Lightweight Network for Real-Time Segmentation of Endoscopic Surgical Instruments

Real-time segmentation of surgical instruments is a fundamental technique in robot-assisted surgical navigation and plays a key role. Despite significant progress in natural image segmentation in recent years, practical realization of accurate real-time semantic segmentation of operative instruments in endoscopy is still a challenging topic. In this paper, we build a lightweight network that can meet real-time requirements while maintaining segmentation accuracy. Specifically, our model adopts an encoder-decoder structure, where the bottleneck for semantic feature extraction is constructed in the encoder part by combining the Ghost module and MobileNetV3, the decoder consists of a depth-separable convolution, an attention module, and a transposed convolution. The depth-separable convolution used reduces the parameters and computation of the model, the attention module captures the global context and encodes the semantics, emphasizes key semantic features and improves the representation of the features, and this module has only a few parameters, which helps to localize the surgical instruments. The model is validated on the EndoVis 2017 dataset to prove the proposed model, and the results show that the method can excellently accomplish the semantic segmentation of surgical instruments, and its inference speed meets the real-time requirement, which provides a reference for further robot-assisted surgical navigation.