Photo credit: www.sciencedaily.com
Quantitative phase imaging (QPI) stands as a significant microscopy method for exploring the intricacies of cells and tissues. Although initial biomedical uses of QPI have emerged, enhancements in both the speed of acquisition and the quality of images are necessary for broader acceptance in the medical field. Researchers from the Center for Advanced Systems Understanding (CASUS) based in Görlitz, alongside colleagues from Imperial College London and University College London, propose an innovative approach that utilizes chromatic aberration—a phenomenon typically responsible for diminishing image quality—to generate high-quality images using standard microscopes. With the application of a generative AI model, they demonstrate that a single exposure is sufficient to achieve the image quality required for promising biomedical applications.
The findings were showcased in late February at the 39th Annual Conference on AI presented by the Association for the Advancement of AI (AAAI) in Philadelphia, with the peer-reviewed conference paper expected to be published later in March.
Labeling biological samples with dyes offers crucial insights but poses challenges that can limit its use in clinical diagnostics due to the time and expense involved. Consequently, there has been increasing interest in label-free microscopy techniques such as QPI. This method investigates not only the amount of light that is absorbed or scattered by the samples but also how the phase of the light is altered as it passes through, changes that correlate with the sample’s thickness, refractive index, and various structural attributes. While conventional QPI equipment can be quite costly, computational QPI techniques significantly reduce these costs.
A key computational QPI approach involves solving the Transport-of-Intensity Equation (TIE), which generates an image based on recorded phase alterations. Although this method integrates well with existing optical microscope setups and typically produces good-quality images, it often necessitates multiple acquisitions at different focal depths to minimize artifacts. The need for through-focus stacks complicates this method, making it technically demanding and less viable in clinical applications.
Utilizing Chromatic Aberration
“Our method operates on principles similar to TIE but requires only a single image, thanks to an innovative synergy between physics and generative AI,” states Prof. Artur Yakimovich, head of a CASUS Young Investigator Group and lead author of the conference presentation. This technique derives phase shift information from a single exposure by leveraging chromatic aberration. Standard microscope lens systems cannot perfectly converge all wavelengths of white light, which results in different focal points for red, green, and blue (RGB) light. By capturing the phase shifts of these wavelengths separately using a standard RGB detector, the researchers can create a through-focus stack that enhances computational QPI, transforming a limitation into a valuable asset, Yakimovich notes.
However, the challenge remains that the focal distance between red and blue light is minimal. Gabriel della Maggiora, a PhD student at CASUS and co-lead author of the research, explains that traditional TIE solutions do not yield meaningful outcomes in this context. “We hypothesized that we could apply artificial intelligence. This idea turned out to be crucial,” says della Maggiora, who further details that after training a generative AI model on an open-access database of 1.2 million images, the model was capable of retrieving phase information even from the limited data offered by the recordings.
Method Validated on Clinical Samples
The researchers built on a generative AI model for enhancing image quality, known as the Conditional Variational Diffusion Model (CVDM), introduced the previous spring. This model, part of a specific category of generative AI called diffusion models, has been shown to require significantly less computational effort during training compared to other models while producing comparable—or superior—results. Leveraging the CVDM strategy, della Maggiora and the team devised a novel diffusion model tailored for quantitative data, allowing them to finally implement computational QPI utilizing chromatic aberrations. They validated this AI-driven approach using standard brightfield microscopes equipped with commercially available color cameras to analyze real-world clinical specimens, such as red blood cells in a human urine sample. This model excelled in capturing the distinctive donut shape of the red blood cells, a feat that established computational TIE methods failed to achieve. An additional benefit of this new generative AI-based quantitative phase imaging is the remarkable reduction of cloud artifacts in the resulting images.
The Yakimovich group’s “Machine Learning for Infection and Disease” initiative is dedicated to developing advanced computational microscopy techniques that can be swiftly integrated into clinical practice. The potential applications, particularly in diagnostics, are expansive. Among the methodologies adopted is generative AI, with a primary focus on minimizing the risk of AI-generated inaccuracies, commonly referred to as hallucinations. The integration of physics-based elements into this AI framework indicates a promising path forward, as demonstrated by the advancements in quantitative phase imaging.
Source
www.sciencedaily.com