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

TL;DR

Self-supervised biomedical image denoising often requires large clean datasets and heavy models, limiting practical use in modalities with scarce clean data. The authors propose Noise2Detail (N2D), an ultra-lightweight, data-free denoising pipeline built on Noise2Noise that uses a 3-layer CNN, pixel-shuffle based refinement, and a detail-recovery retraining to preserve fine structures. Key contributions include a compact 22k-parameter network, a three-stage denoising workflow that disrupts noise correlations and reintroduces detail, and extensive evaluation on CT and fluorescence microscopy demonstrating strong performance with low computational cost. This enables fast, accurate denoising in biomedical settings where both data and compute are at a premium.

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.

Figures (4)

• Figure 1: Overview of our proposed Noise2Detail denoising framework. Initially, $f_\theta$ is trained to denoise an image using a single-image Noise2Noise loss, but residual artifacts persist due to the limitations of the training (see the yellow arrow in the zoomed in insets). We then freeze the weights of the network, apply a sequential refinement to remove the background artifacts, and then retrain $f_{\theta^*}$ to enhance the foreground details, yielding the final denoised image.
• Figure 2: Illustration of the sequential enhancement process with a trained network $f_\theta$. For every $j \in J$, $f_\theta$ is applied twice—initially to denoise the pixel-shuffled downsampled image, and then to correct the grid-like artifacts caused by disrupted spatial correlations when reconstructing the original pixel layout.
• Figure 3: Our method is the only one that maintains the fidelity of the foreground while keeping the background smooth without artifacts.
• Figure 4: Examples of restoration given a noisy image of a zebrafish from a confocal microscope taken from FMD zhang2019poisson. N2V and N2S oversmooth and blur the details while our method maintains the structural fidelity.