Subscribe
With a few keystrokes, anyone with an internet connection can generate fake images through AI-powered platforms like Adobe Firefly or ChatGPT. While some have welcomed this new ability with open arms, others worry about its potential for spreading misinformation. Nefarious actors can easily use it to misrepresent reality with fake images of politicians or events that never happened.
But Rayan Sadri and Ali Rouzbayani, the founders of a young startup called Carez AI, want to do the opposite. They’ve created an AI model to produce synthetic images that can enhance our understanding of the world – and they’re doing it by supporting health research.
“We’re building the infrastructure for image generation in life sciences,” said Sadri on the McGill Delve podcast. He’s the CEO of Carez AI and a McGill University alum. “Our platform enables AI teams in life sciences to create medical images at scale.”
And this is a big deal. Medical images are a critical step in the diagnosis and treatment of life-threatening diseases. X-Rays, MRIs, CT scans, ultrasounds, and other tools have advanced our ability to look into the body to identify problems without ever needing to reach for a scalpel. In the aggregate, researchers can use thousands of these images to paint a picture of how diseases behave and progress – important information for teams creating new products and treatments to support patients.
But for some diseases, limited images are available, which hampers researchers’ ability to study them effectively. You may have collected 200 MRI scans from patients with a rare disease. But when the disease can present in infinitely variable ways, you’ll need a lot more data to accurately understand how to treat it.
This is where AI can help, said Rouzbayani, the Chief Technology Officer at Carez AI.
“Our models learn from real-world data,” he said on the McGill Delve podcast. “Then they can produce samples that are feasible, plausible, high-fidelity, and new.”
In other words, using existing bundles of medical images from real patients, their AI model can accurately create new images of a disease’s countless other variations.
Infrastructure for life science imaging
Carez AI is one of many young companies riding the wave of artificial intelligence. And for both Sadri and Rouzbayani, who are both recent university graduates, this is their first foray into entrepreneurship.
They’re currently courting investors for their company. They believe what sets them apart is their commitment to rigour and supporting research teams.
“We’re pitching part of the future of life sciences,” said Sadri.
So far, the response from investors has been positive, he said. Investors aren’t just buying into a novel use of AI; they’re investing in what could become a critical piece of infrastructure for life sciences research.
That’s because the Carez AI model can be adapted for use with any kind of life sciences dataset that depends on images. Currently, the company is focused on health research, but it’s not a stretch to imagine their tool being useful in botany, zoology, and other life sciences that often depend on images to conduct their research.
This means the company’s growth potential is immense. It’s healthcare now, and limitless other sciences later. Teams from both the private and public sectors can enlist Carez AI to create synthetic biologic images for research and product development.
Avoiding model collapse
The images created through Carez AI’s model are a type of synthetic data – a practice that involves creating artificial data points, based on data already collected from the natural world, to help researchers conduct their work.
The use of synthetic data is not new. Originating in the 1990s, coinciding with the rapid emergence of the Internet and computer sciences, the practice has since become a regular part of conducting health research. But Carez AI’s model significantly accelerates the production of synthetic data. It can process large datasets in very little time, and output meaningful and accurate extrapolations to include in the dataset.
The main challenge with creating an AI model like this one is ensuring scientific validity, explained Rouzbayani. Mainstream AI image-generating tools often feature strange distortions like missing fingers. In medical imaging, these kinds of inaccuracies are unacceptable. Rouzbayani explained how his company is working to prevent these inaccuracies from occurring in their AI-produced synthetic images.
At one level, Carez AI works closely with medical professionals to validate its model’s output. These experts are well-versed in the diseases being studied and can ensure the synthetic images look as they should.
It’s also important that the synthetic images are always based on real-world samples, said Rouzbayani. There can never be a situation where the AI is trained on images itself has created. This would lead to a phenomenon known as “model collapse,” where the model eventually loses touch with reality and produces inaccurate images.
“That’s why we never train our model on synthetic data,” said Rouzbayani.
And even if your model is 100 per cent based in the real world, you must ensure that your data is clean, accurate, and properly tracked and annotated. This helps AI developers understand and control what data their model is using to generate its images.
For Rayan and Rouzbayani, the potential impact of their company is incredibly motivating.
“Five years from now, if they say they cured a rare type of cancer because of our tool – that excites me,” said Sadri.
Rayan Sadri and Ali Rouzbayani also spoke with managing editor Eric Dicaire on the McGill Delve podcast. They explain their AI model, how it can be a game-changer for life sciences research, and what it’s like to run a startup in this unique space. Listen to it here or search “McGill Delve” wherever you get your podcasts.
This article was written by Eric Dicaire, managing editor of McGill Delve.
