Classification of the stages of Alzheimer's disease based on three-dimensional lightweight neural networks

Jun Li, Juntong Liu, Yang Su, Jie Chang, Mingquan YePeerJ Comput. Sci 2025 · DOI

TL;DR

This paper introduces a novel approach for classifying Alzheimer's disease (AD) stages using lightweight 3D convolutional neural networks (CNNs) trained on different types of preprocessed MRI data. The study achieves high accuracy (over 96%) and demonstrates that the choice of input data (standardized vs. segmented gray/white matter) significantly impacts classification performance across different disease stages, offering insights into AD progression.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is a devastating neurodegenerative disorder that significantly impacts the elderly. Early and accurate diagnosis is crucial for managing the disease, as timely interventions can slow its progression. While Magnetic Resonance Imaging (MRI) is a powerful tool for detecting structural brain changes associated with AD, effectively classifying the different stages of the disease (cognitively normal (CN), mild cognitive impairment (MCI), and AD) remains a challenge. Existing methods often rely on 2D imaging or less efficient 3D models, and a deeper understanding of how specific brain tissue changes (gray matter vs. white matter) correlate with disease progression is still needed. This paper aims to address these limitations by developing and evaluating lightweight 3D CNNs for multi-stage AD classification and analyzing the impact of different MRI data representations on performance.

Methods

The researchers employed a suite of lightweight 3D neural networks: 3D-SqueezeNet, 3D-MobileNetV1, and 3D-ShuffleNetV1, along with a 3D version of ResNet18 as a reference. These networks were chosen for their efficiency, characterized by shallower depths and fewer parameters, which helps mitigate overfitting, especially with smaller datasets.

The core of their methodology involved preprocessing T1-weighted MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset using the CAT12 toolbox. This preprocessing pipeline included:

  • Spatial Normalization: Aligning individual brain scans to a standard anatomical template (MNI).
  • Bias Field Correction: Correcting for intensity variations caused by magnetic field inhomogeneities.
  • Tissue Classification: Segmenting the brain into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) based on probabilistic models.
  • Modulation: Adjusting segmented tissue maps to preserve original tissue volumes after spatial normalization.

These preprocessing steps resulted in three types of input images for the neural networks:

  1. Standardized Images (wm): The overall warped and modulated MRI scans.
  2. Segmented Standardized Gray Matter Images (mwp1): Images representing only the gray matter.
  3. Segmented Standardized White Matter Images (mwp2): Images representing only the white matter.

The study then trained each of the four 3D neural network architectures on each of these three image types, evaluating their performance on a held-out test set.

Datasets

The study utilized the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Specifically, they used T1-weighted Magnetization Prepared Rapid Acquisition Gradient Echo (MPRAGE) MRI scans. The dataset comprised 76.34 ± 7.07 years old individuals, with a gender distribution of 1,896 males and 1,627 females. The AD stages were categorized into Cognitively Normal (CN) with 1,530 individuals, Mild Cognitive Impairment (MCI) with 1,275 individuals, and Alzheimer's Disease (AD) with 718 individuals. The data was randomly split into an 80% training set and a 20% testing set, maintaining the original class proportions.

Key Insights and Ideas

  • Input Data Matters: The choice of input data (standardized, segmented GM, or segmented WM) significantly influences the classification performance of the neural networks across different AD stages. This highlights that different brain tissues are affected differently and at different rates during AD progression.
  • GM Dominance in Early Stages: At the cognitively normal (CN) stage, gray matter (GM) features are more prominent and informative for classification. This suggests that early AD-related changes primarily manifest in neuronal structures.
  • WM Becomes Crucial in Later Stages: As the disease progresses to MCI and AD, white matter (WM) features, representing neuronal connections, become increasingly significant. This indicates a shift in the pathological process from neuronal damage to the disruption of neural pathways.
  • Standardized Images Offer a Balanced View: For classifying MCI and AD, using the overall standardized images (wm) often yielded the best results. This suggests that a combination of GM and WM information is most effective for capturing the more complex changes in later disease stages.
  • Lightweight Networks are Effective: The study demonstrates that lightweight 3D CNNs can achieve high classification accuracy for AD stages, proving that complex, deep architectures are not always necessary for this task, especially when dealing with limited data.

Key Results

  • High Classification Accuracy: All trained neural networks achieved an accuracy of at least 96% on the testing set for classifying the three AD stages (CN, MCI, AD).
  • Excellent Precision and Recall: The precisions and recalls for classifying AD stages were above 94% for all networks and input types.
  • Best Performance for CN: For classifying cognitively normal individuals, the segmented standardized gray matter images (mwp1) generally led to the best classification results.
  • Best Performance for MCI and AD: For classifying mild cognitive impairment (MCI) and Alzheimer's disease (AD), the standardized images (wm) often provided the best classification results.
  • Training Speed: Neural networks trained with segmented gray matter images (mwp1) converged fastest, followed by segmented white matter images (mwp2), and then the standardized images (wm).

Limitations and Future Work

  • Dataset Size: The authors acknowledge that the limitation of the dataset size might have influenced the qualitative summary of AD's impact on GM and WM.
  • Future Directions: They suggest that broader datasets and more advanced neural network architectures should be applied to further explore the impact of AD development on brain tissue. Additionally, the study's findings on the differential impact of GM and WM changes across AD stages could inform future research on targeted interventions and diagnostic markers.

Summary updated

BibTeX
@article{li2025classification,
  author = {Jun Li and Juntong Liu and Yang Su and Jie Chang and Mingquan Ye},
  title = {{Classification of the stages of Alzheimer's disease based on three-dimensional lightweight neural networks}},
  journal = {PeerJ Comput. Sci.},
  year = {2025},
  doi = {10.7717/peerj-cs.2897},
}
Figure