Abstract: Because the color of X-ray breast mass image is more monotonous than natural images, the edge of the lump in the breast is blurry, and the textures of breast benign and malignant tumors are similar. All these results in the difficulty of classification of X-ray breast mass image. Based on the characteristics of mammography, this paper proposes a method for the classification of X-ray breast mass image based on Convolutional Neural Network（CNN）. The paper has the following contributions. （1）A scheme to extract feature maps of multiple views of breast image is proposed, and the different sizes of convolution kernels are used to extract different granularity in the model, in order to obtain abundant breast image features. （2）The discriminative method is embedded into the optimized model, via designing the new objective function based on the nuance between breast benign and malignant tumors, and the classification error is therefore amplified to enlarge penalties, further the model pattern can evolved to the smallest error direction. The proposed method is trained on the open data set. Cross validation, indicates that the proposed method can achieve AUC of 0.712 9, which is superior to the state-of-the-art classification algorithm of X-ray breast mass image, and the robustness of the proposed method is stronger.

Key words: breast imaging, breast tumors, deep learning, medical image classification

摘要： X射线乳腺影像与自然图像相比，色彩较为单调，且乳腺肿块边缘模糊，良性肿块与恶性肿块纹理相似，区分度较小。基于卷积深度学习网络提出一种适用于X射线乳腺肿块影像分类的方法，主要贡献如下：（1）提出一种提取乳腺影像多个卷积粒度的特征图的方案，分别使用不同尺寸的卷积核来提取不同粒度的卷积特征图，获得更为丰富的乳腺影像特征；（2）将判别方法嵌入到优化模型中，即设计新的目标函数，对分类误差进行差异化放大，从而加大分类错误的惩罚力度，指导模型向着分类错误最小的方向演进。在公开的乳腺X射线影像数据集上进行训练，通过交叉验证，AUC达到0.712?9，优于最好的乳腺影像分类方法，具有较强的鲁棒性。

关键词: 乳腺影像, 乳腺肿块, 深度学习, 医学图像分类