TL;DR
This paper investigates how to optimize deep learning (DL) models for denoising Arterial Spin Labeling (ASL) MRI images. They found that averaging ASL images before DL processing and using a windowed averaging strategy significantly improves image quality and can reduce scan time, offering practical guidance for clinical applications.
What Problem Does This Paper Address?
Arterial Spin Labeling (ASL) is a powerful MRI technique for measuring blood flow in the brain. However, ASL images inherently suffer from low signal-to-noise ratio (SNR), making them noisy and difficult to interpret. To combat this, researchers often average multiple ASL scans to improve SNR, but this significantly increases scan time. Deep learning (DL) has emerged as a promising tool for denoising these images, but the optimal way to integrate DL with ASL acquisition and processing pipelines (specifically, how to best use averaging strategies during training and inference) has not been systematically explored. This paper aims to fill that gap by investigating how different averaging and training strategies impact DL-based ASL denoising performance.
Methods
The researchers developed a deep learning pipeline for ASL denoising using a 3D U-Net encoder-decoder architecture. They systematically explored several key aspects:
- Averaging Strategies: They compared two pre-averaging methods:
- Windowed Averaging: Consecutive time points within a defined window are averaged.
-
Interleaved Averaging: Time points are selected at regular intervals and averaged. They also investigated averaging before DL processing (pre-averaging) versus after DL processing (post-averaging).
-
Training Strategies:
- Inclusion of Calibration Scans (M₀): They tested models trained with and without the M₀ scan, which provides anatomical reference.
- SNR Matching: They examined the importance of matching the SNR levels of training and testing data.
- Generalizability: They evaluated if models trained on lower-quality data (e.g., 25% or 50% of time points) could effectively denoise higher-quality data.
- Model Architecture: A 3D U-Net was used, and its performance was also validated with a pseudo-3D SwinIR transformer architecture.
- Performance Metrics: Image quality was assessed using Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Normalized Mean Absolute Error (NMAE).
Datasets
The study utilized a diverse dataset comprising 152 single-delay ASL scans from 152 subjects. These scans were sourced from publicly available datasets on OpenNeuro (6 datasets) and an additional dataset from healthy subjects and epilepsy patients. The data included both pseudo-continuous ASL (PCASL) and pulsed ASL (PASL) acquisitions, acquired with various scanners and protocols, ensuring a broad representation of ASL data.
Key Insights and Ideas
- Pre-averaging is King: Averaging ASL images before feeding them into the DL model yielded significantly better results than averaging after denoising. This suggests that reducing noise at the input stage is more effective.
- Windowed Averaging Wins: Windowed averaging consistently outperformed interleaved averaging, indicating that focusing on contiguous time points is more beneficial for DL denoising. This is particularly relevant for scan time reduction.
- M₀ is Generally Helpful: Including the M₀ scan (a measure of tissue magnetization) in the DL model training was almost always beneficial, especially when the input ASL images had lower SNR. It acts as a crucial anatomical reference.
- SNR Matching Matters: Optimal DL performance was achieved when the SNR levels of the training and testing data were similar. This highlights the importance of considering the noise characteristics of the data used for training.
- DL is Generalizable: The study demonstrated that DL models trained on subsets of data (e.g., 25% or 50% of time points) could effectively improve image quality beyond the training "ground truth," showcasing the generalizability of the learned denoising capabilities.
Key Results
- Improved Image Quality: DL denoising, particularly with optimized averaging and training strategies, significantly improved SSIM and PSNR while reducing NMAE across various SNR levels.
- Scan Time Reduction Potential: Windowed averaging before DL denoising showed that a 2-fold acceleration (using half the time points) could be achieved with modest DL denoising without significant quality degradation. 4-fold or even 8-fold acceleration might be feasible with more advanced training strategies.
- Artifact Reduction: The DL models effectively reduced noise and artifacts, including those outside the brain, and importantly, preserved actual ASL signals (like intravascular signals) that could be diagnostically relevant.
- M₀ Impact: Including M₀ provided a significant boost in performance, especially for low-SNR data, with improvements up to 10% in SSIM and PSNR.
Limitations and Future Work
- Limited Architectures: The study primarily focused on 3D U-Net and validated with SwinIR. Further investigation with other DL architectures could confirm the generalizability of the findings.
- Motion Artifacts: While motion correction was applied, residual motion artifacts were excluded. Future work could explore DL's capability to better suppress these artifacts, especially when they are difficult to separate from physiological noise.
- Low SNR GT: The low SNR of the "ground truth" (GT) images for some low-SNR PASL data might have limited the optimal performance of DL training. Grouping input images by SNR levels could reveal additional patterns.
- Single-Shot Readout: The study used single-shot ASL readouts. The trends might differ for multishot readouts, which have fewer averages/time points.
- Adaptive Denoising: The authors suggest that more adaptive DL denoising methods could be beneficial, especially for low-SNR images, potentially by incorporating L1 norm in the loss function to reduce blur and improve anatomical accuracy.
Summary updated