Ultrasound radiomics based on convolutional neural network for predicting effect of neoadjuvant chemotherapy for breast cancer
10.13929/j.issn.1003-3289.2025.03.016
- VernacularTitle:基于卷积神经网络以超声影像组学预测新辅助化疗用于乳腺癌效果
- Author:
Yue YANG
1
;
Xinyan LI
1
;
Wenxin ZHANG
1
;
Fang SUN
1
Author Information
1. 滨州医学院附属医院超声医学科,山东滨州 256603
- Publication Type:Journal Article
- Keywords:
breast neoplasms;
neoadjuvant therapy;
deep learning;
ultrasonography
- From:
Chinese Journal of Medical Imaging Technology
2025;41(3):424-428
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of ultrasound radiomics based on convolutional neural network(CNN)for predicting effect of neoadjuvant chemotherapy(NAC)for breast cancer.Methods Totally 164 women with breast cancer were retrospectively enrolled and divided into effective group(n=68)and ineffective group(n=96)according to the treatment response,also randomly divided into training set(n=131)and validation set(n=33)at the ratio of 8∶2.Based on ultrasound before NAC,radiomics features of breast cancer were extracted and screened with CNN,radiomics models were constructed with logistic regression(LR),support vector machine(SVM),K-nearest neighbor(KNN),random forest(RF)and multilayer perceptron(MLP),respectively.The best radiomics model was selected,deep learning score(DL-Score)was calculated,and the nomogram was drawn combined with clinical features.Results Among 5 radiomics models,MLP model had the best comprehensive efficacy for predicting effect of NAC for breast cancer,and its sensitivity,specificity and area under the curve(AUC)in training set was 77.78％,92.21％and 0.929,respectively,which in validation set was 78.57％,84.21％and 0.921,respectively.The estrogen or progesterone receptor,human epidermal growth factor receptor 2 and DL-Score were all independent predictors of NAC effect for breast cancer(all P＜0.05).The sensitivity,specificity and AUC of nomogram drawn based on the above independent predictors was 83.30％,92.21％and 0.953 in training set,85.71％,94.74％and 0.955 in validation set,respectively.AUC of the nomogram was slightly higher than that of MLP model,but no significant difference was found(both P＞0.05).The integrated discrimination improvement index showed that adding clinical features(i.e.the above-mentioned immunohistochemically indicators)could improve the predictive performance of radiomics models(P＜0.001).Conclusion Ultrasound radiomics based on CNN could be used to predict effect of NAC for breast cancer.Combining with immunohistochemically indicators might improve their efficacy.
