Medical Image Classification Based On Inception Module

In medical field,medical images play a very important role in diagnosing diseases and treating patients,especially in computer-aided diagnosis.Using computers to distinguish healthy and lesion areas in images automatically can give the doctor clearer and more direct information,and the doctor can make more objective and accurate diagnosis.Invasive Ductal Carcinoma(IDC)is the most common subtype of all breast malignancies.To detect IDC and assign an aggressiveness grade of IDC are time-consuming and challenging task,because pathologists need to scan large of benign regions to find the areas of malignancy.Therefore,using computer to diagnose IDC and assign an aggressiveness grade of Invasive Ductal Carcinoma has practical significance.Image classification is the quantitative analysis of the image by computers.The computer distinguishes different targets according to the characteristics of the images,such as pixel colors and texture features.Computers classify the image into one of several classes.However,the complex features of medical images are difficult to extract manually,the preprocessing is cumbersome and complex.The accuracy of classification results is not enough,there is room for improvement.In recent years,there has been arisen a number of technologies which combined the theories of computer vision and deep learning.And convolution neural network has been successful in many areas.Convolution neural network is able to extract the image features automatically,so that it can be suitable for image processing.And using convolution neural network in image classification has become an important research direction.This paper discusses the application of convolution neural network methods in image classification field.For Invasive Ductal Carcinoma in whole slide images,convolution neural network had been proposed for classifying small IDC images.Convolution neural network can automatically extract the features of image.This paper builds a network based on the most used CNN,and changes the images and tags to adapt to tensorflow.So the images and tags can train and test for the network.Then improving the network structure based on the inception module.Inception module can increase the number of layers to improve the effect of fitting.Combine inception module and global average pooling layer in network can decrease the number of parameters.And the network can be transferred and retrained by transfer learning.Transfer learning improves initialization of neural network to raise prediction accuracy.Finally,the advantages of the method used in this article can be demonstrated through comparing the method used in this paper with several traditional algorithms.