Abstract
Surgical instrument segmentation is an essential task in the domain of computer-assisted surgical system. It is critical to increase the context-awareness of surgeons during the operation. We propose a new model based on the U-Net architecture for surgical instrument segmentation, which aggregates multi-scale feature maps and has cascaded dilated convolution layers. The model adopts dense upsampling convolution instead of deconvolution for upsampling. We set the side loss function on each side-output layer. The loss function includes an output loss function and all side loss functions to supervise the training of each layer. To validate our model, we compare our proposed model with advanced architecture U-Net in the dataset consisting of laparoscopy images from multiple surgical operations. Experiment results demonstrate that our model achieves good performance.