Diffusion MRI-based Deep Learning to detect Alzheimer’s Disease in US and Indian Cohorts
Poster No:
221
Submission Type:
Abstract Submission
Authors:
Tamoghna Chattopadhyay1, Neha Joshy1, Yixue Feng2, Julio Villalón-Reina1, Himanshu Joshi3, Ganesan Venkatasubramanian4, John John5, Paul Thompson6
Institutions:
1University of Southern California, Los Angeles, CA, 2University of Southern California, Marina Del Rey, CA, 3Multimodal Brain Image Analysis Laboratory, NIMHANS, Bangalore, Karnataka, 4National Institute of Mental Health & Neurosciences (NIMHANS), Bangalore, India, 5National Institute of Mental Health and Neuro Sciences, Bangalore, Karnataka, 6USC, Marina Del Rey, CA
First Author:
Co-Author(s):
Ganesan Venkatasubramanian, MD, PhD
National Institute of Mental Health & Neurosciences (NIMHANS)
Bangalore, India
National Institute of Mental Health & Neurosciences (NIMHANS)
Bangalore, India
Introduction:
According to the World Health Organization (WHO), around 55 million people worldwide have dementia, and 60-70% of these patients have Alzheimer's disease (AD). The recent advent of new anti-amyloid therapies makes early, accurate AD diagnosis crucial. Automated disease classifiers that analyze vast imaging databases would also help in discovering genetic or environmental factors affecting disease onset and progression. One novel approach to disease classification involves deep learning models, such as convolutional neural networks (CNNs), which can directly analyze raw or minimally processed images, avoiding the lengthy quality control required for traditional, parcellation-based brain morphometry. Here we trained CNNs to detect AD based on diffusion MRI (dMRI), which is sensitive to brain microstructure changes not visible on standard anatomical MRI, based on prior work linking dMRI metrics to age, dementia severity, and brain amyloid levels, a key component of AD pathology [2,3]. Existing CNNs to detect AD mostly use T1-weighted brain MRI data from European or North American cohorts. To address this limitation, we tested our dMRI-based AD classifier in both Indian and North American cohorts.
Methods:
We analyzed two datasets: (1) the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset with 1,195 participants (age: 74.36+/-7.74 years; 600F/595M; 633 healthy controls, CN, 421 with mild cognitive impairment, MCI, and 141 with AD, and (2) the NIMHANS cohort from Bengaluru, India, comprising 301 participants (age: 67.23+/-7.86 years; 169F/132M; 123 CN/88 MCI/90 AD). Preprocessing steps included N4 bias field correction, brain extraction, 6 degree-of-freedom registration to a template with, and resampling to 2 mm isotropic voxels. The T1-weighted (T1w) images were scaled to an intensity range of 0 and 1, aligned to a common template. The DWI were nonlinearly registered to T1w and warped to a common template. The dMRI processing pipeline details may be found in [2, 3].
The 3D CNN architecture is shown in Fig. 1. Training was conducted for 100 epochs, with batch size of 8, an exponentially decaying learning rate of 0.96, Adam optimizer, and mean square error loss function. Dropout and early stopping were used to prevent overfitting. Images were split into independent training, validation, and testing sets (in a 70:20:10 ratio). In the architecture (Fig. 2) after flattening, the layers were concatenated and sent through a dense layer with sigmoid activation function. This Y-shaped architecture merged predictive features distilled from T1w MRI and DTI maps for disease classification, while maintaining previous training parameters.
The 3D CNN architecture is shown in Fig. 1. Training was conducted for 100 epochs, with batch size of 8, an exponentially decaying learning rate of 0.96, Adam optimizer, and mean square error loss function. Dropout and early stopping were used to prevent overfitting. Images were split into independent training, validation, and testing sets (in a 70:20:10 ratio). In the architecture (Fig. 2) after flattening, the layers were concatenated and sent through a dense layer with sigmoid activation function. This Y-shaped architecture merged predictive features distilled from T1w MRI and DTI maps for disease classification, while maintaining previous training parameters.
·Model Architecture
Results:
Overall, DTI-derived metrics performed better on the classification task, with higher balanced accuracy and F1 Score, compared to T1w MRIs. The best balanced accuracy was obtained for DTI-RD maps at 0.896, with F1 Score 0.870. Combining T1w and DTI-MD and DTI-AD for the dual modality experiments gave best results compared to the other two combinations. In most cases, balanced accuracy was higher when T1w and dMRI were combined, relative to using T1 alone.
·Results Table
Conclusions:
We trained 3D CNNs on both diffusion MRI and standard T1w MRI to classify individuals as AD patients vs healthy controls. We tested DTI maps as inputs and found that they outperformed T1w MRI. We evaluated different combinations of maps, but multimodal training did not always work best, because it increases the number of trainable parameters, requiring more data to stabilize the model. Additionally, we evaluated these models on both Indian and North American cohorts, with comparable performance on both. Future work will train methods on larger, more diverse datasets, examining various AD stages and other dementia subtypes. We will also evaluate the added value of quantitative parametric MRI, DAT-SPECT, and resting state fMRI, to enhance AD classification. A robust AD classifier may accelerate the discovery of risk factors for AD in the genome or environment.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Novel Imaging Acquisition Methods:
Diffusion MRI
Keywords:
Cognition
Computational Neuroscience
Data analysis
Machine Learning
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Provide references using author date format
[2] Zavaliangos A., et al., “Diffusion MRI Indices and Their Relation to Cognitive Impairment in Brain Aging: The Updated Multi-Protocol Approach in ADNI3, “Front Neuroinformatics 13:2 (2019).
[3] Thomopoulos S., et al., “Diffusion MRI Metrics and their relation to Dementia Severity: Effect of Harmonization Approaches,” medRxiv (2021).
[4] Lu, B., et al., (2022). A practical Alzheimer’s disease classifier via brain imaging-based deep learning on 85,721 samples. Journal of Big Data, 9(1), Article 101.
[5] Wang D., et al., “Application of multimodal MR imaging on studying Alzheimer's disease: a survey,” Curr. Alzheimer Res. 877-92 (2013).
[6] Knudsen L., et al., “The role of multimodal MRI in mild cognitive impairment and Alzheimer's disease,” J Neuroimaging 148-157 (2022).
[7] Chattopadhyay, T., et al. (2023, March). Predicting dementia severity by merging anatomical and diffusion MRI with deep 3D convolutional neural networks. In 18th International Symposium on Medical Information Processing and Analysis (SIPAIM; Vol. 12567, pp. 90-99). SPIE.
[8] Lam, P., et al., “3-D Grid-Attention Networks for Interpretable Age and Alzheimer’s Disease Prediction from Structural MRI,” arXiv (2020).
[9] Gupta, U., et al. (2023, April). Transferring Models Trained on Natural Images to 3D MRI via Position Encoded Slice Models. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) (pp. 1-5). IEEE.
[10] Nir T., et al., “Fractional anisotropy derived from the diffusion tensor distribution function boosts power to detect Alzheimer's disease deficits, “Magn Reson Med. 78(6):2322-2333 (2017).