NIMG-87. PEDIATRIC BRAIN TUMOR SEGMENTATION INITIATIVE AT CHILDREN’S BRAIN TUMOR NETWORK (CBTN); PROJECT UPDATE ON IMPROVED PERFORMANCE AND FUTURE PROSPECTS

Abstract

Abstract Pediatric brain tumors (PBT) exhibit significant heterogeneity, and accurate assessment of tumor response is crucial for patient management. While bidimensional measurements of tumor size are commonly used, they may underestimate tumor dimensions. Our group previously developed a multi-institutional, and multi-histology model on the Children’s Brain Tumor Network (CBTN) imaging dataset to segment tumor subregions, recommended by the Response Assessment in Pediatric Neuro-Oncology (RAPNO) working group. In the most recent update, we utilized a self-configuring deep learning architecture called nnU-Net on a large dataset of well-annotated PBTs (233 training, and 60 withheld internal and 46 external test sets). The model was trained and validated on multiparametric MRI sequences and differentiated various tumor subregions, including enhancing tumor (ET), nonenhancing tumor (NET), edema (ED), and cystic component (CC) and the whole tumor (WT) regions. Clinical validity was assessed by comparing the volumes of tumor subregions predicted by the model with expert manual segmentations, demonstrating strong agreement, with statistically significant Pearson’s correlation coefficients of 0.93/0.69 for ET, 0.94/0.93 for NET, 0.78/0.93 for CC, and 0.94/0.50 for ED subregions (p< 0.001) for the internal/external test cohorts. The trained model showed excellent performance for the withheld internal test set and decent accuracy for the external test set. Median Dice scores for internal/external test sets were 0.94±0.10/0.90±0.07 for WT, 0.85±0.33/0.84±0.30 for ET, 0.80±0.32/0.64±0.31 for NET, 0.79±0.37/0.67±0.33 for CC, 0.70 ±0.42/0.37±0.43 for ED, and 0.86±0.19/0.80±0.21 for all nonenhancing components (combination of NET, CC, and ED), respectively. The proposed automated segmentation method provides accurate and reproducible volumetric measurements of RAPNO-defined subregions in PBTs. It exhibits robust performance and potential generalization to external datasets and the accuracy is higher compared to the previous DeepMedic model. We will continue to improve the model by constantly adding new training data to further improve the performance and sustain model robustness.