Docbot’s Deep Learning System Aims to Nip Colon Cancer in the Bud

Tech Currents Spotlight
June 6, 2018 By Jackie Connor

By combining a software tool with a massive database, the team hopes to improve polyp detection and colon cancer screenings.

Healthcare is constantly brimming with new technologies dedicated to improving preventive care, expanding treatment options or even finding cures. The American Cancer Society considers colon cancer to be the third most common cancer diagnosed in both men and women in the United States, with more than 97,000 new cases estimated for 2018.

Andrew Ninh, co-founder and chief executive officer of Docbot, a healthcare software for advanced patient data that is captured at the point of care, was motivated to build an artificial intelligence deep learning software system that efficiently captures data in real time after a first-hand experience with systemic frustrations during a hospital visit.

Three months before his 18th birthday, Ninh experienced a spontaneous pneumothorax, or a presence of air in the pleural space when his lung essentially popped. After being rushed to an intensive care unit, Ninh was treated and later informed his heart rate significantly dropped during the night. The computer system alerted nurses, however, they learned to ignore these alarms due to this specific condition’s common occurrence, namely in people who play sports or are of taller stature. Ninh, who stands more than six feet tall, was prompted to consider areas that needed improvement in the healthcare system.

“I didn’t understand that much about healthcare back then, but I knew there was a big opportunity because this was a problem,” said Ninh. “Back then everyone was buying these very expensive medical records, spending billions of dollars, but there were still a lot of inefficiencies.”

Ninh later met William Karnes, M.D., UCI Health gastroenterologist, and UCI associate clinical professor of gastroenterology at the School of Medicine, and the two discovered the potential to develop an AI product trained to identify polyps in real time from UCI’s unique colonoscopy database compiled over several years from its top physicians.

“When I first met Andrew, I realized that all the data we were collecting could also create a report,” said Dr. Karnes.

Dr. Karnes, Docbot co-founder and has worked closely with Ninh to hone Docbot’s capabilities through massive amounts of data collection. Dr. Karnes began with over 10,000 colonoscopies and 190,000 images of polyps to develop the first version of Qualoscopy, a database comprised of quality data, which allowed them to use AI. According to Dr. Karnes, they have collected the largest data set in the world that links polyps to pathogen detection.

“Now Qualoscopy can automatically take in all of that data from the procedure, generate a report and a path requisition,” said Dr. Karnes. “Docbot is a super expanded beautiful version of what I had originally created.”

With Ninh’s “behind-the-scenes” technology for improving medical documentation and record keeping, and Dr. Karnes’ interface that collected data in real time, the marriage of the two aspects created a very sophisticated system that, according to Ninh, is currently the only system that can make self-learning AI possible.

In February, Dr. Karnes received the Institute for Clinical & Translation Science (ICTS) Pilot Award for his abstract “Robust Real-time Polyp Detection and Classification During Colonoscopy Using Deep Learning,” which highlights how AI’s deep learning capabilities for polyp detection improves colon cancer detection rates, namely in low performers. According to the abstract, the AI system detects polyps with 96.4 percent accuracy and predicts precancerous pathology, via images with 94 percent accuracy. The abstract hypothesizes that:
“AI-assisted polyp detection and classification can be effectively utilized in real time during colonoscopy.”

The Docbot team, in addition to Ninh and Dr. Karnes, is now comprised of Tyler Dao, vice president of engineering, and Jason Samarasena, M.D., an interventional gastroenterologist at UCI, who uses polyp photos to teach a deep learning model to recognize an image of polyps in a timely manner.

“It’s up to 98 percent accurate and can take about 10 milliseconds to process an image, which is four times faster than it needs to be run in real time,” said Dr. Karnes. “So, [the clinician] gets the feedback while [they] watch the screen during a colonoscopy.”

To determine if a polyp is precancerous, the current process normally takes about two weeks, whereas Docbot’s system optically detects it in fractions of seconds.

“Detecting the polyp and determining if it’s precancerous in real time has the potential to save a lot of pathology costs,” said Ninh.

A colonoscopy can potentially prevent 90 percent of colon cancers, and according to Dr. Karnes, 90-to-95 percent of colon cancers diagnosed today are from people who are not up-to-date on their colonoscopies.

“The adenoma detection rate should be the same as the adenoma prevalence rate in the community,” said Dr. Karnes. “It’s believed the prevalence of these pre-cancerous polyps is about 50 percent in the screening age population, so ADR should be 50 percent.”

However ADRs can range from as low as seven percent up to about 53 percent, which pinpoints a large gap that Dr. Karnes attributes to variables in the ability of different colonoscopists to find polyps. Automated polyp detection with AI aims to close this gap.

Dr. Karnes, who presented his abstract during a state-of-the-art lecture at Digestive Disease Week in Washington, D.C., emphasized the optimization of quality colonoscopies and how they can minimize barriers, such as coverage and reduce costs.

“It costs the U.S. over $1 billion a year with all these path bottles we are sending to pathology,” said Dr. Karnes. “If we can truly resect and discard or just ignore a polyp and our AI tells us what it is with a high degree of certainty, then we don’t have to send to pathology—therefore saving $1 billion a year in healthcare expenses.”

Docbot’s technology is currently implemented in several facilities around the region, such as the UCI Medical Center. As the team continues to implement their technology into as many ambulatory surgery centers, community hospitals, and teaching centers as possible, their ultimate goal is to introduce the platform to several other medical fields they believe would benefit from real time AI.

“There are a number of other fields that are image intensive,” said Dr. Karnes. “If a pulmonologist were doing a bronchoscopy, this [product] is something that could work really well for them.”