Automated Pediatric Brain Tumor Imaging Assessment Tool from CBTN: Enhancing Suprasellar Region Inclusion and Managing Limited Data with Deep Learning

Background: Fully-automatic skull-stripping and tumor segmentation are crucial for monitoring pediatric brain tumors (PBT). Current methods, however, often lack generalizability, particularly for rare tumors in the sellar/suprasellar regions and when applied to real-world clinical data in limited data scenarios. To address these challenges, we propose AI-driven techniques for skull-stripping and tumor segmentation. Methods: Multi-institutional, multi-parametric MRI scans from 527 pediatric patients (n=336 for skull-stripping, n=489 for tumor segmentation) with various PBT histologies were processed to train separate nnU-Net-based deep learning models for skull-stripping, whole tumor (WT), and enhancing tumor (ET) segmentation. These models utilized single (T2/FLAIR) or multiple (T1-Gd and T2/FLAIR) input imaging sequences. Performance was evaluated using Dice scores, sensitivity, and 95% Hausdorff distances. Statistical comparisons included paired or unpaired two-sample t-tests and Pearsons correlation coefficient based on Dice scores from different models and PBT histologies. Results: Dice scores for the skull-stripping models for whole brain and sellar/suprasellar region segmentation were 0.98±0.01 (median 0.98) for both multi- and single-parametric models, with significant Pearsons correlation coefficient between single- and multi-parametric Dice scores (r > 0.80; p<0.05 for all). WT Dice scores for single-input tumor segmentation models were 0.84±0.17 (median=0.90) for T2 and 0.82±0.19 (median=0.89) for FLAIR inputs. ET Dice scores were 0.65±0.35 (median=0.79) for T1-Gd+FLAIR and 0.64±0.36 (median=0.79) for T1-Gd+T2 inputs. Conclusion: Our skull-stripping models demonstrate excellent performance and include sellar/suprasellar regions, using single- or multi-parametric inputs. Additionally, our automated tumor segmentation models can reliably delineate whole lesions and enhancing tumor regions, adapting to MRI sessions with missing sequences in limited data context.