Application Of Transfer Learning And Convolutional Neural Network In Mammography Image Classification

Breast cancer,as the second most common type of cancer in the world,is one of the causes of increased female mortality and has become the number one killer of women's health.In the clinical diagnosis of breast cancer,mammography has become one of the most advanced breast cancer examination methods due to its advantages such as relatively small pain,simple and repeatable operation,high resolution and high detection rate.Doctors diagnose patients by looking at mammogram,which is usually time-consuming and inefficient.In order to improve the diagnosis efficiency and reduce the risk of misdiagnosis,it is very important to develop a computer-aided diagnosis system based on artificial intelligence for mammogram.However,traditional classification methods need to use a large number of manual features,which are often quite complicated and have limitations.Deep learning can automatically learn features from data,which overcomes the shortcomings of traditional algorithms.But deep learning usually requires a large amount of image data,which requires extremely high cost.For the situation with a small amount of data,transfer learning is a good choice.In addition,most current classification methods of mammogram require based on the detection of image lesions,and the labeling of lesions requires doctors with extensive clinical experience.When the amount of data is large,this work becomes a huge challenge.The classification of whole image can find a breakthrough for this challenge.In view of this,based on a small amount of mammogram data,it is of great significance to use deep learning methods to establish a computer-aided diagnosis system that can realize the classification of whole image,which can not only assist doctors in diagnosis,but also be used in large-scale breast cancer screening.Based on the idea of transfer learning,this paper uses the convolutional neural network method in deep learning to carry out research from the following aspects:(1)First,based on the feature transfer method,a pre-trained convolutional neural network on large-scale natural images and a model pre-trained on relevant public data sets are used as feature extractors to extract the features of the mammogram.The common machine learning algorithms are compared and analyzed.Then based on the model fine-tuning method,the models are fine-tuned on mammogram using a convolutional neural network pre-trained on large-scale natural images and a model pre-trained on relevant public data sets.Through the model fine-tuning method,the accuracy of the model reaches 0.8280,the sensitivity reaches 0.7778,the specificity reaches 0.9000,and the AUC reaches 0.8742.The experimental results prove that the deep transfer learning method can obtain an effective classification model for whole image.(2)Based on the characteristics of mammogram with two views on each side of the breast,two dual-view strategies are used to classify.One is build a mammogram data set containing dual-view information.The other one is based on the idea of parameters sharing between Siamese network's sub-networks,a dual-view model is designed.Among them,the strategy of designing the two-view model was significantly improved in multiple evaluation indexes,with the accuracy exceeding 0.90,the AUC exceeding 0.96,and the sensitivity increasing to 0.8889.Experimental results show that the dual-view strategy can improve the classification performance of the model.(3)Based on the lightweight method in the lightweight convolutional neural network model,reducing network training parameters and computational complexity,and further reducing the model volume.Based on the model integration method,integrating multiple lightweight networks to get more effective classification model.Through the lightweight method,three lightweight models with good performance are obtained.The integration method of model fusion achieves the balance of sensitivity and specificity,and improves the comprehensive performance of the model.Experimental results show that lightweight strategies and model integration methods can help us get more efficient classification model.