Application of Imitation Learning in Path Planning for Percutaneous Cardiac Interventions

Academic Background

Cardiac valve diseases, particularly mitral regurgitation (MR), are the third most common type of valvular heart disease globally and have a higher incidence in the elderly population. MR is characterized by the incomplete closure of the mitral valve during systole, causing blood to flow back from the left ventricle into the left atrium. If untreated, this can lead to severe complications such as heart failure. Traditional open-heart surgery, while effective, is highly invasive and requires a long recovery period. In recent years, minimally invasive percutaneous interventions (such as transcatheter edge-to-edge repair, TEER) have become an alternative to traditional surgery due to their smaller incisions, faster recovery times, and lower complication rates. However, these procedures demand high levels of hand-eye coordination from operators, resulting in steep learning curves, and are typically only performed in specialized centers equipped with cath labs, limiting their accessibility.

To address these challenges, researchers have begun exploring automation techniques to optimize these surgeries, especially focusing on defining safe navigation paths for robotic operations. However, the dynamic environment within the heart (e.g., the periodic motion of the mitral valve) makes traditional static path planning methods inadequate. Therefore, this study aims to develop a learning-based path planning framework specifically designed for mitral valve repair in percutaneous cardiac interventions, addressing both dynamic environments and safety requirements.

Paper Source

This paper was co-authored by Angela Peloso, Rossella Damiano, Xiu Zhang, Anna Bicchi, Emiliano Votta, and Elena De Momi, affiliated with Politecnico di Milano in Italy and the Institute of Biomedical Technologies at the Italian National Research Council (ITB-CNR). The paper was published in 2024 in the IEEE journal IEEE Transactions on Biomedical Engineering. The research was supported by the European Union’s Horizon 2020 program under the project ARTERY, with grant agreement number 101017140.

Research Workflow

1. Research Objectives and Methods

The primary objective of this study was to develop a path planning framework based on imitation learning (IL) for catheter navigation in mitral valve repair surgeries. The study compared the performance of generative adversarial imitation learning (GAIL) and behavioral cloning (BC) against traditional path planning algorithms like rapidly-exploring random trees (RRT). Patient-specific anatomical data were used to create a digital twin model that simulated the dynamic motion of the mitral valve.

2. Construction of the Digital Twin Model

The study began with CT scan data from patients, using 3D Slicer software to segment key cardiac anatomical structures (such as the left atrium, mitral valve, and left ventricle), which were then imported into Unity to construct a static environment. To simulate the dynamic motion of the mitral valve, researchers extracted displacement data of the mitral annular plane during systole (MAPSE) and dynamically updated the target position in Unity via interpolation methods to replicate the periodic motion of the mitral valve.

3. Design and Training of Learning Algorithms

The study employed four learning methods for path planning: Proximal Policy Optimization (PPO), BC, GAIL, and a combination of GAIL+BC. During training, researchers recorded 100 expert-guided catheter paths in a Unity-based simulation environment as demonstration data. The learning algorithms observed the current state (e.g., catheter position, target position, distance) and took actions (e.g., advancing or rotating the catheter) to learn how to plan paths. A task-oriented reward function was also designed to guide the algorithms toward optimizing path planning progressively during training.

4. Parameter Tuning and Validation

To ensure algorithm stability and performance, the researchers conducted a grid search over hyperparameters and selected the best-performing models for validation. They compared the learning algorithms against RRT in static and dynamic environments, assessing metrics such as execution time, path length, target position error, minimum obstacle distance, and average curvature.

Key Results

1. Parameter Tuning Results

In the static environment, the BC algorithm performed best with a strength parameter of 0.1, GAIL with 0.1, and GAIL+BC with 0.1-0.1. In the dynamic environment, BC and GAIL+BC performed best with strength parameters of 0.1-0.1, while GAIL failed to reach the target in any instance.

2. Path Planning Performance Comparison

In the static environment, BC and GAIL+BC algorithms showed significantly shorter path lengths and lower target position errors than RRT, with smoother paths and greater minimum distances from obstacles. In the dynamic environment, GAIL+BC outperformed BC in terms of target position error and smoothness but had slightly longer path lengths than RRT. Overall, the learning algorithms excelled in path repeatability and safety, generating paths similar to those demonstrated by experts without requiring subsequent adjustments.

3. Statistical Analysis and Interpretation of Results

Statistical analysis showed that the learning algorithms achieved significantly lower target position errors than RRT, especially in dynamic environments. Additionally, the paths generated by the learning algorithms performed well in terms of minimum obstacle distance and smoothness, indicating high safety and feasibility in clinical applications.

Conclusion

This study proposes a learning-based path planning framework for catheter navigation in mitral valve repair surgeries. By combining generative adversarial imitation learning and behavioral cloning, the study successfully generated safe, smooth, and repeatable paths in both static and dynamic environments. Compared to traditional RRT algorithms, the learning algorithms demonstrated significant advantages in target position error, path length, and minimum obstacle distance, without requiring subsequent adjustments. This framework lays the groundwork for future developments in robot-assisted cardiac interventions, potentially reducing operator dependency, minimizing risks, and improving patient outcomes.

Research Highlights

1. Innovation: This study is the first to apply generative adversarial imitation learning and behavioral cloning to path planning in percutaneous cardiac interventions, effectively addressing the challenges posed by dynamic environments and safety requirements.
2. Practicality: By embedding expert demonstrations, the learning algorithms generate paths that closely resemble clinical operations, reducing the need for subsequent adjustments and enhancing the standardization and repeatability of surgeries.
3. Clinical Value: This framework provides a new technical pathway for robot-assisted cardiac interventions, potentially promoting the wider adoption and optimization of minimally invasive surgeries.

Other Valuable Information

The study also notes that future work could further expand this framework to accommodate variations in patient anatomy and validate it in real clinical scenarios. Additionally, the researchers proposed the possibility of calibrating between simulation and physical systems to ensure precise navigation of robotic catheters in real-world surgeries.