We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
MRI based classification of mild cognitive impairment in Parkinson's disease using machine learning.
- Authors
Genc, Ozan; Arslan, Dilek Betül; Cengiz, Sevim; Hatay, Gökçe Hale; Kıçik, Ani; Erdoğdu, Emel; Kaplan, Özge Can; Tüfekçioğlu, Zeynep; Bilgiç, Baflar; Hanağası, Haşmet; Gürvit, İ. Hakan; Demiralp, Tamer; Uluğ, Aziz Müfit; Işık, Esin Öztürk
- Abstract
Objective: The aim of this study was to classify Parkinson's disease mild cognitive impairment (PD-MCI), cognitively normal Parkinson's disease (PD-CN) and healthy control (HC) groups based on multimodal magnetic resonance imaging (MRI) using machine learning methods. Methods: 33 PD-MCI, 27 PD-CN and 17 HC participated in this study. The participants were diagnosed by neurologists according to the neuropsychological test and physical examination results. MRI data was obtained at a 3T Philips clinical MR system using a 32-channel head coil. Mean cerebral blood flow (CBF), arterial blood volume (aBV) and bolus arrival time (BAT) maps obtained from arterial spin labeling MRI (ASL-MRI), fractional anisotropy (FA) and mean diffusivity (MD) maps obtained from diffusion tensor imaging (DTI), and metabolite peak ratios obtained from proton MR spectroscopic imaging (1H-MRSI) at various brain regions were used as features. Various machine learning methods were employed with appropriate hyperparameters. In addition, feature selection algorithms and dimension reduction techniques such as principal component analysis (PCA) and non-negative matrix factorization (NNMF) were assessed. Features having high correlation with each other were eliminated. Results: Removing highly correlated features increased the model performance. The subset of features selected by the randomized logistic regression and leave one out cross-validation (RLR-LOOCV) method contained 10% of all features from all the MRI modalities. The classification accuracies were 77% for PD-MCI versus HC, 80% for PD-MCI versus PD-CN, and 71% for PD-CN versus HC. Conclusion: Machine learning based on multimodal MRI might be helpful in early diagnosis of PD-MCI. Future studies aim to improve the classification of PD-MCI in a larger patient cohort. This study has been supported by TÜBITAK #115S219 and Ministry of Development #2010K120330.
- Subjects
MILD cognitive impairment; PARKINSON'S disease; MACHINE learning
- Publication
Anatomy: International Journal of Experimental & Clinical Anatomy, 2018, Vol 12, Issue Supp1, pS46
- ISSN
1307-8798
- Publication type
Abstract