Benign and malignant classification of mammogram images based on deep learning. (May 2019)

• Record Type: Journal Article
• Authors: Li, Hua; Zhuang, Shasha; Li, Deng-ao; Zhao, Jumin; Ma, Yanyun
• Is Part Of: Biomedical signal processing and control. Volume 51(2019)
• Page Start: 347
• Page End: 354
• Publication Date: 2019-05
• Subjects: Breast cancer -- Mammogram images -- Deep learning -- Inception structure -- DenseNet-II neural network model; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.02.017
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: This paper studies and improves a deeper and more complex DenseNet neural network model, inventing a new DenseNet-II neural network model, which can maintain the sparseness of the network structure. It also prevents the model from overfitting and improves computer performance. Through image preprocessing, the quality and quantity of images can be improved, making the recognition rate more accurate. Two processing techniques are used for the mammogram images: zero-mean normalization and data enhancement, which can improve the speed and accuracy of mammogram images classification. mage normalization avoids interference from light, while the adoption of data enhancement prevents over-fitting cause by small data set. The 10-fold cross-validation is used to verify the classification results of five network models. It does not have any assumptions, it is an effective model selection method, which has the universality of application and easy operation. The method proposed in this paper not only improves the benign and malignant classification performance of mammography images, but also provides doctors with more objective and accurate diagnosis results, which has important clinical application value and research significance. Abstract: Breast cancer is one of the most common malignant tumors in women, which seriously affect women's physical and mental health and even threat to life. At present, mammography is an important criterion for doctors to diagnose breast cancer. However, due to the complex structure of mammogram images, it is relatively difficult for doctors to identify breast cancer features. At present, deep learning is the most mainstream image classification algorithm. Therefore, this study proposes an improved DenseNet neural network model, also known as the DenseNet-II neural network model, for the effective and accurate classification of benign and malignant mammography images. Firstly, the mammogram images are preprocessed. Image normalization avoids interference from light, while the adoption of data enhancement prevents over-fitting cause by small data set. Secondly, the DenseNet neural network model is improved, and a new DenseNet-II neural network model is invented, which is to replace the first convolutional layer of the DenseNet neural network model with the Inception structure. Finally, the pre-processed mammogram datasets are input into AlexNet, VGGNet, GoogLeNet, DenseNet network model and DenseNet-II neural network model, and the experimental results are analyzed and compared. According to the 10-fold cross validation method, the results show that the DenseNet-II neural network model has better classification performance than other network models. The average accuracy of the model reaches 94.55%, which improves the accuracy of the benign and malignant classification of mammogram images. At the same time, it also proves that the model has good generalization and robustness.