By Nana O 
Artificial intelligence (AI) is rapidly transforming the landscape of medical diagnostics, with its capacity to analyze vast datasets and discern intricate patterns often imperceptible to the human eye. A groundbreaking study, spearheaded by investigators at Mass General Brigham in collaboration with Boston Children’s Hospital and Dana-Farber/Boston Children’s Cancer and Blood Disorders Center, exemplifies this potential by demonstrating how AI can significantly enhance the management of pediatric brain tumors, specifically gliomas. These tumors, while often treatable, present a significant challenge due to their variable risk of recurrence. The research, published in the esteemed journal The New England Journal of Medicine AI, details the development of deep learning algorithms trained to scrutinize sequential post-treatment brain scans, thereby identifying children at heightened risk of cancer relapse.
The Unmet Need: Predicting Relapse in Pediatric Gliomas
Pediatric gliomas represent a significant proportion of childhood brain tumors, and while advancements in surgical techniques and adjuvant therapies have led to improved survival rates, the specter of recurrence remains a persistent concern. Dr. Benjamin Kann, MD, the corresponding author of the study and a key figure within the Artificial Intelligence in Medicine (AIM) Program at Mass General Brigham and the Department of Radiation Oncology at Brigham and Women’s Hospital, articulated the critical need driving this research. "Many pediatric gliomas are curable with surgery alone, but when relapses occur, they can be devastating," Dr. Kann stated. "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."
This prolonged surveillance, involving regular MR imaging, while essential for early detection of relapse, imposes a considerable emotional and logistical strain on young patients and their families. The anxiety associated with potential recurrence, coupled with the frequent need for hospital visits and scans, can significantly impact a child’s quality of life and disrupt their development and schooling. The development of a more precise predictive tool could therefore not only improve clinical outcomes but also alleviate this significant burden.
Building the Foundation: Data Acquisition and Temporal Learning
The inherent rarity of specific pediatric cancers, including certain types of gliomas, often poses a substantial hurdle for research, as limited data can restrict the ability to train robust AI models. Recognizing this challenge, the researchers strategically leveraged a network of institutional partnerships across the United States. This collaborative effort culminated in the meticulous collection of nearly 4,000 MR scans from a cohort of 715 pediatric patients. This substantial dataset provided the critical mass necessary for training sophisticated AI algorithms.
A pivotal innovation in this study was the application of a technique known as temporal learning. Unlike conventional AI models in medical imaging, which typically analyze single scans in isolation, temporal learning enables the AI to synthesize and learn from multiple scans acquired over a defined period. This approach allows the algorithm to detect subtle changes and evolving patterns within a patient’s brain that might indicate the early signs of tumor recurrence, changes that might be missed when assessing individual scans without temporal context.
A Novel Approach: Chronological Sequencing and Change Association
The development of the temporal learning model was a multi-stage process. Initially, the researchers focused on optimizing the AI’s ability to process chronological information. This involved training the model to accurately sequence a patient’s post-surgery MR scans in their correct temporal order. By learning to arrange scans from earliest to latest, the AI could begin to understand the natural progression of healing and any potential deviations from this norm.
Once the chronological sequencing was established, the researchers fine-tuned the model to specifically associate observed changes in the brain scans with subsequent cancer recurrence. This crucial step involved teaching the AI to identify patterns of change that were statistically correlated with a higher likelihood of relapse. This iterative process of sequencing and association allowed the AI to develop a nuanced understanding of how changes over time in the brain’s appearance on MR scans might predict future tumor behavior. This approach represents a significant departure from previous AI applications in medical imaging, which have largely focused on static image analysis.
Quantifying Success: Predictive Accuracy and the Power of Time
The results of the study were highly encouraging, demonstrating a substantial improvement in predictive accuracy compared to traditional single-image analysis. The temporal learning model demonstrated an accuracy rate of 75-89 percent in predicting the recurrence of either low- or high-grade glioma within one year post-treatment. This level of accuracy is considerably higher than the approximately 50 percent accuracy observed when relying on predictions derived from single MR scans – a performance no better than chance.
Furthermore, the study provided valuable insights into the optimal number of scans required for effective prediction. The researchers found that increasing the number of timepoints post-treatment provided to the AI generally improved its predictive accuracy. However, this improvement plateaued after the model was trained on approximately four to six images. This finding is significant, as it suggests that a manageable number of follow-up scans may be sufficient to achieve a high level of predictive power, potentially streamlining the surveillance process. This data point is crucial for future clinical implementation, as it informs the practicalities of data collection and AI utilization.
Looking Ahead: Clinical Validation and Future Implications
While the findings are promising, the researchers emphasize the need for further validation in diverse clinical settings before widespread implementation. The transition from a research environment to routine clinical practice requires rigorous testing to ensure the AI’s reliability and generalizability across different patient populations and healthcare systems.
The ultimate goal of this research is to translate these AI-driven insights into tangible improvements in patient care. The investigators envision a future where AI-informed risk predictions can guide clinical decision-making in several key ways. For children identified as having a very low risk of recurrence, the AI could potentially lead to a reduction in the frequency of MR imaging, thereby lessening the burden on families and conserving healthcare resources. Conversely, for patients flagged as high-risk, the AI’s predictions could prompt preemptive treatment strategies, such as the timely initiation of targeted adjuvant therapies. This personalized approach to surveillance and treatment holds the promise of optimizing outcomes and minimizing the impact of the disease.
Expert Commentary and Broader Impact
Divyanshu Tak, MS, the first author of the study and also affiliated with the AIM Program at Mass General Brigham and the Department of Radiation Oncology at Brigham, underscored the broader implications of their work. "We have shown that AI is capable of effectively analyzing and making predictions from multiple images, not just single scans," Tak stated. "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." This sentiment highlights the potential of temporal learning to revolutionize the interpretation of longitudinal medical imaging across a spectrum of diseases, not limited to pediatric brain tumors.
The implications of this research extend beyond the immediate application to glioma treatment. The success of temporal learning in this context could pave the way for similar AI-driven predictive models in other fields of medicine where serial imaging is standard practice, such as the monitoring of chronic diseases, the assessment of treatment response in various cancers, or the evaluation of neurological conditions over time. The ability of AI to identify subtle, time-dependent patterns could unlock new avenues for early intervention and personalized medicine.
A Collaborative Endeavor: Authorship and Funding
This significant research was the product of a robust collaborative effort. In addition to the lead authors, Benjamin Kann and Divyanshu Tak, the Mass General Brigham team included Biniam A. Garomsa, Anna Zapaishchykova, Zezhong Ye, Maryam Mahootiha, Tafadzwa Chaunzwa, Hugo JWL Aerts, and Daphne Haas-Kogan. The study also benefited from the contributions of numerous additional authors from Boston Children’s Hospital and Dana-Farber/Boston Children’s Cancer and Blood Disorders Center, including Sridhar Vajapeyam, Juan Carlos Climent Pardo, Ceilidh Smith, Ariana M. Familiar, Kevin X. Liu, Sanjay Prabhu, Pratiti Bandopadhayay, Ali Nabavizadeh, Sabine Mueller, and Tina Y. Poussaint.
The research received crucial financial support from the National Institutes of Health (NIH), specifically the National Cancer Institute (NCI), through grants U54 CA274516 and P50 CA165962. Additional support was provided by the Botha-Chan Low Grade Glioma Consortium. The researchers also expressed gratitude for the invaluable access to imaging and clinical data provided by the Children’s Brain Tumor Network (CBTN), a testament to the power of data-sharing initiatives in advancing pediatric cancer research. The synergy between advanced AI methodologies, comprehensive data resources, and collaborative scientific inquiry represents the future of medical innovation.
