Application Of Transfer Learning In Medical Image Processing

With the gradual maturity of medical imaging technology,medical images have played more important role in clinical diagnosis.Among various types of medical images,magnetic resonance images are widely used in clinical diagnosis due to its characteristics of constructing three-dimensional views and imaging clearly.In recent years,Alzheimer's disease increases has become one of the highest incidence rate diseases,the diagnosis of Alzheimer's patients has become increasingly important.Among them,structural magnetic resonance imaging has been used in the diagnosis of Alzheimer's disease because its advantages of clearly showing changes of tissue structure.It helps doctors to determine treatment methods and prognosis assessment.However,manual diagnosis of structural magnetic resonance imaging is time-consuming and laborious,so the auxiliary diagnosis of magnetic resonance imaging has large application values.However,in clinical practice,structural magnetic resonance images have large number image voxels,small sample sizes,and some even have no labels,which makes it difficult to classify structural magnetic resonance images using machine learning.Therefore,in order to solve the problem of classification on structural magnetic resonance imaging datasets with a small amount of data and no labels,this paper introduces domain adaptation,which is one of transfer learning methods.Domain adaptation is a transfer learning method to solve the problem of lacking of label in target domain.This paper mainly studies the method of deep domain adaptation based on Wasserstein distance and the method of domain adaptation which combining the encoder and generative adversarial network.The main contributions of this article are in the following two aspects.(1)A deep domain adaptation model based on Wasserstein distance used for classification is proposed.We add a domain adaptation layer on the basic classification model.The domain adaptation layer calculates the Wasserstein distance to bridge the gap between the two domain feature distributions and reduce the domain discrepancy,finally the model transfer from source domain to target domain.We test this model on ADNI and OASIS datasets for classification between normal people and patient.Experimental results showed that the proposed deep domain adaptation model based on Wasserstein distance is superior to other metric-based deep domain adaptation methods.(2)A deep domain adaptation model which combining encoder and generative adversarial network is proposed.First,we train an auto-encoder with source domain data.Then,we use the encoder as feature extractor(generator)of the generative adversarial network,the encoder combined with the classifier,gradient reverse layer and discriminator formed the final model.Finally,when the classifier classifies the source domain data,the discriminator performs domain discrimination on the two datasets,which means figure out whether the data comes from the source domain or the target domain.The model can learn shared features of the two domains by making discriminator unable to distinguish between the two domains while classifier do classification on source domain correctly.We verified this model on ADNI and OASIS datasets.The experimental results show that the proposed deep domain adaptive classification model combining encoder and generative adversarial network is superior to other domain adaptive methods.