Summary:
UCLA researchers in the Department of Electrical Engineering have developed a novel deep neural network that generates speckle- and artifact-free high-quality images at different sample depths from a single hologram. The resulting images are equivalent to bright-field images taken throughout a 3D sample.
Background:
Digital holographic microscopy allows imaging and reconstruction of objects in 3D with one single measurement. However, due to the coherent light source and reconstruction methods used in digital holography, the quality of images suffers. Specifically, the contrast and the noise are problematic. Traditional bright-field microscopy uses incoherent light sources and takes many images at different depth of field in order to reconstruct an object in 3D. The image quality, however, is far superior and free of speckles or artifacts. Therefore, novel methods that raise the image quality of digital holography to bright-field standards while still maintaining the simplicity of holographic measurement are needed in practical applications.
Innovation:
A novel method of improving the image quality of digital holographic microscopy was developed using deep learning. A generative adversarial network (GAN) is trained using back-propagated holographic images and their matching bright-field images at each depth of field. After just one training, the resulting deep neural network can output high quality, speckle- and artifact-free images from a single hologram at all depths of field that match the quality of bright-field images at those depths. This approach brings the best of both worlds by fusing the simplicity of holographic measurements and the high image quality of traditional bright-field microscopy.
Pulication:
Cross-Modality Deep Learning Brings Bright-Field Microscopy Contrast to Holography
Applications:
Advantages:
UCLA Case No. 2019-464