DeepLSR: a deep learning approach for laser speckle reduction

Taylor L. Bobrow, Faisal Mahmood, Miguel Inserni, Nicholas J. DurrarXiv 2018 · DOI · arXiv

TL;DR

DeepLSR is a deep learning framework that removes laser speckle noise by learning to map coherent, speckled images to speckle-free, incoherent LED-illuminated ground truths. Using a conditional generative adversarial network (cGAN), it significantly outperforms traditional denoising algorithms and hardware solutions, clearing the path for high-power, miniature laser illumination in applications like medical endoscopy.

What Problem Does This Paper Address?

Laser illumination is highly desirable for advanced imaging modalities due to its high power density, narrow spectral bandwidth, and ability to be delivered through tiny fiber optics. However, the coherence of laser light causes constructive and destructive interference, resulting in a grainy, high-contrast noise pattern known as laser speckle.

This speckle noise severely degrades image quality, preventing lasers from replacing bulky arc lamps or LEDs in clinical devices like endoscopes. Existing hardware solutions (like oscillating diffusers) add cost, size, and limit imaging speed. Meanwhile, traditional software denoising algorithms (like Non-Local Means or BM3D) tend to over-smooth images, wiping out critical high-frequency details and structural edges along with the noise.

Methods

The authors introduce DeepLSR (Deep Laser Speckle Reduction), which frames speckle removal as an image-to-image translation task.

  • Architecture: It utilizes a conditional Generative Adversarial Network (cGAN). A generator network learns to produce a speckle-free version of a laser-illuminated image, while a discriminator network tries to distinguish between the generated image and a real, LED-illuminated ground truth image.
  • Loss Function: The network is optimized using a combination of standard GAN loss and an $\ell_1$ loss. The $\ell_1$ loss forces the generator to maintain spatial correctness and structural accuracy relative to the ground truth without being overly sensitive to outlier pixel values (a common issue with $\ell_2$ loss).
  • Stabilization via Spectral Normalization: To prevent "mode collapse"—a frequent failure mode in GANs where the network fails to learn the full diversity of the data—the authors replaced standard batch normalization with spectral normalization. This allowed a single network to stably learn transformations across vastly different image domains (e.g., plastic toys vs. biological tissue).

Datasets

The authors captured a custom dataset of nearly 3,000 perfectly registered image triads. Each scene was captured under three illumination conditions using identical camera geometry: (1) raw RGB laser, (2) laser with a commercial optical laser speckle reducer (oLSR), and (3) an incoherent white-light LED.

The dataset consists of two main categories:

  1. Assorted Objects: 113 household and laboratory items imaged at various positions (1,533 total images) to provide a wide variety of shapes, textures, and reflectances.
  2. Porcine Tissue: 449 images of ex-vivo porcine gastrointestinal tissue (esophagus, stomach, and intestine) sourced from three different animals (1,347 total images).

For testing, the authors strictly reserved images from 6 completely unseen objects and tissue samples from an entirely separate animal not used during training.

Key Insights and Ideas

  • True Incoherent Ground Truth: Unlike previous deep learning denoising attempts that rely on synthetic noise models or averaged frames, DeepLSR trains on actual physical ground truths (LED illumination), capturing the true physics of the environment.
  • Edge Preservation over Blurring: Traditional filters assume noise is purely high-frequency and stochastic, leading to blurred edges. DeepLSR learns the specific statistical distribution of laser speckle, allowing it to suppress the artifacts while keeping text, lines, and fine tissue structures sharp.
  • Cross-Channel Independence: Analysis of individual color channels revealed that DeepLSR successfully despeckles images even when restricted to a single channel (like the red channel), proving it does not rely on borrowing structural information from adjacent color channels.

Key Results

  • Denoising Performance (Tissue): On gastrointestinal tissue, DeepLSR reduced speckle noise by 6.4 dB, drastically outperforming Non-Local Means (2.9 dB) and BM3D (3.0 dB).
  • Hardware-Software Synergy: When DeepLSR was combined with a physical optical speckle reducer (DeepLSR+oLSR), it achieved a massive 9.4 dB noise reduction.
  • Structural Fidelity: DeepLSR demonstrated a 25-26% improvement in Structural Similarity Index (SSIM) compared to raw laser images.
  • Speed: Once trained, the network can process images at 27 frames per second (fps) on an Nvidia P100 GPU, making it viable for real-time video endoscopy.

Limitations and Future Work

While DeepLSR maintains high-frequency details much better than traditional algorithms, it still causes a minor reduction in spatial resolution (~17% based on a slanted-edge modulation transfer function test).

Additionally, the model's generalizability depends heavily on its training data; changes in laser wavelengths, working distances, or camera lenses might require retraining. The authors suggest that future work could leverage forward physical models to generate synthetic training data or utilize unsupervised networks to eliminate the strict requirement for perfectly registered training image pairs.

Summary updated

Abstract

Speckle artifacts degrade image quality in virtually all modalities that utilize coherent energy, including optical coherence tomography, reflectance confocal microscopy, ultrasound, and widefield imaging with laser illumination. We present an adversarial deep learning framework for laser speckle reduction, called DeepLSR (https://durr.jhu.edu/DeepLSR), that transforms images from a source domain of coherent illumination to a target domain of speckle-free, incoherent illumination. We apply this method to widefield images of objects and tissues illuminated with a multi-wavelength laser, using light emitting diode-illuminated images as ground truth. In images of gastrointestinal tissues, DeepLSR reduces laser speckle noise by 6.4 dB, compared to a 2.9 dB reduction from optimized non-local means processing, a 3.0 dB reduction from BM3D, and a 3.7 dB reduction from an optical speckle reducer utilizing an oscillating diffuser. Further, DeepLSR can be combined with optical speckle reduction to reduce speckle noise by 9.4 dB. This dramatic reduction in speckle noise may enable the use of coherent light sources in applications that require small illumination sources and high-quality imaging, including medical endoscopy.
BibTeX
@misc{bobrow2018deeplsr,
  author = {Taylor L. Bobrow and Faisal Mahmood and Miguel Inserni and Nicholas J. Durr},
  title = {{DeepLSR: a deep learning approach for laser speckle reduction}},
  year = {2018},
  eprint = {1810.10039},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1810.10039},
}
Figure