Pediatric Brain Tumor Segmentation Using Multiresolution Fractal Deep Neural Network

Pediatric (PED) central nervous system tumors are the leading cause of cancer-related deaths in children, with highgrade gliomas having a dismal five-year survival rate below 20%. The rarity of these tumors often leads to delayed diagnosis, difficult treatment regimens, and necessitates multi-institutional collaborations for clinical trials. Automated and precise volumetric measurements can offer a more standardized way to determine the extent of disease and monitor the effectiveness of treatments. Multiresolution Fractal Deep Neural Network (MFDNN) has been proposed earlier for multiscale deep learning fractal-based tissue segmentation. In this work, we hypothesize that a MFDNN can automatically extract intricate multiresolution texture features when analyzing data from pediatric patients with high-grade glioma. We employ MFDNN on a data set of pediatric MRI scans that include diverse tumor subtypes. The training data set consists of multi-parametric MRI (mpMRI) scans from BraTS-PEDs 2023 challenge. The model segmentation efficacy is assessed against separate validation and unseen mpMRI data of high-grade pediatric glioma. The MFDNN model excels in pediatric brain tumor segmentation. It achieves the highest Lesion Wise Dice scores of 0.444 (ET), 0.634 (TC), and 0.783 (WT), with corresponding dice scores of 0.356, 0.672, and 0.830, respectively. The Dice score-based evaluation and uncertainty analysis within MFDNN offer improved brain tumor segmentation compared to the state-of-the-art methods in the literature.