Multi-Scale Attention-Guided Network for mammograms classification. (July 2021)

• Record Type: Journal Article
• Authors: Xu, Chunbo; Lou, Meng; Qi, Yunliang; Wang, Yiming; Pi, Jiande; Ma, Yide
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Mammography -- Classification -- Abnormality -- Multi-Scale -- Self-Adaption; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102730
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A Multi-Scale Attention-Guided Network for recognizing abnormalities in mammograms. The self-adaption and multi-scale contribute to the final classification results. Multiple receptive fields are beneficial for recognizing objects of different sizes. Abstract: For the breast mass segmentation in whole mammograms, in our studies, we observe that there is an enormous performance reduction in the case of considering the normal data during training. Therefore, the mammogram classification (normal vs. abnormal) is essential for boosting the breast mass segmentation performance in whole mammograms and is our research topic in this paper. Due to the breast lesions with a variety of sizes, the mammogram classification (normal vs. abnormal) is a challenging task. To improve the mammogram classification performance, we propose an end-to-end convolutional neural network, namely Multi-Scale Attention-Guided Network (MSANet). Specifically, MSANet can be constructed by stacking several Multi-Scale Attention (MSA) bottlenecks. Each MAS bottleneck consists of a Scale Aggregation (SA) unit and a Multi-Scale Attention Module (MSAM). The SA unit is used to generate multiple feature maps of different scales, and the MSAM is used to allocate the suitable size of receptive field for objects of different sizes. According to the extensive experiments, our proposed MSANet-50 achieves a fully automated classification AUC of 0.942 on the DDSM database, which outperforms several approaches.