Predictive model of endocrine drug resistance in hormone receptor-positive breast cancer based on ultrasound radiomics
10.3760/cma.j.cn131148-20240603-00317
- VernacularTitle:基于超声影像组学构建激素受体阳性乳腺癌内分泌耐药预测模型
- Author:
Xiaoxue LIU
1
;
Lei ZHANG
;
Xudong ZHANG
;
Wei FAN
;
Qingxiang LI
;
Xinran FANG
;
Zihao QIN
;
Junjia WANG
;
Jiawei TIAN
;
Hao CUI
Author Information
1. 哈尔滨医科大学第二附属医院超声医学科，哈尔滨 150086
- Keywords:
Ultrasound examination;
Breast cancer;
Hormone receptor;
Radiomics;
Endocrine resistance
- From:
Chinese Journal of Ultrasonography
2024;33(11):1000-1009
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To establish an ultrasound radiomics model by integrating clinical, pathological, and conventional ultrasound features with radiomics characteristics, and to explore its clinical value in predicting endocrine resistance in hormone receptor(HR)-positive breast cancer.Methods:A retrospective analysis was performed on 478 patients with HR-positive breast cancer from January 2017 to December 2021 in the Second Affiliated Hospital of Harbin Medical University, of which 430 were resistant and 48 were sensitive. The clinical, pathological and immunohistochemical data and ultrasound images were collected.Firstly, the propensity score was used to process and match the data. Secondly, Logistic regression was used to screen clinical, pathological, and conventional ultrasound features associated with endocrine resistance. Then, PyRadiomics was used to extract the radiomic features of grayscale ultrasound images, and a series of methods such as Lasso regression were used to screen the radiomic features related to endocrine resistance. Seven machine learning methods such as random forest were used to build a radiomics model. Finally, clinical, pathological and ultrasound features were added to establish a clinical pathological model, a clinical pathological ultrasound model, a clinical pathological radiomics model and a combined model of the four features, and the model effectiveness was evaluated.Results:①Propensity score matching: 96 patients were matched, including 48 patients in the drug-resistant group and 48 patients in the sensitive group. ②Screening clinical pathological conventional ultrasound features related to endocrine resistance: lymph node metastasis, tumor diameter, posterior echo attenuation, and growth orientation were independent predictors of endocrine resistance (all P<0.05). ③Screening radiomics features related to endocrine resistance: 18 features such as Dependence Entropy. ④Establishing radiomics model: the machine learning model of random forest method (AUC=0.80) performed best. ⑤Radiomics model integrating clinical, pathological and conventional ultrasound features: the AUC of the clinical pathological model was 0.70, the AUC of the clinical pathological ultrasound model was 0.78, the AUC of the clinical pathological radiomics model was 0.82, and the AUC of the combined model was 0.86. Conclusions:The radiomics model established by the random forest method performs best in predicting endocrine resistance in HR-positive breast cancer. The model that integrates multiple features performs best in assessing endocrine resistance.which is expected to provide an objective basis for clinicians to predict endocrine resistance in HR-positive breast cancer.
