NGPS: Structure-Preserving Self-Supervised Denoising via Neighbor-Guided Patch Sampling

In volumetric medical imaging, self-supervised denoising often utilizes adjacent slices as supervision. However, inter-slice anatomical misalignment frequently leads to ghosting and blurred margins when these slices are used naively. Existing methods typically handle this by masking out discrepant regions, which effectively discards valuable anatomical information, especially near high-frequency boundaries. This paper asks: can we recover this displaced anatomical evidence without the computational overhead and potential errors of dense registration?

NGPS (Neighbor-Guided Patch Sampling) introduces a lightweight, misalignment-aware framework that treats displaced anatomical structures as a local retrieval problem. Instead of masking out discrepant pixels, NGPS performs a two-step process:

1. Guide-based Matching: It generates a noise-attenuated guide image using simple filtering (bilateral and median filters) to identify structurally similar patches in neighboring slices.
2. Raw Target Retrieval: Once a match is found, it retrieves the raw, unfiltered pixel values from the original noisy neighbor to serve as the training target.

By decoupling the structural matching (done on the guide) from the signal retrieval (done on the raw data), the model avoids learning from misleading targets while retaining high-frequency information that would otherwise be discarded.

NGPS consistently outperforms existing self-supervised denoising methods across low-dose CT and synthetic-Rician MRI datasets. By effectively utilizing displaced information, the model achieves better structural preservation, as evidenced by improved metrics in HFEN (High-Frequency Error Norm) and GMSD (Gradient Magnitude Similarity Deviation). The framework is computationally efficient, avoiding the need for complex, learned deformation fields or spatial resampling, and remains compatible with standard Noise2Noise-style training objectives.