Transferable Deep Reinforcement Learning With Edge-Contour-Depth Fusion for Autonomous Wireless Capsule Endoscopy...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) offers minimally invasive gastrointestinal visualization but suffers from limited mucosal coverage and navigation method transferability.
• Existing navigation techniques struggle with patient anatomical variations, hindering widespread clinical adoption.

Purpose of the Study:

• To develop a transferable, deep reinforcement learning (DRL) framework for autonomous gastric navigation in WCE.
• To enhance mucosal coverage and procedural efficiency in WCE by overcoming anatomical variability and sim-to-real gaps.

Main Methods:

• A DRL framework utilizing a lightweight edge-contour-depth fusion module for navigation based on anatomical landmarks.
• A two-stage sim-to-real pipeline with an adaptive dynamic programming controller to mitigate physical disturbances like latency and peristalsis.
• Policy operates on low-dimensional landmark coordinates, bridging the sim-to-real gap and reducing computational load.

Main Results:

• Simulations across eight patient models achieved >97% coverage within 50 seconds, outperforming standard DRL agents.
• Ex vivo experiments demonstrated 87% mean coverage and a 53% reduction in procedure time compared to manual control.
• The framework showed robustness across diverse anatomies and stability against physical disturbances.

Conclusions:

• The proposed framework establishes a scalable paradigm for autonomous, high-coverage endoscopic navigation.
• This advancement facilitates the clinical deployment of intelligent WCE systems for improved gastrointestinal diagnostics.
• The landmark-guided DRL approach offers a robust and efficient solution for WCE navigation.