New AI Model Forecasts Pediatric Cancer Recurrence

A deep learning model capable of detecting longitudinal changes in patients' brain imaging, which may signal cancer recurrence, has been developed by a team of experts at Mass General Brigham, Boston Children’s Hospital, and the Dana-Farber/Boston Children’s Cancer and Blood Disorders Center.

The model is the product of a collaboration between experts at Mass General Brigham, Boston Children’s Hospital, and the Dana-Farber/Boston Children’s Cancer and Blood Disorders Center. Supported by funding from the National Institutes of Health, the team built partnerships that provided access to extensive imaging and medical record data from hundreds of pediatric glioma patients.

Although surgery is typically successful in removing gliomas, recurrence can occur in up to half of cases, depending on the tumor grade. Often, such relapses can be deadly, making the ability to identify changes earlier a critical unmet need. 

Benjamin Kann, MD, of the Artificial Intelligence in Medicine Program at Mass General Brigham and the Department of Radiation Oncology at Brigham and Women’s Hospital and team members noted, “It is very difficult to predict who may be at risk of recurrence, so patients undergo frequent follow-up with magnetic resonance (MR) imaging for many years, a process that can be stressful and burdensome for children and families. “We need better tools to identify early which patients are at the highest risk of recurrence.”  

Using a temporal learning technique—which enables algorithms to track longitudinal changes—the data were used to train the model to identify subtle changes on patients’ follow-up MRI scans post-treatment. The model was trained to classify the correct chronological order as a pretext task, and then further refined to predict one-year recurrence for pediatric gliomas from the point of a child’s most recent scan. 

When tested, the model improved recurrence prediction by nearly 60%, significantly outperforming traditional methods. This finding was consistent for both low- and high-grade gliomas. Its predictions improved when more scans were included in its training, though it plateaued between three and six scans.  

With an accuracy ranging from 75%–89%, researchers are optimistic for how the model could improve outcomes in future pediatric brain cancer prognoses. Before deploying and implementing the model in clinical settings, additional studies will be needed to validate the results.

“We have shown that AI is capable of effectively analyzing and making predictions from multiple images, not just single scans,” first author Divyanshu Tak, MS, of the AIM Program at Mass General Brigham, noted. “This technique may be applied in many settings where patients get serial, longitudinal imaging, and we’re excited to see what this project will inspire.”