Duke University AI-powered microscope speeds up imaging and data processing...

29 Oct 2025

...while another Duke project uses embedded GPU for blood profiling in point-of-care diagnostics.

A project at Duke University has demonstrated how commercial AI models could accelerate microscopic examination of 2D materials and analysis of the results.

The Duke Lab of Haozhe "Harry" Wang has developed ATOMIC, or Autonomous Technology for Optical Microscopy & Intelligent Characterization, a platform that "can analyze materials as accurately as a trained graduate student in a fraction of the time," according to Duke.

"The system we've built doesn’t just follow instructions, it understands them," Wang said. "ATOMIC can assess a sample, make decisions on its own and produce results as well as a human expert."

The results, published in ACS Nano, may point to a new era of autonomous research, where AI systems work alongside humans to design experiments, run instruments and interpret data.

Wang's group studies two‑dimensional crystals only a few atoms thick, promising candidates for next-generation semiconductors, sensors and quantum devices. Their exceptional electrical properties and flexibility make them ideal for electronics, but fabrication defects can compromise these advantages.

"To characterize these materials, you usually need someone who understands every nuance of the microscope images," Wang said. "It takes graduate students months to years of high-level science classes and experience to get to that point."

To speed up the process, Wang's team linked an off‑the‑shelf optical microscope to ChatGPT, allowing the model to handle basic operations like moving the sample, focusing the image and adjusting light levels.

Layered on top was Meta's SAM, or Segment Anything Model, an open‑source vision model designed to identify discrete objects, which in the case of these materials would include regions containing defects and pure areas.

Although SAM could recognize regions within the microscopic images, it struggled with overlapping layers, a common issue in materials research. The team added a topological correction algorithm to refine those regions, isolating single-layer areas from multilayer stacks.

In trials across a range of 2D materials, the AI microscope matched or outperformed human analysis, identifying layer regions and subtle defects with up to 99.4 percent accuracy. It maintained this performance even with images captured under imperfect conditions, and in some cases spotted imperfections invisible to the human eye.

"We still need humans to interpret what the AI finds and decide what it means," commented Wang. "But once you have a partner that can complete weeks of analysis in mere seconds, the possibilities for new discoveries are exponential."

Profiling thousands of blood cells at a time

Another Duke group, the BIOS Lab headed by Adam Wax, has developed a real-time and high-throughput quantitative phase microscopy (QPM) processing algorithm deployed on an embedded GPU system, enabling rapid blood profiling for point-of-care diagnostics.

QPM is a holographic imaging technique able to reveal cell morphology, particularly valuable for analyzing blood samples. But the large amount of data collected by high-throughput QPM, potentially from thousands of cells at a time, can take several hours to process on a regular CPU.

The BIOS team created a new real-time pipeline to reconstruct and analyze the high-throughput QPM data of red blood cells at a rate of 1200 cells per second, and implemented its algorithm on a NVIDIA Jetson Orin Nano, a commercial embedded GPU platform that only costs $249.

Results published in Biophotonics Discovery showed that when applied to blood samples the new pipeline achieved high structural similarity with low deviation relative to a traditional processing methods.

The holographic cytometry approach also provided several additional morphological parameters of red blood cells not available from traditional blood screening - parameters well suited to machine learning applications.

"QPM has long held potential to provide detailed information about biological cells, but the technique has yet to find widespread clinical use, often due to the cost or complexity in processing the imaging data," commented Adam Wax.

"We have shown not only a high-throughput means for profiling thousands of cells at a time but also for rapidly processing and analyzing the information. This may be the missing step needed to bring QPM to the clinic."