低成本自动驾驶显微镜或能加速感染诊断 - AI与医疗健康

round the world, diseases such as malaria and tuberculosis (TB) continue to plague millions of people, causing severe illness and death. In recent years, scientists developed more effective drugs against many of these infectious diseases. However, one of the largest hurdles in curbing disease cases remains the ability to detect them.

“Drugs only work in combination with diagnostics because you cannot just take a drug and give it to people without diagnostics,” said Manu Prakash, a biologist and engineer at Stanford University.

Growing up in India, Prakash saw the toll of malaria firsthand. In his research group, he develops tools to address technology limitations in low-resource settings, what Prakash calls “frugal science.”

His team’s most recent project tackles the diagnostic challenges in rural and low-income communities. With collaborators across countries, he and his team developed a self-driving microscope trained on real-world data that is affordable and designed for ease of use.

Improving Upon Gold-Standard Microscopy for Better Diagnosis

Prakash’s focus on low-cost microscopy began with the idea behind the Foldscope—a low-cost, pocket-sized microscope made out of cut and folded paper with optical lenses—that he and this team developed in 2010. With this project proving that developing affordable microscopes was possible, his team turned to applying this concept to diagnostics.

A major choke point in diagnosing infectious diseases are tests that can detect the presence of a pathogen accurately and rapidly and are affordable for the communities that need them. Prakash said that a current problem with many available subsidized tests is that they are not sensitive or specific enough to identify people with low pathogen loads in their samples, leading to these people not receiving treatment and allowing them to spread the infection.

Microscopy, on the other hand, is a gold-standard diagnostic method, and the consumables for it, glass slides, are inexpensive. However, even with experienced community health workers, a person can reasonably review only a couple thousand cells in 30 minutes; a single drop of blood may contain 20 million cells. Additionally, fatigue due to high test volumes reduces microscopists’ identification accuracy.

Automated microscopes, on the other hand, can survey roughly two million cells in two minutes, greatly increasing the diagnostic capacity. Yet existing automated microscopes are too expensive for communities most affected by infectious diseases. To tackle this problem, Prakash and his team created a prototype of a modular, artificial intelligence (AI)-powered self-driving microscope in 2015 that learned how to detect parasites from real-world samples. Although the technology for self-driving capability and AI wasn’t widely available at the time, Prakash was confident that these fields would grow.

The team also decided to take a page from astronomy and incorporate spectroscopy into their design. Normally, Giemsa stain, a fluorescent dye, helps health workers visually discriminate red blood cells from immune cells and parasites, since it latches on to nucleic acids. Yet a different nucleic acid-binding dye, DAPI, emits distinct wavelengths of light depending on whether it is bound to DNA or RNA. In 2016, Prakash and engineer Hongquan Li, who was his student at the time, realized that the different ratio of DNA to RNA in malaria parasites led to a measurable shift in DAPI’s fluorescence spectra. They realized that they could use this shift to distinguish platelets from otherwise similarly sized and similarly stained parasites, improving diagnosis. Incorporating spectroscopy into the diagnostic workflow also allowed the team to use simpler, less expensive lenses in their microscope, further reducing the cost.

However, the team had more lessons to learn as they designed their microscope. For example, initially, the researchers 3D printed the components of their microscopes. In 2016, when they first took these models into clinics in rural parts of Africa—where they experienced everything from sandstorms to intense rain alongside brutal heat and humidity—they realized that they needed to make the microscope more robust to avoid failing in these conditions. So, Prakash and his team switched to using machined parts made from aluminum that provided more durability and were still affordable at scale.

Collaboration Expands the Diagnostic Potential for Automated Microscopy

With these structural advances made to their microscope, Prakash and his team turned to training their self-driving microscope’s AI model. “One of the rules that we apply for all our work is we actually collect all the data in the field, in real-world settings,” he said, explaining that discrepancies in lab-generated samples would make the model not generalizable to field settings. Prakash and his colleagues visited clinics across Africa to collect blood samples and scan them with their microscope to train the model to identify features of the parasite that causes malaria.

As Prakash and his collaborators began testing their parasite-trained microscope in the field, a group in Nepal contacted Prakash about the potential for the technology to detect sickle cell disease. The concept intrigued Prakash, so he and his team trained their microscope on patient samples by the same approach they had used to build their parasite model. They found that the microscope could successfully learn to detect sickle cell disease.

“That was a key ‘ah-ha’ moment for us—that what we are building is a general purpose technology. You don't train a self-driving car to just go on the highway or just go on windy roads,” Prakash said. “You train it for all.”

Today, the researchers are working to validate the microscope to detect five to six different conditions in seven countries across Africa, comparing it to current gold standard microscopy assessment methods, to get regulatory approval. So far, the team has published preprints showing the instrument’s detection of malaria and sickle cell disease, and they are currently developing its ability to diagnose TB. But, Prakash said, it could reasonably learn to detect far more diseases, adding, “That can dramatically change the outcomes for patients.”

Additionally, the team’s decentralized approach to their model means that anyone, anywhere can collaborate to use their system and train a model on their pathogen of interest. “We don't work alone. We work in a codevelopment framework where what we are excited about is getting a large number of people to actually engage in this process of developing self-driving tools with us.”

However, a major limitation to expanding the use of this technology, Prakash said, is collecting the samples to train the model because this requires researchers to be in the field or have partners able to collect this data. This work, though, requires funding, and recent cuts to the US Agency for International Development and other global health funding complicate the research.

“This is a global catastrophe in many ways because many incredible programs are not currently being able to be funded,” Prakash said. “Parasites don’t carry passports.” Diseases spread across countries, so being able to survey and address these pathogens in one part of the world benefits public health everywhere.

References

Li H, et al. Octopi: Open configurable high-throughput imaging platform for infectious disease diagnosis in the field. bioRxiv. 2019.10.1101/684423.
Delahunt CB, et al. Automated microscopy and machine learning for expert-level malaria field diagnosis. Proc IEEE Glob Humanit Technol Conf. 2015:393-399.
Kapuscinski J. Interactions of nucleic acids with fluorescent dyes: Spectral properties of condensed complexes. J Histochem Cytochem. 1990;38(9):1323-1329.
Shrestha P, et al. Morphology-based classification of sickle cell disease and β-thalassemia using a low-cost automated microscope and machine learning. medRxiv. 2024.09.21.24314128.
Li H, et al. Octopi 2.0: Open and scalable microscopy platform for AI-powered diagnostic applications. medRxiv. 2025.03.21.25324364.