3053

Cervical spinal cord atrophy contributes to classification of Alzheimer’s disease and vascular dementia patients
Roberta Maria Lorenzi1, Fulvia Palesi2, Paolo Vitali2, Alfredo Costa3,4, Gloria Castellazzi1,5, Elena Sinforiani6, Giuseppe Micieli7, Egidio D'Angelo4,8, and Claudia A.M. Gandini Wheeler-Kingshott4,5,9

1Department of Electrical, Computer and Biomedical Engineering,University of Pavia, Pavia, Italy, 2Neuroradiology Unit, Brain MRI 3T Research Center, IRCCS Mondino Foundation, Pavia, Italy, 3Unit of Behavioral Neurology, IRCCS Mondino Foundation, Pavia, Italy, 4Department of Brain and Behavioural Sciences, University of Pavia, Pavia, Italy, 5Queen Square MS Centre, Department of Neuroinflammation, UCL Queen Square Institute of Neurology, Faculty of Brain Sciences, University College London, London, United Kingdom, 6Laboratory of Neuropsychology, IRCCS Mondino Foundation, Pavia, Italy, 7Department of Emergency Neurology, IRCCS Mondino Foundation, Pavia, Italy, 8Brain connectivity center (BCC), IRCCS Mondino Foundation, Pavia, Italy, 9Brain MRI 3T Research Center, IRCCS Mondino Foundation, Pavia, Italy

Synopsis

Brain atrophy is an established biomarker for dementia. Here we tested the hypothesis that spinal cord atrophy is also an important in vivo imaging biomarker for neurodegeneration associated with dementia. 3DT1 images of Alzheimer Disease, Vascular Dementia and healthy subjects were processed to calculate spinal cord morphological parameters, such as vertebral spinal cord cross sectional areas and volumes. We confirmed the presence of significant spinal cord atrophy in dementia compared to healthy subjects. In particular, the C2-C3 vertebrae area resulted to have a considerable weight both for discriminating and classifying Alzheimer Disease from Vascular Dementia and Healthy control subjects.

Introduction

Dementia is a degenerative disease that affects the central nervous system. Alzheimer’s disease (AD) is responsible for the largest proportion of cases, whereas vascular dementia (VaD) is the second most common form of dementia. The incidence of dementia is growing, hence it is fundamental to find in vivo non-invasive imaging biomarkers that can help in identifying dementia subtypes. Some studies have shown that white matter, gray matter and specific brain structures such as hippocampi, thalami and amygdalae can be considered markers of dementia because of their significant abnormal atrophy(1). Spinal cord atrophy has been a sensitive imaging biomarker in diseases such as Multiple Sclerosis and Spinal cord injury. Here we question whether spinal cord atrophy could also be significantly different in dementia compared to healthy controls (HC) and therefore contribute to the characterization of dementia patients. In particular, we investigated whether spinal cord atrophy contributes to distinguishing AD, VaD and HC.

Methods

Subjects and MRI acquisition: High-resolution 3DT1 images of 32 HC (69±9.45yrs, 17males), 28 AD (73±7.40yrs, 18males), 19 VaD (76±9.11yrs, 4males) patients were acquired using a Siemens Skyra 3T scanner.

Spinal Cord analysis: 3DT1 images were analyzed with the open source software Spinal Cord Toolbox(2)(3) specifically developed to elaborate spinal cord imaging. Features on C1-C5 vertebrae were analyzed using 3DT1 images. For each subject, 3DT1 image was resized to center the FOV on the spine, the spinal cord was segmented and labelled to identify each spinal cord vertebra separately. Cross-sectional area (CSA) and volume (CSV) were calculated for each single vertebra and C2-C3 pair, given the known sensitivity of this combined level to disease severity in other diseases(4).

Brain atrophy analysis: White (WM) and gray matter (GM) were segmented using SPM12(5) while hippocampi, thalami, and amygdalae using FIRST (FSL(6)). Volume was calculated for intracranial and all other structures to compare spinal cord results with those from literature.

Statistical analysis: A general linear model regression (SPSS(7)) with a statistical threshold of p=0.05 was used to find significant differences between groups of dementia patients (AD and VaD separately) and HCs. Gender, age and total intracranial volume were used as covariates because the three groups were not matched.

Machine learning analysis: Orange(8) was used to find out which spinal cord measures could be relevant in a classification problem between different groups of subjects (AD-HC, VaD-HC and AD-VaD). Features ranking was implemented with ReliefF algorithm in a 5-folds cross validation. Classification accuracy, sensitivity and sensibility of Random Forest (RF) and Logistic Regression (LR) were tested.

Results and discussion

Figure1 reports labelled vertebrae in a randomly chosen subject for each of the three groups. Table1 shows that all CSA values were atrophic in patients compared to HC except for C4 vertebra in AD and C5 vertebra in VaD. Table2 shows that selected brain structures were atrophic in patients compared to HC. No significant differences were found between the two pathological groups. Table3 reports the best features subset to perform classification amongst each pair of groups. Figure3 shows ROC curves highlighting an accuracy of 80% of LR and RF between dementia patients and HC, using only imaging features. In AD-VaD classification the accuracy of LR is about 65%.

Our findings revealed that cervical spinal cord atrophy contributes to the classification of dementia patients, both with respect to HCs and in direct comparison between different dementia subtypes.CSA of C2-C3, C1, C2, C3 vertebrae, volumes of C2, WM, thalami and left hippocampus were significantly decreased in all patients compared to HC. GM and other brain structures were atrophic in AD compared to HC. Furthermore, spinal cord features were generally decreased in VaD with respect to AD patients, although did not reach statistical significance. Among the best features, the C2-C3 CSA was able to contribute to the classification between AD–HC and VaD–HC and both machine learning algorithms with leave-one-out testing had an excellent accuracy in both AD–HC and VaD–HC classifications. It is worth noting that in both classifiers C2-C3 vertebrae were extracted amongst the most informative spinal cord regions able to identify dementia patients. Furthermore, in the AD-VaD classification, C3 volume was included in the best subset and LR had an accuracy of 65%. These classifications are only based on imaging features, and are expected to improve when adding clinical and neuropsychology variables. Finally, studies with larger cohort of subjects will be able to confirm the best features and the classification accuracy.

Acknowledgements

We thank University of Pavia and Mondino Foundation (Pavia,Italy) for funding; UCL-UCLH Biomedical Research Centre (London,UK) for ongoing support.

References

  1. Koikkalainen, J., Rhodius-Meester, H., Tolonen, A. et al. Differential diagnosis of neurodegenerative disease using structural MRI data. NeuroImage Clin(2016), 11:435-449, doi:10.1016/j.nicl.2016.02.019.
  2. De Leener, B., Levy, S., Dupont, S.M. et al. Sct: Spinal cord toolbox, an open source software for processing spinal cord mri data. NeuroImage(2017), 145:24-43, doi:10.1016/j.neuroimage2016.10.009.
  3. http://sourceforge.net /projects/spinalcordtoolbox
  4. Zheng, L. ,Yaldizli, O., Pardini, M. et al. Cervical cord area measurement using volumetric brain magnetic resonance imaging in multiple sclerosis. Multiple Sclerosis and Related Disorder(2015), 4(1): 52-57, doi: 10.1016/j.msard.2014.11.004 5.
  5. https://www.fil.ion.ucl.ac.uk/spm/software/spm12
  6. https://fsl.fmrib.ox.ac.uk/fsl
  7. https://www.spss.it
  8. https://orange.biolab.si

Figures

Figure 1. Labelled vertebrae in HC, AD and VaD subject. Each labelled image is obtained as output of the automatic labelling algorithm implemented in Spinal Cord Toolbox.

Table 1. CSA values with corresponding statistical results between AD-HC, VaD-HC and AD-VaD. Mean(SE) represents the difference between mean value of patients and HC with relative standard error. Significance was set at p<0.05, with Bonferroni correction. * indicates significant values.

Table 2. Brain structures volumes with corresponding statistical results between AD-HC, VaD-HC and AD-VaD. Mean(SE) represents the difference between mean value of patients and HC with relative standard error. Significance was set at p<0.05, with Bonferroni correction. * indicates significant values.

Table 3. Best features subset following the feature selection performed with ReliefF on the entire training set. The algorithm is validated with a 5-fold cross validation and best subset is build up with the most 6 frequent features in each fold.

Figure 2. On the top: ROC curves for AD-HC, VaD-HC, and AD-VaD classification problem using RF and LR. Pathological class (AD = 1 and VaD = 3) is considered as the target class. RF ROC is reported in red, LR in blue. Both classifiers show higher performance (line in bold) than majority algorithm (diagonal). On the bottom: Classifiers performance in dementia patients classification problem using a Leave-one-out procedure, for each ROC curve. RF = Random Forest, LR = Logistic Regression, CA = Classifier Accuracy, Sens = Sensitivity, Spec = Specificity, AUC = Area Under Curve.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
3053