Previous Datasets Performance for Brain Tumor Segmentation of BraTS 2023 Current Dataset

Automated brain tumor segmentation continues to be an exciting challenge. The BraTS 2023 challenge comes with nine tasks, one of which is brain tumor segmentation, with an increasing amount of training data by 1251 data for training and 219 data for validation data. Training using large amounts of data will undoubtedly improve model performance but requires more significant resources and a longer time to train segmentation models. In this study, the author proposes using models trained with previous versions of training datasets, including BraTS 2018, BraTS 2019, and BraTS 2020. The models used in this study include Shallow Dilated with Attention UNet2.5D (SDA-UNet2.5D), Single Level UNet3D, and UNet3D. Each model was tested to segment the BraTS 2023 validation dataset. The best dice segmentation performance supported by the best hausdorff95 performance was achieved by the Single Level UNet3D model, which was trained with the BraTS 2020 training dataset. However, the lesionwise dice performance achieved by the Single Level UNet3D model, a new metric, still needs improvement. The pattern of decreased lesionwise dice performance from ET to TC and WT also needs further research on the causes and ways to treat them.