HCTA-Net: A Hybrid CNN-Transformer Attention Network for Surgical Instrument Segmentation

Abstract

Surgical robots nowadays have an increasingly important role in surgery, and the accurate surgical instrument segmentation is one of important prerequisites for their stable operations. However, this task is against with some challenging factors, such as scaling transformation, specular reflection, etc. Recently, transformer has shown their superior segmentation performance in the field of image segmentation, which has a strong remote dependence detection capability. However, it could not well capture locality and translation invariance. In this paper, taking the advantages of transformer and CNN, a hybrid CNN-Transformer attention network, named HCTA-Net, is proposed for automatic surgical instrument segmentation. To be able to better extract more comprehensive feature information from surgical images, a dual-path encoding unit is proposed for effective feature representation of local detail feature and global contexts. Meanwhile, an attention-based feature enhancement (AFE) module is proposed for feature complementary of dual-path encoding networks. In addition, to mitigate the issue of limited processing capacity associated with simple connections, a multi-dimension attention (MDA) module is built to process the intermediate features from three directions, including width, height and space, to filter the interference features while emphasizing the key feature regions of local feature maps. Further, an additive attention enhancement (AAE) module is introduced for further feature enhancement of local feature maps. Finally, in order to obtain more multi-scale global information, a multi-scale context fusion (MCF) module is proposed at the bottleneck layer to obtain different receptive fields to enrich feature representation. Experimental results show that proposed HCTA-Net network can achieve superior segmentation performance on surgical instruments compared to other state-of-the-art (SOTA) segmentation models.