🧹 Noise2Detail: Lightweight Data-Free Denoising for Detail-Preserving Biomedical Image Restoration

[Paper] [Colab demo] [BibTeX]

Tomáš Chobola*  Julia A. Schnabel  Tingying Peng* 
Technical University of Munich Helmholtz AI King’s College London

* Corresponding author

Accepted early to MICCAI 2025.

Overview

• Ultra-Lightweight Model Design: Introducing a computationally efficient model that achieves fast denoising with minimal memory demands, balancing inference speed and high-quality image restoration.
• Noise2Detail (N2D) Pipeline: Proposes an innovative multistage denoising approach within the Noise2Noise framework, disrupting noise correlations to generate intermediate smooth structures and refining them to recover fine details from noisy inputs.
• Dataset-Free Denoising: Eliminates the need for clean reference images or explicit noise modeling, making it ideal for biomedical imaging where clean training data is scarce due to rare and complex modalities.
• Superior Performance with Low Resources: Outperforms existing dataset-free denoising techniques while requiring significantly fewer computational resources, enabling practical use in real-world applications.

Citing Noise2Detail

Please consider citing our paper if our code are useful:

@inproceedings{Chobola2025,
  title = {Lightweight Data-Free Denoising for Detail-Preserving Biomedical Image Restoration},
  author = {Chobola,  Tomáš and Schnabel,  Julia A. and Peng,  Tingying},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year={2025}
}