Characterizing neurodegeneration in progressive supranuclear palsy using VBM and SVM classification
Karsten Mueller1, Robert Jech2,3, Cecilia Bonnet2,3, Jaroslav Tintěra4, Harald E Möller1, Klaus Fassbender5, Jan Kassubek6, Markus Otto6, Evžen Růžička2,3, and Matthias L Schroeter1,7

1Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany, 2Department of Neurology and Center of Clinical Neuroscience, Charles University in Prague, Prague, Czech Republic, 31st Faculty of Medicine and General University Hospital in Prague, Prague, Czech Republic, 4Institute for Clinical and Experimental Medicine, Prague, Czech Republic, 5Clinic and Polyclinic for Neurology, Saarland University Homburg, Homburg, Germany, 6Clinic and Polyclinic for Neurology, University of Ulm, Ulm, Germany, 7Clinic for Cognitive Neurology, University Hospital Leipzig, Leipzig, Germany


Structural brain differences were investigated between patients with progressive supranuclear palsy (PSP) and healthy controls with T1-weighted images (MP-RAGE) acquired at four centers with different 3T scanners (Siemens). Using voxel-based morphometry, we found a major decline in gray matter density in brainstem, insula, striatum, and frontomedian regions that is in line with the current literature. Support-vector-machine classification provided a high sensitivity of disease detection when using relevant brain regions in feature selection.


Progressive supranuclear palsy (PSP) is a neurodegenerative disease that is characterized clinically by atypical parkinsonism, supranuclear palsy, postural instability, mild dementia, and neuropathologically by the accumulation of tau protein resulting in neurofibrillary tangles.1 It is associated with structural changes in the midbrain (“hummingbird” or “penguin sign”), but recent meta-analyses show that other brain regions, predominantly in striatum and insula, are also affected.2 To further investigate PSP and structural brain changes caused by this rare disease, a set of patients was investigated using a multi-centric approach.


20 PSP patients (7 female, 67.3±7.8 years) were compared to 20 matched healthy controls (8 female, 66.3±7.8 years). T1-weighted images (MP-RAGE) were acquired at four different centers employing Siemens 3T scanners (Trio, Allegra, Skyra, Verio).

Voxel-based morphometry3 was performed with SPM and the VBM8 toolbox. Gray-matter-density (GMD) images4 were generated using the unified segmentation approach and smoothed with an 8mm Gaussian kernel. Voxel-wise statistical analysis (two-sample t-test PSP patients vs. controls) was computed controlling for age, sex, and total intracranial volume. The significance level was set at p<0.05, family-wise error (FWE) corrected threshold on the cluster level.5

To study effects induced by a single center, and to assess between-center variability, statistical analyses were performed separately for Prague, Czech Republic (uni-center approach), and for three German centers (Ulm, Homburg, Leipzig; multi-center approach). A conjunction analysis was performed between the Prague and the German cohorts to investigate the overlap of the results. A further test of between-center variability was performed by merging the participants from Prague and from one of the three German centers to generate three cohorts for another conjunction analysis of the maximum overlap.

Support-vector-machine (SVM) classification (patients vs. controls) was performed with GMD images using the libSVM software package.6 Classification accuracy was assessed by generating a set of 400 models with leaving a patient and a control subject out when building the classifier. Thereafter, it was checked if both remaining data sets were classified correctly. To assess the stability of classification results, two different kernels (linear, polynomial) and two different approaches of feature selection were compared. First, voxels were used with SPM's GM tissue probability map with a minimum GM probabilitiy using various thresholds between 0 and 80%. Second, ROIs were defined with the WFU-Pickatlas selecting striatum, thalamus, caudate, and midbrain, as suggested as PSP’s core network.1,2 Finally, a model was generated from all subjects, and relevant voxels for classification were detected.


Comparing patients and controls, a diminished GMD was observed in brainstem, thalamus, left and right anterior insulae, and in the vicinity of putamen (Fig. 1). Less prominent differences were detected in lateral orbitofrontal regions. Consistent results were obtained when investigating GMD differences in the multi-centric German cohort (Fig. 2). The conjunction analysis comparing the Prague and the German cohort is shown in Fig. 3 with an overlap of reduced GMD in the striatum and insular cortex (yellow). The German cohort showed prominent GMD reductions in brainstem (red) while the Prague cohort also showed diminished GMD in frontomedian regions (blue). Fig. 4 shows the conjunction between three cohorts obtained by merging participants from Prague and one German center with an overlap in striatum, insula, caudate, and frontomedian regions (yellow).

Using SVM with a polynomial kernel and feature selection of voxels with GM tissue probability >0.4, patients and controls were detected with a classification accuracy of 79 and 84%, respectively (sensitivity 0.83; specificity 0.80). Other masks yielded similar accuracies. With a linear kernel, accuracies to classify patients were lower (≈70%). The ROI-based approach focusing on striatum and midbrain outperformed the inclusion of all GM voxels (Fig. 5; accuracies of 84 and 86% for patients and controls, respectively; sensitivity 0.86, specificity 0.85). Both approaches yielded higher sensitivity of disease detection than specificity of correctly classified controls. Relevant voxels for classification were located in brain stem, striatum, and caudate, but also in cerebellar regions (Fig. 5).


In line with previous work,7-9 we found disease-related GMD decrease using a multi-centric approach. The specific pattern of brain atrophy is essential for understanding neuropathological mechanisms in PSP. Brain regions affected by PSP showed a high relevance for correct classification from controls as a precondition for future therapy. Further development of SVM-based classification might complement the radiologist’s MRI-based diagnostics for PSP disease detection and characterization.


Supported by IGA MZ CR NT12282-5/2011 and NT/12288-5/2011, PRVOUK P26/LF1/4; PDF-IRG-1307, BMBF Kompentenznetze Neurodegenerative Demenzen, Michael J Fox Foundation.


1. Williams DR, Lees AJ. Progressive supranucler palsy: Clinicopathological concepts and diagnostic challenges. Lancet Neurol. 2009;8(3):270-279

2. Shi HC, Zhong JG, Pan PL, et al. Gray matter atrophy in progressive supranuclear palsy: meta-analysis of voxel-based morphometry studies. Neurol Sci. 2013;34(7):1049-1055.

3. Ashburner J, Friston KJ. Voxel-based morphometry—the methods. NeuroImage 2000;11(6):805-821.

4. Ashburner J, Friston KJ. Unified segmentation. NeuroImage 2005;26(3):839-851. Unified segmentation.

5. Nichols T. Controlling the familywise error rate in functional neuroimaging: A comparative review. Stat Methods Med Res. 2003;12:419-446.

6. Chang CC, Lin CJ. LIBSVM: A library for support vector machines. ACM TIST. 2011;2(3):e27.

7. Price S, Paviour D, Scahill R, et al. Voxel-based morphometry detects patterns of atrophy that help differentiate progressive supranuclear palsy and Parkinson's disease. NeuroImage 2004;23(2):663-669.

8. Cordato NJ, Duggins AJ, Halliday GM, et al. Clinical deficits correlate with regional cerebral atrophy in progressive supranuclear palsy. Brain 2005;128(6):1259-1266.

9. Ghosh BCP, Calder AJ, Peers PV, et al. Social cognitive deficits and their neural correlates in progressive supranuclear palsy. Brain 2012;135(7):2089-2102.


Significant GMD differences between PSP patients and age- and sex-matched healthy controls (p<0.001, k>1000, controlled for multiple comparisons using FWE-correction with p<0.05 on the cluster-level). In PSP patients, a diminished GMD was observed in the brain stem, insula, and also in wide regions of the striatum predominantly in the putamen.

Orthogonal brain sections showing significant GMD differences between PSP patients and healthy controls (p<0.005, k>1000) in the German cohort. Note that patients were scanned on three different scanners (Siemens, Erlangen) at 3 T using the MPRAGE protocol. Significant clusters are shown using FWE-correction with p<0.05 on the cluster-level.

Conjunction analysis showing both the uni-centric Prague cohort (blue) and the multi-centric German sample (red, see also Figure 2) showing significant GMD differences between PSP patients and healthy controls (p<0.005, k>1000). The overlap (yellow) shows a reduced GMD in patients in the striatum but also in the insular cortex.

Conjunction analysis showing significant GMD differences (p<0.005, k>300) in three groups of participants merging the Prague with one German cohort (PU=Prague and Ulm; PL=Prague and Leipzig; PH=Prague and Homburg). The maximum overlap (yellow) shows major reductions of GMD in patients in the striatum but also in the insular cortex.

Weights of voxels most relevant for SVM classification between PSP patients and healthy controls. Classification accuracy was obtained by cross-validation generating a set of 400 models. SVM classification was performed on all voxels within the SPM’s gray matter mask (top) and within a ROI including striatum and midbrain (bottom).

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)