A Learning-from-noise Dilated Wide Activation Network for denoising Arterial Spin Labeling (ASL) Perfusion Images

Danfeng Xie, Yiran Li, Hanlu Yang, Bai Li, Lei Zhang, Ze Wang

Abstract

Arterial spin labeling (ASL) perfusion MRI provides a non-invasive way to quantify cerebral blood flow (CBF) but it still suffers from a low signal-to-noise-ratio (SNR). Using deep machine learning (DL), several groups have shown encouraging denoising results. Interestingly, the improvement was obtained when the deep neural network was trained using noise-contaminated surrogate reference because of the lack of golden standard high quality ASL CBF images. More strikingly, the output of these DL ASL networks (ASLDN) showed even higher SNR than the surrogate reference. This phenomenon indicates a learning-from-noise capability of deep networks for ASL CBF image denoising, which can be further enhanced by network optimization. In this study, we proposed a new ASLDN to test whether similar or even better ASL CBF image quality can be achieved in the case of highly noisy training reference. Different experiments were performed to validate the learning-from-noise hypothesis. The results showed that the learning-from-noise strategy produced better output quality than ASLDN trained with relatively high SNR reference.

Introduction

Arterial Spin Labeling (ASL) is a powerful, non-invasive MRI technique used to quantify Cerebral Blood Flow (CBF). However, it has a major limitation: an inherently low Signal-to-Noise Ratio (SNR). To obtain usable images, clinicians typically average many individual scans, which increases acquisition time and susceptibility to motion artifacts. While deep learning has shown promise in denoising these images, the lack of a "gold standard" clean image for training remains a significant hurdle. This study explores the "learning-from-noise" (LFN) hypothesis—the idea that a network can learn to find the underlying signal even if both the input and the training target are highly noisy, provided the noise is zero-mean.

Methods

The authors propose two major innovations to tackle ASL denoising: the ASLDN-LFN training strategy and the Dilated Wide Activation Network (DWAN) architecture.

ASLDN-LFN (Learning-from-Noise)

Instead of attempting to create a "clean" reference for training, the model uses two different sets of noisy ASL scans from the same subject. One set acts as the input, and the other as the target. Mathematically, if the network minimizes the difference between two noisy observations of the same underlying signal, it naturally converges toward the mean (or median) of that signal.

DWAN (Dilated Wide Activation Network)

The architecture features two parallel pathways to process information at different scales:

  • Local Pathway: Standard convolutions to capture fine structural details.
  • Global Pathway: Uses dilated convolutions (with increasing rates of 2, 4, 8, and 16) to capture large-scale patterns without increasing the number of parameters.
  • Wide Activation: Unlike standard residual blocks, these blocks expand features before the activation function (ReLU), helping preserve more information during the denoising process.
Figure 1: The architecture of the proposed DWAN, featuring parallel local and global pathways for multi-scale feature extraction.

Loss Function

The study utilizes L1 loss (absolute difference) rather than the standard L2 loss. Because ASL data contains many extreme noise spikes ("outliers"), L1 loss is more robust, targeting the median of the distribution and avoiding the blurring often associated with L2 loss.

Datasets and Evaluation

The study utilized a database of 280 subjects (200 training, 20 validation, 60 testing). For evaluation purposes, a "pseudo gold standard" reference was created by averaging 40 image pairs and applying Gaussian smoothing and outlier cleaning.

Key Results

The proposed ASLDN-LFN method outperformed both traditional averaging and previous deep learning models across multiple metrics:

  • SNR Improvement: The average SNR of the "pseudo gold standard" reference was 5.87. The LFN model achieved an SNR of 8.06, significantly higher than the standard deep learning approach (6.36).
  • Tissue Contrast: The Grey Matter/White Matter (GM/WM) contrast was 2.32 for the LFN model, compared to 2.14 for the reference images, indicating clearer differentiation between brain tissues.
  • Architecture Superiority: The DWAN structure consistently yielded higher PSNR (25.28) and SSIM (0.803) scores compared to standard U-Net or DilatedNet architectures.
Figure 2: Visual comparison of CBF maps across different methods. Row A: Noisy input; Row B: Surrogate reference; Row C: Standard ASLDN; Row D: Proposed ASLDN-LFN.

As shown in Figure 2, the ASLDN-LFN output (Row D) is visibly cleaner and sharper than the noisy input and even the surrogate reference it was trained against. This confirms that deep networks can effectively "filter" noise by seeing enough examples of signal-noise interaction.

Discussion and Conclusion

This study validates the learning-from-noise hypothesis for ASL CBF image denoising. By utilizing dilated convolutions to capture spatially correlated perfusion patterns and a robust L1 loss function, the ASLDN-LFN model produces higher-quality images than traditional surrogate references.

While the results are encouraging, future work is needed to test the model's generalizability across different scanners and clinical populations. Additionally, any denoising process must be carefully monitored to ensure that small, clinically relevant pathological details are not over-smoothed. Overall, the LFN strategy provides a powerful framework for improving ASL image quality when high-quality training data is unavailable.

Summary updated

Abstract

Arterial spin labeling (ASL) perfusion MRI provides a non-invasive way to quantify cerebral blood flow (CBF) but it still suffers from a low signal-to-noise-ratio (SNR). Using deep machine learning (DL), several groups have shown encouraging denoising results. Interestingly, the improvement was obtained when the deep neural network was trained using noise-contaminated surrogate reference because of the lack of golden standard high quality ASL CBF images. More strikingly, the output of these DL ASL networks (ASLDN) showed even higher SNR than the surrogate reference. This phenomenon indicates a learning-from-noise capability of deep networks for ASL CBF image denoising, which can be further enhanced by network optimization. In this study, we proposed a new ASLDN to test whether similar or even better ASL CBF image quality can be achieved in the case of highly noisy training reference. Different experiments were performed to validate the learning-from-noise hypothesis. The results showed that the learning-from-noise strategy produced better output quality than ASLDN trained with relatively high SNR reference.
BibTeX
@misc{xie2020learning,
  author = {Danfeng Xie and Yiran Li and Hanlu Yang and Bai Li and Lei Zhang and Ze Wang},
  title = {{A Learning-from-noise Dilated Wide Activation Network for denoising Arterial Spin Labeling (ASL) Perfusion Images}},
  year = {2020},
  eprint = {2005.07784},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.07784},
}
Figure