Expert-level pediatric brain tumor segmentation in a limited data scenario with stepwise transfer learning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Pediatric Neuro-oncology

Background:

• Automated tumor delineation in pediatric gliomas is crucial for real-time volumetric evaluation and clinical decision-making.
• Existing auto-segmentation algorithms for pediatric tumors are limited by data availability and clinical translation challenges.

Approach:

• Developed and validated deep learning neural networks for pediatric low-grade glioma (pLGG) segmentation using a novel in-domain, stepwise transfer learning approach.
• Leveraged two datasets (n=184 and n=100) for model development and external validation.
• Clinically benchmarked the best AI model against expert segmentations through blinded evaluations, including Likert scales and Turing tests.

Key Points:

• The AI model using stepwise transfer learning achieved superior performance (median DSC: 0.877) compared to baseline models (median DSC: 0.812).
• External validation showed AI accuracy comparable to inter-expert agreement (median DSC: 0.834 vs. 0.861).
• Clinicians rated AI segmentations higher on average (median Likert: 9 vs. 7) and found them more acceptable (80.2% vs. 65.4%).

Conclusions:

• Stepwise transfer learning enables expert-level automated pediatric brain tumor segmentation and volumetric measurement with high clinical acceptability.
• This approach facilitates the development and translation of AI imaging segmentation algorithms, even in data-limited scenarios.
• AI-driven auto-segmentation holds significant potential for improving pediatric glioma diagnosis and treatment monitoring.