Split Learning at the Edge for Live Multi-Organ Segmentation in Robotic Laparoscopic Surgical Imaging

Split Learning is a Machine Learning (ML) technique that consists of running different fractions of a neural network in different processes. It is used to mitigate problems arising from the need of running ML jobs in limited resource environments, e.g., in IoT devices. Studies have combined split learning with other methods, such as early exit (to speed up inference) or dynamic resource allocation, to decrease inference and training latency, adapting the model to constrained devices. An example of low-latency application where split learning is particularly effective is medical imaging. In particular, in laparoscopic robotic surgery ---a minimally invasive procedure typically used in the abdominal and pelvic areas--- such latency gain could be crucial to give immediate feedback to the surgical team. Researchers have proposed the use of split learning in medical imaging to address latency or privacy issues. However, no prior work has focused on achieving a real-time ML edge computing system for processing surgical procedure imaging, while operating on existing proprietary robotic systems. To fill this gap, this thesis presents a live multi-organ segmentation system designed using the split learning paradigm. Our system runs the first neural network segment at the edge and the second subset of layers on a serverless node, for resource efficiency and scalability. Our evaluation shows a latency trade-off analysis during a live video segmentation of human organs, demonstrating the effectiveness of our approach. The proposed architecture can be adapted to other ML at the edge applications, where is desirable to accelerate inference on IoT or embedded devices.