Objective:

To explore the association between magnetic resonance imaging (MRI) features and the 21-gene recurrence score (RS)

and to establish RS prediction models.

Methods:

Clinical and imaging data of estrogen receptor (ER)

+

/ human epidermal growth factor receptor 2 (HER2)

-

breast cancer patients who underwent 21-gene expression assay in Fudan University Shanghai Cancer Center from April 2017 to March 2019 were collected. The patients who underwent preoperative breast MRI were selected. MRI images were evaluated according to the 2013 version of Breast Imaging Reporting and Data System lexicon. The univariate analyses were used to compare differences in MRI imaging features between the high-risk group (RS≥26) and the low-risk group (RS＜26) and multivariate logistic regression was used to construct a model. The patients were divided into the training group and the validation group in a 7∶3 ratio. Pearson correlation coefficient screening method and recursive feature elimination method were used for feature screening

the synthetic minority oversampling technique was used for balancing the training dataset

four different machine learning algorithms (linear support vector machine

random forest

decision tree and K-nearest neighbor) were used to construct the models

and the model performance was evaluated by receiver operating characteristic (ROC) curve.

Results:

A total of 159 patients were enrolled

with 58 in the low-risk group and 101 in the high-risk group. In clinical characteristics

progesterone receptor (PR) status showed difference (

P

=0.017)

and the proportion of patients with positive PR expression was higher in the low- risk group than in the high-risk group. In the MRI characteristics

the distribution of tumor margins showed difference between groups

(

P

=0.008). The tumors in the low-risk group were mostly characterized by spiculated margins (64.8%)

and the tumors in the high-risk group were mostly characterized by irregular margins (54.7%). Incorporating PR status and tumor margin into the multivariate logistic regression model

PR positive patients had a lower risk of recurrence than PR negative patients

with an OR value of 0.110 (

P

=0.038); the recurrence risk of spiculated margins was relatively lower than that of irregular margins

with an OR value of 0.343 (

P

=0.004). The area under curve (AUC) was 0.667. The calibration curve and decision curve indicated that the model had good calibration performance and certain clinical practicability. 111 patients were included in the training group (34 in low- risk group and 77 in high-risk group) and 48 patients were included in the validation group (24 in low-risk group and 24 in high-risk group). The AUC range of the four machine learning models was 0.64-0.69

and the AUCs of support vector machine and random forest models were relatively higher.

Conclusion:

Breast MRI features have a potential role in assessing the recurrence risk of ER

+

/ HER

-

breast cancer patients.