Related Experiment Video
Updated: Jul 13, 2025
05:23
Author Spotlight: Noninvasive Cerebral Blood Flow Determination in Human Functional Brain Region for Diagnosis of Neurological Disorders
Published on: May 31, 2024
575
Transformer-based deep learning denoising of single and multi-delay 3D arterial spin labeling
Qinyang Shou1, Chenyang Zhao1, Xingfeng Shao1
1Laboratory of Functional MRI Technology (LOFT), Stevens Neuro Imaging and Informatics Institute, University of Southern California, Los Angeles, California, USA.
Magnetic Resonance in Medicine
|October 18, 2023
Summary
SwinIR, a deep learning model, effectively denoises 3D arterial spin labeling (ASL) images, outperforming CNNs and other Transformers for improved clinical use.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- 3D Arterial Spin Labeling (ASL) is crucial for non-invasively measuring cerebral blood flow (CBF) and arterial transit time (ATT).
- Image denoising is essential for enhancing the quality and diagnostic accuracy of 3D ASL data.
- Existing denoising methods, including Convolutional Neural Networks (CNNs), face challenges in effectively removing noise from complex ASL data.
Purpose of the Study:
- To introduce SwinIR, a Swin Transformer-based deep learning model for denoising 3D ASL data.
- To compare the denoising performance of SwinIR against CNN-based and other Transformer-based methods for both single-delay and multi-delay ASL.
- To evaluate the impact of model architecture (2D vs. pseudo-3D) and input conditions on denoising efficacy and quantification accuracy.
Main Methods:
- Developed SwinIR and CNN-based spatial denoising models for single-delay 3D ASL.
- Trained and tested models on multi-vendor datasets comprising 66 subjects (119 scans) for single-delay and 6 subjects (10 scans) for multi-delay ASL.
- Evaluated performance using similarity metrics, signal-to-noise ratio (SNR), and quantification accuracy of CBF and ATT.
Main Results:
- SwinIR demonstrated superior denoising performance compared to CNNs and other Transformer models for both single and multi-delay 3D ASL.
- Pseudo-3D models outperformed 2D models, with three slices offering an optimal balance between SNR and quantification accuracy across vendors.
- Spatiotemporal denoising models for multi-delay ASL reduced biases in CBF and ATT quantification compared to spatial-only approaches.
Conclusions:
- SwinIR offers a robust and flexible deep learning solution for denoising 3D ASL data.
- The model significantly improves image quality and has the potential to reduce scan times, facilitating wider clinical adoption of 3D ASL.
- SwinIR's performance surpasses existing CNN and Transformer-based methods, paving the way for enhanced neuroimaging applications.
