An enhanced adaptive non-local means algorithm for Rician noise reduction in magnetic resonance brain images

Kaixin Chen, Xiao Lin, Xing Hu, Jiayao Wang, Han Zhong, Linhua JiangBMC Medical Imaging 2020 · DOI

TL;DR

This paper introduces FANLM, a novel algorithm that significantly improves Rician noise reduction in MRI brain images. By adaptively adjusting the search window size and incorporating fuzzy clustering, FANLM effectively removes noise while preserving crucial image details, outperforming existing methods in both quantitative metrics and visual quality.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a vital tool in medical diagnosis, but the images it produces are often corrupted by Rician noise. This noise degrades image quality, blurring tissue boundaries and making detailed analysis difficult. While many denoising algorithms exist, most are designed for Gaussian noise and are not optimal for the Rician noise characteristic of MRI. Existing non-local means (NLM) algorithms, though better than traditional filters, often rely on fixed parameters that are not ideal for the varying noise levels and image structures found in MRI. This paper aims to address this gap by developing an adaptive algorithm specifically tailored for Rician noise in MRI brain images.

Methods

The core of the proposed method, FANLM (Fuzzy c-means and Adaptive Non-Local Means), combines two key innovations:

  • Adaptive Non-Local Means (NLM): Traditional NLM uses a fixed-size search window. FANLM addresses this by:
  • Adaptive Window Sizing: Instead of a single fixed window, FANLM employs three window sizes (small, medium, large) and adaptively selects the optimal size for each pixel. This selection is guided by the estimated noise variance, which is crucial for handling varying noise levels.
  • Improved Noise Estimation: The algorithm utilizes an improved Median Absolute Deviation (MAD) estimator specifically designed for Rician noise to accurately estimate noise variance.
  • Fuzzy C-Means (FCM) Integration: To further enhance detail preservation, FANLM incorporates an improved fuzzy c-means algorithm (FRFCM) as a pre-processing step. FRFCM helps to divide image regions into fuzzy clusters based on membership degrees, which aids in preserving fine image details during the denoising process. This approach is more robust to noise than traditional FCM.

Datasets

The paper evaluates FANLM using two datasets:

  • Dataset 1 (Simulated Data): T1-weighted brain MR images from BrainWeb were artificially corrupted with Rician noise at levels of 0%, 3%, 5%, 7%, and 9%. This dataset allowed for controlled evaluation of the algorithm's performance under varying noise conditions.
  • Dataset 2 (Real Data): Real T1-weighted brain MR images were acquired using a Siemens MAGNETOM Prisma 3T MRI system. This dataset provided a more realistic assessment of the algorithm's effectiveness on actual clinical data.

Key Insights and Ideas

  • Rician Noise is a Distinct Challenge: The paper highlights that Rician noise, common in MRI, requires specialized denoising techniques beyond those designed for Gaussian noise.
  • Adaptivity is Key: The effectiveness of FANLM stems from its ability to adapt to local image characteristics. By dynamically adjusting the search window size based on estimated noise variance, it avoids over-smoothing in textured areas and under-smoothing in smooth regions.
  • Fuzzy Clustering Enhances Detail Preservation: Integrating FCM before NLM denoising helps to segment image regions more effectively, allowing the NLM algorithm to better preserve fine details and edges that might otherwise be lost.
  • Quantitative and Qualitative Improvements: FANLM not only achieves superior quantitative metrics (PSNR, SSIM) but also demonstrates visually better results, preserving image details more effectively than other methods.

Key Results

  • Superior Noise Reduction: FANLM consistently outperformed traditional NLM, LMMSE, and UWT in terms of both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) across various noise levels and datasets.
  • Effective Detail Preservation: Qualitative assessments and visual comparisons showed that FANLM was significantly better at retaining image details and edge structures compared to the other methods, which often resulted in blurred details or artifacts.
  • Computational Cost: While FANLM showed improved denoising performance, its computational cost was higher than LMMSE and UWT, though generally better than NLM. The paper notes that the adaptive window selection contributes to this increased complexity.

Limitations and Future Work

The authors acknowledge that the computational cost of FANLM is a limitation. The adaptive selection of window size, while beneficial for performance, adds to the processing time. Future work will focus on reducing this computational complexity to make the algorithm more efficient for practical clinical applications.

Summary updated

BibTeX
@article{chen2020enhanced,
  author = {Kaixin Chen and Xiao Lin and Xing Hu and Jiayao Wang and Han Zhong and Linhua Jiang},
  title = {{An enhanced adaptive non-local means algorithm for Rician noise reduction in magnetic resonance brain images}},
  journal = {BMC Medical Imaging},
  year = {2020},
  doi = {10.1186/s12880-019-0407-4},
}
Figure