Deepnoid is betting on Japan’s regulatory process to validate its AI technology before tackling FDA 510(k) clearance and the costly U.S. market. Success in both could determine its place in the global medtech race.
If Deepnoid wants to go global, three things must happen: regulatory alignment, localization, and strategic partnerships.
That’s the assessment of Nishant Nepal, a researcher from Kathmandu, Nepal, who works at the Korean AI-driven medical imaging company. And if Deepnoid can pull it off in Japan first, it could change everything.
The company, known for tools like DEEP:NEURO, an AI model for detecting brain aneurysms, and M4CXR, a large language model that generates radiology reports from chest X-rays, is playing the long game. It has already set up a subsidiary in Dubai to explore Middle Eastern markets and is keeping one eye on eventual U.S. expansion.
But first, Deepnoid wants to establish itself in Japan—a move that, if successful, could serve as both a validation of its AI and a strategic foothold against Western and Japanese competitors.
Japan as the first test
Most AI-driven medtech companies chase FDA approvals and European CE marks as their golden ticket to international markets. Deepnoid is taking a different strategy, placing its chips on Japan first.
The U.S. and Europe present steep regulatory and data-access hurdles.
Japan, though notoriously insular in its healthcare sector, offers a different kind of challenge—one Deepnoid believes it can navigate.
The company sees an opening in Japan’s investment in medical AI and its high adoption rate of imaging technology. There are more than 7,000 MRI machines across Japan, four times as many as in Korea, but the radiologist workforce is stretched thin.
Rather than positioning AI as a replacement for human expertise, Deepnoid is marketing its technology as a second-opinion tool. That distinction matters in Japan, where Choi Hyun-seok, the company’s chief medical officer, said automation is often met with skepticism.
“A successful approval here could set Deepnoid apart from rivals still navigating the looser but more crowded U.S. market,” Choi said in a recent interview.
The company is embedding itself in Japan’s healthcare infrastructure, securing key opinion leaders, launching pilot studies in university hospitals, and aligning with major Japanese medical societies.
“We’ve already built strong relationships with leading neurosurgeons and neuroradiologists in Japan,” Choi said. “Early feedback has been promising. The key is proving we belong in their ecosystem.”
Building AI for the real world
Every morning, Nepal logs in and pulls up the latest scan. He grew up in Kathmandu immersed in code, left for KAIST on a full scholarship in 2019, and spent four years building AI models for real estate and language processing before joining Deepnoid in 2023.
The image flickers on the screen—white matter hyperintensities, tiny lesions in the brain that could signal Alzheimer’s, stroke, or dementia. Until now, doctors have relied on subjective interpretation to assess them. AI could change that, turning vague estimates into precise, quantifiable data.
At Deepnoid, Nepal was handed a problem no one had quite solved: training an AI model to detect white matter hyperintensities in brain scans. According to Nepal, the difficulty is not just in developing a model that can interpret 3D medical images, but in proving that the AI is reliable enough for clinical use.
Currently, radiologists rely on visual estimates to assess white matter hyperintensities. Deepnoid’s model aims to replace approximation with hard numbers, identifying affected areas and calculating precise volumes that could change how these conditions are diagnosed and treated.
For now, the model is not ready for clinical use. According to Nepal, it doesn’t even have a name yet. Before it can move toward FDA 510(k) clearance, Nepal said it must be trained on a more diverse dataset.
Deepnoid plans to integrate it into DEEP:NEURO, its suite of AI-powered diagnostic tools, alongside its cerebral aneurysm detection system. But like every AI in healthcare, its biggest limitation isn’t computing power. It’s data. “We don’t have enough right now,” Nepal said in an interview. “That makes it harder to build something truly robust.”
Deepnoid is sourcing what it can—tapping into hospital partnerships, scouring publicly available datasets, and exploring other avenues—to ensure the model isn’t just trained on Korean patients. The goal is global viability. “We expect to have a working version in a few months,” Nepal said. “Then we’ll start testing on datasets from different countries.”
The cost of breaking into the U.S. market
For AI-driven healthcare companies, proving the technology works is just the beginning. The real challenge is demonstrating that it performs consistently across diverse patient populations and regulatory landscapes.
In the U.S., that means data—specifically, American patient data. The FDA requires that at least half of an AI model’s training data come from American patients. Ideally, Nepal said, that number should be closer to 100 percent.
“That means diversity,” he said. “And diversity means data.” But data isn’t cheap.
“We’re talking about $100,000 to $200,000 for just 400 patient records,” he said. “For a truly comprehensive dataset, it’s more like $200,000 to $300,000.”
That price tag shuts many AI startups out of the U.S. market before they can even apply for regulatory review. Deepnoid is navigating this by working with intermediary vendors like Segment, a medical imaging data provider, to access datasets with broad demographic representation.
The investment is steep, but the potential payoff is substantial. “Healthcare in the U.S. is extremely expensive,” Nepal said. “If our technology can streamline certain processes, it could help reduce costs for patients.” Cost, however, is just one piece of the puzzle. Quality control is another.
“That’s the tricky part,” Nepal said. “Since we don’t handle the data directly during third-party testing, we rely on vendors and external testers to ensure its quality. We trust them to follow best practices and provide accurate reports.”
To maintain that standard, Deepnoid is partnering with institutions like Massachusetts General Hospital in Boston, a benchmark for data quality in medical AI development. “Most companies focus on accuracy,” Nepal said. “But we also focus on compliance validation and detailed technical documentation. That’s what sets us apart.”
Regulatory hurdles do not stop at data. Deepnoid must also meet U.S. cybersecurity regulations, an increasingly scrutinized area for AI-driven healthcare technologies.
“Cybersecurity standards in the U.S. are extremely strict,” Nepal said. To comply, the company has outsourced security testing to an Israeli firm with a track record in FDA approvals. “Because we don’t conduct the cybersecurity tests ourselves, we can confidently state that an external company has verified our product’s security,” he said. That kind of validation strengthens Deepnoid’s FDA submission.
Then comes generalizability. Proving that Deepnoid’s AI is accurate is one thing. Proving it works across different populations, imaging equipment, and clinical environments is another. To ensure objectivity, testing is conducted by an independent team with no involvement in the model’s development.
“We restrict access to test data for different research teams,” Nepal said. That process played out recently with Deepnoid’s DEEP:CHEST model, which detects lung abnormalities. The company purchased U.S. hospital data, ran the AI on it, and compared results with Korean datasets.
“The model performed similarly in both cases,” Nepal said. “That’s the kind of validation we need for every market we enter.”
The long game
For now, Japan is Deepnoid’s proving ground. The company is preparing to launch pilot tests in partnership with university hospitals and expects to announce joint research collaborations soon. It will also participate in Medical Japan 2025, held in Intex Osaka from March 5-7, to showcase its AI models.
The company is betting that a strong foothold in Japan will open doors elsewhere.
Choi believes it’s the right move. “Japanese hospitals are overwhelmed. Radiologists are drowning in workload,” he said. “We’re offering a tool that helps them do more in less time, without sacrificing accuracy.”
But the competition is fierce. Industry giants like Fujifilm and Canon Medical Systems dominate Japan’s AI medical imaging space, backed by deep government ties and an entrenched presence in hospitals. Deepnoid’s strategy is to carve out a niche in generative AI-powered automated radiology reporting, a technology that could ease the burden on radiologists by producing detailed reports in real time.
The demand is there. Now, Deepnoid has to prove it belongs.
![[공지사항] 제 17기 정기주주총회 소집공고](https://static.wixstatic.com/media/b136d4_b5769fdf694342beae4492629180b43c~mv2.png/v1/fill/w_800,h_800,fp_0.50_0.50,q_95,enc_auto/b136d4_b5769fdf694342beae4492629180b43c~mv2.png)
Comentários