February 12, 2019 — Southwest Research Institute engineers and UT Health San Antonio pathologists placed first in an international challenge to develop an automated method to detect breast cancer tumor cells. They trained a computer algorithm previously used for automotive, robotics and defense applications to identify cancer cells for the BreastPathQ: Cancer Cellularity Challenge conducted by the American Association of Physicists in Medicine, the National Cancer Institute and SPIE, the international society for optics and photonics.
“Adapting an autonomous robotics algorithm to solve a health diagnostics problem shows that we really have state-of-the-art techniques,” said Hakima Ibaroudene, SwRI engineer and challenge leader. “Our method has the potential to improve medical imaging, ultimately bolstering healthcare for cancer patients.”
The process of developing the cancer-detecting algorithm began with UT Health San Antonio pathologists teaching the SwRI engineers to recognize breast cancer tumor cells. The engineers then trained the computer algorithm to analyze cell images, looking for defining characteristics that distinguish the cancerous cells from normal ones. Once trained, the SwRI algorithm sorted through images provided for the challenge and matched the findings of human pathologists at the highest rate, making it the top-performing algorithm out of 100 competing submissions.
“The results demonstrated the importance of understanding network design and training the algorithm versus using an ‘out-of-the-box’ model,” said David Chambers, SwRI engineer and challenge participant. “Our approach was driven by subject matter expertise.”
Challenge organizers provided two collections of images: one to train the algorithms, the other to test them. The team analyzed images extracted from breast cancer patients and assigned a score based on the number of cancer cells in each image. Pathologists track tumor response to therapy by determining the percentage of an area that is comprised of tumor cells. Currently, this task is performed manually and relies on experts to interpret complex tissue structures. A reliable automated method, like those developed in this challenge, would produce more consistent data while avoiding human error.
“Artificial intelligence and machine learning approaches to medical image analysis will provide pathologists with a powerful tool to more rapidly identify and quantify important image features,” said Dr. Bradley Brimhall, UT Health San Antonio pathologist and challenge participant. “In doing so, additional diagnostic and prognostic information will be available for providers to guide cancer treatment.”
The challenge team also includes Donald Poole, SwRI engineer, and Dr. Edward Medina, UT Health San Antonio pathologist. Team members will present their winning algorithm at the 2019 SPIE Medical Imaging Conference on February 20 in San Diego.
SwRI’s Human Performance initiative develops machine learning and AI solutions for biomedical and health applications.
For UT Health San Antonio inquiries, contact Will Sansom, (210) 567-2579 or Sansom@uthscsa.edu.