GANscan: continuous scanning microscopy using deep learning deblurring.

Area of Science:

• Microscopy and Imaging Technologies
• Computational Pathology
• Artificial Intelligence in Medical Imaging

Background:

• Whole slide imaging (WSI) traditionally uses a "stop-and-stare" method, ensuring image clarity but resulting in lengthy acquisition times.
• High-speed scanning is crucial for large areas like pathology slides, but motion blur and defocusing are significant challenges.

Purpose of the Study:

• To develop an ultra-fast WSI acquisition method that overcomes the limitations of traditional "stop-and-stare" approaches.
• To leverage generative adversarial networks (GANs) for reconstructing sharp images from high-speed, continuous scans.

Main Methods:

• Developed GANscan, an acquisition method enabling continuous data capture during high-speed stage movement.
• Utilized generative adversarial networks (GANs) to restore sharp images from motion-blurred videos acquired during scanning.
• Implemented GANscan on a Zeiss Axio Observer Z1 microscope without specialized hardware, achieving reconstructions at speeds up to 5000 μm/s.

Main Results:

• GANscan achieves 30x higher throughput compared to conventional "stop-and-stare" WSI systems.
• The method successfully reconstructs crisp images from continuous scans, even with stage speeds up to 5000 μm/s.
• GANscan corrects for defocusing within a +/- 5 μm range and performs inference in under 20 ms/image on a consumer GPU.

Conclusions:

• GANscan offers a viable solution for ultra-fast whole slide imaging, significantly reducing acquisition times.
• The AI-driven approach effectively mitigates motion blur and defocusing artifacts inherent in high-speed microscopy.
• This technology has the potential to accelerate digital pathology workflows and large-scale imaging applications.