⚡ electrical engineering

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

The paper proposes Neighbor-Guided Patch Sampling (NGPS), a lightweight self-supervised framework for volumetric medical imaging denoising that overcomes inter-slice misalignment by decoupling structure matching on a noise-attenuated guide from signal retrieval on raw neighboring slices, thereby constructing effective pseudo-targets without explicit registration or masking.

Original authors: Jaehyun Cho, YoungJoon Yoo

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jaehyun Cho, YoungJoon Yoo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to restore a blurry, noisy photo of a 3D object, like a stack of transparent slices of a loaf of bread. In medical imaging, these "slices" are CT or MRI scans of the human body. To clean up the noise in one slice, a smart computer program often looks at the slices right above and below it for clues. It's like asking your neighbors, "Hey, what does this part of the wall look like?"

The Problem: The "Ghost" Neighbors
The trouble is, human bodies aren't perfectly stacked like a rigid book. If a patient coughs, breathes, or just shifts slightly between scans, the slices don't line up perfectly. The liver in slice 10 might be slightly to the left, while in slice 11, it's slightly to the right.

Old methods tried to fix this in two ways, both of which had flaws:

The "Ignore It" Method: If the slices didn't match, the computer just threw away that part of the data. It was like saying, "I can't ask my neighbor because they are standing in a different spot," so it just left a blank spot. This meant losing a lot of important details, especially around sharp edges like bones or organ boundaries.
The "Stretch It" Method: The computer tried to mathematically stretch and warp the neighbor slice to force it to line up perfectly. But if the noise was too heavy or the shift was too big, this "stretching" would blur the image, making sharp edges look mushy.

The Solution: NGPS (The "Smart Scout")
The authors propose a new method called NGPS (Neighbor-Guided Patch Sampling). Think of NGPS as a smart scout that doesn't just look at the neighbor's front door (the exact same coordinate); it looks around the neighborhood to find the right house.

Here is how it works, step-by-step:

The "Blurry Map" (The Guide): First, the computer takes a quick, rough look at the noisy slices and blurs them slightly to create a "guide map." This map ignores the static (noise) but keeps the general shape of the buildings (anatomy). It's like looking at a low-resolution sketch to see where the walls are.
The "Spot the Difference" Check: The computer compares the sketch of the current slice with the sketch of the neighbor slice. If they look different in a specific spot, it flags that spot as "misaligned."
The "Search Party" (Patch Sampling): Instead of giving up or stretching the image, the computer sends a search party into the neighbor slice. It looks at a small neighborhood (a 15x15 grid) around the flagged spot. It asks: "Which little patch in the neighbor slice looks most like the patch in my current slice?"
The "Raw Treasure" (Retrieval): Once it finds the matching patch in the neighbor's sketch, it goes back to the neighbor's original, raw, noisy photo at that exact matching spot and grabs the pixel value.
- Crucial Detail: It uses the blurry sketch to find the location, but it grabs the real data from the original noisy photo. This ensures it doesn't accidentally smooth out the sharp edges while it's searching.
The "Group Vote" (Top-K): To be extra sure, it doesn't just pick the single best match. It picks the top 4 best matches, averages their values, and uses that as the answer. This is like asking four different neighbors for the time and taking the average to get a more accurate result.

Why It's Better

No Blurring: Because it grabs the raw data from the correct location instead of stretching the image, sharp edges (like the rim of a kidney or a bone) stay sharp.
No Wasted Data: It doesn't throw away the "misaligned" parts of the image. It finds the information that was just displaced.
Fast: It doesn't need to do heavy, complex math to warp the whole image. It just does a quick local search.

The Results
The paper tested this on real medical data (low-dose CT scans of the abdomen and chest, and MRI scans of the brain).

On CT scans: It removed the "quantum noise" (graininess) better than previous methods, especially in very low-dose scans where the images are usually very grainy.
On MRI scans: It handled synthetic noise well and kept the brain structures clear.
Speed: It was significantly faster than other advanced methods that try to align images.

In a Nutshell
NGPS is like a detective who realizes that if a witness (the neighbor slice) is standing in a slightly different spot, you don't ignore them or force them to stand where they aren't. Instead, you look around their feet to find the exact spot where they are standing, ask them what they see, and use that answer to clean up your own picture. This keeps the picture clear, sharp, and full of detail.

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

1. Problem Statement

In volumetric medical imaging (e.g., CT and MRI), self-supervised learning (SSL) using neighboring slices offers a promising avenue for denoising without requiring clean ground-truth targets. However, a critical challenge arises from inter-slice misalignment. Anatomical structures often shift between adjacent slices due to patient motion or physiological variability.

When standard SSL methods naively use adjacent slices as supervision targets at the same coordinates, this misalignment causes the network to learn from displaced anatomical evidence. This results in "ghosting" artifacts and blurred anatomical boundaries. Existing remedies typically fall into two categories, both with significant drawbacks:

Masking-based approaches: These exclude pixels with high discrepancy between slices from the training loss. While this stabilizes training, it discards a substantial portion of neighboring evidence (approximately 20% in clinically relevant regions), particularly around high-frequency anatomical boundaries where structural information is most critical.
Registration-based approaches: These estimate dense deformation fields to warp neighboring slices into alignment before supervision. However, under severe noise or large inter-slice gaps, these alignment estimates become unreliable, and the spatial resampling required can smooth edges and introduce additional artifacts.

The core problem is how to retrieve displaced anatomical evidence from a local neighborhood without relying on dense deformation estimation or discarding valuable supervision signals.

2. Methodology: Neighbor-Guided Patch Sampling (NGPS)

The authors propose NGPS, a lightweight framework that reframes misalignment handling as a training-time supervision construction problem. Instead of discarding misaligned regions or warping entire slices, NGPS decouples structure matching from signal retrieval.

The pipeline operates in three main steps:

Guide Generation and Conservative Masking:
- A noise-attenuated "guide" volume is generated by applying a 2D bilateral filter followed by a median filter to the noisy input slices. This preserves structural boundaries while suppressing quantum noise.
- A direction-aware misalignment mask is created by comparing the guide slices. Pixels where the guide difference exceeds a threshold $\tau$ are flagged as misaligned.
Decoupled Patch Matching and Raw Retrieval:
- For pixels flagged as misaligned, NGPS searches a local spatial window (e.g., $15 \times 15$ ) in the adjacent slice for structurally similar patches. This search is performed on the filtered guide to ensure robust structural matching.
- Crucially, once a structurally similar patch is identified, the supervision signal is retrieved directly from the raw, unfiltered noisy pixel at the matched coordinates in the adjacent slice. This separation ensures that the target value remains unfiltered (preserving the noise statistics required for Noise2Noise-style training) while the correspondence is established on a cleaner representation.
Top-K Ensemble Target Construction:
- To reduce variance associated with a single best match, NGPS aggregates the top- $K$ candidate pixels (based on patch similarity) and averages their raw values to form the pseudo-target.

Training Objective:
The model is trained using a hybrid loss function:

Static Regions: For unmasked pixels, standard same-coordinate neighboring supervision (Noise2Noise) is applied.
Misaligned Regions: For masked pixels, the loss is computed against the NGPS-retrieved targets.
Regional Consistency: An additional regularization term encourages consistency between the model's output on adjacent slices in static regions, reducing slice-to-slice flicker without constraining misaligned areas.

3. Key Contributions

Reframing Misalignment: The authors identify the "structural supervision gap" in neighboring-slice SSL, clarifying the trade-off between masking (discarding data) and registration (relying on potentially unreliable warping).
Neighbor-Guided Patch Sampling (NGPS): A novel mechanism that recovers displaced supervision via local guide-based patch search and raw-target retrieval. It avoids dense deformation-field estimation and spatial resampling.
Comprehensive Validation: The method is validated across diverse benchmarks, including realistic low-dose CT (AAPM-Mayo), simulated ultra-low-dose CT with streak artifacts (LIDC-IDRI), and synthetic Rician noise MRI (IXI).

4. Experimental Results

Quantitative Performance:
- On the AAPM-Mayo (Quarter-Dose) dataset, NGPS achieved the best results across all five metrics (PSNR, SSIM, FSIM, HFEN, GMSD), outperforming strong baselines like Deformed2Self and NS-N2N.
- On the LIDC-IDRI (Ultra-Low-Dose) dataset, NGPS showed improved fidelity and boundary-sensitive restoration, achieving higher PSNR, FSIM, and HFEN compared to NS-N2N. While SSIM gains were marginal, the method demonstrated superior preservation of fine anatomical margins.
- On the IXI (MRI) dataset with synthetic Rician noise, NGPS consistently maintained the highest PSNR and SSIM across noise levels (5%, 7%, 9%), outperforming both single-image and volumetric baselines.
Robustness to Slice Thickness:
- NGPS demonstrated superior stability when slice thickness increased from 1.25 mm to 2.5 mm. The PSNR drop was only 0.16 dB for NGPS, compared to 0.63 dB for NS-N2N and significantly larger drops for registration-based methods (Deformed2Self) and masking-based methods (Noise2Sim).
Computational Efficiency:
- NGPS is highly efficient, requiring approximately 0.72 seconds to prepare targets for a 100-slice volume. This is roughly 19 times faster than NS-N2N and 7.5 times faster than Pixel2Pixel, primarily due to vectorized patch search restricted to masked pixels and the use of lightweight filtering.

5. Significance and Claims

The paper claims that NGPS offers a practical and effective strategy for structure-preserving volumetric self-supervised denoising. By retrieving local supervision rather than discarding it or relying on complex warping, NGPS successfully recovers displaced anatomical evidence.

The authors emphasize that their approach:

Preserves High-Frequency Details: Unlike masking methods that blur boundaries or registration methods that smooth edges, NGPS retains sharp anatomical margins.
Avoids Heavy Computation: It eliminates the need for dense deformation-field estimation, making it lightweight and fast.
Generalizes Well: It performs consistently across different modalities (CT and MRI) and noise regimes (realistic Poisson and synthetic Rician).

Limitations Acknowledged:
The authors note that NGPS assumes a locally corresponding structure exists within the search window. Performance may degrade under large through-plane gaps, abrupt anatomical changes, or non-geometric appearance variations (e.g., staining). Additionally, the current validation relies on synthetic noise for MRI, and future work is needed for acquisition-realistic low-field MRI and clinical downstream tasks.