Median Filtering of Tensor-Valued Images

Martin Welk, Christian Feddern, Bernhard Burgeth, Joachim WeickertDAGM-Symposium 2003 · DOI

TL;DR

This paper introduces a novel median filter designed specifically for tensor-valued images, such as those produced by Diffusion Tensor Magnetic Resonance Imaging (DT-MRI). By extending the robust, edge-preserving properties of scalar median filters to the matrix domain, this new method offers a powerful tool for denoising and smoothing complex tensor data while preserving crucial structural information.

What Problem Does This Paper Address?

Modern imaging techniques like DT-MRI generate data not as simple pixel intensities, but as matrices (tensors) at each spatial location. These tensors capture complex directional information, such as the diffusion of water molecules in biological tissues, which is vital for understanding brain connectivity and diagnosing neurological conditions. However, this rich data is often corrupted by noise, which can obscure important details and lead to misinterpretations. Existing noise reduction methods, while effective for scalar images, often struggle to handle the inherent structure and anisotropy of tensor-valued data. Specifically, they may blur out important discontinuities or fail to preserve the positive-semidefinite nature of these matrices, which is a fundamental property of physical tensors. This paper addresses the need for a nonlinear, structure-preserving filter that can effectively denoise these tensor-valued images without sacrificing critical information.

Methods

The authors propose a median filter for tensor-valued images by generalizing the fundamental property of the scalar median: it minimizes the sum of absolute deviations from the data points. For matrices, this translates to finding a matrix $\mathbf{X}$ that minimizes the sum of Frobenius norms of the differences between $\mathbf{X}$ and each data matrix $\mathbf{A}_i$ in a local neighborhood:

$$ \mathbf{A} := \underset{\mathbf{X}}{\operatorname{argmin}} \sum_{i=1}^n ||\mathbf{A}_i - \mathbf{X}|| $$

This minimization problem is solved using a gradient descent approach. A key aspect of their algorithm is a step-size control mechanism that adapts based on over- and undershoots detected during the iteration process. This ensures stable convergence to the solution. The paper also highlights that this definition of the median for matrices is rotationally invariant, a desirable property for physical tensors.

Datasets

The paper evaluates its proposed median filter on two types of data:

  • Synthetic data: A synthetic tensor field with a known discontinuity is used to demonstrate the edge-preserving capabilities of the filter.
  • Real-world data: A 2D frame from a 3D DT-MRI dataset of a human head is used to showcase the filter's performance on actual medical imaging data. This dataset is augmented by replacing 30% of the data with noise matrices to test robustness.

Key Insights and Ideas

  • Generalizing the Median: The core idea of extending the median's "least absolute deviations" property to the matrix domain using the Frobenius norm is a powerful conceptual leap. This provides a principled way to define a "central" tensor in a neighborhood.
  • Robustness and Edge Preservation: The filter inherits the excellent robustness against outliers and the ability to preserve sharp edges/discontinuities that are hallmarks of scalar median filters. This is crucial for applications where structural integrity is paramount.
  • Rotational Invariance: The proposed median definition is inherently rotationally invariant, meaning the result doesn't depend on the coordinate system used to represent the tensors. This aligns with the physical interpretation of tensors.
  • Convergence of Tensor Fields: The experiments suggest that, similar to scalar median filtering, tensor-valued median filtering can also reach "root signals" (stationary states) after a sufficient number of iterations, indicating a well-behaved convergence property.

Key Results

The paper demonstrates the effectiveness of the matrix-valued median filter through qualitative results presented in Figures 1 and 2.

  • Figure 1 shows that the filter effectively preserves discontinuities in a synthetic tensor field while smoothing out noise.
  • Figure 2 illustrates the filter's application to real DT-MRI data. It shows that the filter can significantly reduce noise (up to 30% of data replaced by noise) while maintaining the structural integrity of the tensor field, including discontinuities at image boundaries. The experiments also show that increasing the filter's neighborhood size (e.g., from 3x3 to 9x9) leads to more pronounced filtering, as expected.

Limitations and Future Work

The authors acknowledge that their current definition of the median is restricted to real 2x2 matrices, although they state that extension to larger matrices is straightforward. They also note that the filter's behavior at image boundaries can lead to shifts in discontinuities due to the use of reflecting boundary conditions.

Future work planned by the authors includes generalizing other nonlinear filters (beyond median filtering) to be applicable to tensor field filtering. This suggests a broader research agenda focused on developing a comprehensive suite of advanced image processing tools for tensor-valued data.

Summary updated

BibTeX
@inproceedings{welk2003median,
  author = {Martin Welk and Christian Feddern and Bernhard Burgeth and Joachim Weickert},
  title = {{Median Filtering of Tensor-Valued Images}},
  booktitle = {DAGM-Symposium},
  year = {2003},
  doi = {10.1007/978-3-540-45243-0\_3},
}
Figure