Abstract
INTRODUCTION: T1w/T2w ratio mapping, combining voxel-wise signal intensities in T1-weighted (T1w) and T2-weighted (T2w) structural MRI, has been used to investigate cortical architecture in the brain, but has also shown promise in tissue discrimination, even in tumor tissue. Given this, we investigate whether the inclusion of these established T1w/T2w ratio maps, or a similar T1w - T2w combined map, can improve performance on a novel task; automated segmentation of tumor tissue in pediatric brain tumor cases from the BraTS-PED 2024 dataset. METHODS: Using the BraTS-PED 2024 dataset (n = 261 pediatric brain tumor patients), we trained and evaluated (with a five-fold cross validation approach) segmentation performance across tumor subregions with nnU-Net, a state-of-the-art deep learning framework. Multiple model configurations were compared; a) a standard baseline model using typical multiparametric MRI (mpMRI, including T1w, T2w, FLAIR and contrast-enhanced T1w MRI) as input modalities and b) an experimental configuration using standard mpMRI inputs plus a T1w/T2w ratio map. Performance was assessed using Dice scores and statistical comparisons with Bonferroni correction to assess he direct 'added benefit' of the T1w/T2w ratio maps. RESULTS: Inclusion of T1w/T2w ratio or the combined maps did not significantly improve segmentation accuracy across any tumor subregion. While minor increases in ET segmentation were observed with the ratio map, these were not statistically significant. Combined maps showed marginal improvements in ET and NET segmentation but reduced performance in CC and ED regions. CONCLUSIONS: Overall, we demonstrate that T1w/T2w ratio maps do not improve deep learning models for segmenting pediatric brain tumor subregions using nnU-Net, despite their strong biophysical basis for tissue discrimination. T hese findings suggest that such data augmentation strategies may not provide added value and highlight the importance of rigorous validation in medical imaging research.