Source-linked AI summary
Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network
Dongsheng Jiang, Weiqiang Dou, Luc Vosters, Xiayu Xu, Yue Sun, Tao Tan
TL;DR
MRI denoising is needed because acquisition noise degrades image quality and can affect diagnosis and quantitative processing, while existing methods require optimization and noise-level estimation. The paper proposes a multi-channel residual DnCNN for Rician-noise removal and evaluates specific- and general-noise training on real and synthetic 3D MR data. The specific-noise model outperforms compared methods across the reported test datasets, while the general model performs better than the comparisons in two datasets and can be applied without specifying noise level.
Problem
MRI acquisition noise can impair image quality, diagnosis, and quantitative processing, while existing denoising methods depend on optimization and estimated noise levels.
Method
The paper uses a ten-layer multi-channel residual DnCNN to denoise Rician-noise-corrupted 3D MR images, training models for specific and general noise levels.
Results
The noise-specific model outperforms compared methods on PSNR and global SSIM across the reported test datasets; the general model is better than comparisons in two datasets.
Takeaways & Limitations
The general model can denoise MR images directly without assuming or estimating noise level in advance, supporting its reported general applicability.
Takeaways & Limitations
The general model may require retraining or tuning for a corresponding dataset, and its performance is slightly worse than the noise-specific model.
Abstract
from arXiv · showhide
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.
1. Introduction
MRI denoising is important because acquisition noise can impair diagnosis and quantitative image processing. The paper proposes a multi-channel residual CNN to denoise 3D MR images without test-time optimization or prior noise-level estimation.
- Acquisition noise can degrade MRI quality and reduce the reliability of diagnosis, segmentation, registration, and classification.
- The study extends DnCNNs to restore noise-free MR images while avoiding optimization during testing.
- A multi-channel approach adapts the network to 3D volumes and is hypothesized to improve training stability, speed, and denoising robustness.
- The proposed method is evaluated against four state-of-the-art denoising methods.
- The paper presents what it identifies as the first deep-learning method for denoising Rician-noise-corrupted MR images.
2.1. MCDnCNN
MCDnCNN adapts DnCNN to 3D MR volumes using multi-channel residual learning. It predicts residual noise rather than the clean image, then reconstructs the noise-free volume.
- MCDnCNN adapts DnCNN to 3D data through a multi-channel architecture targeting noise-free MR-image recovery.
- Residual learning is used because residual networks are easier to optimize and can benefit from increased depth.
- The network has ten convolutional layers: one ReLU input layer, eight batch-normalized ReLU layers, and one output layer.
- The output preserves the input dimensions because the architecture omits max-pooling.
- Training minimizes averaged mean square error between the desired residual image and the estimated residual.
- After slice-wise noise estimation across the volume, the method reconstructs the original noise-free volume.
- The study trains both a specific-noise model, MCDnCNNs, and a general-noise model to address unknown Rician noise levels.
2.2. Comparison with other methods
The study compares MCDnCNN with four established 3D denoising filters. These baselines extend nonlocal-means or DCT methods with structural, rotational, or computational improvements.
- Four well-established denoising methods are tested alongside the proposed approach on 3D MR images.
- The optimized block-wise NLM filter combines automated smoothing-parameter tuning, relevant-voxel selection, blockwise processing, and parallel computation.
- The wavelet-based extension mixes 3D sub-bands into the optimized block-wise NLM filter while maintaining comparable computation time.
- The local DCT filter adapts to Rician noise using a pseudo-oracle principle and is reported to denoise faster than the two preceding filters.
- PRI NLM3D is a rotationally invariant NLM variant that uses a prefiltered image and was reported to provide the best denoising performance among these filters in under one minute.
2.3. MR image acquisition and evaluation
The evaluation uses multiple real and simulated MRI datasets to test general applicability across acquisition settings. Performance is measured across noise levels with PSNR and global SSIM.
- Training and testing use three subsets from the IXI and Brainweb public datasets to investigate cross-dataset general applicability.
- IXI-Hammersmith dataset: The IXI-Hammersmith data comprise 30 healthy-subject T1-weighted brain-image sets, with 20 for training and 10 for evaluation.
- IXI-Guys dataset: The Guys dataset contains 10 T1-weighted images acquired at Guys Hospital with a Philips 1.5T system.
- Brainweb dataset: Brainweb provides realistic simulated MRI volumes used to evaluate denoising approaches.
- Evaluation measures: Methods are evaluated at noise levels from 1% to 15% in 2% increments.
- Evaluation measures: PSNR and global SSIM are used as quantitative measures, with global SSIM defined as the average local SSIM.
- Evaluation measures: Local SSIM uses a 3x3x3 voxel kernel and incorporates local means, variances, covariance, and constants c1 and c2.
3. Results
Across IXI-Hammersmith, Guys, and Brainweb datasets, the noise-specific MCDnCNNs generally achieved the strongest denoising results, while the general-noise MCDnCNNg also outperformed competing methods across substantial noise ranges.
- IXI-Hammersmith: MCDnCNNs significantly outperformed all compared methods in PSNR and SSIM on the IXI-Hammersmith dataset.MCDnCNNg ranked second, exceeding other methods in PSNR from 1% to 15% noise and SSIM from 5% to 15%.
- IXI-Hammersmith: At 15% noise, MCDnCNNs produced the cleanest visual result on IXI-Hammersmith, while competing methods either retained background noise or lost image details.Both MCDnCNNs and MCDnCNNg preserved visually good brain-foreground results.
- IXI-Guys: MCDnCNNs showed the most robust PSNR and SSIM performance on the Guys dataset despite neither model being retrained there.MCDnCNNg exceeded other methods in PSNR from 3% to 13%, but ODCT3D and PRI NLM3D exceeded it in SSIM from 3% to 15%.
- Visual comparisons: Figures 3, 5, and 7 provide representative 15% noise denoising examples across the evaluated datasets.The Guys and Brainweb examples were reported to show effects similar to those observed for IXI-Hammersmith.
- Brainweb: MCDnCNNs significantly outperformed all other methods in PSNR and SSIM on Brainweb.MCDnCNNg was second only to MCDnCNNs and significantly exceeded competing methods in both measures from 5% to 15% noise.
4. Discussion and Conclusion
The proposed multi-channel residual CNN denoises MR images within a unified framework, with noise-specific and general models showing favorable performance and cross-dataset applicability. The general model can be applied without prior noise-level estimation, although dataset-specific retraining may improve performance.
- Conclusion: The noise-specific model outperforms other methods on PSNR and global SSIM across all three test datasets.The reported datasets are IXI-Hammersmith, IXI-Guys, and Brainweb.
- Conclusion: The general noise-applicable model performs better than compared methods in two datasets and shows good general applicability.Its performance is described as unqualified on one dataset, probably because retraining is needed.
- Practical advantage: The general model is trained across a wide range of noise levels and can be applied without assuming or estimating the noise level in advance.This contrasts with the compared methods, which require parameter tuning for a specific noise level.
- Cross-dataset applicability: The model trained on one 3T MR dataset can be directly applied to two datasets with different magnetic field strengths.The authors note that Brainweb uses simulated noise-free data and describe its results as most convincing.
- Future work: The network uses a deep convolutional architecture, while future work proposes exploring DCGANs, U-net, complete 3D CNNs, and additional MRI applications.The proposed network avoids max-pooling to preserve output size, and future work considers other architectures and anatomical studies.