Conditional Relativistic GAN for Fast Part Segmentation of Surgical Instruments

Generative Adversarial Network (GAN) provides holistic structure-level semantic masks for the part segmentation of instruments in robotic surgery. Most discriminators in GANs judge real and fake synthetic masks independently, which may ignore their relationships. Therefore, we introduce a Conditional Relativistic GAN (CR-GAN) that concurrently distinguishes fake outputs from ground-truth labels. Then we formulate the pixel-wise semantic segmentation task into a min-max optimization problem and optimize our model with a Wasserstein divergence objective function. Experiments on datasets from MICCAI Endoscopic Vision Challenges show that our method successfully balances performances and running speed compared to other large cap-acity models.