Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network (1712.08726v2)

Abstract: The denoising of magnetic resonance (MR) images is a task of great importance for improving the acquired image quality. Many methods have been proposed in the literature to retrieve noise free images with good performances. Howerever, the state-of-the-art denoising methods, all needs a time-consuming optimization processes and their performance strongly depend on the estimated noise level parameter. Within this manuscript we propose the idea of denoising MRI Rician noise using a convolutional neural network. The advantage of the proposed methodology is that the learning based model can be directly used in the denosing process without optimization and even without the noise level parameter. Specifically, a ten convolutional layers neural network combined with residual learning and multi-channel strategy was proposed. Two training ways: training on a specific noise level and training on a general level were conducted to demonstrate the capability of our methods. Experimental results over synthetic and real 3D MR data demonstrate our proposed network can achieve superior performance compared with other methods in term of both of the peak signal to noise ratio and the global of structure similarity index. Without noise level parameter, our general noise-applicable model is also better than the other compared methods in two datasets. Furthermore, our training model shows good general applicability.

• The paper proposes a novel ten-layer multi-channel CNN that predicts pixel-level noise without needing explicit noise parameters.
• It demonstrates superior performance over traditional methods with improved PSNR and SSIM across varying noise intensities.
• Results highlight the method's robustness and its potential for efficient MRI data preprocessing in clinical diagnostics.

An Evaluation of Deep Learning-Based Denoising of 3D Magnetic Resonance Images

The paper "Denoising of 3D Magnetic Resonance Images with Multi-Channel Residual Learning of Convolutional Neural Network" introduces a novel approach to the denoising of MRI images that diverges from traditional denoising methodologies by employing a convolutional neural network (CNN). This work addresses the limitations of existing state-of-the-art methods, which typically require knowledge of the noise level parameter and are dependent on time-consuming optimization processes. The proposed CNN methodology contrasts with these methods by eliminating the need for such prior information and optimizations, thus improving both efficiency and applicability.

Methodology Overview

The researchers developed a ten-layer CNN incorporating residual learning and a multi-channel strategy, specifically tailored to handle the intrinsic 3D nature of MRI data. This network's architecture is distinctly characterized by its capability to predict noise at each pixel, rather than attempting to directly compute noise-free images, thus maintaining essential input details and ensuring accurate denoising.

Two training paradigms are explored in this paper: training on a specific noise level and training on a general noise level. The first approach allows the model to perform optimally when noise levels are known beforehand, while the latter assesses the network's robustness across varying noise intensities without estimating noise levels, offering a more universally applicable solution.

Comparative Analysis

The efficacy of the proposed network was evaluated against several established denoising methods, including the optimized block-wise non-local means and Markov random field approaches, among others. Results demonstrated that both the noise-specific and general noise models of the CNN outperformed traditional methods when examining key metrics such as the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM). Notably, the evidence suggests the CNN's results held across various datasets, highlighting the model's resilience and general applicability.

Numerical Results

A significant aspect of the paper lies in the quantitative superiority of the proposed method. On the IXI-Hammersmith dataset, the noise-specific model (MCDnCNNs) achieved consistently higher PSNR and SSIM scores across varying levels of noise, affirmatively illustrating its robust denoising capacity. The general-noise model (MCDnCNNg) also exhibited competitive performance despite the absence of noise parameterization, further asserting the model's potential in practical applications where noise levels are unknown.

Implications and Future Prospects

The implications of this methodology are noteworthy in the context of clinical MRI diagnostics. By providing a reliable denoising tool that does not rely on pre-estimates of noise, the framework advances toward more efficient MRI data preprocessing, which is crucial for accurate diagnosis, segmentation, and classification in medical imaging applications.

Future work as outlined by the authors includes exploring deeper architectures such as full 3D convolutional networks and generative adversarial networks (GANs) to handle data more effectively. Additionally, expanding the application of this denoising technique to other imaging modalities such as lung or cardiac MRI provides exciting new directions for further research that could potentially enhance the utility of this approach across diverse medical imaging challenges.

In conclusion, this research paper provides a comprehensive evaluation of a CNN-based denoising approach that could significantly impact MRI data preprocessing by improving efficiency and applicability while maintaining high-performance denoising capabilities.