Year: 2026
Type: conference-paper
Abstract: Ultrasound-guided minimally invasive puncture is widely applied in clinical biopsy and interventional therapy. However, it still suffers from low needle tip positioning accuracy, severe inter-frame trajectory jitter, weak dynamic anti-interference capability, and poor adaptability to nonlinear physiological motion. To address the positioning issues, this paper proposes a high-precision needle tip tracking framework combining Convolutional Neural Network and Mamba (CNN-Mamba) temporal fusion with adaptive Kalman filtering (KF). EfficientNetV2-B0 is adopted as the backbone to achieve robust single-frame needle tip coordinate regression. The Mamba selective state space model is introduced to enforce temporal sequence smoothing and jitter suppression while retaining subtle motion characteristics. An adaptive filtering strategy is further designed to switch dynamically between the 4D Constant Velocity (CV4) and 6D Constant Acceleration (CA6) motion models, employing Extended Kalman Filter (EKF) for linear motion and Unscented Kalman Filter (UKF) for nonlinear motion with adaptive observation noise adjustment. Experimental results show that, compared with the traditional CNN + KF baseline method, the proposed method reduces the mean absolute error (MAE) and root mean square error (RMSE) by 0.12% and 24.61%, respectively, while achieving superior tracking stability and dynamic adaptability. This method is highly robust to ultrasonic noise and artifacts, enabling real-time visual navigation of ultrasound-guided biopsy robots.
Authors: Yangning Shi, Yuan Xu, Erli Lyu
Institution: Macao Polytechnic University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7204789699