NIMG-88. A TRANSFER LEARNING APPROACH FOR AUTOMATIC SEGMENTATION OF TUMOR SUB-COMPARTMENTS IN PEDIATRIC MEDULLOBLASTOMA USING MULTIPARAMETRIC MRI: PRELIMINARY FINDINGS

Abstract

Abstract PURPOSE Superior outcomes for medulloblastoma (MB) requires precise surgical resection which can be guided by tumor segmentation. We present the first attempt at automatic segmentation of MB tumors via a hierarchical transfer-learning model that (1) segments the entire tumor habitat (enhancing tumor (ET), necrosis/non-enhancing tumor (NET), edema), followed by (2) training separate models for each of the sub-compartments. Transfer learning from adult brain tumors is used to optimize segmentation of tumor sub-compartments for pediatric MB. METHODS We evaluated 300 adult glioma studies (BRATS) and 49 pediatric MB studies (2-18 years old), both consisting of Gd-T1w, T2w, FLAIR sequences. The MB cohort was collected from Children's Hospital of Los Angeles (Nf19) and Cincinnati Children’s Hospital Medical Center (Nf30). Scans were registered to age-specific pediatric atlases, followed by bias correction and skull-stripping. Ground truth for the tumor sub-compartments was generated via consensus across two experts. We employed a 3D nn-Unet segmentation model on BRATS dataset using initial learning rate of 0.01, stochastic gradient descent as optimizer, and an average of dice loss and cross-entropy loss as the loss function. A hierarchical transfer learning model with Models Genesis was then applied, which allowed for fine tuning every layer on the pediatric MB dataset, across 5-fold cross validation. Dice score was used as performance metric, such that a perfect overlap between ground truth and prediction would yield a Dice score of 1. RESULTS Our 3D hierarchical segmentation model yielded mean dice scores of 0.85±0.03 for the entire tumor habitat; 0.77±0.048 for ET, 0.73±0.09 for edema, and 0.56±0.09 for NET + necrosis segmentation, across cross-validation runs. Overall, tumor outline and segmentation matched well with the ground truth, especially for the entire tumor, ET and enhancing tumor sub-compartments. CONCLUSIONS Our segmentation approach holds promise for accurate automated delineation of the tumor sub-compartments in pediatric Medulloblastoma.