Phase imaging catches sight of obscured objects

26 Jul 2023

UCLA diffractive optical network could allow clearer views of cells in tissues.

Imaging objects that are partially or completely obscured by structures that diffuse the light from those objects, as is the case in many biological systems, is a challenge that engineers have met with both optical and computational solutions.

Quantitative phase imaging (QPI), in which the key visual information is revealed by the change in optical path-length and hence phase shift that takes place when photons pass through biological tissue, is one solution. QPI has successfully been applied to non-invasive imaging of certain specimens with high sensitivity and resolution.

However, QPI traditionally involves large-scale computational resources for image reconstruction and phase retrieval algorithms, factors which ultimately inhibit the frame rates achievable. QPI can also have difficulties accounting for random scattering, often a feature of tissue systems.

A project at the UCLA lab of Aydogan Ozcan has now developed a new methodology for quantitative phase imaging of objects that are completely covered by random, unknown phase diffusers. The work was reported in Light: Advanced Manufacturing.

The Ozcan lab has previously carried out research into how sets of fabricated diffractive surfaces can act on light arriving from behind scattering media.

At SPIE Photonics West 2023 (as reported in Optics.org's Show Daily pdf) Ozcan described how this approach can allow all-optical image reconstruction. An algorithm processing light that arrives after passing through sequence of diffractive surfaces can be trained, by observing how those surfaces scramble known targets and objects, to take unfamiliar inputs and reconstruct the original unaltered target.

The new study created a QPI diffractive network able to convert the phase information of an input sample into a quantitative output intensity distribution even in the presence of materials or objects diffusing the light from the target.

On-chip integrated deep neural network for biomedicine

UCLA's device involves a 4-layer QPI diffractive deep neural network (D²NN), with the layers separated by a distance of 2.67 x the light's illumination wavelength.

"During its training, various randomly generated phase diffusers were utilized to build resilience against phase perturbations created by random unknown diffusers," commented the UCLA team. "After the training, which is a one-time effort, the diffractive layers can then perform all-optical phase recovery and quantitative phase imaging of unknown objects that are entirely hidden by unknown random diffusers."

After being trained using standard representations of handwritten numerals, the new QPI diffractive network could successfully image different numerals through new random phase diffusers, without the need for a digital image reconstruction algorithm.

UCLA also investigated the impact of factors such as the number of diffractive layers and the trade-off between image quality and output energy efficiency, finding that deeper diffractive optical networks could generally outperform shallower designs. Its QPI diffractive network can be physically scaled to operate at different parts of the electromagnetic spectrum without redesigning or retraining its layers, according to the team.

"The QPI designs can potentially be integrated with existing CCD/CMOS image sensors, by fabricating the thin diffractive layers on top of the active area of an image sensor array," said the project in its published paper.

"Such an on-chip integrated D²NN can be placed at the image plane of a standard microscope to convert it into a diffractive QPI microscope. This diffractive computing framework for phase retrieval and QPI through random unknown diffusers can potentially advance label-free microscopy and sensing applications in biomedical sciences, among other fields."