DCT-CenterMask: A Real-Time Instance Segmentation Network for Kidney Ultrasound Images.

Ultrasound is the preferred imaging method to detect chronic kidney disease. It has the characteristics of real-time dynamic imaging. The existing real-time instance segmentation model can meet the real-time requirements of 30fps of segmentation speed, but there is still a certain gap in segmentation accuracy compared with the instance segmentation model. In this paper, a real-time instance segmentation model based on CenterMask is proposed. Using the characteristics that discrete cosine transform (DCT) can reduce the energy loss in the process of down sampling the ground truth mask, a new mask representation is trained on the premise of maintaining the real-time segmentation speed. We also integrate deformable convolution and a new attention mechanism into the network to improve the overall receptive field range of the network and the ability to learn detailed features. For further research, we made a COCO format kidney instance segmentation dataset. Through experiments, our model has improved the mask AP accuracy by about 6% compared with the original model.