Machine learning models based on ultrasound radiomics for preoperatively distinguishing atypical parathyroid tumors/parathyroid carcinoma and parathyroid adenoma
10.13929/j.issn.1003-3289.2025.06.013
- VernacularTitle:超声组学机器学习模型术前鉴别非典型甲状旁腺肿瘤/甲状旁腺癌与甲状旁腺腺瘤
- Author:
Chunrui LIU
1
;
Peng WAN
;
Haiyan XUE
;
Yidan ZHANG
;
Wenxian LI
;
Jian HE
;
Zhengyang ZHOU
;
Jing YAO
Author Information
1. 南京大学医学院附属鼓楼医院超声医学科,江苏南京 210008
- Publication Type:Journal Article
- Keywords:
parathyroid neoplasms;
ultrasonography;
artificial intelligence;
diagnosis,differential
- From:
Chinese Journal of Medical Imaging Technology
2025;41(6):908-913
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of machine learning(ML)models based on ultrasound radiomics for preoperatively distinguishing atypical parathyroid tumor(APT)/parathyroid carcinoma(PC)and parathyroid adenoma(PA).Methods Totally 330 primary hyperparathyroidism patients who underwent surgical treatments were retrospectively enrolled and categorized into APT/PC group(n=78)and PA group(n=252)according to surgical pathology and clinical follow-up results,also divided into training set(n=231)and test set(n=99)at the ratio of 7∶3.Based on preoperative ultrasound,545 radiomics features were extracted,and recursive feature elimination(RFE),Kruskal-Wallis or analysis of variance methods were used to screen the features,respectively.Support vector machine(SVM),linear discriminant analysis(LDA),least absolute shrinkage and selection operator logistic regression(LRLASSO),also random forest(RF)and decision tree(DT)algorithms were adopted to construct ML models for differentiating APT/PC and PA,respectively.Then the models were trained in training set,their performance were verified in test set,and a 5-fold cross-validation was adopted to screen out the better combinations.Results Compared with Kruskal-Wallis and analysis of variance methods,the distinguishing efficacy of SVM,LDA,LRLASSO,RF and DT models constructed based on features screened out using RFE method in training set(area under the curve[AUC]=0.870,0.878,0.850,0.847,1.000)and test set(AUC=0.856,0.842,0.827,0.847 and 0.704)were all relatively higher.In test set,the AUC of SVM,LDA,LRLASSO and RF models constructed based on the features screened out using RFE method(included 25,23,17 and 23 features)were all higher than that of DT model(8 features)(all P＜0.001).No significant difference of AUC was found between SVM,LRLASSO or RF models and LDA model(all P＞0.05).The AUC of SVM and RF models were higher than that of LRLASSO model(both P＜0.05),while of SVM and RF models were not significantly different(P＞0.05),indicating that SVM,LDA and RF models were better ones.Conclusion SVM,LDA,LRLASSO,RF and DT models based on ultrasound radiomics could effectively distinguish APT/PC and PA preoperatively,among which SVM,LDA and RF models had better diagnostic efficacy.
