Related Experiment Video
Updated: Dec 18, 2025
09:47
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
1.6K
Generalization of deep learning models for ultra-low-count amyloid PET/MRI using transfer learning
Kevin T Chen1, Matti Schürer2, Jiahong Ouyang3
1Department of Radiology, Stanford University, Stanford, CA, United States. ktchen@stanford.edu.
Summary
Deep learning successfully creates diagnostic amyloid PET/MRI images from ultra-low-count data, improving image quality and accuracy. Transfer learning (method B) showed the best performance, but data bias must be considered for network generalization.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Amyloid Positron Emission Tomography/Magnetic Resonance Imaging (PET/MRI) is crucial for diagnosing Alzheimer's disease.
- Acquisition of high-quality PET/MRI scans can be limited by scan time and radiation dose.
- Deep learning offers potential for enhancing low-quality imaging data.
Purpose of the Study:
- To evaluate the performance of deep learning models for generalizing ultra-low-count amyloid PET/MRI enhancement.
- To assess the impact of different scanning hardware, protocols, and transfer learning strategies on image quality and diagnostic accuracy.
Main Methods:
- Eighty simultaneous [18F]florbetaben PET/MRI studies were acquired across two sites with differing hardware and protocols.
- Ultra-low-count PET data (1% dose/5% time) were reconstructed and enhanced using a deep convolution neural network.
- Network performance was compared using direct application (A), transfer learning (B), training from scratch on site-specific data (C), or all data (D).
Main Results:
- Network-synthesized images exhibited reduced noise compared to ultra-low-count reconstructions.
- Transfer learning (method B) yielded the best quantitative metrics, including standardized uptake value ratios (SUVRs) with minimal variability and high diagnostic effect size.
- Methods B-D produced images scoring similarly or better than ground-truth, while method A showed lower accuracy and image quality.
Conclusions:
- Deep learning effectively generates diagnostic amyloid PET images from short-frame reconstructions, significantly reducing noise and improving quality.
- Transfer learning enhances generalization across different scanner types and protocols, but careful consideration of data bias is essential for reliable application.
