Automatic Segmentation of Rare Pediatric Brain Tumors Using Knowledge Transfer From Adult Data
Xinyang Liu, Erin R. Bonner, Zhifan Jiang, Holger R. Roth, Syed Muhammad Anwar, Roger J. Packer, Miriam Bornhorst, Marius George Linguraru
Abstract
Diffuse midline glioma (DMG) is a rare but fatal pediatric brain tumor. An automatic pipeline to analyze patient MRI can help monitor tumor progression and predict overall survival. Clinical implementation requires automatic segmentations of subregions of DMG. Given the rarity of data, we investigated how pretraining state-of-the-art deep learning models on adult brain tumor data would allow for a knowledge transfer to pediatric data and improve overall segmentation performance. We retrospectively collected multisequence MRI of 45 children diagnosed with DMG (a total of 82 scans with different timepoints). Five-fold cross-validations were performed on the DMG dataset using SegResNet and nnU-Net, each with and without pretraining on the BraTS2021 dataset of 1,251 glioblastoma multiform subjects. Best segmentation results were achieved using nnU-Net with pretraining (Dice scores of 0.859±0.229 and 0.880±0.072 for the enhancing region and the whole tumor, respectively). Our results suggest knowledge transfer from adult brain tumor images can improve pediatric brain tumor segmentation performance. Using pretraining also helped in speeding up training convergence for downstream tasks.