DeepVinci: A Semantic Segmentation Model with Edge Super-vision and Densely Multi-scale Pyramid Module for DaVinci Gynecological Surgery

Abstract

Abstract The successful development of self-driving cars has opened the door for the possibility of automated surgery, which may help alleviate the problem of limited access to quality surgical care in underserved areas. Automated surgical navigation typically involves three stages: 1) identifying and localizing organs, 2) identifying organs that require further surgical attention, and 3) automatically planning the surgical path and steps. This study focuses on the first stage, which is organ identification and localization. The daVinci surgical system offers a promising platform for automated surgical navigation due to its advanced visual and semi-automatic operating capabilities. This paper proposes a deep learning-based semantic segmentation method for identifying organs in gynecological surgery. We introduce a novel end-to-end high-performance encoder-decoder network called DeepVinci, which includes two modules (the Densely Multi-scale Pyramid Module (DMPM) and the Feature Fusion Module (FFM)) to overcome the limited Field of View (FoV) issue and enhance global context information. Additionally, we integrate an edge-supervised network to refine the segmentation results during decoding. Experimental results show that DeepVinci is superior to mainstream semantic segmentation models (including UNet, FCN, DeepLabV3, and MaskRCNN) on our collected test dataset. The Dice Similarity Coefficient (DSC) and Mean Pixel Accuracy (MPA) values are 0.684 and 0.700, respectively. As collecting daVinci gynecological endoscopy data is challenging, we also introduce a new dataset of 110 gynecological surgery videos from Tri-Service General Hospital in Taipei, Taiwan. This dataset provides valuable video data for further research in daVinci gynecological surgery.