AI can transform medical image analysis into a high-precision tool, available to doctors at the click of a button. Anthony Paek and Brandon Suh from Lunit explain how they want to make that happen.

Chest X-rays and mammograms are the most common images taken in medicine, yet they are much harder to interpret than 3-dimensional images such as CT scans. To Anthony Paek, this challenge sounded like the perfect area to apply his expertise in computer vision and deep neural networks.

Anthony had led a research group on deep learning during graduate studies and was looking for a field where he could apply his knowledge and experience to do something meaningful. After finishing his Ph.D. at KAIST in Korea, he recruited a few colleagues from his research group and founded Lunit, a start-up company that develops AI-based image analysis for various medical imaging techniques.

Helping doctors with cloud-based services

Lunit’s goal is to provide an easy-to-use image analysis service to clinics and doctors. The company is currently rolling out its first product, Lunit INSIGHT for chest X-rays. According to Brandon Suh, Lunit’s Chief Medical Officer: “Any user can create an account [at our website] and upload 10 DICOM images per day for free.”

DICOM is a medical imaging format that is very common in radiology. “Doctors and radiologists are very familiar with this format,” Brandon says. Once the analysis is complete, the clinician can access it on Lunit’s website.

Brandon and Anthony emphasize that because their service is cloud-based, clinics don’t need any additional infrastructure to use Lunit INSIGHT. “Hospitals have a broad range of infrastructure, including the PACS system they use,” Anthony explains. This diversity makes it difficult to develop one-size-fits-all solutions that can be deployed at the hospital. “Cloud-based integration make our efforts much easier.”

Privacy matters, especially in healthcare

Anthony tells me there is a trend in the healthcare industry to move towards cloud services. However, he admits, data privacy concerns are major issues in the industry. Lunit is sensitive to these concerns and has adopted a strict policy on privacy.

“All images are anonymized in the browser,” Brandon explains. “This happens before the images are sent to our server for analysis.” The image data is stored on the Lunit servers only for the duration of the analysis. “All images are immediately deleted from our servers right after the analysis,” Anthony says.


Complex approval procedures slow down development

Another issue the traditionally fast-moving start-up industry is facing in the medical sector is complex regulatory approval procedures. Currently, Lunit INSIGHT is termed a research-only product that cannot be used in a clinical context. Although Lunit offers a cloud-based service that could theoretically be used anywhere in the world, the company needs to be approved by the regulatory body of every country they want to offer their services in.

Anthony and Brandon point to Korea and the United States as Lunit’s primary target markets. “Our algorithms are not FDA-approved yet,” Brandon says. “We’re currently going through the regulatory process and expect our first product to be approved by the end of 2018.”

For now, Lunit has published INSIGHT mainly for promotional purposes, Brandon explains. He says that they’re one of the first companies in the world to release a functioning AI-based medical application online for public use. “We see a lot of value in that.”

The process to approve a new technology for clinical use often takes years. For startups with a limited budget, this is a very long time until they can monetize their product. “It takes [a lot of] time and resources to get a product approved by the FDA,” Brandon agrees.

While clinical trials are a way to get access to new technologies and drugs early, they can only benefit a small number of patients.

Brandon believes that developing new medical technology standards could help speed up the approval process. With access to standardized components that are already approved by regulatory bodies such as the FDA, startup companies could get their innovative products approved much faster, Brandon hopes.

Distributed training counters data acquisition challenges

Training data is a highly valuable resource in the AI sector, and Lunit is no exception. Especially in the medical sector, the large numbers of samples required to train a neural network are difficult to acquire.

Anthony and Brandon describe data acquisition as one of their major challenges. “Over the last 2 years, we have put a lot of time and effort in establishing partnerships with various hospitals,” Brandon says. He explains that hospitals usually want to keep their images in-house.

Lunit has adapted to this situation, however. “We go into the hospital, we take active part in the collection and curation of the data,“ Brandon says. The company even brings its equipment into the hospital to train their AI algorithms right there in the clinic.


“Because the scale of the data is so important, we usually work with multiple hospitals for the development of a single algorithm,” Anthony explains. This segmentation created a new challenge: Lunit had to find a way to train several instances of the same algorithm independently and later merge the results into one. “We are trying to develop smart AI systems that allow only the [training] parameters to be shared [outside the hospital] during a collective training session with multiple sites,” he says.

Using biomarkers to discover what humans can’t see

Technologies like Lunit INSIGHT could significantly reduce the time it takes to analyze X-ray images, as well as increase the accuracy of the analyses. In the future, Anthony and Brandon want Lunit to go even further. The company is developing imaging biomarkers that can go beyond what humans are capable of detecting in medical images today.

“We [want to] create an analysis that humans were not able to do at all,” Brandon explains. “We’re developing imaging biomarkers that can accurately predict the disease type or disease outcome through AI-powered image analytics. For example, we want to predict tumor response [as well as] identify the population of patients that would benefit the most from receiving invasive tests or treatment.”

Using invasive procedures only where they are really required could significantly reduce the risk for patients, as well as costs for the healthcare sector in general. Anthony mentions breast biopsies as an example: “Accurately detect[ing] breast cancer lesions in mammograms could prevent unnecessary biopsies.”

Medical imaging is one of the most promising sectors where AI-powered technology can directly improve people’s lives. Anthony and Brandon agree that many challenges still lay ahead, but they are sure their product will make a difference in the world. “We’re working on creating something unprecedented in the medical sector,” Brandon says.