Self-Supervised Deep Learning-Based Denoising for Diffusion Tensor MRI

Medical Imaging Life Sciences Medicine Bioengineering
self-supervised learning deep learning convolutional neural network diffusion tensor imaging denoising

Background Introduction

Diffusion Tensor Magnetic Resonance Imaging (DTI) is a widely used neuroimaging technique for imaging the microstructure of brain tissues and white matter tracts. However, noise in Diffusion-Weighted Images (DWI) can reduce the accuracy of microstructural parameters derived from DTI data and also necessitate longer acquisition times to improve the Signal-to-Noise Ratio (SNR). Although deep learning methods based on Convolutional Neural Networks (CNNs) have shown outstanding performance in image denoising, they typically require additional high SNR data to supervise the CNN training, limiting the practical application of supervised learning methods in denoising. The Denoising Effect of Diffusion Tensor Imaging

Paper Source

The title of this paper is “SDnDTI: Self-Supervised Deep Learning-Based Denoising for Diffusion Tensor MRI,” with the main authors including Qiyuan Tian, Ziyu Li, Qiuyun Fan, Jonathan R. Polimeni, Berkin Bilgic, David H. Salat, and Susie Y. Huang, from institutions such as the Athinoula A. Martinos Center for Biomedical Imaging at Massachusetts General Hospital, Department of Radiology at Harvard Medical School, Department of Biomedical Engineering at Tsinghua University, and the Harvard-MIT Health Sciences and Technology Program. The paper has been published in the journal “Neuroimage.”

Research Methodology

Workflow

This paper proposes a self-supervised deep learning method named SDnDTI, which can denoise DTI data without requiring additional high SNR data. The specific workflow is as follows: 1. Data Grouping: The multi-directional DTI data is divided into multiple subsets, each containing six DWI volume data. 2. Signal Transformation: Using the diffusion tensor model, DWI volume data in each subset is transformed to the same diffusion encoding direction, generating multiple repeated DWI volume data with the same image contrast but different noise observations. 3. Denoising: First, a deep three-dimensional convolutional neural network is used to denoise each repeated DWI volume data. The training objective of the CNN is the average of all repeated DWI, which has a higher SNR. Then, the denoised images from the CNN are averaged to obtain a higher SNR. 4. Validation and Analysis: The denoising effect of SDnDTI is compared by output images and their generated DTI metrics with the ground truth values generated using a large number of DWI volume data.

Data Source

This paper utilizes two sets of data with different spatial resolutions, b-values, and numbers of DWI volumes, from the Human Connectome Project (HCP) and the Lifespan HCP in Aging project.

Research Results

Key Findings at Each Step

In the experiments, SDnDTI showed significant SNR improvement on both the original and synthesized denoised data. The DWI after SDnDTI denoising not only appeared clearer visually but also demonstrated superior performance in multiple quantitative indices.

  • SNR Improvement: Compared with raw data, the MAE of the denoised DWI image is about one-third of the original data, PSNR increased by about 7 dB, and SSIM increased by about 0.1.
  • Retention of Image Details: The denoised images by SDnDTI retained more texture details, especially in the internal capsule region, showing better performance compared to BM4D and AONLM methods.

Conclusions and Significance

The SDnDTI method achieves efficient denoising without requiring additional high SNR data, enhancing the feasibility of deep learning and CNN denoising methods in practical research and clinical applications. This method not only improves the quality of DTI data but also speeds up the acquisition of DTI data, significantly benefiting the mapping of brain tissue microstructure, fiber tracts, and structural connectivity.

Moreover, the SDnDTI method demonstrated its generalization ability across different datasets and further improved denoising performance and shortened training time through finetuning. This offers broad prospects for applications in clinical and neuroscience research involving rapid DTI acquisition and high-quality data requirements.

Highlights of the Study

  • No Need for High SNR Data: Unlike traditional deep learning denoising methods, SDnDTI does not require additional high SNR data for training.
  • Retention of Image Details: This method effectively retains the detailed texture information of brain tissues, outperforming other traditional denoising methods.
  • Broad Application Prospects: Due to its excellent generalization capability and training efficiency, SDnDTI is suitable for more research and clinical application scenarios.

Other Valuable Information

The SDnDTI code will be made publicly available on the GitHub platform (https://github.com/qiyuantian/sdndti), which will help researchers and clinicians apply this approach in their own studies. Additionally, the paper provides some public datasets and open-source software tools for comparative experiments, further advancing research in this field.

The SDnDTI method is a powerful self-supervised denoising tool with significant implications for improving DTI data quality and accelerating DTI data acquisition. This method not only exhibits technical innovation but also shows extensive potential and value in practical applications.