What matters in reinforcement learning for tractography.

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Deep reinforcement learning (RL) offers a novel approach to white matter tractography, aiming to reconstruct neural pathways without manual reference streamlines.
• Existing RL frameworks for tractography are complex, with limited understanding of individual component impacts.

Purpose of the Study:

• To thoroughly investigate the influence of various components within deep RL frameworks for tractography.
• To provide evidence-based recommendations for optimizing RL algorithms, seeding strategies, input signals, and reward functions in tractography.

Main Methods:

• Trained approximately 7,400 deep reinforcement learning models.
• Conducted extensive computational analysis, accumulating nearly 41,000 hours of GPU time.
• Systematically evaluated the impact of different RL algorithms, seeding strategies, input signals, and reward functions.

Main Results:

• Identified critical factors influencing the performance of deep RL in tractography.
• Determined which RL components are most effective for accurate white matter reconstruction.
• Quantified the impact of design choices on tractography outcomes.

Conclusions:

• Deep reinforcement learning is a promising avenue for automated white matter tractography.
• Specific recommendations are provided to enhance the efficacy and efficiency of RL-based tractography methods.
• An open-source codebase, trained models, and datasets are released to facilitate further research and development.