Evaluating STU-Net for Brain Tumor Segmentation

Brain tumor segmentation is vital in addressing the tumor’s high heterogeneity, enhancing accurate diagnosis, guiding effective treatment, and improving prognosis predictions. In recent years, state-of-the-art methods in this domain have primarily evolved from U-Net-based architectures, demonstrating notable advancements in glioma radiographic segmentation. While large-scale models pre-trained on extensive datasets have significantly propelled deep learning progress, most current medical image segmentation models remain small-scale, encompassing only tens of millions of parameters. As the BraTS competition amasses more data and with the advancements of large-scale models in the past year, there emerges a compelling need to investigate the potential benefits of larger network architectures for adult glioma segmentation. In this context, we utilize the Scalable and Transferable U-Net (STU-Net) and its pre-trained variants on the RSNA-ASNR-MICCAI Brain Tumor Segmentation (BraTS) 2023 dataset. Being one of the most extensive medical image segmentation models, STU-Net’s sizes range between 14 million to 1.4 billion parameters. Our research seeks to evaluate STU-Net’s efficacy and transferability on the BraTS23 dataset, encapsulating diverse MRI scans of brain tumor patients. The code and pre-trained models are available at https://github.com/uni-medical/STU-Net .