Article Preview
Top1. Introduction
Dementia is a brain disorder that incorporates constant declination of mental ability due to damaged neurons of the cerebrum. Dementia patients experience cognitive decline, prompting disturbance of their everyday existence. As verified by the World Health Organization (WHO), around fifty million population across the globe are encountering dementia while around ten million new instances are anticipated to arise annually (World Health Organization, n.d.). The disease does not have any cure apart from controlling its progression. Subsequently, early treatment of this neurodegenerative disease is a critical part of improving the patient’s condition. Various cognitive, clinical, and genetic tests are available for diagnosing dementia. Magnetic Resonance Imaging (MRI) images are utilized for diagnosing the disease, as they hold a connection with the topology of the brain, and also, the modifications in the morphological structure of the brain are easily visible (Castro et al., 2020).
Recently researchers and neurologists are contributing to the early diagnosis of dementia and have achieved encouraging results (Tabaton et al., 2010). Computer-Aided Decision (CAD) systems have gained popularity over the past decade with more than 1000 papers in the literature of the medical field. The design of CAD usually follows the given steps: (1) Acquisition of Image, (2) Pre-processing, (3) Extracting features, (4) Classification, and (5) Validation of results (Doi 2005, 2007). The above steps are important for the correct classification, but the extraction of a good set of features is of fundamental importance. (Mazurowski et al., 2008). Artificial Neural Networks (ANN), Machine Learning (ML), and Deep Neural Networks (DNN) techniques have shown great potential for providing aid in the diagnosis of Dementia. The majority of the studies proposed in the literature for the detection of the disease are based on supervised learning approaches such as Decision Tree (DT), ANN, Bayes Classifiers, or Support Vector Machine (SVM) (Bishop, 2006; Chaplot et al.,2006; Stonnington et al., 2010; Li et al., 2007; Vieira et al., 2017; Jiang et al., 2020).
The ANNs are classified into 2 categories on consideration of the hidden layers, shallow neural networks, and deep neural networks (Nielsen et al., 2015). Shallow Networks are networks having a single layer and deep networks have more than one layer, which is represented by Fig. 1.
Figure 1.
Architectural Representation of (a) Shallow Neural Networks and (b) Deep Neural Networks
There are few studies available that assess the differences between shallow neural networks and deeper neural networks (Schmidhuber, 2015; Winkler and Le, 2017). It is clear from the literature, that by introducing multiple layers, ANN becomes capable of learning features at a different level of abstraction. This capability of DNN makes easy generalization in comparison to shallow architectures. DL methods have arisen as the greatest development in machine learning. DL automatically extracts the discriminative image features, which helps in avoiding the errors while feature engineering. Different researchers implemented DL using Convolutional Neural Network (CNN) (Payan and Montana, 2015; Martinez et al., 2018, dense networks (Ortiz et al., 2017),), residual nets (Litjens et al., 2017), and many other networks (Hjelm et al., 2014; Vieira et al., 2017) in the field of dementia.