Nonlocal transform-domain filter for volumetric data denoising and reconstruction

Area of Science:

• Medical Imaging
• Signal Processing
• Computer Vision

Background:

• Volumetric data, common in medical imaging (e.g., MRI, CT), often suffers from noise.
• Existing denoising methods like BM3D are effective for 2D images but not directly applicable to 3D volumetric data.

Purpose of the Study:

• To extend the powerful BM3D filtering technique to handle 3D volumetric data.
• To develop an algorithm (BM4D) for effective denoising and reconstruction of volumetric datasets.

Main Methods:

• BM4D utilizes a grouping and collaborative filtering approach, extending BM3D's paradigm to 4D.
• Mutually similar 3D voxel patches (cubes) are stacked into a 4D 'group' for joint filtering in a transform domain.
• A 4D transform exploits both local voxel correlations within cubes and non-local correlations between cubes, enabling sparse representation and noise separation via coefficient shrinkage.

Main Results:

• BM4D demonstrates state-of-the-art performance in denoising volumetric data corrupted by Gaussian and Rician noise.
• The algorithm effectively reconstructs volumetric phantom data from noisy and incomplete k-space measurements.
• BM4D shows significant promise as a regularizer in volumetric data reconstruction tasks.

Conclusions:

• BM4D offers a powerful and effective solution for denoising and reconstructing volumetric data.
• The 4D collaborative filtering approach significantly improves upon existing 2D methods when applied to 3D datasets.
• BM4D's applicability extends to various fields requiring high-fidelity volumetric data processing, including medical imaging.