Using Interpretable Deep Learning to Predict Age with Cortico-Hippocampal Functional Connectivity
Poster No:
1418
Submission Type:
Abstract Submission
Authors:
Yifei Sun1, Marshall Dalton2,3, Fernando Calamante1,3,4, Jinglei Lv1,3
Institutions:
1School of Biomedical Engineering, The University of Sydney, Sydney, New South Wales, 2School of Psychology, The University of Sydney, Sydney, New South Wales, 3Brain and Mind Center, The University of Sydney, Sydney, New South Wales, Australia, 4Sydney Imaging, The University of Sydney, Sydney, New South Wales, Australia
First Author:
Co-Author(s):
Marshall Dalton
School of Psychology, The University of Sydney|Brain and Mind Center, The University of Sydney
Sydney, New South Wales|Sydney, New South Wales, Australia
School of Psychology, The University of Sydney|Brain and Mind Center, The University of Sydney
Sydney, New South Wales|Sydney, New South Wales, Australia
Fernando Calamante
School of Biomedical Engineering, The University of Sydney|Brain and Mind Center, The University of Sydney|Sydney Imaging, The University of Sydney
Sydney, New South Wales|Sydney, New South Wales, Australia|Sydney, New South Wales, Australia
School of Biomedical Engineering, The University of Sydney|Brain and Mind Center, The University of Sydney|Sydney Imaging, The University of Sydney
Sydney, New South Wales|Sydney, New South Wales, Australia|Sydney, New South Wales, Australia
Jinglei Lv
School of Biomedical Engineering, The University of Sydney|Brain and Mind Center, The University of Sydney
Sydney, New South Wales|Sydney, New South Wales, Australia
School of Biomedical Engineering, The University of Sydney|Brain and Mind Center, The University of Sydney
Sydney, New South Wales|Sydney, New South Wales, Australia
Introduction:
Exploring healthy brain aging sets a baseline to understand aging related brain diseases and can help develop targeted interventions. Among various brain circuits, cortico-hippocampal connectivity is reported to be associated with age-related cognitive decline (Dennis et al. 2014). However, the impact of aging on the cortico-hippocampal connectivity requires more research. Our study uses convolutional neural networks (CNN) to predict age with the seed-based cortico-hippocampal functional connectivity (FC). We also generated saliency maps of CNN on the cortex to explain the regional variability of contribution to age prediction.
Methods:
We used the minimally preprocessed (Glasser et al. 2013) resting-state functional magnetic resonance imaging (rs-fMRI) data from the Human Connectome Project Aging dataset (Bookheimer et al. 2019) (720 subjects, aged 36-100 years). After applying spatial smoothing and temporal filtering, we calculated the seed-based cortico-hippocampal FC. To test our method specificity, we divided the hippocampus into anterior and posterior seeds and computed two additional sets of FC maps.
We trained three 3D CNN models to predict age with the same architecture to minimize the mean absolute error (MAE) using the whole, anterior, and posterior hippocampal cortical FC, respectively. 5-fold cross-validation was used to avoid overfitting. We then used Layer Class Activation Mapping (LayerCAM) (Jiang et al. 2021) to interpret models and identify age-predictive brain regions by generating high-resolution saliency maps. We also explored differences between the anterior and posterior hippocampal FC models through one-way t-test on saliency maps.
We trained three 3D CNN models to predict age with the same architecture to minimize the mean absolute error (MAE) using the whole, anterior, and posterior hippocampal cortical FC, respectively. 5-fold cross-validation was used to avoid overfitting. We then used Layer Class Activation Mapping (LayerCAM) (Jiang et al. 2021) to interpret models and identify age-predictive brain regions by generating high-resolution saliency maps. We also explored differences between the anterior and posterior hippocampal FC models through one-way t-test on saliency maps.
Results:
CNN models predicted age with MAEs of 6.9, 7.1, and 6.8 years for whole, anterior, and posterior hippocampal FC, respectively. Peak prediction accuracy (MAE of 4.4 years) was observed in subjects aged 55-60 (Fig. 1A). While another study, focusing on a narrow age range (40-69), achieved a prediction MAE of 4.8 years using whole brain FC matrices (He et al. 2020), our study involves a wider age span and whole brain voxel-wise hippocampal FC for finer spatial resolution.
The mean saliency map from the whole hippocampal cortical FC based model (Fig. 1B) presents hotter (redder) regions with higher contribution of the connectivity to the prediction, e.g. precuneus, retrosplenial cortex, and occipital lobe. These areas align with known aging relevant regions. The precuneus is vulnerable to Alzheimer's disease (AD) (Zhang et al. 2021) and hippocampal-retrosplenial cortex FC relates to tau accumulation in the medial parietal region (Ziontz et al. 2021). The occipital lobe atrophy was also found in healthy old adults (Harrison et al. 2019).
Mean saliency maps from anterior and posterior hippocampal cortical FC models (Fig. 2A-B)) show distinct patterns. One-way t-tests identify significant differences (Fig. 2C-D, z-score>1.645). Specifically, FCs between the anterior hippocampus and lateral occipital cortex, precuneus, and medial prefrontal cortex were more influential on age prediction than their posterior hippocampus connections. Conversely, posterior hippocampal FC with medial occipital and posterior parahippocampal had a greater prediction effect. These disparities may due to different functional roles of the anterior and posterior hippocampus (Moser et al.1998), and aging may differently affect anterior and posterior FC.
The mean saliency map from the whole hippocampal cortical FC based model (Fig. 1B) presents hotter (redder) regions with higher contribution of the connectivity to the prediction, e.g. precuneus, retrosplenial cortex, and occipital lobe. These areas align with known aging relevant regions. The precuneus is vulnerable to Alzheimer's disease (AD) (Zhang et al. 2021) and hippocampal-retrosplenial cortex FC relates to tau accumulation in the medial parietal region (Ziontz et al. 2021). The occipital lobe atrophy was also found in healthy old adults (Harrison et al. 2019).
Mean saliency maps from anterior and posterior hippocampal cortical FC models (Fig. 2A-B)) show distinct patterns. One-way t-tests identify significant differences (Fig. 2C-D, z-score>1.645). Specifically, FCs between the anterior hippocampus and lateral occipital cortex, precuneus, and medial prefrontal cortex were more influential on age prediction than their posterior hippocampus connections. Conversely, posterior hippocampal FC with medial occipital and posterior parahippocampal had a greater prediction effect. These disparities may due to different functional roles of the anterior and posterior hippocampus (Moser et al.1998), and aging may differently affect anterior and posterior FC.
Conclusions:
We trained a 3D CNN model that can predict age using cortico-hippocampal FC, with LayerCAM to aid interpretation. Key regions impacting predictions are consistent with brain regions known to be affected during aging. Further fine-grain analysis showed differences in the contributions from anterior and posterior hippocampal FC to age prediction, indicating our method's specificity to hippocampal subregions. Future efforts will aim to enhancing performance, validate results, and explore clinical applications.
Lifespan Development:
Aging 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural)
Keywords:
Aging
FUNCTIONAL MRI
Machine Learning
Other - Functional Connectivity
1|2Indicates the priority used for review
Provide references using author date format
Dennis, E. L. and P. M. Thompson (2014). "Functional brain connectivity using fMRI in aging and Alzheimer’s disease." Neuropsychology review 24: 49-62.
Glasser, M. F., S. N. Sotiropoulos, J. A. Wilson, T. S. Coalson, B. Fischl, J. L. Andersson, J. Xu, S. Jbabdi, M. Webster and J. R. Polimeni (2013). "The minimal preprocessing pipelines for the Human Connectome Project." Neuroimage 80: 105-124.
Harrison, T. M., R. La Joie, A. Maass, S. L. Baker, K. Swinnerton, L. Fenton, T. J. Mellinger, L. Edwards, J. Pham and B. L. Miller (2019). "Longitudinal tau accumulation and atrophy in aging and alzheimer disease." Annals of neurology 85(2): 229-240.
He, T., R. Kong, A. J. Holmes, M. Nguyen, M. R. Sabuncu, S. B. Eickhoff, D. Bzdok, J. Feng and B. T. Yeo (2020). "Deep neural networks and kernel regression achieve comparable accuracies for functional connectivity prediction of behavior and demographics." NeuroImage 206: 116276.
Jiang, P.-T., C.-B. Zhang, Q. Hou, M.-M. Cheng and Y. Wei (2021). "Layercam: Exploring hierarchical class activation maps for localization." IEEE Transactions on Image Processing 30: 5875-5888.
Moser, M. B. and E. I. Moser (1998). "Functional differentiation in the hippocampus." Hippocampus 8(6): 608-619.
Zhang, K., H. Mizuma, X. Zhang, K. Takahashi, C. Jin, F. Song, Y. Gao, Y. Kanayama, Y. Wu and Y. Li (2021). "PET imaging of neural activity, β-amyloid, and tau in normal brain aging." European Journal of Nuclear Medicine and Molecular Imaging: 1-13.
Ziontz, J., J. N. Adams, T. M. Harrison, S. L. Baker and W. J. Jagust (2021). "Hippocampal connectivity with retrosplenial cortex is linked to neocortical tau accumulation and memory function." Journal of Neuroscience 41(42): 8839-8847.