An Unet Boosting Training Strategy for the BraTS-ISBI 2024 Goat Challenge

Tumor segmentation represent a major challenge in current radiological practice. Morevoer, the high heterogeneity of clinical presentation of brain tumors introduce even more complexity to perform a precise evaluation of the lesions. This paper introduces an iterative training method for domain generalization in brain tumor segmentation, focusing on the BraTS-ISBI 2024 GoAT challenge. Our UNet-based model achieves average dice scores of 0.743, 0.783, and 0.787 for enhancing tumor (ET), whole tumor (WT) and tumore core (TC) segmentation, respectively. The approach shows significant improvement on the validation set through iterative refinement, outperforming a baseline model trained on the entire dataset. The code is publicity available at https://github.com/rikhy967/UNetboost.