Objective：

To explore the feasibility of automatic classification of breast magnetic resonance imaging (MRI) background parenchymal enhancement (BPE) based on deep learning and thresholding segmentation.

Methods:

Breast MRI dataof 547 breast cancer patients were r

etrospectively collected from January 2010 to October 2018 in Peking University First Hospital. According to the Breast Imaging Reporting and Data System (BI-RADS)

each image was classified into four categories by two senior radiologists

which was used as the reference standard. Three steps were performed to acquire the BPE category automatically. Firstly

the region of fibroglandular tissue (FGT) was segmented by using a U-Net 3D model. The Dice similarity coefficient (DSC) was used to evaluate the efficacy of the model. Secondly

the region of BPE was segmented by the thresholding technology. Thirdly

the ratio of the BPE volume to the FGT volume was calculated and the BPE category was obtained. The BPE category was automatically input into the structured report. The multi-class confusion matrix was used to analyze the performance of the calculated BPE

with the metrics of accuracy

F1 score

and Kappa.

Results:

The average DSC of the U-Net model was 0.902. The Macro accuracy

F1 Macro

F1 Micro

and Kappa was 0.95 (95% CI: 0.88-0.93)

0.84

0.90

and 0.81 (95% CI: 0.76-0.86)

respectively.

Conclusion:

The automated classification of breast MRI background parenchymal enhancement is feasible and has potential clinical application value.