MAESTRO: Supporting Real-Time Segmentation Networks on Surgical Systems via Edge Computing

Deep neural networks (DNNs) enable accurate segmentation of surgical video streams, but their high computational and memory demands pose challenges for deployment on resource-constrained surgical systems. We present MAESTRO, an adaptive edge comput- ing architecture that supports real-time execution of segmentation networks on surgical platforms. MAESTRO uses split learning to partition inference between the surgical de- vice and an edge server, dynamically selecting the optimal cut layer to balance latency, energy consumption, and data privacy. We evaluate MAESTRO using a YOLOv11 model trained on the Dresden Surgical Anatomy Dataset (DSAD) and tested on Da Vinci robotic surgery videos. Experiments demonstrate up to 43% latency reduction and 56% energy savings compared to full offloading, while maintaining low data leakage risk. MAESTRO provides a flexible and efficient solution for deploying segmentation networks in real-time, privacy-sensitive surgical environments, and generalizes to other low-resource applications.