Deep learning architectures for Parkinson's disease detection by using multi-modal features
Introduction
Parkinson's disease (PD) is a neurological degenerative disorder that is affecting millions of people globally. It is very important to diagnose PD in its early stages as the delay in its diagnosis may lead to harmful results including death of the patient. Hence, early diagnosis of PD is important for better management and treatment planning of PD. A list of acronyms used in this research is provided in Table 1.
In recent years, researchers [1,2] have provided models for PD classification (i.e., for classifying the given subjects into PD and healthy) based on machine learning and statistical approaches and by integrating the features available from different modalities. The current research focuses on integrating the features from two different modalities (a) neuroimaging (T1 weighted MRI scans and SPECT) and (b) biological (CSF), and applying deep learning architectures on these integrated features. Since the features from these multiple modalities are combined or integrated together to form a heterogeneous dataset, the terms multi-modal and heterogeneous are used interchangeably in this research.
CAD systems, based on the deep learning architectures AE and CNN, are used not only for diagnosing PD [3,4] and tracking its progression but also for diagnosing other neurodegenerative diseases such as Alzheimer's disease [5,6] and Creutzfeldt-Jakob disease [7]. Morabito et al. [7] achieved an average accuracy of 89%, an average sensitivity of 92%, and an average specificity of 89% in differentiating early-stage Creutzfeldt-Jakob disease from rapidly progressive dementia by using deep learning concept. Zhang et al. [8] reviewed deep learning methods for two neurodegenerative disorders (Alzheimer's disease and PD), one neurodevelopmental (Autism spectrum disorder), and one psychiatric disorder (Schizophrenia) and discussed the limitations of the existing studies. Tufail et al. [9] applied 3D CNN for multiclass classification of Alzheimer's and PD by using PET and SPECT neuroimaging modalities.
Deep learning architectures have shown better results than traditional methods that are based on AI [10,11] because of their ability to learn the latent representation of data automatically. Also, deep learning-based architectures have the potential to model the non-linear relationships between data, thus making the deep learning-based architectures more appropriate for classification tasks. Inspired by earlier researches [3,[12], [13], [14], [15]], in the current research, we used deep learning architectures and developed a model that can perform PD classification more accurately.
For PD classification, in the current research, we presented a FL framework and a ML framework. The FL framework is based on the deep learning models SSAE (i.e., SAE1 as well as SAE1 + SAE2) and CNN, while the ML framework is based on CNN alone. An overview of these deep learning models is provided in Fig. 1.
Deep learning architectures provide more accurate results, for early detection of PD, than conventional AI [15], because deep learning architectures can extract features from the input set without a need for preprocessing the input. The terms deep architectures, deep learning architectures, deep learning models, and deep neural networks are used interchangeably in the current research. Deep architectures consist of simple structures that can do nonlinear operations and these simple structures are arranged to form a stack.
AE is an unsupervised learning-based ANN that uses backpropagation algorithm to provide the output value closest to the input value. That is, the goal of AE is to reduce the deformation between input and output data. The encoder and decoder are two parts of AE [5,16]. One of the variations of AE is SAE that comprises of a hidden layer connected to the input values by a weight matrix known as encoder. The hidden layer provides the output to a reconstruction vector by using a weight matrix to form the decoder [4].
Depending on the complexity of the problem to be solved, any number of SAEs can be layered or stacked, where the output of each hidden layer is connected to the input of the successive hidden layer to form a SSAE [4]. To enhance the accuracy of PD classification, the assumed deep learning architecture is optimized in a supervised manner by stacking another layer (as the last layer in SSAE) called softmax classifier (Fig. 2) that can predict the output labels efficiently. Softmax classifier is a generalization of logistic regression and this classifier provides a different probability for each of the labeled data or category, where the sum of all the probabilities is equal to one [4].
In the current research, both SAE and SSAE architectures are used for PD classification.
A CNN comprises three types of layers namely convolution layers, pooling layers, and fully-connected layers that can be stacked according to their functionality [17]. The convolution layer is connected to the input layer and determines the output from the neurons by convoluting the scalar product between the weights and the inputs, by applying different sized filters. The output of the convolution layer is then passed through the activation function followed by the pooling layer that further reduces the dimensions of the input data within that activation, thus reducing the computational complexity of the model. The last layer, i.e., the fully connected layer, produces the scores of each class from the activations that are to be used for classification. There are various hyper parameters that are employed by CNN. They are filter size, stride, and pooling type. More details about CNN can be found in Liu et al. [18]. In the current research, CNN is also used for PD classification.
The contributions of the current research include the following:
- (1)
Based on the authors' knowledge, this is the first model, for PD classification, that employs deep learning architectures on a heterogeneous dataset made up of neuroimaging (features from MRI scans; SBR values from SPECT) and biological (CSF) markers.
- (2)
Only those subjects for which the values of the above specified features (i.e., neuroimaging as well as CSF markers) are available are used in the current research. If any of those features' values are not available for a subject then that subject is not used in the current research.
The rest of the paper is outlined as follows: Section 2 describes an overview of the existing PD classification studies. Section 3 provides the details of the dataset used in the current research. Section 4 discusses FL and ML frameworks, for PD classification, by using heterogeneous datasets. Section 5 discusses the current research results and their comparison with existing research. Finally, Section 6 provides conclusion.
Section snippets
Related work
This section provides an overview of some recent studies that used deep learning architectures, for PD classification, by using neuroimaging or other biomarkers.
Dataset details
The values of (a) MRI scans, (b) SBR features of four striatal regions i.e., LC, LP, RC, RP, and (c) four CSF markers i.e., α-synuclein, Total-tau, P-tau181P, and Aβ42 for the 132 subjects (consist of 59 healthy and 73 PD) are downloaded from PPMI database (www.ppmi-info.org/data) in January 2021. In the current research, only those subjects for which the data from both the modalities i.e., neuroimaging (MRI scans; and SBR values from SPECT) and biological (CSF) markers is available are used.
FL and ML frameworks, for PD classification, by using heterogeneous dataset
In the current research, two frameworks, FL and ML, are presented for PD classification. These frameworks are based on deep learning architectures, as explained later on in 4.1 FL framework, for PD classification, by using heterogeneous dataset, 4.2 ML framework, for PD classification, by using heterogeneous dataset. These two frameworks are different from each other in terms of combining and using the features from two modalities i.e., from neuroimaging (MRI scans and SBR) and biological
Performance results and analysis
Since the dataset used for both FL and ML frameworks is imbalanced (i.e., the number of PD subjects (the majority class) is not same as the number of healthy subjects (the minority class) considered in this study), accuracy parameter alone is not enough to properly measure the overall performance [31] as the accuracy parameter tends to bias towards the majority class. To overcome this issue, for both FL and ML frameworks, other parameters, sensitivity, specificity, F1-Score, and geometric-mean [
Conclusion
In the current research, we presented FL and ML frameworks, both are based on deep learning architectures and heterogeneous dataset, for PD classification. ML is based on CNN alone, while FL is based on SSAE and CNN. The way the heterogeneous dataset is formed for FL is different from that for ML although both the heterogeneous datasets are made up of neuroimaging and biological markers.
The value of the accuracy parameter obtained in the FL framework (by using CNN) is better than the value of
Funding
The current research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Declaration of competing interest
The authors do not have any conflict of interest.
Acknowledgement
Data used in the preparation of this article were obtained from the Parkinson's Progression Markers Initiative (PPMI) database (http://www.ppmi-info.org/data). For up-to-date information on the study, visit http://www.ppmi-info.org. Data biospecimens used in the analyses presented in this article were obtained from the Parkinson's Progression Markers Initiative (PPMI) (http://www.ppmi-info.org/specimens). As such, the investigators within PPMI contributed to the design and implementation of
References (31)
- et al.
High-accuracy detection of early Parkinson's disease through multimodal features and machine learning
Int. J. Med. Inf.
(2016) - et al.
A survey of deep neural network architectures and their applications
Neurocomputing
(2017) - et al.
Predicting severity of Parkinson's disease using deep learning
Procedia Comput. Sci.
(2018) - et al.
Handwritten dynamics assessment through convolutional neural networks: an application to Parkinson's disease identification
Artif. Intell. Med.
(2018) Computational anatomy with the SPM software
Magn. Reson. Imag.
(2009)- et al.
Voxel-based morphometry—the methods
Neuroimage
(2000) - et al.
Relief-based feature selection: introduction and review
J. Biomed. Inf.
(2018) - et al.
A practical approach to feature selection
- et al.
Early detection of Parkinson's disease through multimodal features using machine learning approaches
Int. J. Signal Imag. Syst. Eng.
(2018) - et al.
Deep convolutional autoencoders vs PCA in a highly-unbalanced Parkinson's disease dataset: a DaTSCAN study