Robotic Instrument Segmentation With Image-to-Image Translation

The semantic segmentation of robotic surgery video and the delineation instruments are important for enabling automation. Despite major recent progresses, majority latest deep learning models instrument detection rely on large datasets with ground truth labels. While demonstrating capability, reliance labelled data is a problem practical applications because systems would need to be re-trained domain variations such as procedure type or sets. In this letter, we propose alleviate by training that synthesised using image-to-image translation techniques investigate different methods perform process optimally. Experimentally, demonstrate same network architecture can trained both real our proposed synthetic without affecting quality output models’ performance. We show several approaches provide experimental support publicly available datasets, which highlight potential value approach.