Developing a weakly supervised deep learning framework for breast cancer diagnosis with HR status based on mammography images

Abstract

The status of hormone receptors (HR) at the molecular level is crucial for accurate diagnosis and effective treatment of breast cancer. Meanwhile, mammography is an effective screening method for detecting breast cancer, which significantly improve survival. However, diagnosing the molecular status of breast cancer involves a pathological biopsy, which can affect the accuracy of the diagnosis. To non-invasively diagnose the hormone receptor (HR) status of breast cancer and reduced manual annotation, we proposed a weakly supervised deep learning framework BSNet which detected breast cancer with HR status and benign tumors. BSNet was trained on 2321 multi-view mammography cases from female undergoing digital mammography for the general population at Harbin Medical University Cancer Hospital in Heilongjiang Province during the period 2017-2018 and was validated on the external cohort. The average AUCs of BSNet on the test set and the external validation set were 0.89 and 0.92, respectively. BSNet demonstrated excellent performance in non-invasive breast cancer diagnosis with HR status, using multiple mammography views without pixel annotation. Furthermore, we developed a web server (http://bsnet.edbc.org) for easy use. BSNet described high-dimensional mammography of breast cancer subtypes, which helped inform early management options.

Keywords: Breast cancer diagnosis; Deep learning; HR status; Mammography images; Weakly supervision.

Graphical abstract

Fig. 1

Mammography image data collection (A) The number of HR+ /HR- breast cancer and Benign cases. (B) The data distribution for training, validating, and testing.

Fig. 2

The baselines model architectures. (A) The modified architecture of AlexNet. (B) The modified architecture of Vgg16. (C) The modified architecture of ResNet18. (D) The modified architecture of ResNet34. (E) The modified architecture of ResNet50.

Fig. 3

BSNet architecture.

Fig. 4

The performance comparison with feature integration methods of BSNet model. (A) The performance of global average pooling on the test set and the external validation set. (B) The performance of max pooling on the test set and the external validation set. (C) The performance of mean pooling on the test set and the external validation set. (D) Global average pooling of BSNet model generated feature heatmap. The upper right is the characteristic heatmap of the benign, the middle right is HR+ feature heatmap. the Bottom right is the HR- feature heatmap. class 0 represents Benign, class1 represents HR+ , class 2 represents HR-.

Fig. 5

The content of BSNet webserver.