20,000 Leagues Into Your Bowels: How GPUs Are Helping to Diagnose Gastrointestinal Anomalies

by Tony Kontzer

In the classic 1966 science fiction film “Fantastic Voyage,” a team of doctors and a space-aged submarine are shrunken and injected into an injured scientist to repair a brain clot.

While microscopic medical teams are not yet journeying into the human body, a combination of AI, deep learning and GPUs is giving us a similar perspective, starting with a detailed view of our guts.

Poland-based startup CTA.ai has developed software called GastroView that can analyze video of the gastrointestinal tract taken with tiny cameras swallowed in the form of a capsule. This work has significantly accelerated the process of diagnosing anomalies in the colon and intestines, while improving the accuracy of those diagnoses.

And then there’s the matter of relieving patients from much more unpleasant diagnostic methods.

GastroView “enables examination of the gastrointestinal tract that is far more comfortable for the patient than the traditional endoscopy,” said Mateusz Marmolowski, CTA.ai co-founder and CEO.

When Marmolowski and co-founder Marek Trojanowicz established the company in 2013, they originally focused on virtual and augmented reality-related technologies. But that soon shifted to applying their machine learning and image processing expertise to solving medical imaging challenges.

Hence, GastroView was born, with machine learning development occurring at the company’s headquarters in Gdansk, while colleagues at MIT in Cambridge, Mass., worked on image identification.

Bottoms Up!

GastroView works by having a patient swallow a pill-sized capsule equipped with two cameras, illuminating LEDs, a CMOS image sensor, an on-board battery, a transmitter and antennae. Over the ensuing 8 hours, the cameras collect video of the gastrointestinal tract, and that video is encoded and wirelessly transmitted to a data recorder worn by the patient.

Once uploaded into GastroView, the video is broken into a series of 50,000 to 100,000 endoscopic images, and deep learning algorithms are applied to automatically detect polyps, bleeding and other anomalies.

The results, compared to a traditional endoscopy, are a 70 percent reduction in the time it takes to analyze images, a 50 percent reduction in related costs, an increase in the number of anomalies detected and diagnosed, and a much more pleasant experience for the patient.

“Automatic diagnosis supporting tools can not only reduce a doctor’s effort during the examination, but also increase the sensitivity of the examination,” said Marmolowski.

Enter Deep Learning and GPUs

To train the deep learning algorithms and convolutional neural networks behind GastroView’s ability to detect and identify diseases and anomalies, CTA.ai relied on a server running four NVIDIA TITAN X Pascal GPUs, which Marmolowski said do the job 10 times faster than CPUs. Most of its training uses a custom-built wrapper for the Caffe open source framework, employing both CUDA and cuDNN libraries.

Trojanowicz said the company eventually wants to increase the number of these computing setups for increased flexibility and scalability, as well as better economy of scale, as it seeks to attract medical technology companies as customers.

“The role of GPUs is essential,” said Marmolowski.

CTA.ai also has its eyes on additional medical applications for the technology behind GastroView. Specifically, Trojanowicz said, the company is working on projects that will apply its deep learning and machine learning capabilities to CT scan images of the brain, as well as to abdominal ultrasonography.