An artificial neural network can transform low-resolution microscopic images of samples into high-resolution images, revealing more details of the sample, which could be crucial for pathology and medical diagnostics. [Image: Ozcan Research Group/UCLA]
Since the advent of digital image processing, scientists have sought new ways to automate the process and gain new insights into large data sets. In two related studies, a team of researchers at the University of California at Los Angeles (UCLA), USA, has taught artificial neural networks how to reconstruct holographic images and enhance the spatial resolution of optical microscopy.
According to OSA Fellow Aydogan Ozcan and his UCLA colleagues, the deep-learning technique results in sharper holographic images with less computational effort, and could lead to various types of automated medical diagnostics.
Convolutional neural nets
Both sets of experiments involved so-called convolutional neural networks, a type of machine learning inspired by the organization of the brain’s visual cortex. In the first study (Light Sci. Appl., doi: 10.1038/lsa.2017.141), Ozcan’s group taught the neural networks how to recover phase information and reconstruct complex-valued images based on intensity-only holograms. The training of the networks involved comparing a statistical transformation of a complex-valued image to a standard phase-retrieval algorithm.
Once the neural networks had learned their tasks, the UCLA group performed holographic imaging of three types of tissue samples commonly tested in a medical laboratory: red blood cells, Pap smears for cervical cancer screening, and breast tissue. The neural networks reconstructed the images with significant reduction of out-of-focus interference artifacts.
Sharpening up bright-field microscopy
Ozcan and his colleagues also applied these convolutional neural networks to conventional bright-field optical microscopy (Optica, doi: 10.1364/OPTICA.4.001437). In this case, the researchers trained the neural network to relate low-resolution images of specimens to high-resolution images. After a single session of training, the team had the network process microscopic images of breast and lung tissue. The network returned enhanced images in less than a second of processing time for each image.
These neural networks run on a regular laptop computer, not a specialized supercomputer. The researchers hope such deep-learning tools can help pathologists detect tiny tissue abnormalities in medical specimens.