Deep Learning For Medical Image Analysis

Medical image is crucial for the clinical diagnosis.As the imaging equipment isbecoming popular,the number of medical image,which has better quality thanbefore,grows fast.Deep convolutional network has achieve remarkable success,more and more research is trying to apply deep learning on medical image.In the task of classification and segmentation,recent deep learning models performs better than the traditional method.Today the number of radiologist is insufficient to cope with increasing data.On the other hand,doctors are susceptible to fatigue and subjective factors,hardly able to work for a long time.In conclusion,it is important to research algorithms for medical image.The main work of this paper is listed bellow.(1)Researched the application of transfer learning on blood-cell image classification.By transferring pre-trained model parameters,fine-tuning the VGG,GoogleNet and ResNet framework,comparing the results of different fine-tuning methods.This paper carry out the classification of 4 type of blood-cell image samples,all the used models have trained by end-to-end way,required no the complex preprocessing and artificial feature extract pipeline.The experimental results show that by making used of transferring learning the models have reached a relatively high accuracy,reduced the training time.(2)Aiming to the problems that the excessive parameters of 3D convolution model and the lack of continuous features of 2D model in existing AD(Alzheimer's Disease)classification methods,this paper proposed a brain MRI image classification algorithm which combines 2D convolutional neural networks and long short-term memory network.To generate magnetic resonance images and extract image features,it trained a deep convolutional generative adversarial networks.That adding regularization to the loss function is to make the generated image sharper,resulting in the convolution network converges faster.A stacked long short-term memory network is made to perform classification task by using the input feature sequence and the hidden states of the LSTM layers.A 3D MRI image can also be viewed as a sequence of 2D images,hence it is possible to transform a image sequence to a feature sequence by a convolutional network and classify by a LSTM network,which allows extracting the 2D image features and the temporal dimension features,reducing the demand of memory compared with a 3D convolutional network.The experimental results show that the proposed pipeline can be effectively used to classify MRI image,achieved 0.9 for the area under the receiver operating characteristic curves both NC/MCI and NC/AD categories.(3)Based on the U-Net model,introduce prior knowledge to accelerate the convergence of the segmentation algorithm.Used auto-encoder to obtain the prior knowledge related to image features by pre-training a U-Net model.Merged the mask and the current input retina image as the new input to carry out the final segment.