Swin UNETR: A Novel Model of 3D Tumor Semantic Segmentation Using Vision Transformers

In medical imaging, brain tumor segmentation is essential for precise diagnosis and therapy planning. Tumor appearance and site variations frequently require assistance with traditional segmentation approaches. In this work, we present a novel method that leverages Swin UNet transformers, which combine the advantages of transformer designs and convolutional neural networks (CNNs). The hierarchical vision transformer model Swin UNet is especially well-suited to managing the multi-scale and complicated nature of medical pictures. On a variety of 2D and 3D semantic tasks as well as across a range of imaging modalities, the popular "U-shaped" network topology has achieved state-of-the-art performance benchmarks. However, FCNNs perform poorly in modeling long-range information due to the convolution layers' limited kernel size, which results in flaws in growths of different sizes. With our model surpassing current standards with a Dice coefficient of 0.882 and an Intersection over Union (IoU) of 0.822, the results show notable gains in segmentation accuracy. Additionally, the Swin UNet model demonstrated better generalization ability across various tumor types and MRI modalities. In order to compute self-attention, the Swin transformer encoder uses shifting to extract features at five different resolutions. At every resolution, a skip link is made to an FCNN-based decoder.