See-Mode Improves Stroke Detection and Prevention with Machine Learning on AWS

Dr. Milad Mohammadzadeh and Dr. Sadaf Monajemi were doctoral students in Singapore when they discovered they could make a significant impact in the healthcare industry. Mohammadzadeh was researching the application of fluid dynamics in solving medical problems, and Monajemi was studying how to apply artificial intelligence (AI) and machine learning (ML) to medical imaging. After completing their PhDs, they learned through discussions with clinicians that the current methods of predicting and preventing strokes based on medical images weren’t particularly effective.

After completing their degrees, the duo started See-Mode in 2017. Their mission: a world without strokes. The MedTech company uses AI-based medical software to improve the analysis of medical images such as ultrasounds, CT, or MRI scans to streamline the image analysis workflow and ultimately predict critical risk factors of stroke. Its first product, Augmented Vascular Analysis (AVA), is a medical AI software for automated analysis and reporting of vascular ultrasound scans. Aside from Australia, See-Mode’s AVA is regulatory approved within Europe, United States of America (US), and Singapore. The startup is working with hospitals and imaging centers such as the Royal Melbourne Hospital in Australia, National University Hospital in Singapore, and the University of Washington in the US.

See-Mode decided to launch AVA on the Amazon Web Services (AWS) Cloud, choosing AWS for its broad set of capabilities and the ability to manage its ML workflows on an end-to-end platform.

“Having a centralized location for billing, security, compliance, and a scalable infrastructure for building and deploying ML models means we can spend more time focusing on our core mission of improving stroke prediction,” says Mohammadzadeh, cofounder and director of See-Mode.

In particular, See-Mode takes advantage of Amazon SageMaker to train the ML models that underpin AVA. It runs the models on ultrasound images that clinics provide and produces a detailed report identifying the locations of these dangerous plaques. Improving the speed and accuracy of stroke detection helps clinicians offer better treatment plans for their patients.

As a fully managed service, Amazon SageMaker enables See-Mode to easily deploy the ML models it builds in-house and scales them automatically. When the company has a peak in load, it rapidly deploys more ML models while maintaining high performance. With these highly available, quick-scaling models, AVA software can process a large amount of ultrasound images and generate an accurate report in under a minute.

“An issue we wanted to solve was how we could analyze 50–100 images that would come in continuously, versus a single image from each patient at any given time. With AVA running on AWS, we can analyze that load within seconds and return immediate results to the user. This enables clinicians to save time on reporting their medical studies and spend that time with patients who need them. Using AWS has been a major component to solving that issue,” Mohammadzadeh says.

In addition to scalability, See-Mode required a highly elastic infrastructure to meet AVA’s large and complex requirements. “We have double-digit ML models deployed in production every day, with rapid spikes in traffic. Our underlying infrastructure has to be dynamic,” says Oliver Howden, technical lead at See-Mode. To meet this requirement, See-Mode uses Amazon Elastic Compute Cloud (Amazon EC2) instances coupled with AWS Inferentia to accelerate deep learning workloads at a low cost.

“With Amazon EC2 and AWS Inferentia, we can quickly adjust compute requirements accordingly and pay only for what we use instead of overprovisioning for hardware upfront, which is costly and time-consuming,” Howden adds.

See-Mode’s technology foundation on AWS also provides the startup with high uptime. This is a particularly important requirement for the company’s global healthcare customers, where obtaining timely information is critical.

To serve markets beyond Australia, See-Mode leverages multiple AWS Regions and Availability Zones. This enables the startup to connect from its offices in Melbourne to a customer database in a different location. Clinicians can then receive and perform analyses on ultrasound images taken only seconds earlier. Images can also stay inside sovereign geographies using AWS Local Zones to comply with regional privacy regulations.

Data from the different hospitals and imaging centers are transferred to See-Mode’s servers in batches. With the AWS Cloud, See-Mode can perform batch processing tasks using its own ML models and, at times, complements them with other AWS services such as Amazon Rekognition.

In dealing with multiple medical and imaging centers, See-Mode realized that the ultrasound images from its healthcare partners were of varying quality. See-Mode uses Amazon SageMaker to incorporate this diversity into the datasets it uses to train and continually improve the AVA software.

And, with an international customer base, See-Mode must adhere to the standards set by regulatory agencies around the globe. “We take advantage of the comprehensive AWS compliance controls, including support for HIPAA and GDPR,” says Howden. “For example, patient data must be anonymized before leaving Australia, and AWS enables us to easily adhere to this requirement. Furthermore, with the centralized platform that AWS provides, satisfying compliance requirements is straightforward. This is great for us and reassuring to our customers.”

The company uses Amazon GuardDuty to deliver continuous monitoring and detect threats by logging each time the server is accessed. It also relies on AWS CloudTrail to provide event history, which See-Mode uses for auditing purposes.

Following the success of AVA, See-Mode is exploring ways to innovate further with ML. Currently under research and development, See-Mode is validating its plaque composition and blood flow analysis to improve the prediction and prevention of strokes. “We have an ambitious mission and a team set to deliver on that,” says Mohammadzadeh. “And we’re just getting started.”

 To learn more, visit aws.amazon.com/machine-learning.