Breast Disease Image Classification Based On Improved Convolutional Neural Network And Multi-scale Feature Fusion

In women,breast cancer,as the highest incidence of cancer,seriously affects women’s health.However,the cure rate is very high if detected early by medical means.So early screening and treatment is an important means to improve the survival rate of breast cancer patients.Mammography is very suitable for early breast examination because of its minimal harm to the patient’s body,convenient operation and cheap price.Therefore,mammography is regarded as the main medical image in academic research and clinical practice.The effect of convolutional neural network in image classification is greatly improved compared with other classification algorithms,which provides a new idea for breast X-ray image processing.Aiming at the disadvantages of time-consuming and low accuracy of computer aided diagnosis（CAD）for breast tumor,this paper proposed a method based on improved convolutional neural network and feature fusion to classify breast molybdenum target images.The contributions made in this paper are as follows:（1）An improved Alexnet image classification model for breast diseases based on attention mechanism was proposed.Considering that the proportion of breast lesions in mammogram images is very small,the network structure of traditional Alexnet model is improved in this paper.The large convolution kernels at the lower level were changed to several small 3*3 convolution kernels to increase the nonlinear capability of the model and reduce the training parameters.Meanwhile,the inter channel attention mechanism is applied to the network model.By weighting each feature channel,the key information with higher relevance to the current task is gathered in the input data,and other unimportant information is reduced,so that the model can make more accurate judgment.The experimental results show that the accuracy of breast tumor classification based on attention mechanism proposed in this paper is improved.（2）A breast disease image classification model based on improved residual network was proposed.For Res Net model,the residual block structure is widened.Compared with the original residual block,the convolution kernel of the broadened residual block is one-dimensional,and one more residual channel is added.We replace the original 3×3 convolution kernel with 1×3 convolution kernel in the residual block,and the parameter can be reduced to 1/3 of the original.The number of parameters of a broadened one-dimensional residual block is only 2/3 of that of an original residual block.Therefore,the residual block can effectively reduce the number of network parameters and improve the efficiency of network training.In addition,the broadened network structure can make each layer of the network obtain more features,so as to improve the classification accuracy.Experimental results show that this method can improve the accuracy of classification.（3）A multi-scale feature fusion model for breast disease image classification was proposed.The two improved convolutional neural networks were used to extract different features from breast molybdenum target images of different scales respectively,and then transfer learning and fusion features were used for classification,therefore,the generalization and feature expression ability of the model are improved.The experimental results are as follows: The accuracy of the improved Resnet network on the test set is 96.17%,and that of the improved Alexnet on the test set is 96.28%.Compared with single-channel network,the accuracy of multi-scale feature fusion model optimized by genetic algorithm reaches 98.12%,while only 291 parameters need to be trained for transfer learning.Experimental results show that the network model designed in this paper shows high accuracy and less parameter redundancy.