Machine learning approach to classify Alzheimer disease and Vascular Dementia using MRI quantitative metrics
Elia Tagliani1, Gloria Castellazzi1,2, Andrea De Rinaldis1,2, Fulvia Palesi2,3, Letizia Casiraghi2,4, Elena Sinforiani5, Paolo Vitali6, Nicoletta Anzalone7, Giovanni Magenes1, Claudia AM Gandini Wheeler-Kingshott2,8, Giuseppe Micieli9, and Egidio D'Angelo2,4

1Department of Electrical, Computer and Biomedical Engineering, University of Pavia, Pavia, Italy, 2Brain Connectivity Center, C. Mondino National Neurological Institute, Pavia, Italy, 3Department of Physics, University of Pavia, Pavia, Italy, 4Department of Brain and Behavioral Sciences, University of Pavia, Pavia, Italy, 5Neurology Unit, C. Mondino National Neurological Institute, Pavia, Italy, 6Brain MRI 3T research center, C. Mondino National Neurological Institute, Pavia, Italy, 7Scientific Institute H. S. Raffaele, Milan, Italy, 8NMR Research Unit, Queen Square MS Centre Department of Neuroinflammation, UCL Institute of Neurology, London, United Kingdom, 9Department of Emergency Neurology, C. Mondino National Neurological Institute, Pavia, Italy

Synopsis

Despite the large number of studies on dementia, efforts to define a clear profile of cognitive impairment for Vascular Dementia (VaD),as well as its differentiation from Alzheimer Disease (AD), are still poor. In this study we tested the power of imaging metrics and adopted a data mining approach, based on Diffusion Tensor Imaging and resting state fMRI, to assess the reliability of machine learning approaches for the automated diagnosis of AD and VaD. Our results show that machine learning algorithms are able to discriminate VaD from AD, representing a suitable approach to build an automated diagnostic system for dementia-like diseases.

Target

Researchers and clinicians who are interested in data mining approach to classify dementia-like diseases.

Introduction

Alzheimer Disease (AD) and Vascular Dementia (VaD) are two of the most common neurodegenerative kinds of dementia disease. Over the years, advanced structural and functional MRI techniques, such as Diffusion Tensor Imaging (DTI) and resting state functional MRI (rs-fMRI) have emerged as powerful methods to investigate the signs of neurodegeneration1,2. However, despite a large number of studies, efforts to define a clear profile of cognitive impairment for VaD, as well as its differentiation from AD, are still poor3. In this study we tested the power of imaging metrics and adopted a data mining approach, based on DTI and rs-fMRI, to assess the reliability of machine learning approaches for the automated diagnosis of AD and VaD.

Methods

31 subjects affected by AD (age 72.88±7.33, MMSE=15.97±6.38) and 27 subjects diagnosed with VaD (age 76.67±7.77, MMSE=17.94±4.22) underwent MRI examination using a 3T Siemens Skyra MR scanner (Siemens, Erlangen, Germany) with a 32-channels head-coil. All patients were selected based on their neuropsychological examination. 35 healthy controls (HC), age 69.43±9.65 and MMSE=28.56±1.45, were also studied. MRI acquisitions: 1) DTI: twice refocused SE-EPI (TR/TE=10000/97 ms, 70 slices with no gap, acquisition matrix=120×120, voxel size=2x2x2mm3, 64 directions, b=1200s/mm2, 10 volumes with no diffusion weighting); 2) rs-fMRI: EPI 2D BOLD sequence (TR/TE=3010/20ms, voxel size=2.5mm isotropic, FOV=224mm, 60 slices for a total of 120 volumes); 3) a 3DT1-weighted scan was also collected with a MPRAGE sequence (TR/TE/TI=2300/2.95/900ms, flip angle 9°, 256 slices, slice, voxel size=1x1x1.2mm3, FOV=270mm) for anatomical reference. From DTI we calculated fractional anisotropy (FA) and mean diffusivity (MD) maps using FSL4. Using these maps we extracted mean FA and MD values of 10 brain areas (Fig.1) that resulted as being particularly relevant from a previous Tract Based Spatial Statistics (TBSS) analysis.5 For each subject, rs-fMRI images were preprocessed using FSL4 and then parcelled using the automated anatomical labelling (AAL6) atlas into 116 areas. For each AAL area we calculated both mean FC value (by averaging the mean value of signals of all the voxels within the region) and graph metrics using the BCT7 toolbox. Extracted measures from DTI (mean FA and MD of areas in Fig.1) and from rs-fMRI (mean FC values and graph metrics of the AAL regions) were used as input features to compare 4 different classifiers, all implemented in Matlab: Support Vector Machine (SVM)8, Multilayer Perceptron (MLP)8, Radial Basis Function network (RBF network)8 and Adaptive Neuro-Fuzzy Inference System (ANFIS)9. Each classifier was run testing multiple input feature sets: i) DTI indexes, ii) FC indexes, iii) graph metrics, iv) DTI+FC, v)DTI+graph metrics. This step allowed us to identify the optimal input set of features to maximise the classification accuracy. In order to further optimise the classification performance we applied a feature selection algorithm (ReliefF8) to rank and then select the features with higher discrimination ability to construct the classifier.

Results

The machine learning algorithms were able to classify AD and HC with accuracy (ACC)=94%, VaD and HC with ACC=96%, AD and VaD with ACC=92% (Fig.2). In particular, comparing AD vs HC the best classification performance was obtained using ANFIS using a combined input set of features of DTI indexes and graph metrics (Fig.2,3), while SVM (with RBF kernel) showed high classification accuracy using only graph metrics as input features (AC=93%). Also, results from this comparison confirmed the key role of the hippocampi and default mode network (DMN) impairment in AD. SVM gave the highest classification performances when using combined DTI+graph metrics (AC=96%) and DTI+FC indexes (AC=93%) in discriminating VaD vs HC (Fig.2,4). Furthermore, DTI+graph metrics used as feature dataset for ANFIS obtained the best classification accuracy when comparing VaD vs AD (AC=92%, Fig.2,5).

Discussion and Conclusions

Our results showed that machine learning approaches were able to accurately discriminate both AD and VaD from HC as well as AD from VaD, while also confirming a differential involvement of the hippocampi and DMN regions in AD, compared to HC. The achieved accuracy (ACC>90%) is superior to reported values based on volumetric measurements10. SVM and ANFIS emerged as the most efficient classifiers. Moreover, these algorithms significantly improved their classification accuracy when using multimodal feature sets (e.g. DTI+graph metrics) rather than single modality feature sets (e.g DTI indexes). In particular, the feature set composed by DTI indexes and graph theoretical measures allowed to accurately discriminate AD, VaD and HC with the best classification performances (>90%). Results suggest therefore that machine learning algorithms represent a suitable approach to build an automated image-based system for advising on clinical diagnosis of dementia-like diseases.

Acknowledgements

The C. Mondino National Neurological institute (Pavia, Italy), the University of Pavia and the UCL-UCLH Biomedical Research Centre (London, UK) for ongoing support.

References

1 - Binnewijzend M (2012) Resting-state fMRI changes in Alzheimer's disease and mild cognitive impairment. Neurobiol aging 33(9): 2018-28

2 - Zhang D (2013) Determination of vascular dementia brain in distinct frequency bands with whole brain functional connectivity patterns. PLoS One 8(1): e54512

3 - Graham NL (2004) Distinctive cognitive profiles in Alzheimer's disease and subcortical vascular dementia. J Neurol Neurosurg Psychiatry 75(1): 61-71

4 - FSL, www.fmrib.ox.ac.uk/fsl

5 - De Rinaldis A (2015) Fractional Anisotropy reductions differentiate between Alzheimer Disease and Vascular Dementia. Proc OHBM abstact n.7483

6 - Tzourio-Mazoyer N (2002) Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 15:273-89

7 - Rubinov and Sporns O (2010) Complex network measures of brain connectivity: uses and interpretations. Neuroimage 52:1059-1069

8 - Witten IH (2011) Data mining, practical machine learning tools and techniques. Morgan Kaufmann Ed.

9 - Hosseini MS (2012) Review of medical image classification using the Adaptative Neuro-Fuzzy Inference System. J Med Signals Sens 2(1):49-60

10 - Wolz R (2011) Multi-method analysis of MRI images in early diagnostics of Alzheimer's disease. PLoS One 6(10):e25446

Figures

Fig.1 - Picture showing the 10 brain areas resulted from a previous TBSS analysis5 as the most relevant when comparing VaD, AD and HC. For each region mean FA and MD values were extracted and used as input features to construct the classifiers.

Fig.2 - Table showing the best classification performances obtained using single (FC indexes,graph metrics,DTI indexes) and multimodal (DTI+FC, DTI+graph metrics) feature sets. For each comparison (HC vs AD, HC vs VaD, AD vs VaD) the DTI+graph metrics feature set (in green) resulted the most informative when used in ANFIS and SVM.

Fig.3 - Picture showing the most informative features (according to ReliefF) discriminating HC vs AD. Discriminant DTI indexes (on the left) are located in the hippocampi areas, while discriminant graph nodes (on the right) are all areas classically considered part of Default Mode Network, making our results in line with literature.

Fig.4 - Picture showing the most informative features (according to the ReliefF) discriminating HC vs VaD (AC=96% using SVM classifier with RBF kernel). Location of discriminant DTI indexes and graph nodes is consistent with the hypothesis that age-related reduction in occipital activity might be coupled with age-related increased frontal activity.2

Fig.5 - Picture showing the three most informative graph nodes and features (according to the ReliefF ranking) to discriminate AD vs VaD.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
3765