Neural Wavefront Shaping for Aberration Correction and Diffraction-Limited Imaging Through Scattering Media

NeuWS uses untrained neural networks to recover both an object’s brightness and the optical aberrations from random phase‐modulated intensity measurements, enabling guidestar-free, diffraction-limited imaging through static or dynamic scattering media in minutes.

Background

In many applications—ranging from biological microscopy and ophthalmology to industrial inspection—the ability to see through blurry or cloudy materials is critical. When light passes through scattering layers or imperfect lenses, it becomes distorted, leading to blurred or low-contrast images. High-resolution imaging in these scenarios enables researchers and clinicians to observe fine details, diagnose conditions, and perform precise measurements. Traditional correction methods often rely on hardware solutions or invasive markers, but they can be impractical for delicate samples or rapidly changing environments. As a result, there is a growing need for flexible approaches that can restore image clarity without disrupting the scene or relying on external guide targets.

Existing techniques that exploit precise light modulation or use reference beacons face several drawbacks. Many require a known target—often a fluorescent bead or reflective point—and sequentially scan each position to infer distortions, which can be slow and intrusive. Others assume the scene or distortion is static, making them unsuitable for moving objects or dynamically changing materials. Moreover, complex calibration procedures and specialized illumination setups can limit their use outside controlled laboratory settings. These constraints result in long acquisition times and reduced applicability when real-time imaging or gentle handling of samples is needed.

Technology

NeuWS captures a series of intensity images by projecting random phase patterns onto a spatial light modulator and imaging through a scattering medium. Each snapshot is viewed as the object’s brightness blurred by the effect of the phase modulation in the aperture. Instead of using physical guidestars or raster scanning, it relies on two untrained coordinate-based neural networks—one that maps spatial (and time) coordinates to phase corrections and another that predicts object brightness and motion via a neural texture map. By matching predicted and measured images and automatically tuning the networks, the system produces diffraction-limited views of hidden objects in minutes on a GPU.

Traditional methods that try to correct light distortions often require controlled illumination, reference targets or point-by-point scanning and struggle with moving scenes. This approach reframes the problem as a joint estimation of both the scene and the distortions, using coordinate-based neural networks to create smooth, continuous maps of phase errors and object appearance that adapt over time without external calibration. As a result, it cuts the number of measurements, handles changing scattering media and moving objects, and yields high-resolution reconstructions through complex materials—representing a major step forward in non-invasive optical imaging.

Benefits

  • Guidestar-free operation, removing the need for fiducial markers or point scanning
  • Diffraction-limited imaging through both static and dynamic scattering media
  • Joint estimation of object brightness and optical aberrations via untrained neural networks
  • Reduced measurement count compared to conventional wavefront-shaping methods
  • Capability to reconstruct dynamic scenes under time-varying aberrations
  • Post-capture computational correction that bypasses real-time optical latency
  • Generalizable framework applicable to diverse aberration types and scene complexities
  • Reconstruction achieved in minutes using GPU-accelerated processing

Applications

  • Biomedical tissue imaging
  • Retinal diagnostic imaging
  • Scattering sample microscopy
  • Industrial optical inspection
  • Underwater vision enhancement