Imitation learning for supervised autonomous tumor resection in central airway obstruction

Area of Science:

• Robotics
• Artificial Intelligence
• Surgical Automation

Background:

• Central airway obstruction (CAO) procedures demand high precision and coordination.
• Existing robotic surgery methods face challenges with tissue variability and real-time constraints.
• Current data-driven approaches are limited to short, single-skill tasks, not cyclic surgical workflows.

Purpose of the Study:

• To investigate the potential of imitation learning for fully autonomous task-level and high-level control in CAO tumor resection using the da Vinci Research Kit.
• To develop a system capable of handling the complexities of robotic surgery, including degraded visibility and deformable tissues.

Main Methods:

• A hierarchical framework combining low-level action chunking with transformers (ACT) policies for autonomous task execution.
• Finite state machine (FSM)-based high-level coordination for procedure-level sequencing, decomposing the workflow into five recurring tasks.
• Training policies on synchronized multi-view video and kinematics data, with analysis of different data regimes via ablation studies.

Main Results:

• Low-level policies achieved autonomous execution with 0.938mm RMSE, comparable to human operators.
• High-level coordination enabled supervised autonomous full procedure execution with 1.232mm accuracy.
• Expert surgeons confirmed comparable resection quality and superior retraction performance compared to human operators.

Conclusions:

• This research demonstrates the first learning-based supervised autonomy for CAO removal, achieving expert-level accuracy.
• The hierarchical framework allows for full procedure execution with minimal supervision required only at task transitions.
• Public release of training datasets (multi-view imagery, robot kinematics) aims to advance surgical automation research.