基于改进DenseNet的乳腺钼靶肿块分类方法

Mass Classification of Breast Mammogram Based on Improved DenseNet

乳腺X线摄影技术是目前乳腺癌早期发现和诊断的重要手段。然而乳腺X线图像中肿块边缘模糊,分类相对困难,因此提升乳腺肿块的诊断精度从而及早预防和治疗仍是医学领域的一大挑战。针对乳腺肿块的特点,提出了一种结合密集卷积神经网络(DenseNet)和压缩激励(SE)模块的新网络(DSAMNet),该网络融合了二者优势,既加强特征重用,又实现特征提取过程中的特征重标定。根据SE模块嵌入DenseNet的不同位置,提出了模型SE-DenseNet-A、SE-DenseNet-B和SE-DenseNet-C。对SE-DenseNet的池化函数进行改进,提出了模型DSAMNet-A、DSAMNet-B和DSAMNet-C。综合不同结构和不同深度的网络模型在公开数据集CBIS-DDSM上进行训练和测试。实验结果表明,DSAMNet-B有更加优异的性能,其准确率比DenseNet模型的准确率提高了10.8%,AUC达到了0.929。

Mammography technology is currently an important method for early detection and diagnosis of breast cancer.However,it is relatively difficult to classify tumor masses due to the blurred edges of mammograms.Therefore,it is still a major challenge in the medical field to improve the diagnostic accuracy of breast masses so as to prevent and treat them as early as possible.According to the characteristics of breast masses,a new network(DSAMNet)combining dense convolu-tional neural network(DenseNet)and Squeeze-and-Excitation(SE)modules is proposed.This network combines the advantages of the two,which not only strengthens feature reuse,but also realizes feature recalibration in the feature extraction process.In this paper,the models SE-DenseNet-A,SE-DenseNet-B,and SE-DenseNet-C are proposed in consideration of the different positions where the SE module is embedded in DenseNet.The pooling function of SE-DenseNet is improved,and models DSAMNet-A,DSAMNet-B and DSAMNet-C are proposed.Network models with different structures and different depths are trained and tested on the public dataset CBIS-DDSM.The experimental results show that DSAMNet-B has more excellent performance,whose accuracy rate is 10.8%higher than that of the DenseNet model,and the area under the curve(AUC)reaches 0.929.