Deep Reinforcement Learning for Small Bowel Path Tracking using Different Types of Annotations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Accurate small bowel path tracking in 3D is difficult due to anatomical complexity and high ground-truth annotation costs.
• Existing methods often require extensive, precise annotations, limiting their practical application.

Purpose of the Study:

• To develop a deep reinforcement learning (DRL) tracker for small bowel path tracking that accommodates varied annotation types.
• To reduce the cost and effort associated with obtaining ground-truth data for small bowel path annotation.

Main Methods:

• A novel DRL environment was designed to process CT scans with both full ground-truth paths and partial segmentations.
• A reward system was implemented that functions effectively even when complete path annotations are unavailable.
• The DRL tracker was trained using a mixed dataset incorporating different levels of annotation fidelity.

Main Results:

• Experimental validation confirmed the efficacy of the proposed DRL tracking method.
• The approach demonstrated robustness across datasets with varying annotation quality.
• The method successfully tracked small bowel paths using less detailed annotations.

Conclusions:

• The proposed DRL tracker offers a practical solution for small bowel path tracking, significantly reducing annotation costs.
• This method enhances the usability of medical imaging datasets with weak or incomplete annotations.
• The study highlights the potential of DRL in medical image analysis for complex anatomical structures.