A Study On Classification Of Multimodal Alzheimer’s Disease Based On Transfer Learning

Alzheimer’s disease(AD)is a complex neurological disease.The etiology and pathological mechanism of AD are still not fully understood,and there is no reliable diagnosis and treatment to help patients recover.In the early stage of AD,that is,Mild Cognitive Impairment(MCI),early detection and diagnosis can slow down the progression of the disease.In recent years,brain imaging techniques such as magnetic resonance imaging(MRI)and positron emission tomography(PET)have provided new ideas and methods for the early diagnosis and prediction of AD.In the study of AD course classification based on machine learning,single-modal image data is often used,and the extraction of image information is not comprehensive enough.Moreover,in the combined study of MRI and PET images,less data can be obtained for the protection of patient privacy.Due to the high price and wide popularity of PET images,the data volume of these two modal images will be quite different,which will cause great difficulties in constructing the diagnostic model of AD.Therefore,in this paper,the generative adversarial network is first used to generate corresponding PET images based on MRI images to complete the missing PET images,so as to prepare for multimodal image fusion.Then,the paired MRI and PET data are fused at the pixel level to obtain rich data information.Finally,the lightweight convolutional neural network is improved by using transfer learning technology to classify AD more accurately.The main research contents of this paper are as follows:(1)PET image generation method based on generative adversarial networks.For the problem of unequal number of MRI and PET images,this paper uses a generative adversarial network to generate the missing PET images to achieve the number matching of PET images and MRI images.The specific method of improving the Cycle-GAN network is to use the UNet network structure to replace the generator,add the GCT attention mechanism module,and introduce the hyperparameter β in the cycle generation.Changing the generator to U-Net can fully learn independent features,and adding attention mechanism can pay more attention to the complex areas in PET images,which helps to improve the quality of PET image generation.(2)MRI and PET fusion method based on DWT-PCA.The amount of data information of single modal data is less and it is difficult to meet the requirements of model training.Therefore,this paper fuses MRI and PET images to obtain complementary information of the two imaging images to a large extent.The specific method is to use the pixel-level image fusion method,mainly combining the wavelet transform algorithm with the PCA algorithm,which can maximize the retention of the original image information,reduce the data dimension,and obtain the MRI-PET fusion image with better fusion effect.(3)Multimodal AD classification method based on transfer learning.Due to the small number of related image data sets of AD,if the traditional model training method is adopted,the model will not be fully fitted or over-fitted,and the training results will be poor.Therefore,this paper uses transfer learning technology to improve the classification network Mobile NetV3,and introduces the Fused-MBConv module,which can make better use of fewer data sets for experiments and reduce resource consumption to a certain extent.Through the experimental test of ADNI database,the method proposed in this paper can achieve 97.73 % accuracy in four classifications,and also has good performance in classification accuracy,recall rate and F1-score.