Breast Image Mass Detection Model Based on Multi-Scale Feature Reinforcement Network
Xianzhe Wang,Ying Li,Hongxu Guo,Renjie He,Jie Quan,Lingyue Wang,Lei Guo
DOI: https://doi.org/10.1016/j.bspc.2024.106655
IF: 5.1
2024-01-01
Biomedical Signal Processing and Control
Abstract:Analyzing X-ray images is a critical method for diagnosing breast cancer. The prompt and accurate detection of masses from extensive image data is vital in reducing breast cancer incidence and mortality rates. While deep learning-based approaches show promise in effectively detecting cancer, they still face challenges. Specifically, these approaches struggle to extract features from small objects and lack multi-scale feature fusion. Addressing these issues is crucial for more robust detection. In this paper, we propose the Multi-scale Reinforcement Feature Pyramid Network (MRFPN) to tackle these challenges. Firstly, we introduce the Feedback Connections Enhance Module (FCEM), which integrates additional feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. We also design the Symmetry Complement Module (SCM) as a feedback connection structure, leveraging the complementarity of convolution and pooling operations to significantly enhance the feature representation capability of the feature maps, particularly the edge information in images. Additionally, we propose the Multi-scale Fusion Strengthen Module (MFSM) to enhance the expressive power of each layer’s feature map by utilizing feature map information from multiple scales. Within the MFSM, we design the Receptive Expansion Module (REM), which employs dilated convolution to broaden the receptive field, thereby efficiently improving the network’s capability to detect objects of arbitrary sizes. Furthermore, by incorporating MRFPN into Faster RCNN, we design the Multi-scale Reinforcement Breast Mass Detection Network (MRBMDNet). Extensive experiments demonstrate that MRBMDNet outperforms the baseline on datasets of different qualities. Our method achieved 67.84％ (in AP@50) and 96％ (in TPR@20) on the commonly used INbreast dataset, 67.60％ (in AP@50) and 95％ (in TPR@20) on the DDSM dataset, and 87.5％ and 63.2％ (in TPR@20) on the publicly available BCS-DBT and MIAS datasets, respectively. In conclusion, our proposed FCEM and MFSM are effective and essential for achieving the outstanding performance of MRFPN in small object detection. The SCM module within the FCEM and the REM module within the MFSM are highly effective and play a crucial role in enhancing the overall detection capabilities of the network. Code is available at https://github.com/xianzhe-wang/MRFPN.