Independent VIP: Unveiling spatial patterns in resting-state functional MRI networks

1910 

Abstract Submission 

Vaibhavi Itkyal¹, Theodore LaGrow², Anees Abrol³, Vince Calhoun⁴

¹Emory University, Decatur, GA, ²Georgia Institute of Technology, Beaverton, OR, ³Georgia State University, Atlanta, GA, ⁴GSU/GATech/Emory, Decatur, GA

Vaibhavi Itkyal  
Emory University
Decatur, GA

Theodore LaGrow  
Georgia Institute of Technology
Beaverton, OR

Anees Abrol  
Georgia State University
Atlanta, GA

Vince Calhoun  
GSU/GATech/Emory
Decatur, GA

This study addresses the imperative need for accurate early diagnosis of Alzheimer's Disease (AD) by exploring the synergies between structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI). The novel approach integrates rs-fMRI networks which are computed using independent component analysis followed by voxelwise intensity projections (i.e., iVIP), with sMRI, surpassing traditional metrics like amplitude of low-frequency fluctuations (ALFF) and fractional ALFF (fALFF). Inspired by AlexNet, a multi-channel convolutional neural network effectively captures both spatial and temporal dependencies, achieving a 93.31% test accuracy and a 97.79 AUC score on the classification task of AD vs cognitively normal (CN) from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Our multimodal deep learning results outperform unimodal approaches, highlighting crucial differences in neurobiologically relevant regions. Saliency visualizations emphasize distinctions in the hippocampus, amygdala, caudate nucleus, and thalamus, aligning with existing literature. This innovative multimodal deep learning model validated on publicly available data demonstrates superior diagnostic performance, offering valuable insights into AD-related alterations through spatiotemporal information integration.

We used ADNI data, comprising sMRI and rs-fMRI scans from 466 subjects, including 383 CN and 83 AD individuals. The rigorous preprocessing includes spatial normalization, tissue segmentation, and transformations using SPM 12, with stringent quality control measures applied. The fMRI preprocessing involves motion correction, slice-timing correction, standardization to MNI space, resampling, and Gaussian smoothing. All pre-processing information for this dataset can be found in Du et al. 2020. Spatially constrained independent component analysis generated 53 intrinsic connectivity networks (ICNs) using Neuromark (Du et al. 2020), from which independent voxelwise intensity projection images (iVIPs) are computed. These iVIPs, representing max, abs min, and max abs of ICNs, serve as input for a specialized 3D multimodal deep learning architecture inspired by AlexNet (Abrol et al. 2021; Krizhevsky, Sutskever, and Hinton 2017). The model, operating on rs-fMRI or sMRI data, captures spatial patterns through convolutional layers, adaptive pooling, and dropout for classification. The study achieves a 93.31% test accuracy in discriminating AD and CN, surpassing traditional metrics like ALFF (Turner et al. 2013) and fALFF. Training and saliency maps, generated through guided backpropagation, provide insights into the model's decision-making processes. We did 8-fold cross-validation and used Adam algorithm.

iVIP compared to traditional metrics like ALFF/fALFF, iVIP exhibits quantitative superiority in test accuracy and balanced accuracy, particularly excelling in two-way classification with max(ICN) and max(abs(ICN)). Deep learning models based on iVIP outperform those derived from ALFF/fALFF, achieving an 85.02% peak test accuracy for AD vs CN. Combining sMRI and iVIP in a multimodal model (sMRI + iVIP) achieves the highest accuracy at 93.31%, surpassing unimodal models. Saliency maps highlight distinct spatial patterns, with sMRI emphasizing the hippocampus, amygdala, and caudate nucleus, while max(ICN) reveals variations in the cingulate cortex, thalamus, and caudate nucleus. These findings provide valuable insights into AD-related brain alterations, emphasizing the synergistic benefits of multimodal approaches in neuroimaging studies (Abbott et al. 2014; Wang et al. 2016; Lee et al. 2020).

The study integrates iVIP with sMRI into a 3D CNN, achieving 93.31% accuracy for early AD diagnosis, surpassing traditional metrics - ALFF/fALFF. Our multimodal approach provides insights into neurobiological changes, emphasizing its potential for effective clinical strategies in AD diagnosis and intervention.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)

Connectivity (eg. functional, effective, structural)

Methods Development ¹

Task-Independent and Resting-State Analysis ²

Data analysis

FUNCTIONAL MRI

STRUCTURAL MRI