Multi-Modal Neuroimaging Neural Network-Based Feature Detection for Diagnosis of Alzheimer’s Disease

Xianglian Meng, Junlong Liu, Xiang Fan, Chenyuan Bian, Qingpeng Wei, Ziwei Wang, Wenjie Liu, Zhuqing JiaoFrontiers in Aging Neuroscience 2022 · DOI

TL;DR

This paper introduces a novel multi-modal LassoNet framework that integrates resting-state fMRI and diffusion tensor imaging (DTI) data to improve the accuracy of Alzheimer's disease (AD) diagnosis. By combining structural and functional brain network information, their method achieves superior classification performance compared to single-modality or other multi-modal approaches, offering a promising tool for early AD detection and understanding its underlying mechanisms.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is a devastating neurodegenerative condition that poses a significant diagnostic challenge, especially in its early stages (mild cognitive impairment, MCI). Current diagnostic methods often rely on single imaging modalities, which can be insufficient to capture the complex, multi-faceted nature of AD's impact on the brain. Single-modality approaches, like those using only fMRI or DTI, often lack the comprehensive spatiotemporal resolution needed to fully understand brain network alterations. This limitation hinders the ability to accurately distinguish between healthy controls (HC), individuals with MCI, and those with AD, and to pinpoint the specific brain regions most affected. The paper aims to bridge this gap by developing a more robust and accurate diagnostic tool through the fusion of multiple neuroimaging data types.

Methods

The core of this research is the development of a Multi-Modal LassoNet framework. This approach combines the strengths of two neuroimaging techniques:

  • Resting-State Functional Magnetic Resonance Imaging (rs-fMRI): This modality captures the brain's functional connectivity by measuring temporal correlations in blood-oxygen-level-dependent (BOLD) signals between different brain regions.
  • Diffusion Tensor Imaging (DTI): This technique maps the structural connectivity of the brain by tracking the diffusion of water molecules along white matter tracts.

The proposed framework integrates these two data sources within a LassoNet architecture. LassoNet is a type of neural network that incorporates Lasso regression, a penalized regression method known for its ability to perform feature selection by shrinking the coefficients of less important features to zero. In this multi-modal context, LassoNet is used for:

  1. Feature Extraction: Extracting relevant information from both rs-fMRI and DTI data.
  2. Feature Selection: Identifying the most discriminative brain regions and connections associated with AD. The DTI data is used to create a "penalty matrix" that helps to regularize and guide the feature selection process, leveraging the coupling between structural and functional brain networks.
  3. Classification: Building a model to classify subjects into HC, MCI, or AD categories.

The framework also incorporates a residual connection and an arbitrary feed-forward neural network to enhance its modeling capabilities. The authors also compare their method against other popular regularization techniques like Group Lasso, Sparse Group Lasso, and ElasticNet to demonstrate the superiority of their multi-modal LassoNet approach.

Datasets

The study utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. This publicly available dataset provided neuroimaging data for:

  • 33 Healthy Controls (HC)
  • 29 Early Mild Cognitive Impairment (EMCI) subjects
  • 23 Alzheimer's Disease (AD) patients

A total of 85 subjects' rs-fMRI and DTI data were used. The paper does not introduce a new benchmark dataset but rather leverages the existing ADNI data for validation.

Key Insights and Ideas

  • Synergy of Modalities: The fusion of rs-fMRI and DTI data significantly enhances diagnostic accuracy compared to using either modality alone. This highlights that AD affects both the functional communication and structural integrity of the brain, and capturing both is crucial.
  • LassoNet for Feature Selection: The LassoNet framework, with its built-in feature selection capabilities, is highly effective in identifying the most relevant brain regions for AD diagnosis, mitigating the "curse of dimensionality" often encountered with high-dimensional neuroimaging data.
  • Structural-Functional Coupling: The incorporation of DTI-derived structural information as a penalty matrix in the LassoNet model effectively leverages the relationship between structural connectivity and functional activity, leading to a more robust model.
  • Identification of Key Brain Regions: The study successfully identified specific brain regions (e.g., Hippocampus, Frontal_Inf_Orb_L, Parietal_Sup_L, Putamen_L, Fusiform_R) that are discriminative across different diagnostic groups, providing insights into the pathological mechanisms of AD.
  • Early Diagnosis Potential: The framework's strong performance, particularly in distinguishing between AD and HC, suggests its potential for earlier and more accurate diagnosis of AD and its precursor stages.

Key Results

The Multi-Modal LassoNet framework demonstrated superior performance across all three comparison groups (AD-HC, AD-EMCI, EMCI-HC):

  • AD-HC Group: Achieved the highest classification accuracy (ACC) of 90.68% ± 0.34%, with excellent sensitivity (SEN) of 88.81% ± 0.68% and specificity (SPE) of 91.91% ± 0.55%.
  • AD-EMCI Group: Showed an ACC of 83.63% ± 0.74%, with SEN of 87.32% ± 1.22% and SPE of 79.00% ± 1.33%.
  • EMCI-HC Group: Achieved an ACC of 88.77% ± 0.70%, with SEN of 90.87% ± 1.05% and SPE of 87.06% ± 0.95%.

In all groups, the Multi-Modal LassoNet consistently outperformed other methods (Lasso+SVM, GroupLasso+SVM, Sparse Group Lasso+SVM, ElasticNet+SVM) in terms of ACC, SEN, SPE, GMean, and F1 scores. The Receiver Operating Characteristic (ROC) curves also showed significantly higher Area Under the Curve (AUC) values for the proposed method, with AUCs of 0.9120 (AD-HC), 0.8478 (AD-EMCI), and 0.8975 (EMCI-HC). The study also noted that the gap in accuracy between experiments for the Multi-Modal LassoNet was less than 2%, indicating good stability.

Limitations and Future Work

  • Segmentation Template Sensitivity: The authors acknowledge that brain network analysis results can be sensitive to the choice of segmentation template. They used the AAL atlas (90 regions) but suggest that future work should verify the robustness of their method across different segmentation scales.
  • Fiber Tracking Limitations: The deterministic fiber tracking technique used for DTI analysis has limitations in accurately capturing crossing or diverging white matter fibers. Future research could benefit from more advanced probabilistic fiber tract-tracing strategies to improve structural network construction accuracy.
  • Generalizability: While the study used a well-established dataset (ADNI), further validation on independent datasets would be beneficial to confirm the generalizability of the proposed framework.
  • Multi-Modal Integration: The current framework integrates two modalities (rs-fMRI and DTI). Future work could explore the integration of additional modalities (e.g., PET, structural MRI) to further enhance diagnostic capabilities.
  • Understanding Mechanisms: While the study identifies discriminative regions, further research is needed to fully elucidate the dynamic mechanisms underlying AD development and how these identified regions contribute to the disease process.

Summary updated

Abstract

Alzheimer's disease (AD) is a neurodegenerative brain disease, and it is challenging to mine features that distinguish AD and healthy control (HC) from multiple datasets. Brain network modeling technology in AD using single-modal images often lacks supplementary information regarding multi-source resolution and has poor spatiotemporal sensitivity. In this study, we proposed a novel multi-modal LassoNet framework with a neural network for AD-related feature detection and classification. Specifically, data including two modalities of resting-state functional magnetic resonance imaging (rs-fMRI) and diffusion tensor imaging (DTI) were adopted for predicting pathological brain areas related to AD. The results of 10 repeated experiments and validation experiments in three groups prove that our proposed framework outperforms well in classification performance, generalization, and reproducibility. Also, we found discriminative brain regions, such as Hippocampus, Frontal_Inf_Orb_L, Parietal_Sup_L, Putamen_L, Fusiform_R, etc. These discoveries provide a novel method for AD research, and the experimental study demonstrates that the framework will further improve our understanding of the mechanisms underlying the development of AD.
BibTeX
@inproceedings{meng2022multi,
  author = {Xianglian Meng and Junlong Liu and Xiang Fan and Chenyuan Bian and Qingpeng Wei and Ziwei Wang and Wenjie Liu and Zhuqing Jiao},
  title = {{Multi-Modal Neuroimaging Neural Network-Based Feature Detection for Diagnosis of Alzheimer’s Disease}},
  booktitle = {Frontiers in Aging Neuroscience},
  year = {2022},
  doi = {10.3389/fnagi.2022.911220},
}
Figure