Abstract

Current self-supervised denoising techniques achieve impressive results, yet their real-world application is frequently constrained by substantial computational and memory demands, necessitating a compromise between inference speed and reconstruction quality. In this paper, we present an ultra-lightweight model that addresses this challenge, achieving both fast denoising and high quality image restoration. Built upon the Noise2Noise training framework—which removes the reliance on clean reference images or explicit noise modeling—we introduce an innovative multistage denoising pipeline named Noise2Detail (N2D). During inference, this approach disrupts the spatial correlations of noise patterns to produce intermediate smooth structures, which are subsequently refined to recapture fine details directly from the noisy input. Extensive testing reveals that Noise2Detail surpasses existing dataset-free techniques in performance, while requiring only a fraction of the computational resources. This combination of efficiency, low computational cost, and data-free approach make it a valuable tool for biomedical imaging, overcoming the challenges of scarce clean training data—due to rare and complex imaging modalities—while enabling fast inference for practical use.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3959_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ctom2/noise2detail

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ChoTom_Lightweight_MICCAI2025,
        author = { Chobola, Tomáš and Schnabel, Julia A. and Peng, Tingying},
        title = { { Lightweight Data-Free Denoising for Detail-Preserving Biomedical Image Restoration } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},
        page = {317 -- 326}
}

Reviews

Review #1

Please describe the contribution of the paper

The authors present Noise2Detail (N2D), a novel denoising pipeline tailored for biomedical imaging scenarios, where access to clean reference data and computational resources is often limited.

The approach leverages a lightweight, three-layer CNN that operates without any paired training data and is capable of effectively denoising real-world noisy images.

A key innovation lies in the integration of zero-shot Noise2Noise training with pixel-shuffle downsampling, enabling the method to suppress noise while preserving fine structural details in the signal.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

Zero-shot denoising: A denoising approach that eliminates the need for paired clean and noisy images, enabling training solely on noisy data.

Novel training formulation: One of the paper’s most innovative contributions is its unconventional training strategy, in which a denoised image serves as the input and the original noisy image is treated as the target. Given the unstructured and unpredictable nature of the noise, this setup encourages the model to prioritize the reconstruction of missing high-frequency signal components rather than learning to replicate the noise itself.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

Lack of comparison with supervised methods: The authors emphasize that obtaining paired noisy and clean images is challenging and that synthetically adding noise to clean images may not accurately reflect real-world noise characteristics. However, this strengthens the case for including a comparison with supervised denoising methods trained on synthetic noise. If such methods yield competitive performance, it would raise questions about the necessity of exclusively training on noisy data. A comparative analysis would clarify the relative advantages of the proposed zero-shot approach.

Limited evaluation metrics: The paper reports only Peak Signal-to-Noise Ratio (PSNR) as the evaluation metric. While PSNR is widely used for natural images, it is known to be suboptimal for medical imaging, where structural and perceptual fidelity is crucial. Incorporating additional metrics such as Structural Similarity Index (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS) would provide a more comprehensive and clinically relevant assessment of denoising quality.

Clinical relevance of visual results: The visual examples presented should emphasize clinically significant structures, such as lesions or subtle pathological findings. It is essential to demonstrate that the denoising process does not oversmooth diagnostically important features. Including cases such as lesions or conditions like emphysema would strengthen the argument that the method preserves medically relevant details.

Architectural justification: While the simplicity of the proposed architecture is noted as a design strength, additional justification is needed. The authors should explore how increasing network depth or introducing more complex architectures affects both denoising performance and inference time. Furthermore, the rationale for choosing exactly three layers is unclear—comparisons with alternatives (e.g., two or four layers) would help clarify whether this choice is empirically motivated or arbitrary.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The authors claimed to release the source code and/or dataset upon acceptance of the submission.
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

I would consider increasing my evaluation score if the authors include additional quality metrics such as SSIM and LPIPS, provide visual results that highlight clinically important details, and present at least a minimal ablation study analyzing the impact of network depth on performance versus inference time.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(3) Weak Reject — could be rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The paper introduces an interesting engineering novelty and addresses an important problem in medical imaging—the lack of clean and noisy image pairs. However, the current experiments are not enough to clearly show how well the method works. More results are needed to support the claims, and overall, the work is not yet mature enough for acceptance at MICCAI.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #2

Please describe the contribution of the paper

The paper presents a lightweight neural network based self-supervised denoising methods with a two stage pixel-shuffle based refinement framework.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

The proposed denoising method achieve a good balance between performance and computational burden.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

1) The paper does not compare the methods with the state-of-the-art conventional denoising methods for example BM3D and some more latest denoisng algorithms. 2) The paper does not give enough comparison on the implementation difference between the proposed methods and the peer methods. 3) The paper should give details on the “single image adaptations” mentioned at section 3.1.
Please rate the clarity and organization of this paper

Satisfactory
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

N/A
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(5) Accept — should be accepted, independent of rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The paper does not give enough introductions and comparisons on the implementation difference between the proposed methods and the peer methods to justify its contributions.
Reviewer confidence

Somewhat confident (2)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #3

Please describe the contribution of the paper

This paper presents a new self-supervised and lightweight image denoising method called Noise2Detail. The method is designed to improve the balance between image quality and processing time, and its effectiveness is demonstrated on both synthetic and real datasets.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1. The paper is clearly written, and the proposed method is well-motivated. It builds upon ZS-N2N (zero-shot noise2noise) by introducing two key steps—background correction and foreground enhancement—that lead to improvements in contrast and resolution.
2. The approach is fully self-supervised, which removes the need for paired clean/noisy data. This is particularly promising for applications in medical imaging.
3. The model is lightweight, with only around 22k parameters. The authors also provide a comparison of inference time against other state-of-the-art methods, which helps demonstrate its practical value.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1. The final step, which trains the network to simply match the original noisy image, appears to reintroduce some noise. This may lead to a slight drop in visual quality, as observed in the figures. It might be helpful to consider adding a regularization term—such as an L1 norm—to encourage sparsity and potentially suppress residual noise.
2. The experimental evaluation would be strengthened by additional metrics beyond PSNR, such as SSIM or perceptual quality metrics, to better capture visual improvements.
3. The paper would benefit from a brief discussion on limitations and potential directions for future work.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

N/A
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(5) Accept — should be accepted, independent of rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The method is novel, well-presented, and shows solid experimental results. While there are some minor issues, such as a somewhat coarse approach to resolution enhancement and limited evaluation metrics, the overall contribution is clear and valuable.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Author Feedback

Dear Area Chair, We thank the reviewers for their insightful and valuable feedback on our manuscript. We are grateful for the opportunity to address their comments and provide further clarifications. We are delighted that the reviewers recognize our work as “a novel denoising pipeline tailored for biomedical imaging scenarios” (R2) that achieves “a good balance between performance and computational burden” (R1) with “its effectiveness demonstrated on both synthetic and real datasets” (R3). They commend Noise2Detail as “innovative” (R2) and highlight its “effectiveness” (R3), “lightweight” design (R1, R2, R3), “unconventional training strategy” (R2), and ability to “[eliminate] the need for paired clean and noisy images” (R2, R3). However, the reviewers raised concerns regarding (i) comparisons with conventional denoising algorithms (e.g., BM3D) or supervised methods trained on synthetic noise (R1, R2), (ii) justification for the three-layer CNN architecture (R2), and (iii) limitations of the method (R3). (i) Our method is a strictly data-free denoising algorithm that relies solely on the noisy input image, eliminating the need for prior training data or additional information during inference. To ensure fair comparisons, we evaluated Noise2Detail against other self-supervised methods. Conventional denoising algorithms, such as BM3D, typically require the noise level as input, which provides explicit guidance on noise characteristics and simplifies the task but is often difficult or impossible to obtain in practice. In contrast, supervised methods depend on large datasets to generalize effectively, a requirement that is frequently impractical in biomedical imaging due to the scarcity and high cost of paired data. Furthermore, training on data artificially degraded with synthetic noise is often infeasible, as clean data is rarely available, and simulations frequently fail to capture the complexities of biological data. Prior work has discussed that dataset-based methods are sensitive to training data and underperform on out-of-domain data (ZS-N2N, Mansour and Heckel; Self2Self, Quan et al.), underscoring the value of our approach, which operates independently of external information beyond the input image. (ii) The Noise2Detail architecture builds on the ZS-N2N model, which also employs a three-layer design for lightweight denoising. It has been shown the model size can still be reduced at the cost of slight performance drop (ZS-N2N, Mansour and Heckel), while increasing depth risks overfitting and degrades performance. Our empirical analysis confirmed this pattern, revealing performance saturation with additional layers. Consequently, we selected the three-layer design as the optimal trade-off between computational efficiency and denoising quality, ensuring a lightweight model suitable for resource-constrained applications. (iii) While Noise2Detail delivers robust denoising across both synthetic and real noise, it may not achieve the same structural smoothness as Noise2Self in certain cases, where smoothness often results from blurring the input image: a common trait of overparameterized methods (see Fig. 4). However, Noise2Detail’s significantly lower computational cost makes it well-suited for resource-constrained environments, such as microscopy imaging, where near-real-time processing is critical (e.g., smart microscopy). By operating without prior training data or noise information, Noise2Detail enhances generalizability but sacrifices additional regularization that could further refine image quality. In scenarios where datasets with consistent degradation patterns are available, dataset-based methods employing Noise2Noise training may yield superior results through pre-training of larger models. We will follow the suggestions and evaluate the model with a wider variety of metrics and tasks in the journal extension of the manuscript.

Meta-Review

Meta-review #1

Your recommendation

Provisional Accept
If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

N/A

back to top

Lightweight Data-Free Denoising for Detail-Preserving Biomedical Image Restoration

Author(s):