New algorithm boosts accuracy, speed of lung tumor identification

UAlberta computing scientists develop algorithm through a neural network to track tumors in real-time that outperforms recent state-of-the-art methods.

Andrew Lyle - 21 February 2019

Imagine an operating room in the near future. A doctor consults a MRI scan-a tumor highlighted on the display, deciding that it requires treatment. A neural network creates a model of how the tumor moves with the patient's breathing, a complex motion of soft tissue. With the model as a guide, a radiotherapy machine matches the movement of a narrowly-focused radiation treatment beam to the patient's breathing, minimizing exposure to healthy tissue and targeting the tumor as precisely as possible.

That vision of precision health and patient-specific modelling for medical care drives the research of Pierre Boulanger, a professor in the Department of Computing Science-and is now one step closer to reality, with a new study that uses an algorithm to identify lung tumors in real-time from MRI scans.

"A patient-specific model helps with surgical planning. But in order to create such a model, one needs to take medical imaging data and turn it into something one can simulate," explained Boulanger. "Algorithms like the one developed in our laboratory can be used to generate a patient specific model for diagnosis and surgical treatment."

The new algorithm was developed in collaboration with Nazanin Tahmasebi, graduate in the Department of Computing Science, and Kumaradevan Punithakumar, from the Faculty of Medicine and Dentistry'sDepartment of Radiology and Diagnostic Imaging.

The algorithm was "trained" on a set of scans in which doctors have identified lung tumors. It then processes an enormous set of images, learning what tumors look like and the properties they share, and is then tested against scans that may or may not contain tumors.

The result? A tool that outperforms other state-of-the-art methods of identifying tumor boundaries and can quickly and accurately delineate healthy tissue from unhealthy tissue even in real-time, according to the study.

Getting a clearer picture

Game, data set, match

Boulanger explains that while neural networks are a technology that has been around since the 70s, their effectiveness at analyzing data has grown in leaps and bounds due to technology from a surprising source: computer games.

"The problem before, especially in the 90s, was that we just didn't have the computing power. Thanks to gamers, we now have GPUs that provide huge amounts of computing power for graphics," explained Boulanger, referring to the graphical processing units (GPUs) used to create the visuals of games. "On a traditional computer, it might take a month to train a network. With a GPU, it can take minutes."

Lung tumors are particularly challenging because they move significantly as the patient breathes. Beyond that challenge, MRI scans can also be difficult to interpret.

"The tumor regions in scan results can be very subtle, and even once identified, need to be tracked over time as the tumor moves with breathing," said Boulanger. "The new algorithm is able to combine many possibilities to find the best descriptors to identify unhealthy regions in a scan."

Surmounting these challenges to get a clearer picture of the tumor boundaries, the algorithm designed by Boulanger and his team could help adaptive radiation therapy reduce damage to nearby healthy tissue during radiation treatment.

While the networks like this one could prove critical in identifying tumors, Boulanger notes that it will in no way replace the need for doctors and the importance of human high-level thinking in the full treatment of patients.

"Most importantly, these tools are designed to improve medical outcomes alongside a skilled professional, and to help to make the process faster," said Boulanger. "Medicine, as a field, is always looking to go further and improve the quality of care for patients. Neural networks are a tool that can help that goal."

The paper, "A Fully Convolutional Deep Neural Network for Lung Tumor Boundary Tracking in MRI", was published in the 40th Annual International Conference Proceedings of the IEEE Engineering in Medicine and Biology.


Want to learn more about the world-class artificial intelligence research being conducted at the University of Alberta? Check out the Faculty of Science AI hubartificial-intelligence/index to find out how our students and researchers are advancing the field on the global stage.