Deep Learning Based DCE-MRI Method Of Breast Fibroglandular Tissue Segmentation

Breast cancer has surpassed lung cancer as the most common cancer.Studies have shown that the mortality rate of breast cancer can be significantly reduced through high-quality prevention,early detection,and treatment services.Early screening of breast cancer based on DCE-MRI is an effective method,and the Background Parenchymal Enhancement(BPE)is an important indicator for early screening of breast cancer.The quantitative calculation of BPE relies on the accurate segmentation of fibroglandular tissue(FGT)in breast MRI images.However,due to the uneven grayscale of the FGT region,fuzzy boundaries,varying sizes,and the presence of dual targets,as well as the class imbalance problem associated with small-sized FGT,accurate segmentation of breast FGT in DCE-MRI poses significant challenges.Segmentation of the entire breast region(breast mask)can reduce the computational workload of subsequent segmentation of breast fibroglandular tissue and tumors,while improving segmentation accuracy.To address these issues,this study proposes a two-stage deep convolutional neural network for breast FGT segmentation task,with a multi-scale feature attention network used for the first stage segmentation of DCE-MRI breast ROI(region of interest,referring to the breast mask in this paper),and a pyramid multi-scale attention network used for the second stage segmentation of DCE-MRI breast FGT area based on the breast ROI segmentation network.First,the breast DCE-MRI three-dimensional data is preprocessed by enhancement,denoising,standardization,and normalization,and then a multi-scale feature attention network is designed.The network uses four parallel feature extraction subnetworks as the basic framework to improve its edge recognition capability.The network model is lightweight using depth-wise separable convolution,and multiple scale attention fusion modules and Non-local modules are introduced to enhance the network’s multi-scale feature global capture and fusion capabilities,improving the robustness of breast ROI segmentation.Secondly,a pyramid multi-scale attention network is proposed,which uses dense connected depth-wise separable convolution modules as the basic building block to enhance the recognition capability of dual-target objects.A pyramid feature fusion module is used to alleviate the problem of under-segmentation of the target area,and the loss function is improved to improve the optimization direction of the network model,achieving accurate segmentation of breast FGT.The effectiveness of the network model and related modules are verified through comparative experiments,and the experimental results are analyzed.Comparative experimental results on the breast dataset of the affiliated Ruijin Hospital of Shanghai Jiaotong University School of Medicine show that the method proposed in this paper achieves a Dice index of 94.32% in the DCE-MRI breast ROI segmentation task and a Dice index of 85.74% in the DCE-MRI breast FGT segmentation task,which are superior to other methods.The proposed modules and structures have effectively improved the network segmentation performance,and have certain clinical application value and reference significance.