Exploring the robustness of TractOracle methods in RL-based tractography.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Diffusion MRI enables reconstruction of white matter's fibrous architecture.
• Reinforcement learning (RL) shows promise in tractography, surpassing traditional methods.
• TractOracle-RL uses anatomical priors to reduce false positives in RL-based tractography.

Purpose of the Study:

• Investigate extensions of the TractOracle-RL framework using recent RL advancements.
• Evaluate performance across diverse diffusion MRI datasets.
• Introduce and assess a novel RL training scheme, Iterative Reward Training (IRT).

Main Methods:

• Extended TractOracle-RL framework with four novel approaches.
• Evaluated performance on five distinct diffusion MRI datasets.
• Developed and implemented Iterative Reward Training (IRT), inspired by RLHF, using bundle filtering for oracle refinement.

Main Results:

• Combining an oracle with RL consistently yields robust tractography across methods and datasets.
• RL methods trained with oracle feedback significantly outperform standard tractography techniques.
• IRT demonstrated superior accuracy and anatomical validity compared to existing methods.

Conclusions:

• Oracle integration is key for reliable RL-based tractography.
• Iterative Reward Training (IRT) offers a novel and effective approach for enhancing tractography accuracy.
• RL methods, particularly with oracle guidance, represent a significant advancement in neuroimaging analysis.