Evaluating the impact of reinforcement learning on automatic deep brain stimulation planning

Area of Science:

• Neurosurgery
• Artificial Intelligence
• Medical Imaging

Background:

• Traditional automated planning for brain electrode placement faces limitations, including sensitivity to local optima.
• Machine learning approaches are increasingly replacing conventional optimization techniques.
• Deep brain stimulation (DBS) electrode placement planning is a complex, multi-objective optimization problem.

Purpose of the Study:

• To explore the feasibility of using deep reinforcement learning (DRL) for automated planning of deep brain stimulation (DBS) electrode placement.
• To evaluate DRL's performance in a single-electrode placement scenario.

Main Methods:

• A deep Q-learning approach was developed, defining states based on electrode trajectory and associated information, and actions as possible motions.
• Deep neural networks were utilized to navigate the complex state space derived from MRI data.
• A reward function was designed to prioritize safety and accuracy in reaching target brain structures.

Main Results:

• The DRL approach demonstrated superior navigation of complex anatomy, yielding safer and more precise electrode placements compared to a segmented electrode reference.
• Compared to conventional techniques, DRL improved accuracy by 2.3% in average proximity to obstacles and 19.4% in average orientation angle.
• Computation times increased significantly, from 2 to 18 minutes.

Conclusions:

• DRL shows significant potential for DBS electrode trajectory planning, particularly in complex scenarios with high-dimensional state and action spaces.
• While modest accuracy gains were observed in the single-electrode case, DRL's resilience against local optima is a key advantage.
• This study serves as a foundational step towards addressing the more complex challenge of planning multiple-electrode placements.