As families waited in long lines at a clinic in Masafu, rural Uganda, Paul Lebel was helping deploy the Remoscope, a compact blood-testing system powered by on-device AI. It can analyze up to 2 million red blood cells in one to 12 minutes, compared with at least 45 minutes for a standard Giemsa-stained blood-smear test. During clinical studies in Uganda, Lebel analyzed blood samples on two Remoscopes, while technicians smeared blood onto glass slides, chemically fixed and stained samples, and manually counted parasites under a microscope.
“Our technician, James [Emorut], and I were running samples from children whose blood was being drawn right beside us,” Lebel said, explaining they were not allowed to provide research results to patients. “We knew the answer before they left the room, and yet they were sent to wait for hours. That made a big impact on me.”
Diagnostics remain out of reach for almost half of the world’s population, as poor, marginalized and rural communities often depend on last-mile services — the context Sweden-based N23 Health and Oxalis Imaging, the Chan Zuckerberg Biohub spin-off commercializing the Remoscope, are looking to serve.
To bring the devices to market, the companies are building early sources of revenue while pursuing regulatory approval and the long-term goal of enabling early triage at the point of care. “You need to be financially sustainable to reach as many patients as possible,” said Lisa Falco, CEO at N23 Health, “and our goal is to become the No. 1 screening method for breast cancer in the Global South.”
The companies are taking different approaches: Oxalis uses high-throughput cell imaging and models that can expand across diseases, while N23 is building a fixed, single-purpose model for breast-cancer triage.
2 million cells, analyzed in 12 minutes
The team built the Remoscope to relieve technician fatigue, one of the main bottlenecks in malaria diagnosis. A person can read the images it captures, just not at the pace a model can. “People fall through the cracks because technicians get overburdened. They work 12 or 14 hours a day,” said Lebel, who recently stepped into the role of founder and CEO of Oxalis Imaging.
Paul Lebel. Credit: personal archive
At the CZ Biohub in San Francisco, where Lebel continues as a senior engineering manager, the Remoscope started taking shape when a deep-UV microscope he developed revealed malaria-infected cells in an unmodified blood sample, allowing real-time analysis without expensive reagents or the biases and errors of sample preparation. An in-house microfluidic chip lines up rows of blood cells flowing in parallel, “flattened like pancakes,” as Lebel puts it. This lets the device image hundreds of cells rather than one at a time, as conventional imaging flow cytometers do, boosting throughput to 2 million cells in 10 minutes.
The Remoscope’s first neural networks were trained to detect parasites on images annotated by hand. Lebel annotated hundreds of thousands of cells, an effort he estimated amounted to weeks of full-time work. In a clinical study in Tororo, one of Uganda’s highest-malaria-burden districts, the Remoscope detected 95 parasites per microliter of diluted blood.
Looking beyond malaria
As the team continues to study the Remoscope’s accuracy for diagnosing malaria, it plans to expand to sepsis and to noncommunicable conditions such as sickle cell disease. To get there, company’s researchers are using AI to find features scientists may not yet know to look for.
Lisa Falco. Credits: N23 Health
With multiple-instance learning, they’re training models to analyze all the cell images from a patient’s sample at once, labeled only by patient diagnosis. In a clinical cohort in San Francisco, the model, Lebel said, picked up well-known signs of sepsis while also pointing to new potential biomarkers. “The AI model can integrate all these different features and say: ‘If you have A, B, C features in your sample, together those are highly predictive of a condition.’ And you can discover that as you go,” he said.
Expanding screening with single-indication AI
According to the World Health Organization, the odds of a woman surviving breast cancer in low-income countries are almost half those in high-income countries. N23 Health is addressing the same shortage of trained staff with a narrower approach to AI: a single-indication device with a small, fixed neural network trained to reproduce radiologist-level pattern recognition where experts are not available. “Static, single-purpose models are less costly to train, safer and easier to validate, and perform just as well,” Falco said.
Clinics without reliable connectivity rarely send field data back to improve a model. Falco said that matters less for N23 Health, because its models are frozen once submitted for regulatory approval and only monitored for changes in performance. “You still need some kind of continuous evaluation to make sure there are no data or performance drifts, but to a much lesser extent than for generative AI,” she said.
N23 Health’s moat, Falco said, is the annotated breast-imaging dataset developed over years of research by Kristina Lång, the company’s chief medical officer and an associate professor at Lund University. Extending that expertise through a portable device became viable more recently, as phone-connected ultrasound probes advanced enough to deliver higher-quality images. The system identifies suspicious lumps at the point of care using a model that runs locally, without an internet connection, and an app designed to guide healthcare workers to find the right spot on the breast for AI analysis.
Training with N23 Health’s device. Credits: N23 Health
In a clinical trial in Ethiopia, local healthcare workers trained for one day screened 5,300 women, and preliminary results shared by Falco suggest the device doesn’t require a specialist radiologist to conduct the examination. The solution works as well as mammography for initial triage, with equipment costing 25 times less, according to early results from a previous trial in Sweden. “Maybe our solution is not dependent on the latest AI, but without AI, nothing of this would be possible," Falco said.
While the company initially focuses on East Africa, where it sees a quicker time to market and more openness to AI in healthcare, Falco believes the clinical evidence could also back adoption in high-income countries. “But here we already have well-functioning care pathways. It takes longer to change something that is already working,” she said. “When you're solving a problem that doesn't have a solution, that's an easier way.”
Even with support from governments, early adopters, distributors, third-party buyers, and funders, Lebel does not take success for granted. “There is no guarantee we will succeed where others have failed,” he reflects. “Fundamentally, however, our technology was conceived out of an effort to create something accessible for low- and middle-income countries, and I think it's tractable.”
Disclosure
Lebel spoke with The Infinite Loop at the Nebius AI Discovery Awards ceremony in London, where the Remoscope took first place in the Medical Devices category, under the temporary name Real Time Imaging Systems (RTIS). The Infinite Loop is a Nebius editorial project. Editorial decisions are made independently.
