0.632 self-sampling method was used to further select the 7 best combinations.Logistic regression analysis was used to obtain the regression coefficients and establish the final diabetic foot classification prediction model.ROC curve was drawn,and AUC,sensitivity,specificity,and accuracy of the training and test sets were calculated to verify its prediction performance.Results We screened 7 fundus radiomics markers for diabetic foot patients,and based on this established a DF/DM classification prediction model.The AUC,sensitivity,specificity,and accuracy of the model were 0.9586,0.9840,0.9200,and 0.9280 in the training set,and 0.9271,0.9889,0.8810,and 0.8969 in the test set,respectively.Conclusion In this study,seven DF fundus markers were screened using radiomics technology.Based on this,a highly accurate and easy-to-use DF/DM classification model was constructed.This technology has the potential to increase the efficiency of DF screening programs.