Objective:

To investigate the predictive value of radiomics based on digital breast tomosynthesis (DBT) in molecular subtypes of breast cancer.

Methods:

The data of 380 patients with inva

sive breast cancer confirmed by pathology after DBT examination in the Fudan University Shanghai Cancer Center from January 2019 to August 2020 were retrospectively analyzed. The DBT images of each patient included craniocaudal (CC) position and mediolateral oblique (MLO) position. And 380 whole- tumor features of lesions based on DBT were extracted

after performing dimensionality reduction and screening

the final retained features were put into three different machine learning models

including logistic regression (LR)

support vector machine (SVM) and random forest (RF)

respectively. Receiver operating characteristic (ROC) curve was used to evaluate the predictive efficacy of three models based on DBT images for the four molecular classification of breast cancer.

Results:

Of the 380 lesions confirmed by pathology

72 were Luminal A type

175 were Luminal B type

54 were human epidermal growth factor receptor 2 (HER2) over- expression type

and 79 were triple-negative breast cancer (TNBC). The three models all can predict the molecular subtypes of breast cancer efficiently. Among the three different models

RF model had a better performance

and in the test set

AUC values predicted by dichotomy were 0.82

0.71

0.70

and 0.71 for Luminal A type

Luminal B type

HER2 over-expression type

and TNBC respectively. Among DBT radiomics features

the entropy and entropy-related features

as well as the morphological features are related to the molecular subtypes of breast cancer.

Conclusion:

The radiomics model based on DBT imaging can predict the molecular subtypes of breast cancer

and the radiomics features that represent the heterogeneity and morphology are helpful for the differentiation of breast cancer molecular subtypes.