Deep-3D microscope: 3D volumetric microscopy of thick scattering samples using a wide-field microscope and machine...

Area of Science:

• Biomedical Imaging
• Microscopy
• Machine Learning

Background:

• Confocal microscopy offers high-resolution 3D imaging but is costly and slow.
• Traditional wide-field microscopes are accessible but lack confocal's resolution and depth.
• Limitations include high cost, slow speed, and photobleaching in confocal systems.

Purpose of the Study:

• To develop a cost-effective method for achieving confocal microscopy image quality.
• To leverage machine learning to enhance wide-field microscopy capabilities.
• To enable high-resolution 3D volumetric imaging using standard laboratory microscopes.

Main Methods:

• Utilized a 3D generative adversarial network (GAN) for image reconstruction.
• Trained the neural network using image pairs from wide-field and confocal microscopes.
• Acquired multiple images with varying focus settings from a wide-field microscope.

Main Results:

• Reconstructed 3D volumetric images rivaling confocal microscopy in resolution and contrast.
• Demonstrated accurate reconstruction of thick (∼40 microns) and scattering samples.
• Showcased generalization ability to unseen data and stable reconstruction results.

Conclusions:

• Machine learning can empower traditional wide-field microscopes to achieve confocal imaging performance.
• This approach democratizes high-quality 3D imaging, making it accessible to more laboratories.
• Learning-based microscopy offers a viable alternative to expensive, specialized confocal systems.