AI Tool Predicts Pediatric Brain Cancer Recurrence with High Accuracy

AI Breakthrough in Predicting Pediatric Brain Cancer Recurrence

Researchers from Mass General Brigham, in collaboration with Boston Children's Hospital and Dana-Farber/Boston Children's Cancer and Blood Disorders Center, have developed a groundbreaking artificial intelligence (AI) tool that significantly improves the prediction of brain cancer recurrence in children. The study, published in The New England Journal of Medicine AI, demonstrates the potential of AI in revolutionizing pediatric cancer care 1

1

.

The Challenge of Pediatric Gliomas

Pediatric gliomas, while often curable with surgery, pose a significant challenge due to the unpredictable nature of their recurrence. Dr. Benjamin Kann, the study's corresponding author, explains, "Many pediatric gliomas are curable with surgery alone, but when relapses occur, they can be devastating" 2

2

. The current practice involves frequent follow-up magnetic resonance (MR) imaging for years, which can be stressful and burdensome for children and their families.

Innovative Temporal Learning Approach

The researchers employed a novel technique called temporal learning to train deep learning algorithms. This method analyzes sequential, post-treatment brain scans to identify subtle changes over time. Unlike traditional AI models that focus on single scans, this approach synthesizes findings from multiple brain scans taken over several months post-surgery 3

3

.

Impressive Accuracy in Predicting Recurrence

The temporal learning model demonstrated remarkable accuracy in predicting glioma recurrence:

1. Prediction accuracy of 75-89% for both low- and high-grade gliomas by one-year post-treatment.
2. Significantly outperformed single-image based predictions, which had an accuracy of about 50% (no better than chance).
3. Optimal performance achieved with 4-6 images, after which improvement plateaued 1

   1

   .

Data Collection and Model Development

The study leveraged institutional partnerships across the country to collect nearly 4,000 MR scans from 715 pediatric patients. This extensive dataset was crucial in training the AI model effectively. The researchers first trained the model to sequence post-surgery MR scans chronologically, enabling it to recognize subtle changes. Subsequently, they fine-tuned the model to associate these changes with cancer recurrence 2

2

.

Potential Impact on Patient Care

While further validation is necessary before clinical application, the researchers are optimistic about the potential impact of this AI tool:

1. Possibility of reducing imaging frequency for low-risk patients.
2. Enabling preemptive treatment with targeted adjuvant therapies for high-risk patients.
3. Alleviating the stress and burden of frequent scans on children and families 3

   3

   .

Future Directions

The success of this AI model opens up new possibilities in medical imaging analysis. Divyanshu Tak, the study's first author, notes, "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" 1

1

. The researchers aim to launch clinical trials to evaluate how AI-informed risk predictions can improve patient care and outcomes.