FATUnetr:fully attention Transformer for 3D medical image segmentation

Recent studies have explored the applicability of the Transformer model in 3D medical segmentation tasks. The core attention mechanism of transformer, has the property of capturing remote dependencies, which is highly robust when segmenting medical images. Inspired by the emerging visual groupings in Vision Transformers (ViTs), self-attention can be made more robust by improving the intermediate representation. We propose a novel framework called FATUnetr that aims to take advantage of the full attention module to improve accuracy. We also reduce the number of parameters by applying deeply separable convolution at the bottleneck layer. We have verified the effectiveness in the extracranial vault multiple atlas labelling (BTCV) dataset for multi-organ segmentation.