Feasibility of nnUNet Model in MR Image Segmentation: A Case of Glioma Segmentation in Sub-Saharan Africa
Keywords: Brain Tumor Segmentation, nnUNet, Generative style transfer, self-supervised
TL;DR: This paper is an effort towards making the nnUNet model generalizable for Brain Tumor segmentation on Sub-Saharan African dataset
Abstract: The study looks into Brain Tumor Segmentation with Optimized UNet as its baseline model. The study also employs both versions 1 and 2 of the nnUNet model for 2, 5, 10, 30, and 300 epochs across the BraTS 2021 dataset and the BraTS 2023 African dataset. The results have shown a slight reduction in performance on the African dataset though this has been improved by further fine-tuning of the model. In fact, the study achieved a 92.56\% dice similarity score on the African dataset for 300 epoch size. We also did a 5-fold training of the models for all the epoch sizes stated above and had the best results for folds 3 and 4 on both datasets at 0.9198 and 0.9256 respectively for 300 epochs.
Submission Category: Machine learning algorithms
Submission Number: 89
Loading