A Study On The Early Diagnosis Of Alzheimer’s Disease Through Neuroimaging And Deep Learning

Alzheimer’s disease(AD)is an incurable neurological illness that affects the elderly globally and has a substantial social and economic impact.Neurofibrillary tangles and amyloid plaques are hallmarks of AD,which are directly linked to neurodegeneration in the brain.Early detection and identification of AD ensures optimal patient care and monitoring of disease-modifying therapeutic interventions.For better diagnosis and monitoring of AD,imaging biomarkers like Positron Emission Tomography(PET)and Magnetic Resonance Imaging(MRI)are useful.Clinicians benefit greatly from automated image-based patient categorization.Deep learning(DL)methods such as Convolutional Neural Networks(CNNs)are state-of-the-art feature extraction and classification systems that take raw features from a large,annotated data set,such as a collection of images or genomes,and utilise them to produce a predicting tool based on patterns hidden within the data.Once trained,the algorithms may use that knowledge to interpret data from a variety of different sources.In this doctoral thesis,an attempt has been made to investigate DL techniques for early detection of AD utilising MRI and PET neuroimaging modalities.Data from The Open Access Series of Imaging Studies(OASIS)and the Alzheimer’s Disease Neuroimaging Initiative(ADNI)is used.There is a need to examine novel approaches for efficient CNN architectures for different binary and multiclass classification tasks,impact of data filtering approaches,impact of data augmentation techniques as well as the impact of disharmony between batch normalization and dropout techniques,which are two of the most widely used elements of modern CNN architectures.Different CNN versions,such as transfer learning architectures and custom 2D and3 D designs,have been used to examine binary and multiclass classification issues.Cross-domain transfer learning,image filtering,data augmentation,and disharmony between batch normalisation and dropout approaches,among other novel concepts published in the literature,have all been explored and their impacts have been monitored closely.The purpose of this doctoral thesis is to look into the implications of these ideas for the early detection of AD.This thesis is divided into four sections,each of which examines the influence of a different parameter on a CNN’s diagnostic capacity to discriminate between three classes: AD,Mild Cognitive Impairment(MCI),and Normal Control(NC).MCI is a stage in between AD and NC that exists on a continuum between the two,with the potential to evolve to advanced AD.The first section looks at how representation learning,such as 2D and 3D-CNN transfer and non-transfer learning architectures,affects the early detection of AD utilising MRI and PET modalities.The goal of this study is to discover the best feasible representation for differentiating between the AD,MCI,and NC classes.The results show that custom 3D-CNN architectures trained using PET modality data outperform2 D and 3D-CNN architectures trained with MRI and PET modalities.Issues like class imbalance and data leakage have been looked into.A key issue affecting the performance of these designs has been discovered to be class imbalance.Xception transfer learning architecture is used for the first time in this thesis to investigate the impact of representation learning on the early diagnosis of AD,and improved performances on the AD-NC binary,MCI-AD binary,and AD-NC-MCI multiclass classification tasks are achieved.Xception distant domain transfer learning architecture is proposed along with custom 3D-CNN architectures using both MRI and PET neuroimaging modalities.The proposed model outperformed its 3D counterpart on MCI-NC and AD-NC binary classification tasks.To the best of our knowledge,in the literature,Xception transfer learning model has not been specifically deployed to study these four classification problems,i.e.,AD-NC,NC-MCI,AD-MCI and AD-NC-MCI using PET neuroimaging modality and our study is unique in that it utilized the scans from this modality to report performance on these four classification tasks,i.e.,AD-NC,NC-MCI,AD-MCI and AD-NC-MCI.The second section of the thesis examines the influence of three data augmentation approaches on the early detection of AD utilising MRI and PET modalities,including random weak Gaussian blurring,random zoomed in/out,and random width/height shift.The goal of this study is to see how data augmentation strategies impact DL architecture performance.The best results were obtained using CNN architectures that used random zoomed in/out,random weak Gaussian blurring augmentation,and random width/height shift.It should also be highlighted that these augmentation approaches do not solve the problem of class imbalance in discriminating tasks employing the MRI neuroimaging modality.Random weak Gaussian blurring augmentation for multiclass classification is used for the first time in this thesis,and better performances on AD-NC binary,MCI-AD binary,and AD-NC-MCI multiclass classification tasks are achieved.The third section of the thesis looks at the impact of image filtering techniques on the early detection of AD utilising the PET neuroimaging modality.The influence of four strategies is examined in this thesis:(1)box filtering,(2)Gaussian filtering,(3)modified Gaussian filtering,and(4)median filtering are all examples of filtering methods.The goal of this study is to use box filtered,Gaussian filtered,modified Gaussian filtered,and median filtered data to measure the performance of 3D-CNN architectures.For AD-NC,MCI-AD,NC-MCI,and AD-NC-MCI classification tasks,3D-CNN architectures trained using box filtered,modified Gaussian filtered,median filtered,and Gaussian filtered data performed the best.For these classification tasks,no single filtering strategy outperforms others.These filtering algorithms are used for the first time in this thesis and enhanced performances on the AD-NC binary,MCI-AD binary,and AD-NC-MCI multiclass classification tasks are achieved.Finally,in the fourth section of the thesis,the influence of discrepancies between batch normalisation and dropout approaches on the early detection of AD utilising PET and MRI neuroimaging modalities is investigated.The influence of these strategies is examined under three scenarios:(1)training with no dropout in the presence of batch normalisation,(2)training with only one dropout layer directly before the softmax layer,and(3)training with a single convolutional layer between a dropout layer and a batch normalisation layer.The architectures trained under scenario(2)were found to be the best for the AD-NC binary classification task,AD-MCI binary classification task,and AD-NC-MCI multiclass classification task,while the architecture trained under scenario(1)was found to be the best for the NC-MCI binary classification task.This phenomena has been investigated for the early detection of AD for the first time in this doctoral thesis and better performance on the AD-NC binary,MCI-AD binary,and AD-NC-MCI multiclass classification tasks are achieved.Performance of architecture trained under scenario(3)was found to be inferior because of variation in the network activations distribution due to the change in network parameters during training resulting in saturated regimes where gradients are bounded to the extremes.Overall,it is found that image and volume-based classifications might aid in the early detection of AD and may serve as a potential biomarker for further study in this intriguing subject.