[Paper Review] Predicting Alzheimer's Disease Using 3DMgNet

Yelu Gao, Huang Huang+1 more|arXiv (Cornell University)|Jan 12, 2022

AI in cancer detection4 citations

TL;DR

This paper proposes 3DMgNet, a novel 3D multigrid convolutional neural network architecture that integrates multigrid processing with deep learning to improve early diagnosis of Alzheimer's disease (AD) from MRI scans. Trained on the ADNI dataset and tested on a private dataset, the model achieves 92.133% accuracy in classifying AD versus normal control (NC), while significantly reducing model parameters.

ABSTRACT

Alzheimer's disease (AD) is an irreversible neurode generative disease of the brain.The disease may causes memory loss, difficulty communicating and disorientation. For the diagnosis of Alzheimer's disease, a series of scales are often needed to evaluate the diagnosis clinically, which not only increases the workload of doctors, but also makes the results of diagnosis highly subjective. Therefore, for Alzheimer's disease, imaging means to find early diagnostic markers has become a top priority. In this paper, we propose a novel 3DMgNet architecture which is a unified framework of multigrid and convolutional neural network to diagnose Alzheimer's disease (AD). The model is trained using an open dataset (ADNI dataset) and then test with a smaller dataset of ours. Finally, the model achieved 92.133% accuracy for AD vs NC classification and significantly reduced the model parameters.

Motivation & Objective

• To develop a unified deep learning framework that enhances early diagnosis of Alzheimer’s disease using neuroimaging data.
• To reduce the subjectivity and workload associated with clinical diagnostic scales for AD by automating classification using MRI scans.
• To design a parameter-efficient architecture that maintains high performance in AD vs. NC classification.
• To leverage multigrid processing within a 3D convolutional neural network to improve feature extraction from volumetric brain MRI data.
• To validate the model on both open (ADNI) and private datasets for robustness and generalization.

Proposed method

• The proposed 3DMgNet integrates multigrid techniques with 3D convolutional neural networks to process volumetric MRI data efficiently.
• The architecture uses a hierarchy of grids at multiple resolutions to capture both local and global brain structure patterns.
• Skip connections are incorporated to facilitate feature flow and gradient propagation across different scales.
• The model is trained end-to-end on the ADNI dataset using standard cross-entropy loss for binary classification (AD vs. NC).
• A smaller, private test dataset is used to evaluate generalization performance and avoid data leakage.
• Model efficiency is improved by reducing the number of parameters through multigrid-based feature learning.

Experimental results

Research questions

• RQ1 Can a 3D multigrid convolutional neural network architecture improve the accuracy of Alzheimer’s disease classification from MRI scans?
• RQ2 To what extent does integrating multigrid processing enhance feature representation in 3D brain MRI data?
• RQ3 How does 3DMgNet compare in performance and parameter efficiency to standard 3D CNNs in AD vs. NC classification?
• RQ4 Does the model generalize well to independent, private datasets beyond the ADNI training data?
• RQ5 Can the proposed architecture reduce the need for subjective clinical scales in early AD diagnosis?

Key findings

• The 3DMgNet model achieved a classification accuracy of 92.133% in distinguishing Alzheimer’s disease from normal control subjects on the test dataset.
• The model significantly reduced the number of parameters compared to standard 3D CNNs, improving computational efficiency.
• The integration of multigrid processing enhanced feature learning across multiple spatial scales in 3D MRI volumes.
• The use of skip connections contributed to stable training and improved performance.
• The model demonstrated strong generalization capability when evaluated on a private, independent test dataset.
• The results suggest that 3DMgNet can serve as a reliable, automated tool for early AD diagnosis, reducing reliance on subjective clinical assessments.