Hu Ningxuan, Gao Zhizhen, Xie Zongyu, Li Lei
School of Medical Imaging, Bengbu Medical University, Anhui, China.
Department of Radiology, The First Affiliated Hospital of Bengbu Medical University, Anhui, China.
Front Oncol. 2025 Jul 17;15:1600057. doi: 10.3389/fonc.2025.1600057. eCollection 2025.
Worldwide, breast cancer ranks among the cancers with the highest incidence rate. Early diagnosis is crucial to improve the survival rate of patients. Digital Mammography (DM) is widely used for breast cancer diagnosis. The disadvantage is that DM relies too much on the doctor's experience, which can easily lead to missed diagnosis and misdiagnosis. In order to address the shortcomings of traditional methods, a CLGB-Net deep learning model integrating local and global information is proposed for the early screening of breast cancer. Four network architectures are integrated into the CLGB-Net model: ResNet-50, Swin Transformer, Feature Pyramid Network (FPN), and Class Activation Mapping (CAM). ResNet-50 is used to extract local features. The Swin Transformer is utilized to capture global contextual information and extract global features. FPN achieves efficient fusion of multi-scale features. CAM generates a class activation weight matrix to weight the feature map, thereby enhancing the sensitivity and classification performance of the model to key regions. In breast cancer early screening, the CLGB-Net demonstrates the following performance metrics: a precision of 0.900, recall of 0.935, F1-score of 0.900, and final accuracy of 0.904. Experimental data from 3,552 samples, including normal, benign, and malignant cases, support these results. The precision of this model was improved by 0.182, 0.038, 0.023, and 0.021 compared to ResNet-50, ResNet-101, Vit Transformer, and Swin Transformer, respectively. The CLGB-Net model is capable of capturing both local and global information, particularly in terms of sensitivity to subtle details. It significantly improves the accuracy and robustness of identifying lesions in mammography images and reduces the risk of missed diagnosis and misdiagnosis.
在全球范围内，乳腺癌是发病率最高的癌症之一。早期诊断对于提高患者生存率至关重要。数字乳腺摄影（DM）被广泛用于乳腺癌诊断。其缺点是DM过于依赖医生的经验，容易导致漏诊和误诊。为了解决传统方法的不足，提出了一种集成局部和全局信息的CLGB-Net深度学习模型用于乳腺癌的早期筛查。CLGB-Net模型集成了四种网络架构：ResNet-50、Swin Transformer、特征金字塔网络（FPN）和类激活映射（CAM）。ResNet-50用于提取局部特征。Swin Transformer用于捕获全局上下文信息并提取全局特征。FPN实现多尺度特征的高效融合。CAM生成类激活权重矩阵对特征图进行加权，从而提高模型对关键区域的敏感性和分类性能。在乳腺癌早期筛查中，CLGB-Net表现出以下性能指标：精确率为0.900，召回率为0.935，F1分数为0.900，最终准确率为0.904。来自3552个样本（包括正常、良性和恶性病例）的实验数据支持了这些结果。与ResNet-50、ResNet-101、Vit Transformer和Swin Transformer相比，该模型的精确率分别提高了0.182、0.038、0.023和0.021。CLGB-Net模型能够捕获局部和全局信息，特别是对细微细节的敏感性。它显著提高了乳腺钼靶图像中病变识别的准确性和鲁棒性，降低了漏诊和误诊的风险。
