From handicap to assets: AI approach leverages optical phenomena to generate better microscopic images

Quantitative phase imaging (QPI) is a microscopy that is widely used to investigate cells. Previous biomedical applications based on QPI have been developed, but both acquisition speed and image quality need to be improved to ensure widespread reception.

Scientists at the Görlitz-based Centre for Advanced Systems Understanding (CASUS) at Helmholtz-Sentulm Dresden-Rossendorf (HZDR), as well as Imperial College London and University College London, suggest that they utilize an optical phenomenon known as chromatic anomalies, usually degrades image quality.

By using the generated AI model, only one exposure is required to obtain the image quality required to make QPI attractive for biomedicine applications. The team presented work in late February at the 39th Annual Meeting on AI by the Association for Progressive AI in AI (AAAI 2025) held in Philadelphia this year. The corresponding meeting forms are available on the ARXIV preprint server.

Labeling biological samples with dyes or other agents reveals valuable insights. However, this approach has several drawbacks that hinder widespread use in clinical diagnosis. It takes time and requires expensive equipment and reagents.

Therefore, research over the past few years has focused on microscopy without specific labels, such as QPI. Here it is not only interesting, but also the size of light absorbed or scattered from the sample. Using scattering information, QPI captures how a sample shifts the phase of light passing through it. This is a change directly related to its thickness, refractive index, and other structural properties. QPI requires very expensive equipment, but no calculation QPI is required.

One of the most prominent computational QPI approaches is to solve the Strength Transport Equation (TIE). This differential equation allows you to calculate an image of a sample based on the recorded phase changes. This approach can be easily integrated into existing optical microscope setups, resulting in good quality images.

On the downside, TIE methods often require multiple acquisitions with different focus distances to remove artifacts. This type of tie-based QPI is often unfeasible in clinical settings, as it is time-consuming and technically demanding to work with through-focus stacks.

Uses color abnormalities

“Our approach relies on similar principles, but only one image is needed for the clever combination of physics and generative AI,” says Professor Artur Yakimovich, leader of the Casus Young Investigator Group and author of the work presented at the AAAI conference.

Information about phase shifts caused by biological specimens is not from additional exposure to other focal lengths. Thanks to a phenomenon known as chromatic anomalies, it is also possible to generate a focused stack from a single exposure.

Most lens systems in microscopes cannot fully bring white light at all wavelengths (polychromatic) to one point of convergence. This is a handicap that can only be modified by highly specialized lenses. This means that, for example, the focal lengths of red, green, and blue (RGB) light differ slightly in focal lengths.

“By using traditional RGB detectors, we can individually record the phase shifts of these three wavelengths, we can build a through focus stack that facilitates the calculation QPI that transforms handicap into assets,” explains Yakimovich.

“There is one challenge when achieving QPI using chromatic anomalies: the distance between the focus of red light and blue light is very small,” says Dr. Gabriel Della Maggiola. One of two lead authors of CASUS student and publication. Resolving a tie using standard methods will not yield meaningful results.

“We then reasoned that we could use artificial intelligence. After all, this idea proved to be critical,” adds Della Maggiora. “After training the generated AI model on an open access dataset consisting of 1.2 million images, the model was able to obtain phase information despite relying on very limited data input from the records.”

Methods verified in real-world clinical specimens

The team utilized the Conditional Variation Spreading Model (CVDM), a generation AI model for image quality improvements presented last spring. It belongs to a specific family of generated AI models named diffusion models. Developers emphasize that training CVDM requires significantly less computational effort than training other diffusion models, but the results are the same or even better.

Using the CVDM strategy, Della Maggiora and her colleagues have developed a new diffusion model that can be applied to quantitative data. This model allows you to finally achieve computational QPI based on chromatic anomalies.

They validated the generation AI-based approach, for example, using a common Brightfield microscope equipped with a commercially available color camera to create microscopic images from real-world clinical specimens. Analyzing red blood cells in human urine samples, this method was able to reveal the donut-like shape of these cells, but another computational kinetic-based approach was not.

An additional advantage was that cloud artifacts were not virtual in images calculated with the new generation AI-based quantitative phase imaging variant.

Yakimovich Group’s Machine Learning for Infection and Diseases is developing new computational techniques for microscopy that can be applied immediately in clinical settings. For example, the diagnostic potential is enormous. The technique used is generation AI. Generating AI tends to generate hallucinations, so the main focus of the group is to reduce them.

Incorporating physically-based elements is an important approach here. As the AI-based quantitative imaging example shows, this approach is extremely promising.