0619

Predicting Progression from Mild Cognitive Impairment to Alzheimer’s Disease
Anees Abrol1, Zening Fu1, and Vince D. Calhoun1

1The Mind Research Network, Albuquerque, NM, United States

Synopsis

This exploratory analysis tests the suitability of deep residual networks to learn neuroanatomical abnormalities from the structural MRI (sMRI) modality, and utility of dynamic (i.e. time-varying) functional connectivity approaches in delineating discriminative functional MRI (fMRI) features to predict progression of individuals with mild cognitive impairment to Alzheimer’s disease. Results demonstrate better than state-of-the-art prediction performance using the structural MRI modality alone. Multimodal prediction performed significantly better than unimodal sMRI or fMRI predictions, thus corroborating the benefits of predicting in the augmented space. Results also corroborate the diagnostic utility of the sMRI and fMRI features used to make the predictions.

Introduction

Alzheimer’s disease (AD) is an irreversible, progressive dementia that demands early diagnosis and therapeutic intervention, while mild cognitive impairment (MCI) is an intermediate condition between typical age-related cognitive deterioration and dementia. MCI individuals can progress to some form of dementia (not necessarily AD); however, in absence of a narrower prodrome for AD, MCI is often used as a prodromal stage of AD to predict and characterize progression to AD. This work focuses on predicting progression of MCI to AD by evaluating the structural MRI (sMRI) and functional MRI (fMRI) data of the cognitively normal (CN), MCI and AD elderly populations. More specifically, the aim of this work is to predict the subset of MCI individuals who would progress to AD within a period of three years (progressive MCI or pMCI) and the other subset of MCI individuals that do not progress to AD within this period (stable MCI or sMCI).

Methods

The first part of this work tests the suitability of a deep residual network1 (Figure 1) to learn neuroanatomical abnormalities in the gray matter volume (GMV) estimated from ADNI sMRI data (n = 828; CN: 237, sMCI: 245, pMCI: 189 and AD: 157). We test this network using a rigorous, repeated (n = 10) k-fold (k = 5) cross-validation procedure, where we first estimate the predictive (diagnostic/prognostic classification) power by training and testing on the MCI population only, followed by use of domain transfer learning (DTL) approach to perform network training additionally on the CN and AD groups. The second part of this work proposes a novel, multimodal (sMRI-fMRI) data fusion framework to predict progression of MCI to AD. This part featured unimodal (separate sMRI and fMRI) as well as multimodal prediction analyses on a smaller subset of subjects of the earlier used dataset for which both modalities were available (n = 132; CN: 34, sMCI: 36, pMCI: 24 and AD: 38). For the fMRI modality, features based on dynamic (i.e. time-varying) functional network connectivity (dFNC) were estimated (Figure 2) using a previously proposed method 2. The estimated sMRI and fMRI features were fused using canonical correlation analysis (CCA) in the multimodal prediction analysis (Figure 3). All three prediction analyses were performed using a rigorous stratified, repeated (n = 10) k-fold (k = 3) cross-validation procedure on the same training and test folds, following which a three-way performance comparison was conducted.

Results

In the first part of this work, our initial sMRI (n = 828) prediction analysis on the MCI group alone achieved a prediction accuracy of 77.2%, which was further enhanced in the DTL case to 82.7% (Figure 4). The DTL case performed significantly better than a classical SVM classifier (p = 2.57e-8). Additionally, both of these reported accuracies are a significant improvement over the state-of-the-art performance (75.44% as reported in 3) in making similar predictions using the sMRI modality alone. Next, in the second part of this work on the smaller multimodal data (n=132), the fMRI, sMRI and the multimodal prediction analyses resulted in cross-validated prediction accuracies of 70%, 75% and 78% respectively (Figure 5). Significant improvement was observed with multimodal prediction as compared to the unimodal fMRI (p = 1.0330e-6) and sMRI prediction analyses (p = 6.72e-4). Additionally, the prediction with sMRI features was found to be significantly better than that from the fMRI features (p = 0.0016).

Discussion

A significant improvement in the cross-validated prediction accuracy from the augmented (i.e. fused) feature space over those from the separate sMRI and fMRI feature spaces corroborates the presence of complimentary diagnostic information available in both modalities and validates the benefits of making such predictions in the augmented space. Results also evince that deep learning networks can be considered well-suited and further explored to evaluate neuroanatomical aberrations from the sMRI data, whereas dFNC approaches could be undertaken to capture diagnostic biomarkers from the fMRI data.

Conclusion and Future Work

Our results highlight the possibility of early identification of modifiable risk factors for understanding progression to AD using similar advanced deep learning and dFNC based frameworks. Such delineation is significant for early identification of individuals with high disease risk, who could also be reliably recruited for testing preventive treatments of AD. While, in this work, we train the two modalities separately and fuse their resultant feature spaces, future work would target development of a deep learning network that enables end-to-end training in both modalities in a parallel fashion for better optimization of the network’s weights, a feature that is highly likely to enhance the prediction performance.

Acknowledgements

This work was supported by NIH grant numbers 2R01EB005846, P20GM103472, P30GM122734, and R01REB020407 as well as NSF grant 1539067 to Dr. Vince D. Calhoun.

References

1. He, K., Zhang, X., Ren, S. & Sun, J. in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 770-778 (2016).

2. Rashid, B. et al. Classification of schizophrenia and bipolar patients using static and dynamic resting-state fMRI brain connectivity. NeuroImage 134, 645-657, doi:10.1016/j.neuroimage.2016.04.051 (2016).

3. Lu, D., Popuri, K., Ding, G. W., Balachandar, R. & Beg, M. F. Multimodal and Multiscale Deep Neural Networks for the Early Diagnosis of Alzheimer’s Disease using structural MR and FDG-PET images. Scientific Reports 8, 5697, doi:10.1038/s41598-018-22871-z (2018).

Figures

Figure 1: Structural MRI features were extracted using a deep residual network. These networks are composed of multiple residual blocks, each of which stack small stacks of convolutional and batch normalization layers followed by non-linear activation functions such as rectified linear units. In this study, as suggested by the data, we use a model with 3 residual layers for evaluating predictive (diagnostic/prognostic classification) performance of progression to AD.

Figure 2: Functional MRI features were extracted using a combination of time-varying FNC using a sliding window approach and a FC state regression pipeline 2. The windowed correlation (WCs) data were fed into a FC state regression pipeline that involved a fold-and-diagnostic-group-specific clustering (k = 5) to estimate the FC states, followed by regression of the WCs onto the estimated states to determine the contribution of each of these states in each of the WCs. These contributions were then averaged in the window dimension for each subject and used to train the linear SVM in a nested cross-validated setup.

Figure 3: An overview of the prediction performance comparison of the unimodal and multimodal analyses conducted in this work.

Figure 4: The domain transfer learning (DTL) task additionally trained the ResNet with all datasets from the CN and AD. (A) and (B) A significant improvement for all studied classification metrics (6% in accuracy, 7% in sensitivity, 5% in specificity and 7% in AUC) was observed for the DTL task as compared to the standard inter-MCI class classification task. The reported accuracy in the DTL task is also a substantial improvement of over the state-of-the-art performance as reported in 3. (C) The DTL task reported a significant performance improvement over the classical SVM model with a p-value of 2.5762e-8.

Figure 5: A comparison of accuracies from the two unimodal and multimodal prediction analyses. Multimodal prediction performance was found to be significantly better than the performance from the unimodal prediction analyses. The mean normalized accuracies for the fMRI, sMRI and multimodal prediction were 0.70, 0.75 and 0.78 respectively. A comparison of the cross-validation accuracies demonstrated significant differences (p < 0.005) between all three groups, with the p-values increasing the following order: fMRI vs. sMRI, sMRI vs. fMRI+sMRI, and fMRI vs. fMRI+sMCI.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
0619