Improving CNN Cardiac MRI Segmentation by Transformer and Direction Feature

As the development of medical image processing, automatic cardiac magnetic resonance images segmentation has been one of the most important issues. However, the unclear boundaries and inhomogeneous intensity in MRI images usually cause indistinction for pixels from different class around boundaries, limiting accuracy of the segmentation result. In this paper, Transformer which is common in natural language processing is introduced into UNet, which is common in medical image segmentation to capture global information; meanwhile, direction feature is used enhance comparison for inter-classes pixels for better performance near the boundaries. The network can be divided into segmentation module and direction feature module: self-attention mechanism is used to replace the convolution module in building blocks for long-range dependence; result of segmentation module is taken as input of direction feature module, which utilize direction vector from pixel to its nearest boundary pixel to improve segmentation accuracy nearby boundaries. Performance of the network is evaluated and got satisfied result.