This paper presents a new privacy-preserving synthetic continual semantic segmentation framework for robotic surgery. The researchers address the problem of ‘catastrophic forgetting’ in Deep Neural Networks (DNNs), which refers to the decline in performance on previously learned tasks after learning new ones. This is a particular issue in situations where data scarcity exists, and the dataset of the old instruments for the old model cannot be released due to privacy concerns. The presented framework blends and harmonizes open-source old instrument data with the synthesized background and new instrument data with an extensively augmented real background. The study demonstrates the effectiveness of the framework on the EndoVis 2017 and 2018 instrument segmentation dataset.
Publication date: 9 Feb 2024
Project Page: https://github.com/XuMengyaAmy/Synthetic_CAT_SD
Paper: https://arxiv.org/pdf/2402.05860
