Abstract:
Neurodegenerative disorders represent a major global health challenge
characterized by the progressive degeneration of neurons in the central nervous
system. These conditions primarily impact neural networks and can manifest in a
variety of symptoms, including cognitive decline, behavioral changes, and movement
disorders. The degeneration associated with these diseases often leads to
complications that may result in disability and, ultimately, death. Some of the most
common neurodegenerative disorders include Alzheimer’s disease (AD), Parkinson’s
disease (PD), Huntington’s disease, amyotrophic lateral sclerosis, and multiple
sclerosis. Out of all neurodegenerative disorders, Alzheimer’s and Parkinson’s disease
are the most prominent because they affect a significant portion of the population.
Their high prevalence, combined with the profound effects on cognitive and motor
functions, makes them critical concerns for public health and caregiving systems.
Recent advancements in neuroimaging techniques have enhanced the ability to
visualize brain changes associated with neurodegeneration. These techniques allow for
the observation of atrophy in brain tissue and deposition of abnormal proteins.
Additionally, neuroimaging techniques help assess changes in brain activity and
connectivity, offering valuable information about the progression of the disease. In
recent literature, numerous machine learning methods have been proposed to enhance
the identification of neurodegenerative disorders through neuroimaging modalities.
However, these advances are combined with several research gaps. These include
small sample sizes, limited incorporation of white matter features, and insufficient
integration of clinical test scores into classifiers. Many studies focus on
region-of-interest features, neglecting the potential of whole-brain analysis, and often
fail to adequately identify prodromal stages of disease. Additionally, the use of
separate algorithms for feature selection and classification can lead to the loss of
critical information. Distinguishing neurodegenerative diseases from others with
similar symptoms is challenging, compounded by the lack of standardized
preprocessing pipelines for imaging data. Addressing these issues could significantly
enhance the diagnostic accuracy of predictive models.
This thesis explores the application of machine learning techniques to analyze
neuroimaging data, aiming to enhance the identification and understanding of
neurodegenerative diseases. It presents three proposed mechanisms for detecting
disorders such as Alzheimer’s and Parkinson’s disease, specifically addressing the gaps
identified in the existing literature. The first proposed mechanism presents an approach
for detecting brain regions that contribute to Alzheimer’s disease using support vector
machine classifiers and the recently developed Self Regulating Particle Swarm
Optimization (SRPSO) algorithm. The classifiers for distinguishing subjects into AD
iv
patients and Cognitively Normal (CN) individuals were built using Gray Matter (GM)
and White Matter (WM) volumetric features extracted from structural magnetic
resonance images. It could be observed from results that the classifier built using both
GM and WM features provided better accuracy than the performance of classifier built
using either GM or WM features only. Moreover, consideration of clinical features in
addition to volumetric features further improves the accuracy of the classifier. In order
to identify the brain regions that are important for AD vs CN classification problem,
SRPSO is used to extract GM and WM features that are important for better
classification performance. This helped in reducing the number of features by a factor
of hundred. The features identified by SRPSO were also mapped back to the brain
using automatic anatomic labeling template to identify brain regions that exhibit
degeneration in AD. In addition to identifying areas known to be involved in AD like
cerebellum, hippocampus, this helped in finding newer areas that might contribute
towards AD.
The second proposed mechanism presents a new approach to distinguish
progressive Mild Cognitively Impaired (pMCI) subjects who eventually develop
Alzheimer’s disease from stable MCI (sMCI) subjects whose situation does not
deteriorate into AD. The proposed approach combines the discriminating capabilities
of classifiers and representation learning capacities of autoencoders into a unified
architecture, and is hence termed as Joint Autoencoder and Classifier Deep Neural
Network (JACDNN). JACDNN employs a single classifier and multiple autoencoders
that are trained together to perform pattern classification. The autoencoders in
JACDNN, regularizes individual layers in the network used for classification to learn
representations useful for reconstructing a given input. Performance of JACDNN has
been evaluated on several machine learning problems pertaining to dementia, namely
AD vs CN, AD vs sMCI, CN vs pMCI and pMCI vs sMCI. These problems are
targeted using two datasets. The first dataset consist of GM features of subjects and the
second dataset consist of combination of GM and WM features. It is observed that
better classification results are obtained when the classifier is built on GM and WM as
compared to GM features alone. Performance comparison of JACDNN with other
existing approaches has been conducted for these problems. The results clearly
indicate that JACDNN performs better than other existing approaches for these
problems.
The third proposed mechanism presents a novel deep neural network architecture
for distinguishing subjects of Parkinson’s disease from cognitively normal cohort. The
architecture combines the representation learning capacities and discriminating
capabilities of convolutional neural network to improve classification performance of
the network. For this purpose, the network consists of a classification sub-network and
multiple reconstruction sub-networks. The classification sub-network discriminates
v
between PD and CN subjects and the reconstruction sub-networks are used to
regularize the individual layers of the classification sub-network. All sub-networks are
trained simultaneously to optimize the customized objective function of the network.
The network is trained on Single Photon Emission Computerized Tomography
(SPECT) images obtained from baseline study of Parkinson’s Progression Marker
Initiative (PPMI). When compared to recent state of the art methods for PD detection,
the proposed mechanism reports upto 10% improvement in the classification accuracy