Laparoscopic tubal sterilization is a widely used method for permanent contraception, where accurately identifying the uterus and fallopian tubes is crucial. However, this process becomes more difficult when these structures are obscured or partially hidden during surgery. Conventional imaging techniques often struggle to accurately identify anatomical features in such scenarios. In contrast, recent progress in deep learning, particularly with the You Only Look Once (YOLO) algorithm, has shown promising results by enabling real-time object detection in medical imaging. By incorporating KeyPoint detection with the newest YOLOv10 version, the system’s ability to detect and pinpoint anatomical landmarks is significantly improved, even when occlusions are present. This study presents a novel approach that merges YOLOv10 with KeyPoint detection to reliably identify the uterus and fallopian tubes in real-time laparoscopic footage, even when these structures are partially hidden. The method also identifies the Ampulla and Isthmus at the Uterine-Tubal junction, even in cases where visualization is obstructed. Additionally, an ellipse estimation technique is applied to further enhance detection under occluded conditions by estimating the uterus’s position and orientation when it is not fully visible. The results highlight the potential for real-time application of this system in clinical settings, marking a step forward in AI-assisted laparoscopic surgery.

In gynecologic surgery, a uterine manipulator is one of the instruments used to perform the laparoscopy. Throughout the past decade, a number of robotic technology applications used for uterine manipulation during surgery have been designed with the aim of increasing the efficiency, improving the precision, and reducing the workload of medical assistants. Although the RCM (Remote Center of Motion) mechanism is one of the key features in a Minimally Invasive Surgical (MIS) robot, the preliminary result in this study, in which the RCM mechanism was applied in a uterine manipulation robot, proved that this may cause unpleasant sensations such as irritation or harm to the nearby area during such manipulation. Therefore, a design of a non-RCM 2-DoF (Degree of Freedom) Robotic Uterine Manipulation System, in cooperation with an existing, reusable and tiltable-tip uterine manipulator, for laparoscopic gynecologic surgery has been proposed and evaluated via a mathematical model along with numerical analysis, a 3D uterus model, and a 1:1 uterus manikin model in order to demonstrate the use of the essential functions. According to the experimental results, the maximum load of 500 g has been handled well by the prototype, with the movement ranges of ±150° in the roll panel and ±90° in the pitch panel (0∼90° for anteversion and 0∼−90° for retroversion, if needed, which can be achieved by rotating the instrument to the other side). Furthermore, to verify this new design prior to its use on patients, and also in consideration of the ethics of human experimentation, through extensive testing on five donated soft-tissue cadavers, the proposed robot received positive feedback from all five surgeons performing the experiments and could offer effective uterine manipulation at the angular velocity of 4 °/s (0.67 RPM) with steady delineation of the vaginal fornices to create necessary motions in the pitch and roll panels of 30∼80° and ±15°, respectively, providing efficient visualization of the uterus. These features make this robot a valuable addition to the surgical instruments available to gynecologic surgeons.

Robotic-assisted Minimally Invasive Surgery (RMIS) has advanced laparoscopic gynecological procedures by improving precision and reducing invasiveness. However, the lack of direct force sensing during uterine manipulation remains a challenge, potentially increasing the risk of tissue damage and compromising surgical accuracy. This study presents an indirect force estimation method for robotic uterine manipulators using machine learning techniques to predict tip forces from handle force measurements. A dataset was developed by systematically varying uterine weight and ligament resistance, simulated through elastic bands of different colors and lengths. Three regression models—Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM) networks, and Support Vector Regression (SVR)—were evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R²). All models demonstrated reliable performance across diverse testing conditions, with XGBoost and SVR achieving a balance of predictive accuracy and computational efficiency, while LSTM effectively leveraged temporal dependencies. These findings underscore the viability of machine learning-based indirect force sensing as a practical alternative in scenarios where direct measurement is infeasible, paving the way for improved haptic feedback and surgical safety.