Brain Tumor Segmentation using 3D Swin UNETR

The application of a vision transformer (ViT) to medical image analysis has been promising recently, particularly in brain tumor segmentation from MRI images. Following this achievement, we present a new architecture, Swin UNETR, which synergistically integrates the Swin Transformer and a U-shaped network architecture for enhanced segmentation. The Swin Transformer, with its hierarchical structure and shifted windowing mechanism, has the capability to inherently extract local and global contextual information. In our approach, we utilize the Swin Transformer as the encoder to capture multi-scale features and the decoder to reconstruct the segmentation map with skip connections to maintain spatial information. The architecture enables the model to effectively deal with the inherent morphology of the tumor and the variability of the size. Experimental results on benchmark datasets, including BraTS 2024, demonstrate that Swin UNETR surpasses conventional convolutional neural network-based approaches, confirming its feasibility as a useful tool for assisting clinicians in accurate brain tumor delineation.