AI SolutionsBlogNVIDIAConsulting ServicesApplication ServicesHealthcareData AnalyticsWWT ServicesAI & Data
Blog • • 4 minute read

Hackathon Event Advances AI in Medical Research

In a successful two-day hackathon, WWT and NVIDIA experts trained AI models to detect pancreatic tumors through 3D image segmentation, promising advancements in medical research.

In this blog

  1. About the event
  2. The problem and solution
  3. The impact

 About the event

Our experts partnered with NVIDIA to host a two-day hackathon focused on advancing AI in medical research. We chose an impactful medical problem related – tumor segmentation – and leveraged the open-source MONAI framework using NVIDIA's GPU technology. The hackathon provided a great proof-of-concept validation of the power and potential of AI tools to create a life-saving impact in the medical field.

With the help of WWT's medical expert Dr. Erin Jospe and NVIDIA's Michael Zephyr and Michael Levy, we selected pancreatic tumor segmentation as the problem statement for the hackathon. Pancreatic cancer is one of the deadliest forms of cancer, with a five-year survival rate of about 10 percent. This is partly because pancreatic cancer often goes undetected until it has reached an advanced stage, making it difficult to treat. Better screening could often mean earlier detection and better prognosis.

Screening for pancreatic cancer typically involves imaging tests such as computed tomography (CT), magnetic resonance imaging (MRI) or endoscopic ultrasound (EUS). These tests can detect abnormalities in the pancreas, but requires highly-trained professionals to interpret those medical images, especially in the early stages of the disease when the abnormalities may be small and difficult to distinguish from normal tissue.

3D image segmentation identifies and separates specific structures or regions within three-dimensional image volumes and could be a valuable decision-making tool for better and earlier detection in pancreatic cancer treatment. For example, 3D image segmentation can be used to identify and separate the pancreas from surrounding tissues in a CT scan, and it can be trained to identify lesions in the pancreas. These algorithms make it easier for radiologists to detect abnormalities within the pancreas and monitor changes in the size and shape of the pancreas over time.

 The problem and solution

The purpose of the hackathon was to explore the use of AI in medical imaging segmentation and potential innovations in enabling valuable medical research. Experts were tasked with building a segmentation model using the MONAI framework to segment the pancreas and tumor over a short period of time. The teams were given a dataset with 282 3D volumes of portal venous phase CT scans from Memorial Sloan Kettering Cancer Center to train and test their models.

Two CT scan images after using the MONAI framework. CT scan on the left shows MONAI without label. CT scan on the right shows MONAI with label.
Figure 1: Sample training data after a few transformations done using MONAI with (right) and without label (left), where the yellow and purple areas correspond to pancreas and tumor, respectively.

Teams across consulting services, application services and AI-defined networking (AIDN) gathered at WWT's global headquarters to solve the problem using NVIDIA GPUs and state-of-the-art AI algorithms. The event team set up a dedicated development environment in the WWT Advanced Technology Center. 

Image of WWT and NVIDIA experts at WWT's Advanced Technology Center. Participants are sitting at desks in a conference room.

Each team had access to an NVIDIA A100 GPU and was provided an isolated sandbox virtual environment to work and develop the solution. The NVIDIA team provided support with MONAI. They also provided NVIDIA AI for enterprise software licenses to allow a smooth onboarding to the development environment. Throughout the course of the event, mentors from our data science team and NVIDIA's MONAI team provided support for participants.

The teams processed the images using transformation pipelines including operations, such as scaling, cropping, orienting and spacing, and random transformations, such as random cropping. Most teams selected the U-Net architecture, which was introduced in 2015 and developed for biomedical image segmentation. The teams built a PyTorch training process that utilized data caches on Nvidia A100 GPUs, enabling fast training of the complex model. The teams were able to complete hundreds of epochs of training within a day and obtain satisfactory results on identifying the pancreas and tumor in 3D CT scan images.

 The impact

WWT and NVIDIA experts share their findings on a PowerPoint presentation.

The teams yielded promising results from the two-day event, which strongly validate that AI segmentation is a valid tool for increasing the efficacy in pancreas disease screening. The different model designs could improve the use of MONAI and other AI algorithms, and be packaged into scalable solutions to improve patient and radiology experiences in field medicine.

3D image segmentation shows great promise in improving the accuracy and efficiency of pancreatic cancer screening and treatment. The partnership between WWT and NVIDIA continues to improve research in this field, and given more data and training time, the models and results from this event could become essential tools in the fight against pancreatic cancer.

Technologies