Related Experiment Video
Updated: Jun 1, 2025
09:47
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
947
Adapting to evolving MRI data: A transfer learning approach for Alzheimer's disease prediction
Rosanna Turrisi1, Sarthak Pati2, Giovanni Pioggia3
1Institute for Biomedical Research and Innovation (IRIB), National Research Council of Italy (CNR), Messina, Italy; Machine Learning Genoa Center (MaLGa), University of Genoa, Genoa, Italy.
Neuroimage
|January 18, 2025
Summary
Transfer learning significantly improves Alzheimer's Disease (AD) detection using 3D MRI scans. This approach enhances diagnostic accuracy, even with limited or varied imaging data, by adapting existing models.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Integrating 3D magnetic resonance imaging (MRI) with machine learning shows promise for Alzheimer's Disease (AD) detection.
- Limited data due to evolving MRI protocols can lead to model overfitting in AD diagnosis.
Purpose of the Study:
- To explore Transfer Learning (TL) strategies for improving AD diagnosis accuracy.
- To address challenges of limited data and changing MRI acquisition protocols.
Main Methods:
- A Baseline 3D-Convolutional Neural Network model was trained on 3T MRI scans.
- Two TL scenarios were investigated: (A) adapting historical 1.5T to 3T MRI data, and (B) adapting 2D models pre-trained on ImageNet for 3D MRI.
- Both General (feature extraction + classification) and Deep (model fine-tuning) approaches were tested.
Main Results:
- In scenario (A), TL boosted accuracy from 63% to 99%.
- In scenario (B), fine-tuning natural image models improved baseline accuracy by up to 12 percentage points (83% overall).
- Radiomic features outperformed TL features in the General Approach for scenario (B).
Conclusions:
- Transfer learning is highly effective for enhancing AD diagnosis with 3D MRI, particularly when dealing with data variations.
- Fine-tuning pre-trained models offers a viable strategy for AD detection when historical data is unavailable.
