Alzheimer's disease identification using joint mutual information based feature selection and extreme learning machine: Structural MRI, CSF and cognitive score

Background: Earlier detection of Alzheimer's disease (AD) can help with proper treatment and prevent brain tissue damage.Recently, advanced deep learning techniques have successfully demonstrated human-level performance in numerous fields including medical image analysis.So in this study, we propose a 3D deep neural (3D-DN) network for early diagnosis of the AD using structural MR images. Methods:We used the T1-weighted sMRI scans from the NACC dataset (https://www.alz.washington.edu/).A total of 2353 subjects (498 patients with AD, 218 patients with MCIs (stable MCI, who had not converted to AD within a 36-month time period), 440 patients with MCIc (MCI caused by AD, who had converted to AD within a 36-month time period), and remaining 1197 patients with healthy control (HC).Moreover, before passing this 3D scans to the 3D-DN network, we applied some image preprocessing method on them, like, at first 3D images from the NACC datasets were normalized to the MNI space using SPM12 and the DARTEL registration method and also a data augmentation technique was used to increase the number of the samples.Moreover, these normalized images were later segmented into three tissues (gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF)).Furthermore, extracted GM images were passed to the 3D-DN network which uses pre-trained 3D VGG16 architecture for the training and testing purpose.Here, we have used 10-fold stratified cross-validation (SKF-CV) technique to overcome the problem of overfitting.The input of the model is three-dimensional images of 145 × 121 × 145 size.Results: In our study, we have validated our proposed method on three binary classification problem (AD vs HC, MCIs vs MCIc, and HC vs MCIc).Here in this study, we have modified 2D VGG16 CNN architecture to 3D VGG16 model.The network was pre-trained on ImageNet, an everyday image dataset containing 14 million images of 1000 classes, before being fine-tuned using 90% of the NACC dataset.Moreover, the dataset was divided into three part within the ratio of 80:10:10, whereas, 80% data is used for training and 10% used for validation and remaining 10% is used for testing purpose.Highest accuracy, sensitivity, and specificity (higher than 98%) were obtained for the AD vs NC group.3D VGG16 architecture was also able to discriminate MCIs vs MCIc group with an optimal performance value higher than 90%.Conclusion: 3D DNN model show promises for building a model for the automated, individual and early detection of AD and its classification accuracy also reach significantly higher than other machine learning strategies.For this dataset, our 3D VGG16 modal has performed well compared to other recently published state-ofthe-art methods.