A fully automatic CT images segmentation method for organs at risk in head and neck

Radiotherapy is one of most treatments for tumors. To accurately control the radiation dose distribution and lessen the radiation damage to normal tissues and organs in radiotherapy, it is necessary to delineate organs at risk precisely. In this paper, a novel automatic segmentation method for organs in head and neck, named SAU-Net, was proposed, the architecture of which was based on 3D U-Net with residual connection. Non-local spatial attention implemented by squeeze-and-attention (SA) module was introduced to solve unbalanced segmentation accuracy caused by the huge difference of organs' volume, which increased the ability to aggregate multi-scale context information by encoding the global features. In order to avoid extra convolution operations stacking too much local information and reduce the number of parameters, the model reduced the number of convolution kernels. Evaluating the dice score of the model performance, the segmentation accuracy of SAU-Net was 13.7% and 8.2% higher than 3D U-Net's and 3D ResU-Net's respectively. It demonstrates that the SAU-Net delineates organs at risk in head and neck faster and more accurately.