Importance of the Data in the Surgical Environment
摘要
This concise review explores advanced strategies for realistic data augmentation in surgical data science, covering diverse methodologies. It discusses augmenting real data through techniques like unpaired translation from indocyanine-dyed colonoscopy videos and controlled experimental data from surgical training setups. The effectiveness of these methods for downstream tasks, such as segmentation and generalization, is evaluated. Beyond traditional GAN-based approaches, the review introduces emerging models like diffusion models and neural radiance fields. Emphasizing the delicate balance between realism and consistency, the review addresses challenges associated with content-preserving synthesis. It also highlights the importance of collaborative efforts between surgical and data science expertise to guide annotation processes and overcome obstacles to clinical translation.