Swin-UNETR

Swin UNETR is a state-of-the-art 3D medical image segmentation architecture: it pairs a hierarchical Swin Transformer encoder with a CNN-based decoder to capture long-range dependencies.

This model, Swin UNEt TRansformers (Swin UNETR), delivers superior performance in 3D medical segmentation tasks. Its core design leverages a Swin Transformer as the encoder, efficiently computing self-attention within shifted windows to model global context (long-range dependencies). This encoder connects via skip connections to a CNN-based decoder, which handles the high-resolution feature recovery necessary for precise segmentation. The architecture achieved top-ranking benchmarks in the BraTS 2021 challenge for brain tumor segmentation and is a state-of-the-art performer on the Medical Segmentation Decathlon (MSD) dataset.

https://github.com/Project-MONAI/research-contributions/tree/main/SwinUNETR