Related Experiment Video
Updated: May 4, 2026
09:47
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
2.0K
Latent feature representation with stacked auto-encoder for AD/MCI diagnosis
Heung-Il Suk1, Seong-Whan Lee, Dinggang Shen
1Biomedical Research Imaging Center (BRIC) and Department of Radiology, University of North Carolina, Chapel Hill, NC, 27599, USA, hsuk@med.unc.edu.
Brain Structure & Function
|December 24, 2013
Summary
This study introduces a deep learning approach using stacked auto-encoders for diagnosing Alzheimer's disease (AD) and mild cognitive impairment (MCI). The method achieves high accuracy by analyzing complex patterns in neuroimaging data.
Area of Science:
- Neuroimaging analysis
- Artificial intelligence in medicine
- Biomedical data science
Background:
- Alzheimer's disease (AD) and mild cognitive impairment (MCI) diagnosis relies on identifying subtle patterns in neuroimaging data.
- Traditional methods often use simple low-level features, potentially missing complex relationships.
- Deep learning offers a promising avenue for uncovering intricate patterns in medical data.
Purpose of the Study:
- To develop a computer-aided diagnosis system for AD and MCI using deep learning.
- To leverage latent feature representation with stacked auto-encoders (SAE) for enhanced classification accuracy.
- To explore the effectiveness of SAE in identifying complex, non-linear patterns within neuroimaging features.
Main Methods:
- Utilized a stacked auto-encoder (SAE) for unsupervised pre-training on target-unrelated samples to initialize parameters.
- Fine-tuned the SAE model using target-related samples, combining latent features with original low-level features (MRI, PET).
- Evaluated the model on four binary classification tasks: AD vs. healthy normal control (HC), MCI vs. HC, AD vs. MCI, and MCI converter (MCI-C) vs. MCI non-converter (MCI-NC).
Main Results:
- Achieved high diagnostic accuracies on the ADNI dataset: 98.8% for AD/HC, 90.7% for MCI/HC, 83.7% for AD/MCI, and 83.3% for MCI-C/MCI-NC.
- Demonstrated the effectiveness of combining latent features with original features for robust classification.
- Validated the deep learning approach's superior performance over traditional methods relying on simple features.
Conclusions:
- Deep learning, specifically SAE, offers a powerful tool for analyzing complex patterns in neuroimaging data for brain disease diagnosis.
- The proposed method significantly enhances diagnostic accuracy for Alzheimer's disease and mild cognitive impairment.
- This work highlights the potential of deep learning in advancing neuroimaging data analysis for clinical applications.