Abstract
PURPOSE
Accurate delineation of pediatric medulloblastoma (MB) tumors is required for accurate surgical resection and efficient treatment planning. However, manual delineation is time consuming and prone to errors and inter-reader variability. We present the first attempt at automatic segmentation of MB tumors via a transfer learning approach that utilizes adult brain tumor segmentations to optimize segmentation of 1) the pediatric tumor habitat, comprising enhancing tumor (ET), necrosis/non-enhancing tumor (NET), and edema sub-compartments, and 2) the individual tumor sub-compartments.
METHODS
Our cohort consisted of 300 adult tumor MRI scans (BRATS) and 78 pediatric MB scans (46 training, 32 testing) with Gd-T1w, T2w, and FLAIR protocols. Training set subjects were collected from Children’s Hospital of Los Angeles (N=18) and Cincinnati Children’s Hospital Medical Center (N=28), whereas test set subjects were collected from Children’s Hospital of Philadelphia. Preprocessing included age-specific atlas registration, skull stripping, and bias correction. Then, using nnUnet framework, we trained 3D- deep learning U-net models on BRATS dataset for the tumor sub-compartments: ET, edema, and NET + necrosis, as well as the tumor habitat. Our initial learning rate was 0.01, with stochastic gradient descent as optimizer, and an average of dice loss and cross-entropy loss as the loss function. We then performed transfer learning using Models Genesis on the pediatric subjects.
RESULTS
Our segmentation model yielded mean dice scores of 0.87± .02 for tumor habitat, .83± .04 for ET, .742±.05 for edema, and .54±.11 for NET + necrosis, across fivefold cross-validation runs. For test set, our model yielded mean dice scores of 0.80 for the tumor habitat, 0.67 for ET, 0.54 for edema, and 0.28 for NET + necrosis.
CONCLUSION
Our transfer learning model shows promise in accurate automatic delineation of the MB tumor habitat and its individual sub-compartments, towards efficient surgical and treatment planning in MB tumors.