BTSKifaru/SPARK_BTS_KIFARU

MICCA_BRATS_Africa_2023 Kifaru Team: Generative Style Transfer for MR Image Segmentation: A Case of Glioma Segmentation in Sub-Saharan Africa # Generative Style Transfer for MR Image Segmentation: A Case of Glioma Segmentation in Sub-Saharan Africa Abstract. In Sub-SaharanAfrica (SSA), the utilization of lower-quality Magnetic Resonance Imaging (MRI) technology raises questions about the applicability of machine learning (ML) methods for clinical tasks. This study aims to provide a robust deep learning-based brain tumor segmentation (BraTS) method tailored for the SSA population using a threefold approach. Firstly, the impact of domain shift from the SSA training data on model efficacy was examined, revealing no significant effect.Secondly, a comparative analysis of 3D and 2D full-resolution models using the nnU-Net framework indicates similar performance of both the models trained for 300 epochs achieving a five-fold cross-validation score of 0.93. Lastly, addressing the performance gap observed in SSA validation as opposed to the relatively larger BraTS glioma (GLI) validation set, two strategies are proposed: fine-tuning SSA cases using the GLI + SSA best-pretrained 2D fullres model at 300 epochs, and introducing a novel neural style transfer-based data augmentation technique for the SSA cases. This investigation underscores the potential of enhancing brain tumor prediction within SSA’s unique healthcare landscape. Keywords: Brain Tumor Segmentation · Neural style transfer · nnU-Net ## Citation If you use this research and/or software, please cite it using the following: ```bibtex @article{Chepchirchir2025GenerativeStyleTransfer, author = {Chepchirchir, Rancy and Sunday, Jill and Confidence, Raymond and Zhang, Dong and Chaudhry, Talha and Annazodo, Udunna and Muchungi, Kendi and Zou, Yujing}, title = {Generative Style Transfer for MR Image Segmentation: A Case of Glioma Segmentation in Sub-Saharan Africa}, year = {2025}, month = {January}, version = {1.0.0}, doi = {XXX}, url = {github.com, }