Self-Supervised Elimination of Non-Independent Noise in Hyperspectral Imaging.

Guangrui Ding, Chang Liu, Jiaze Yin, Xinyan Teng, Yuying Tan, Hongjian He, Haonan Lin, Lei Tian, Ji-Xin ChengNewton 2026 · OpenReview

Sure, here's a blog-style summary of the paper "Self-Supervised Elimination of Non-Independent Noise in Hyperspectral Imaging":

TL;DR

This paper introduces SPEND, a novel self-supervised deep learning method for denoising hyperspectral images, specifically addressing the challenge of non-independent noise that plagues advanced imaging techniques like SRS and MIP microscopy. SPEND achieves significant signal-to-noise ratio improvements, enabling more accurate chemical analysis and visualization of biological samples.

What Problem Does This Paper Address?

Hyperspectral imaging is a powerful tool for identifying molecules based on their unique spectral fingerprints. However, the raw data from advanced techniques like Stimulated Raman Scattering (SRS) and Mid-Infrared Photothermal (MIP) microscopy is often contaminated by sophisticated noise. Traditional denoising methods struggle with this noise because they assume it's independent and identically distributed (i.i.d.). In reality, the complex signal acquisition processes in these microscopes lead to spatially correlated and spectrally varied noise, which significantly degrades image quality and hinders accurate chemical analysis, especially for low-concentration molecules. This paper aims to tackle this challenge by developing a denoising method that can effectively handle this non-i.i.d. noise.

Methods

The core of this work is the Self-Permutation Noise2Noise Denoising (SPEND) framework. SPEND is a self-supervised deep learning approach that doesn't require ground truth data. Here's how it works:

  • Permutation Strategy: SPEND identifies an axis (spectral or spatial) where the noise exhibits the highest fluctuation (lowest correlation). By permuting the data along this axis, it creates two substacks that are treated as independent measurements of the same sample. This is crucial for the "noise-to-noise" training paradigm.
  • Noise-to-Noise Training: Using the permuted substacks, SPEND trains a U-Net-based neural network to learn the noise characteristics and remove them. The key idea is that by having two independently generated noisy versions of the same underlying signal, the network can learn to predict and subtract the noise.
  • U-Net Architecture: A standard U-Net architecture is employed for its effectiveness in image segmentation and restoration tasks, adapted for 3D hyperspectral data.

Datasets

The paper utilizes hyperspectral SRS and MIP microscopy datasets of biological samples, including:

  • OVCAR5 cancer cells
  • U87 cancer cells
  • HPG-treated SJSA-1 cancer cells
  • C. albicans fungal cells

These datasets were used for both training the SPEND model and evaluating its performance against existing methods. The authors do not explicitly mention creating new benchmark datasets, but their work demonstrates the effectiveness of SPEND on these complex biological imaging datasets.

Key Insights and Ideas

  • Non-Independent Noise is Ubiquitous: The paper convincingly demonstrates that noise in advanced hyperspectral imaging techniques (SRS, MIP) is far from i.i.d. It's spatially correlated (due to lock-in amplifier delays and sample perturbations) and spectrally varied (due to thermal effects and laser noise). This finding is critical for understanding the limitations of existing denoising methods.
  • Permutation is Key to Self-Supervision: The clever use of data permutation along the axis of highest noise fluctuation is the linchpin of SPEND. It effectively creates the "independent" views needed for noise-to-noise training without requiring multiple acquisitions of the same sample.
  • SPEND Unlocks Deeper Chemical Analysis: By significantly improving the signal-to-noise ratio, SPEND enables more accurate chemical unmixing, allowing for the identification and mapping of low-concentration biomolecules in both "fingerprint" (crowded spectral regions) and "silent" (less crowded spectral regions) spectral windows.
  • Universality Across Modalities: The success of SPEND on both SRS and MIP microscopy suggests its broad applicability to other imaging modalities that suffer from similar complex noise characteristics.

Key Results

  • 8-fold Signal-to-Noise Ratio Improvement: SPEND achieves a substantial increase in SNR, making previously unresolvable features visible.
  • Reduced Spectral Distortion: SPEND demonstrates less spectral distortion compared to other methods like BM4D, preserving spectral fidelity crucial for chemical identification.
  • Enhanced Chemical Unmixing: SPEND significantly improves the accuracy of chemical unmixing, enabling better differentiation of lipids, proteins, and cholesterol, even in complex biological environments.
  • Improved Spatial Resolution: In MIP imaging, SPEND helps achieve a resolution close to that of high-power imaging, surpassing traditional methods.
  • Fast Inference: The denoising process for an entire image stack takes approximately 14 seconds, making it practical for real-time or near-real-time applications.

Limitations and Future Work

  • Training Data Acquisition: The authors acknowledge that achieving the "independent measurements" required for N2N-based frameworks can pose specific demands on data acquisition strategies. They recommend adhering to Nyquist sampling rates tailored to the specific features of interest.
  • Misfitting Problems: While SPEND is robust, misfitting can still occur. The authors suggest that expanding the training dataset size and complexity could help refine the model's accuracy and generalizability.
  • Future Directions: The paper suggests that SPEND's capabilities could extend beyond hyperspectral imaging to enhance SNR in videos and other lock-in-based 3D imaging modalities that face similar complex noise challenges. This includes techniques like stimulated Raman photothermal microscopy and Brillouin scattering microscopy.

Summary updated

BibTeX
@inproceedings{ding2026self,
  author = {Guangrui Ding and Chang Liu and Jiaze Yin and Xinyan Teng and Yuying Tan and Hongjian He and Haonan Lin and Lei Tian and Ji-Xin Cheng},
  title = {{Self-Supervised Elimination of Non-Independent Noise in Hyperspectral Imaging.}},
  booktitle = {Newton},
  year = {2026},
  url = {https://openreview.net/forum?id=qWBUv1nZZE},
}
Figure