RANT: A cascade reverse attention segmentation framework with hybrid transformer for laryngeal endoscope images

Automatic segmentation of organizations and organs is a prerequisite for medical image analysis and computer-assisted diagnosis and treatment. The larynx is an important part of the human body, and scholars have paid less attention to the segmentation research in the larynx. For electronic laryngoscope images, a cascaded reverse attention network with hybrid transformer (RANT) is presented. Firstly, the RANT combines transformer and CNN in a serial way to capture the global dependency feature based on the low-level spatial detail feature. Secondly, combined with the reverse attention and receptive field block module (RRM), connects features of different scale in a cascade way to gradually mined the target. Finally, the segmentation results are optimized by convolutional conditional random fields (ConvCRFs). The experimental results on two laryngoscopy datasets show that the RANT achieves the best balance of multi-organ segmentation compared with other the state-of-the-art segmentation networks. On the two datasets, the mIoU of RANT reaches 76.63% and 88.77%, the mDSC reaches 83.45% and 93.49%, which is greatly improved compared with the benchmark. The RANT network can segment each class of laryngoscope image effectively. This method is of great significance in clinical application.