Dual Attention Guided R2 U-Net Architecture for Right Ventricle Segmentation in MRI Images

Lei Jiang, Hengfei Cui, Chang Yuwen, Yanning Zhang

Abstract: Right ventricle segmentation plays an important role in the computer-aided diagnosis of heart diseases. However, due to the small area of right ventricle and limited dataset, the performances of the existing deep learning segmentation methods are not good enough. For some small areas of right ventricle that are difficult to segment, we apply a novel dual attention module on the decoding path of Dilated R2 U-net to extract better feature representations in this work. The dual attention module in this work is divided into position attention module and channel attention module. The positional attention module suppresses the irrelevant feature representations in the feature map and enhances the useful feature representations to improve the sensitivity and prediction accuracy of the model. The channel attention module enhances the interdependence of the feature representation of channels by gathering the information of the associated channels in the feature map. We use dilated convolutions to expand the receptive field of the model. By adding dual attention modules, our model shows higher precision than Dilated U-net on the Right Ventricle Segmentation Challenge (RVSC) test dataset.