Medical Image Assistetd Diagnosis Based On Convolutional Neural Network And Transfer Learning

Recent years,domestic and foreign research institutes have made significant progress in artificial intelligence and related fields.In particular,deep learning and other technologies have made great progress in many practical engineering applications compared to previous methods,caused widespread Concern around the world,aroused the interest of corporate employees,college students,and related researchers.The study of deep learning has a long history,but the real beginning of large-scale application of engineering practice began when Geoffrey Hinton and his team repeatedly made breakthroughs in the ImageNet Image Recognition Contest,which attracted the judges and people all over the world.Since then,more and more people have applied deep learning technology to the field of image analysis,and have continuously improved their algorithms,designed many more practical network structures,and made deep learning technology more and more mature in the field of image analysis,thanks to the improvement of hardware computing capabilities and the popularity of GPUs,it is also inseparable from the efforts of researchers.At present,deep learning technology has also been applied to the field of medical image analysis.CNN can abstract a large number of features,but training a CNN from scratch is time consuming and computational.First of all,training CNN requires a lot of labeled data.These labels are difficult to obtain.Especially in the field of medical image processing,these labels need to spend a lot of time for professional doctors to label,and the number of related pathologies is scarce.Training a deep CNN often encounters problems such as overfitting and non-convergence of the loss functions.The solution is to repeatedly adjust the network structure and learning parameters,This requires a lot of computing resources,memory resources,and it is extremely time consuming.This paper aims at the problems in the existing technologies,combines the advantages of the existing technologies,adopts the method of clustering and then fine-tuning,and uses the method of assisted domain transfer learning to improve the accuracy of CNN model recognition and compares the experimental results of other methods.The experiment was done on the OASIS medical image dataset,which solved some problems in the current medical image processing field.The details are as follows:(1)This paper proposes an image-aided diagnosis method for Alzheimer's disease based on the assisted domain transfer learning method.A natural image subset with high correlation is clustered between the natural image and the medical image,and the subset is used as a medical image transfer learning.This method first uses Alzheimer's disease datasets and natural image datasets as clusters,selects natural images that are close to the AD images,uses this natural image to fine-tune the VGG16 network,and convolution-pooling layer pairs in the network.The parameters were retained,and CNN classification AD images were initialized with these parameters,and experimental results were obtained after 5 cross-validation.(2)In the lung medical image dataset experiments,several factors affecting the prediction accuracy of CNN model were discussed,including data enhancement,transfer learning and scratch training.Different network structures,this paper compared the two network structures of AlexNet and VGG16.Two methods of training from scratch and transfer learning were compared in the processing of BRATS2015 dataset.