Machine Learning-Based Ultrasound Radiomics Signature: A Retrospective Analysis for Differentiating Benign from Malignant Non-Mass Lesions in Dense Breasts.
Authors
Affiliations (1)
Affiliations (1)
- Breast Cancer Center, Hubei Cancer Hospital, Tongji Medical College, Huazhong University of Science and Technology, National Key Clinical Specialty Discipline Construction Program, Hubei Provincial Clinical Research Center for Breast Cancer, Wuhan Clinical Research Center for Breast Cancer, Wuhan, Hubei, 430079, People's Republic of China.
Abstract
The purpose of this study is to establish and validate a machine learning based ultrasound radiomics feature that combines clinical and ultrasound features, which can be used to identify benign and malignant non mass lesions (NML) of dense breast and evaluate its diagnostic value. This study is a retrospective single center study. We included 619 patients with dense breast NML diagnosed by ultrasound from January 2017 to January 2023. The patients were randomly divided into training group (n=434) and validation group (n=185) according to the ratio of 7:3. About 848 radiomics features were extracted from two-dimensional ultrasound images and screened by lasso regression. The clinical model, ultrasound model, radiomics model and combined model were established. The diagnostic performance was evaluated by ROC curve, correction curve and decision curve analysis (DCA), and the model differences were compared by Delong test. There were 304 cases of malignant lesions and 315 cases of benign lesions confirmed by postoperative pathology. Multivariate logistic regression analysis showed that age, lesion length, microcalcification, surrounding structure distortion and blood flow were independent predictors of malignancy. Twelve non-zero coefficient radiomics features were selected to construct the radiomics features. The AUC of the combined model was the highest, the training set was 0.89 (95% CI: 0.86-0.92), and the validation set was 0.83 (95% CI: 0.78-0.89). Delong test showed that there were significant differences between the combined model and the other three single models (all p<0.05). The calibration curve showed good consistency in predicting the actual pathology, and DCA verified its best clinical application value. We successfully combine ultrasound imaging, clinical and ultrasound features to build a prediction model, which has a good diagnostic effect on the differentiation of benign and malignant breast dense non-small cell lymphoma, and provides a reliable basis for clinical treatment decision-making of breast cancer.