Diagnosis Of Alzheimer’s Disease Based On Patch Attention

Alzheimer’s disease(AD)is a neurodegenerative disease that frequently occurs in the elderly.It is often manifested as memory and cognitive dysfunction.In the later stage of AD,it can be life-threatening due to physical comorbidities.Since Alzheimer’s disease has an incubation period of several years or even decades before the onset of symptoms,and there is currently no cure for the disease.Therefore,accurate diagnosis and prediction of AD patients is crucial.Deep learning technology has been widely used in the field of image processing.The current methods used to study AD mainly include: voxel-based methods,region of interest(ROI)-based methods,and the whole image-based methods.Because the three-dimensional structural magnetic resonance imaging(s MRI)image contains too many voxels,and the number of model training images is relatively small,the voxel-based method may cause over-fitting.The process of manually labeling ROI features and the classification network model are independent of each other,causing it to not be well coordinated with the classifier.Therefore,the diagnostic performance of this method may be poor;the method based on the whole image is to use the whole brain image as the input of the network.Make it difficult to detect subtle structural abnormalities.Based on the above problems,the article uses a hierarchical full convolutional network model to extract image features at patch level,region level,and individual level,and then connect the corresponding network modules in series to achieve the integration of feature extraction and classifier construction.Since the channel-wise attention(CA)can assign different weights according to the importance of the information contained in different channels when extracting the advanced features of the image,based on the CA module,this paper proposes a patch attention(PA)module,assign different weights to different input patches,and build a hierarchical full convolutional network model PA-H-3DFCN based on PA.The specific work is as follows:(1)Select the s MRI of 231 NC(normal control),198 AD,238 s MCI(stable mild cognitive impairment)and 167 p MCI(progressive mild cognitive impairment)subjects in the ADNI data set and preprocess them.Based on the Alex Net,2D and 3D fully convolutional neural network models A-FCN and A-3DFCN are designed.The input images are the coronal slices of the s MRI image and the entire three-dimensional s MRI image.(2)A fully convolutional network H-3DFCN for hierarchical extraction of image features is proposed.The network input is a lot of 25×25×25 patches containing anatomical landmarks.The input layer,patch-level sub-network,region-level subnetwork and subject-level sub-network are connected in series to extract image features hierarchically,avoiding the problem of incomplete brain image information contained in a single patch or ROI.The channel-wise attention module plays an important role in extracting the advanced features of the image.The paper proposes a patch attention module based on the channel-wise attention,which assigns different weights to patches containing different amounts of AD information;Then the patch attention is added to the patch-level sub-network(P-net),and the Alzheimer’s disease diagnosis module PAH-FCN based on PA is constructed.(3)Due to the high correlation between MCI and AD brain image structure changes,the migration of the network model trained by AD vs NC is used in the classification of s MCI vs p MCI.According to the network performance evaluation indicators selected in the experiment,the diagnostic performance of the four network models of A-FCN,A-3DFCN,H-3DFCN and PA-H-FCN are analyzed and compared with other algorithms for diagnosing AD.The robustness of the Alzheimer’s disease diagnosis model based on patch attention proposed in this paper is verified.Among them,the accuracy of AD vs NC is 0.9320 and the specificity is 0.9670;the accuracy of s MCI vs p MCI is 0.8110 and the specificity is 0.8470.Compared with other algorithms,the AD vs NC model in this paper has the highest specificity,and the s MCI vs p MCI model has the highest accuracy.