EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Curiosity-Driven Reinforcement Learning for Autonomous Navigation in Magnetically Controlled Capsule Endoscopy

Soheil Dehghani, Parastoo Mohsenipour, Farhan Pahlevani, Mahshad Mohammadi, Raha Farahani, Borhan Beigzadeh

IEEE Access · 2026 · 14 · 74221-74240

DOI: 10.1109/access.2026.3693588 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Soheil Dehghani und 5 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Gastrointestinal disorders affect approximately 40% of the global population, presenting diagnostic challenges that conventional endoscopic procedures struggle to address due to invasiveness and miss rates ranging from 6% to 28%. While magnetically controlled capsule endoscopy (MCCE) offers a minimally invasive alternative, achieving comprehensive anatomical coverage remains challenging due to sample inefficiency and limited robustness in reinforcement learning-based navigation systems. This paper introduces Adaptive Spectral Exploration - Proximal Policy Optimization (ASE-PPO), a curiosity-driven reinforcement learning framework combining Random Network Distillation (RND) with spectral parameter space noise for autonomous gastric navigation. The primary novelty lies in the synergistic integration of three contributions: spectrally structured colored noise enabling coherent directional exploration, direct coupling between RND-derived novelty and parameter perturbation amplitude, and an O(1) incremental normalization framework for scalable deployment. Validation in VR-Caps demonstrates that ASE-PPO achieves 95.4% gastric coverage at 120 seconds — representing 4.15–6.17% improvements over baseline PPO — with ablation studies confirming RND’s contribution beyond spectral noise alone. ASE-PPO additionally demonstrates 18.6% standard deviation reduction and 21.8% coefficient of variation improvement, with statistically significant coverage gains at 120 seconds ( $p=0.035$ ), establishing curiosity-driven reinforcement learning with structured exploration as a promising paradigm for autonomous gastrointestinal examination systems.

Zitieren

APA

Dehghani, S., Mohsenipour, P., Pahlevani, F., Mohammadi, M., Farahani, R., & Beigzadeh, B. (2026). Curiosity-Driven Reinforcement Learning for Autonomous Navigation in Magnetically Controlled Capsule Endoscopy. IEEE Access, 14, 74221-74240. https://doi.org/10.1109/access.2026.3693588

Vancouver

Dehghani S, Mohsenipour P, Pahlevani F, Mohammadi M, Farahani R, Beigzadeh B. Curiosity-Driven Reinforcement Learning for Autonomous Navigation in Magnetically Controlled Capsule Endoscopy. IEEE Access. 2026;14:74221-74240. doi:10.1109/access.2026.3693588

BibTeX
@article{Dehghani2026,
  author = {Dehghani, Soheil and Mohsenipour, Parastoo and Pahlevani, Farhan and Mohammadi, Mahshad and Farahani, Raha and Beigzadeh, Borhan},
  title = {Curiosity-Driven Reinforcement Learning for Autonomous Navigation in Magnetically Controlled Capsule Endoscopy},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {74221--74240},
  doi = {10.1109/access.2026.3693588}
}
Kostenlos starten →