Deep learning architectures for Parkinson's disease detection by using multi-modal features

• Published in: Computers in biology and medicine Vol. 146; p. 105610
• Main Authors: Pahuja, Gunjan, Prasad, Bhanu
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.07.2022; Elsevier Limited
• Subjects: Accuracy; Alzheimer's disease; Biological effects; Biomarkers; Cerebrospinal fluid; Classification; CNN; Creutzfeldt-Jakob disease; CSF; Datasets; Deep learning; Diagnosis; Internal Medicine; Machine learning; Magnetic resonance imaging; Medical imaging; Movement disorders; MRI; Neural networks; Neurodegenerative diseases; Neuroimaging; Other; Parkinson's disease; Single photon emission computed tomography; SPECT; SSAE; Deep learning; CNN; SSAE; PD; AE; MRI; CSF; SPECT
• DOI: 10.1016/j.compbiomed.2022.105610
• ISSN: 0010-4825; 1879-0534

The use of multi-modal features for improving the diagnosing accuracy of Parkinson's disease (PD) is still under consideration. Early diagnosis of PD is very crucial for better management and treatment planning of PD, as the delay in the diagnosis may even lead to death of the patient. Two frameworks, feature-level and modal-level, both of which are based on deep learning, are presented to classify the given subjects into PD and healthy by using neuroimaging (T1 weighted MRI scans and SPECT) and biological (CSF) features as the dataset. In the feature-level framework, all these features are integrated to form a heterogeneous dataset which is then supplied to two deep learning models to diagnose PD. In the modal-level framework, the number of features from T1 weighted MRI scans is reduced first, by using the filter feature selection method ReliefF. Those reduced number of features from the MRI scans are integrated with SPECT and CSF features to form another heterogeneous dataset, which is then fed to a deep learning model. Due to imbalanced nature of the dataset (consists of 73 PD and 59 healthy subjects), F1-score, geometric-mean, sensitivity, and specificity are measured, in addition to measuring the accuracy, to evaluate the performance of the developed models. A maximum accuracy of 93.33% and 92.38% is observed, for CNN, in the feature-level framework and modal-level framework, respectively. Though the complexity of the approach based on multi-modal features is high as compared to an approach that uses only one type of feature, i.e., either neuroimaging or biological; the results prove that the approach based on multi-modal features is useful for classifying the given subjects into PD and healthy and can help the clinicians in accurately diagnosing the PD. •Feature-level framework and modal-level framework are presented for classifying the subjects into PD and healthy.•These frameworks are based on deep learning, and use neuroimaging and biological features.•The results prove that the approach based on multi-modal features is useful for classifying the subjects into PD and healthy.•A maximum accuracy of 93.33% and 92.38% is observed, for CNN, in the feature-level framework and modal-level framework, respectively.