IMIDEX: Achieving 83% sensitivity in lung nodule detection with Google Vertex AI

Every year in the United States, nearly a quarter of a million people are diagnosed with lung cancer . The majority of these cases are discovered at stage four when the chance of survival is 5-10% . This is due to a lethal combination of factors: lung nodules are notoriously difficult to detect and diagnose, the population is increasingly aging, and radiologists are not equipped with the time, tools, or resources to keep up.

But Richard Vlasimsky, co-founder and CEO of IMIDEX , imagined a different outlook for lung cancer diagnoses, a solution that could detect potentially cancerous nodules at an earlier stage to get patients into treatment faster. With a mission to save lives from the world's most deadly cancer, Vlasimsky, together with IMIDEX co-founder and COO Kris Zeschin, created VisiRad™XR , one of the first FDA-cleared medical devices to run on Google Cloud Vertex AI .

Leveraging artificial intelligence (AI) and machine learning (ML) capabilities, IMIDEX harnesses the power of data to inform medical care and create better outcomes for patients.

From homegrown solution to out-of-the-box tools from Google Cloud

A chest x-ray is the most common radiologic procedure in the world, typically used to detect health concerns, such as rib fractures and pneumonia. However, chest x-rays fall short when it comes to identifying suspicious lung tissue. Radiologists today are tasked with manually analyzing hundreds of these x-rays every day, resulting in fatigue and burnout. Currently, radiologists can only detect lung nodules on x-rays with 30-50% sensitivity .

"Patients with lung cancer usually don't present symptoms until it's too late," Vlasimsky says. "There's a lot of suffering in a battle with very grim odds. The goal of our technology is to use data science to help downstage the diagnosis of lung cancer and improve the chance of survival."

Using a custom-built computer vision solution, IMIDEX developed a high quality dataset of chest x-rays that precisely identified lung nodules and masses. These images were analyzed and validated by US board-certified radiologists to create a baseline model. But the homegrown solution required manual input and optimizations to choose the correct hyperparameters and reliably measure the model's effectiveness, which slowed the team's progress and hindered results.

"We had all this data just waiting to be unlocked, but we were spending most of our engineering cycles on improving processes, rather than getting better models," says Erik Allen, Chief Technology Officer at Echelon , a firm that helps startups accelerate growth through data science. "There were too many hurdles to overcome on our own, so we turned to Vertex AI to augment that benchmark for us."

Creating the first Vertex AI-powered nodule detection device

IMIDEX's cloud environment was already built on Google Cloud , so implementing an out-of-the-box ML system that integrated seamlessly with its existing infrastructure was a natural step forward. The team spent nine months of trial and error with its custom solution, then finalized its preliminary model with Vertex AI in just two weeks. Implementing a trusted AutoML process enhanced IMIDEX's ability to automate previous workflows while still achieving 83% detection sensitivity.

With minimal input, Vertex AI automatically adjusts hyperparameters, performs processes en masse through Google Cloud Pipeline Components, and validates results. This allows teams with excellence in data science to focus their efforts on higher value tasks, such as identifying and correcting anomalies in the data, and pinpointing subsets of data that need more attention. IMIDEX stores its Digital Imaging and Communications in Medicine (DICOM) data in Cloud Storage and BigQuery , allowing the team to easily query datasets when needed.

"We quickly saw the value in going all-in on our Vertex AI approach, instead of pursuing our custom model," Allen says. "By allowing Vertex AI to be our modeling partner and pushing the system in a lot of different ways, we effectively produced the accuracy we needed."

FDA clearance unlocks the doors to downstage lung cancer diagnosis

Operating in the heavily regulated medical device industry, IMIDEX could only go so far with its technology without FDA clearance. To earn clearance and take its product to market, the team ran its models through validation studies using data from 16 hospitals and health centers in the United States, then ran follow-up clinical studies using six data sites to determine effectiveness and safety.

Backed by Vertex AI, IMIDEX's models are 30% more sensitive in lung cancer detection than radiologists on their own. This gives IMIDEX's technology the potential to detect lung cancer in stage one or two, when the rate of survival is 80-90%.

IMIDEX is now on its way to deploying its technology with a goal to accelerate the diagnosis of lung cancer. With its Google Cloud stack, the company has the compliance, security, and infrastructure it needs to ensure a smooth integration into hospital systems across the country.

"IMIDEX's technology has the potential to identify hundreds of thousands of patients in the United States that need additional treatment," says Dr. Keith Hearon, an IMIDEX investor and former postdoctoral fellow at the MIT Langer Lab. "Not only does this provide an opportunity to improve patient outcomes, but it can also alleviate increasing burdens on physicians and complement the growing AI ecosystem around cancer care. We're just scratching the surface of the impact of this technology."