News

Researchers at UC San Diego have developed advanced deep learning techniques that could revolutionize treatment planning for breast cancer radiotherapy – making it faster and improving its quality. The team sought to reduce inconsistencies in treatment plans and improve patient outcomes by leveraging artificial intelligence (AI) using the Expanse system at the San Diego Supercomputer Center (SDSC), which is a pillar of the School of Computing, Information and Data Sciences (SCIDS) at UC San Diego.

“Our study focused on two innovative methods for training AI models to predict radiation doses: a ‘glowing’ mask algorithm and a gradient-weighted loss function,” said Lance Moore, a computational data science research specialist at UC San Diego School of Medicine Department of Radiation Medicine & Applied Sciences. “The glowing mask encodes distance data into treatment images while the gradient-weighted function prioritizes accuracy at the borders of the high-dose regions.”

These techniques were tested via U.S. National Science Foundation ACCESS allocations on SDSC’s Expanse using three-dimensional U-Net deep learning models trained on data from more than 300 breast cancer treatment plans.

“The results were promising as the combination of glowing masks and the gradient-weighted loss function produced the most accurate predictions – achieving a high level of precision in dose distribution,” Moore said. “For example, the error in mean dose to critical organs like the heart and lung was very small, and dose comparisons to clinical plans showed high agreement.”

Kelly Kisling, an associate clinical professor at UC San Diego School of Medicine’s Department of Radiation and Applied Sciences and principal investigator of the study, explained that the team’s findings suggest that incorporating the new AI methods into automated radiotherapy planning systems could significantly reduce the time and effort required to develop high-quality treatment plans for patients with breast cancer.

A depiction of the model, called a 3D U-Net, used for this project. The model uses the images of the patient from the treatment plan to predict what dose should be delivered. Credit: Lance Moore

“Thanks to ACCESS allocations on Expanse, we were able to create a large model using resources that are beyond the scope of an individual research lab. This large model is more accurate and precise than other comparable models – making it a major step forward in personalized cancer care,” Kisling said.

The study was published in the International Journal of Medical Physics Research and Practice.

Computational support was provided by U.S. National Science Foundation ACCESS (allocation no. MED210003).