Development of new neuro-radiological markers of Parkinson's disease using pattern recognition

Background and objectives: Automatic classification of Parkinson’s disease (PD) versus healthy controls (HC) based on structural MRI has so far focused on unimodal approaches. However, this method is subject to a poor temporal and spatial resolution leading to low classification accuracy. To overcome this limitation we propose to integrate different modalities by generating a single decision function based on a multi-kernel method, exploiting the complementary information it offers. We predict that the integration of multiple modalities produces greater classification enhancement. Materials and methods: 3Tesla MRI was acquired in 42 patients with PD and 42 age and gender matched healthy controls. We relied on Unified Parkinson’s Disease Rating Scale (UPDRS) for evaluating the clinical status. We used structural and quantitative maps of T1, T2*, proton density (PD), magnetization transfer (MT), Multi-parameter (MT magnetization transfer, proton density (A), Iron Deposit (R2 *), mixing water content, iron, and the fraction of macromolecules tissues (R1) at 1 × 1 × 1 mm3 resolution. We identified cortical and subcortical brain regions (cortex, putamen, globus pallidus, substantia nigra), and cortical grey matter. We applied existing classification algorithms in the field of neuroscience using a classification algorithm based on Support Vector Machines (SVMs) [1], executed using the Pattern Recognition for Neuroimaging Toolbox (PRoNTo) [2]. The processes of classification was the following, data were mean centered and leave one subject out cross-validation was performed, making the test set independent from the training set. Analyses were restricted to voxels where all subjects had non-zero values. Statistical significance of the classifications was tested using permutation testing (1000 permutations) with random assignment of group class to the input image. Subsequently, we combined different modalities (MT, A, R1, R2) and identified the combination giving the highest sensitivity and sensibility in PD classification. As classifier we used support vector machines that are inspired by statistical learning theory Vladimir Vapnik and Multiple Kernel Learning approach, introduced by Lanckriet [3],[4]. Our approach can be seen as an analogue of MKL with SVMs. Conclusion & Future work: Identification of brain areas with affected intensity in the Parkinson’s group compared to Healthy Controls in single modalities using pronto is helpful. However, the subsequent multi-kernel approach utilizes unimodal information in a combined fashion so that emergent information is obtained, transcending effectiveness unimodal approaches. In conclusion, our findings suggest that combining different imaging modalities and different regions of interest increase classification accuracy significantly. These results are promising for objective diagnosis in medical practice.Development of new neuro-radiological markers of Parkinson's disease using pattern recognitio