Research On Deep Learning Methods For Medical Image Analysis

Medical image analysis is an indispensable technology in medical research,diagnosis and treatment.Its main role is to extract hidden important physiological and pathological information or knowledge from medical images.Deep learning has the outstanding ability to mine the internal structure of large-scale,high-dimensional data.It has achieved unprecedented success in many areas of pattern recognition and artificial intelligence.Therefore,we take several important and typical medical images such as melanoma derma mirror images,2D Hela fluorescence microscope images,peripheral blood smear white blood cell microscope images,and cervical colposcopy images as the research object.Research is carried out from the aspects of images preprocessing,training set amplification,deep neural network structure selection and improvement,and images analysis effect evaluation.A systematic and comprehensive study on the deep learning method for medical image analysis is carried out.The main work and achievements are as follows:1)We studied an improved CNN(convolutional neural network)structure for automated analysis of melanoma images(characteristic learning and segmentation)to distinguish between indistinguishable lesions such as pigmented spots and clinically-determined lesions.It can reduce screening errors and improve the precision of skin cancer diagnosis.We use the nonlinear activation functions ReLU and ELU to effectively alleviate the gradient disappearance problem,and use the RMSprop and Adam optimization and loss algorithms to achieve faster convergence.Finally,a batch normalization(BN)layer is added between the convolutional layer and the active layer to solve the problem of gradient disappearance and explosion.In addition,the segmentation method can use a smaller training set to obtain better segmentation results with shorter training and inference time.2)We studied the dermatological image classification method based on GoogleNet convolutional neural network.The model is capable of performing classification tasks for skin lesion images.In order to solve the imbalance of the training data set,we mainly use the preprocessing method at the data level,and the undersampling method has been adopted,thereby achieving a basic balance between classes.Finally,the average precision of skin lesion image classification reached 0.89,with an average recall rate of 0.88 and f1-score of 0.88.This method is significantly improved compared to the SegNet method and the FCN method based on deep learning.3)We studied a method for sub-cell image classification and prediction based on the CapsNet-Hela network model.The capsule layer we use consists of 36 capsules so that there is more room to store additional useful information in order to extract features.Consistent routing can pass useful information to the next layer while discarding data that might cause noise in the results.Experiments show that the classification effect using 3 route iterations is better.Correspondingly,the accuracy of classifying image dataset using the CapsNet-Hela network model reached 93.08%.This method is significantly improved compared to the SVM method based on Local Binary Patterns and the SVM method based on Haralick.4)We studied the method of segmenting and classifying the four main types of white blood cells(eosophilic cells,lymphocytes,monocytes,neutrophils)in a large number of blood images based on CapsNet-WBC network model.First,a white blood cell(WBC)image is segmented using a convolutional neural network.Then,the CapsNet-WBC network model is established and utilized to classify and predict the segmented WBC images.The experimental results show that our method has a classification accuracy of 85% for the WBC image test set and 99% for the training set.This method is significantly improved compared to other methods,and thus is expected to be applied to the automatic processing of larger-scale WBC images.5)We studied two deep learning models to achieve automatic classification method of cervical images.The experiment was divided into three steps: First,6992 cervical images were segmented and preprocessed using a convolutional neural network similar to U-Net.Secondly,we expand the number of cervical images after segmentation.Third,the 40152 cervical images after amplification were classified using the CapsNet-Cervix network model.An additional reconstruction loss is used to facilitate the encoding of the input image by the DigitCaps layer.In the experiment,the above method obtained higher classification accuracy than other traditional machine learning methods: the test set classification accuracy was 80.1%,and the training set accuracy was 99%.Moreover,after completing the model training,the instant segmentation and classification prediction of massive images can be completed in a short time.