超声影像组学在肝包虫病分型中的应用研究

目的 构建基于超声影像组学的肝包虫病（hepatic echinococcosis, HE）分型的机器学习模型。方法 回顾性纳入2005年至2022年于四川省甘孜州石渠县及四川大学华西医院HE患者的超声声像图。本研究为回顾性、大样本、两中心的诊断准确性研究,采用超声影像组学方案构建HE分型的机器学习（machine learning, ML）模型。将图像按病灶类型以8: 2分层随机划分为训练集与独立测试集,通过10折交叉验证策略训练模型。ML模型构建流程包括,病灶分割（手工勾画并分割）、影像组学特征提取、特征预处理、基于8种分类器构建HE分型的ML模型,包括支持向量机、自编码器、线性判别分析、随机森林（random forest, RF）、逻辑回归、自适应增强、决策树、朴素贝叶斯。采用受试者工作特征（receiver operating characteristic, ROC）曲线分析评估ML模型的诊断性能。结果 共纳入4976例HE患者,其中男性2157例,女性2819例,年龄8-95（43.4±16.9）岁;囊型肝包虫病（cystic echinococcosis, CE）患者1641例,泡型肝包虫病（alveolar echinococcosis, AE）患者2981例,混合型患者354例。使用23452张超声图像进行模型训练、验证和测试,其中囊型肝包虫病图像8557张,泡型肝包虫病图像14895张。RF模型在交叉验证集和独立测试集中均表现最佳,交叉验证集及独立测试集下的敏感度、特异度、准确率、曲线下面积分别为0.71、0.76、0.73、0.82和0.62、0.89、0.76、0.86。结论 RF模型具有较高的准确率和健壮性,有益于提高基层工作者在流行区对肝包虫病亚型的超声诊断水平,减少误诊发生。

Application of Ultrasound Radiomics in the Classification of Hepatic Echinococcosis

Objective To construct a machine learning model for the classification of hepatic echinococcosis (HE) based on ultrasound radiomics.Methods A retrospective review was conducted on ultrasound images of HE patients from 2005 to 2022 in Shiqu County, Ganzi Tibetan Autonomous Prefecture, Sichuan Province, and West China Hospital of Sichuan University. This retrospective, large-scale, two-center study aimed at assessing diagnostic accuracy utilized a machine learning (ML) approach based on ultrasound radiomics for HE classification. Images were stratified into training and independent test sets in an 8: 2 ratio according to lesion type, and the model was trained using a 10-fold cross-validation strategy. The ML model construction process included lesion segmentation (manual delineation and segmentation), radiomics feature extraction, feature preprocessing, and the construction of the HE classification ML models using eight classifiers: support vector machine, auto-encoder, linear discriminant analysis, random forest (RF), logistic regression, adaboost, decision tree, and naive Bayes. The diagnostic performance of the ML model was evaluated using receiver operating characteristic (ROC) curve analysis.Results A total of 4,976 HE patients were included, comprising 2,157 males and 2,819 females, with ages ranging from 8 to 95 years (43.4 ± 16.9). Among them, 1,641 were cystic echinococcosis (CE) patients, 2,981 were alveolar echinococcosis (AE) patients, and 354 were mixed-type patients. A total of 23,452 ultrasound images were used for model training, validation, and testing, including 8,557 images of CE and 14,895 images of AE. The RF model demonstrated the best performance in both the cross-validation and independent test sets, with sensitivity, specificity, accuracy, and area under the curve of 0.71, 0.76, 0.73, 0.82, and 0.62, 0.89, 0.76, 0.86, respectively.Conclusion The RF model exhibits high accuracy and robustness, contributing to the improvement of ultrasound diagnostic capabilities for different subtypes of hepatic echinococcosis in endemic areas, thereby reducing the occurrence of misdiagnosis.