نوع مقاله : مقاله پژوهشی
نویسندگان
1 دانشجوی دکترای دانشکده مهندسی برق و کامپیوتر، دانشگاه تربیت مدرس، تهران، ایران
2 دانشیار دانشکده مهندسی برق و کامپیوتر، دانشگاه تربیت مدرس، تهران، ایران
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Given widespread concerns about the complications associated with traditional endoscopy, research into the use of less invasive endoscopic capsules has gained significant attention. However, the passive movement of these capsules often prevents access to specific angles and areas of interest to the clinician. To overcome this limitation, we propose a novel autonomous navigation approach based on deep reinforcement learning, utilizing a proximal policy optimization (PPO) algorithm. This method automates the capsule's positioning, pathfinding, and motion control. Our approach integrates multi-modal sensor data to estimate the target point over time. Recognizing that initial algorithm training requires substantial data, we developed a near-realistic virtual environment. This environment facilitates the training of an intelligent agent and includes an actuator model with a magnetic coil structure, a capsule equipped with a dipole magnet, a camera, an inertial sensor, and a 3D model of the large intestine. The primary objective of this research is to reduce operator intervention, allowing clinicians to focus more on the clinical and medical aspects of endoscopy. The proposed method was trained with various hyperparameters, and its performance was evaluated based on metrics such as movement and alignment toward the target, as well as entropy. The evaluation results demonstrate that by optimally adjusting the buffer size and batch size, the algorithm achieves effective tracking and stability.
کلیدواژه‌ها [English]
)