AI Tool Predicts Relapse of Pediatric Brain Cancer from Brain MRI Scans

Predicting which patients are at risk for recurrence remains challenging, leading to frequent follow-ups with magnetic resonance (MR) imaging for several years. This process can be both stressful and burdensome for children and their families. There is a pressing need for better tools to identify early which patients are most likely to experience a relapse. Artificial intelligence (AI) holds great potential for analyzing large medical imaging datasets and identifying patterns that may go unnoticed by human observers. Now, AI-assisted analysis of brain scans may help improve the care of pediatric gliomas.

Research on rare diseases, such as pediatric cancers, often faces the hurdle of limited data. This study, conducted by Mass General Brigham (Somerville, MA, USA) and their collaborators, used institutional partnerships across the United States to gather nearly 4,000 MR scans from 715 pediatric patients. To maximize AI's ability to "learn" from a patient's brain scans and better predict recurrence, the researchers employed a technique called temporal learning. This method trains the AI model to synthesize findings from multiple brain scans taken over several months following surgery. In most medical imaging AI models, the algorithm is trained to draw conclusions from single scans, but temporal learning — which had not previously been applied to medical imaging AI research — uses images collected over time to predict cancer recurrence.

To build the temporal learning model, the researchers first trained the system to sequence a patient’s post-surgery MR scans in chronological order, enabling the model to detect subtle changes in the scans. They then refined the model to correctly link those changes to future cancer recurrence when applicable. The results, published in The New England Journal of Medicine AI, revealed that the temporal learning model was able to predict the recurrence of either low- or high-grade gliomas by one year after treatment, with an accuracy of 75-89%. This was significantly more accurate than predictions based on single images, which showed an accuracy of approximately 50%, comparable to random chance. Providing the AI with images from additional timepoints post-treatment further improved the model’s accuracy, with only four to six images required before the improvement plateaued. However, the researchers emphasized that further validation across different settings is necessary before clinical use. Ultimately, they aim to launch clinical trials to determine whether AI-driven risk predictions can enhance care, such as by reducing imaging frequency for low-risk patients or by preemptively administering targeted therapies to high-risk patients.

“We have shown that AI is capable of effectively analyzing and making predictions from multiple images, not just single scans,” said first author Divyanshu Tak, MS, of the AIM Program at Mass General Brigham and the Department of Radiation Oncology at the Brigham. “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.”