Affiliations: [a] School of Information Science and Technology, Fudan University, Shanghai, China | [b] Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China | [c] School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China
Correspondence:
[*]
Corresponding author: Yi Guo, School of Information Science and Technology, Fudan University, Shanghai, China. E-mail: guoyi@fudan.edu.cn.
Abstract: BACKGROUND: High-precision detection for individual and clustered microcalcifications in mammograms is important for the early diagnosis of breast cancer. Large-scale differences between the two types and low-contrast images are major difficulties faced by radiologists when performing diagnoses. OBJECTIVE: Deep learning-based methods can provide end-to-end solutions for efficient detection. However, multicenter data bias, the low resolution of network inputs, and scale differences between microcalcifications lead to low detection rates. Aiming to overcome the aforementioned limitations, we propose a pyramid feature network for microcalcification detection in mammograms, MicroDMa, with adaptive image adjustment and shortcut connections. METHODS: First, mammograms from multiple centers are represented as histograms and cropped by adaptive image adjustment, which mitigates the impact of dataset bias. Second, the proposed shortcut connection pyramid network ensures that the feature map contains more information for multiscale objects, while a shortcut path that jumps over layers enhances the efficiency of feature propagation from bottom to top. Third, the weights of each feature map at different scales in the fusion are trainable; thus, the network can automatically learn the contributions of all feature maps in the fusion stage. RESULT: Experiments were conducted on our in-house dataset and the public dataset INbreast. When the average number of positives per image is one on the in-house dataset, the recall rates of MicroDMa are the 96.8% for individual microcalcification and 98.9% for clustered microcalcification, which are higher than 69.1% and 91.2% achieved by recent deep learning model. Free-response receiver operating characteristic curve of MicroDMa is also higher than other methods when models are performed on INbreast. CONCLUSION: MicroDMa network is better than other methods and it can effectively help radiologists detect and identify two types of microcalcifications in clinical applications.
Keywords: Deep learning, detection, pyramid network, mammography, breast microcalcification, convolutional neural network
