Pathological Image Classification Based On Convolutional Neural Network

Liver cancer is a malignant tumor with high morbidity and mortality,which is extremely harmful to human health and life.Therefore,it is very important for cancer patients to quickly and accurately identify and classify liver tissue sections.For the qualitative diagnosis of liver cancer,a variety of examination methods have been applied in the clinic,such as alpha-fetoprotein examination,pathological section,noninvasive ultrasound,CT,MRI examination,blood enzymology examination,etc.,to provide more orientation assistance for the diagnosis of liver cancer.According to the diagnosis,the specific inspection needs to be selected according to the actual situation.For the liver tissue section,the traditional classification method is to identify and predict by extracting image features.Because the image features are not obvious and need manual extraction,the external factors are too large and the classification method is unstable.Therefore,it is time-consuming and labor-intensive to classify pathological sections by conventional methods and the recognition rate is low.To this end,a deep learning based convolutional neural network method is used to identify classifications.Convolutional neural networks have better fault tolerance and can learn and process problems in parallel.In this thesis,we explore the pathological slice detection,establish a convolutional neural network model for training and recognition of the data set,and achieve a higher recognition rate by fine-tuning the model.The main work has the following three aspects: First,the Inception V3 model is improved,the full connection layer is removed,and the flatten layer,the dense layer,the Targeted Dropout layer,and the dense layer are sequentially added,and the sample image is used as input data through the convolutional neural network.The training results can obtain the experimental results,eliminating the cumbersome feature extraction link,and the experimental results show that the recognition rate is greatly improved.Second,the lightweight model MobileNet is improved,the full connection layer and Softmax activation function are removed and the flatten layer,the dense layer,the Targeted dropout layer and the dense layer are added in sequence.Compared with the previous experiment,the training time is greatly reduced.The third is to improve the VGG16 model,add the dilated convolutions on the original VGG16 model,increase the receptive field,and then remove the full connection layer,followed by the flatten layer,the fully connected layer,the Targeted dropout,and the fully connected layer..The accuracy of the experimental results is also relatively improved.Through the above experiments,the accuracy of verification by convolutional neural network is significantly higher than the current traditional method,and the training time is reduced by more than ten hours,which improves the efficiency.This is of great value to most medical research.