Research On The Classification Method Of Skin Tumors Based On Histopathological Images

In the current clinical diagnosis of skin tumor medicine,pathological diagnosis is the absolute standard for doctors.But in the early stage of tumor,the diagnosis is difficult and the number of patients is increasing much faster than the training of professional pathologists.So how to improve the efficiency of doctors while enhancing the objectivity of diagnosis is a problem to be solved.Therefore,in this thesis,four types of histopathological images of skin tumors are used as the research object to explore the multi-classification problem of histopathological images in depth,and the main research work includes the following aspects.In the pre-processing stage,to address the problem of how to feed the pathological images into the convolutional neural network without losing the spatial information and reducing the detail information,the Grab Cut algorithm is firstly adopted to quickly extract the region of interest from the pathological images to reduce the size of the images;then the region of interest is extracted for patch processing,so as to crop it into the size suitable for the model input;finally,since the annotation of pathological images takes a lot of time and effort of professional doctors and the number of different kinds of skin tumors is extremely unbalanced,a combination of traditional image processing techniques and deep learning is adopted to perform data enhancement for how to overcome the problem of insufficient and unbalanced data of pathological images.The optimization of the generative adversarial network is completed by using the self-attention mechanism in the enhancement processing of deep learning.Lastly,a better model input data set is derived through experimental comparison and analysis,which lays a good data foundation for further classification studies.In the classification stage,the parameters are shared and optimized by the transfer learning method to reduce the model training time,and the comparison experiments of three different series of classification models,Dense Net,Mobile Net,and Efficient Net,are conducted on the better dataset obtained.Finally the more effective Efficient Net-B4 network is selected for model optimization and improvement.To further improve the classification accuracy,the Efficient Net-B4 model with the residual structure are combined and the different convolution kernel size of the residual structure is compared with the experimental analysis,then the effect of different optimizer is compared.Finally,a classification network combination based on the residual structure of Efficient Net-B4 model with Adam optimizer.Compared with the original Efficient Net-B4 classification model,this method resulted in a 2.5% increase in network sensitivity and accuracy and a 0.6% increase in specificity.