Fast Tip Tracking via Needle Detection and Estimation during Ultrasound-guided Prostate Biopsy

Objective: Accurate and real-time needle tip tracking is critical in image-guided prostate biopsy procedures. Ultrasound (US)-based tracking methods, though, are afflicted by suboptimal image quality, low needle visibility, and dynamic, variable appearance of the needle tip. Therefore, this task requires a lightweight and robust approach, capable of handling complex US environments in real-time. Methods: We propose a Needle Detection and Estimation (NDE) framework for US-guided needle-tip tracking in real time. Instead of directly detecting the needle tip, we first detect the more consistently represented needle axis using a lightweight line detection model. The needle tip is then estimated as one of the endpoints of such a detected axis. A Kalman-filter-based needle estimation module is added to take advantage of the last detection to improve and guide the current detection, which is particularly helpful within frames where the needle is not visible. Segment of Line (SoL) and Copy-Paste object augmentation strategies are used to increase robustness while reducing dependency on large training datasets. Results: Experiments demonstrate that our method outperforms existing tracking techniques, achieving success rates of 95.27% and 95.62%, and average tracking errors of 1.94 mm and 2.53 mm for the In vitro and In vivo datasets, respectively. Furthermore, our method achieves a tracking speed of 102.13 FPS, making it suitable for real-time scenarios. Conclusions: Our NDE framework can significantly enhance the reliability and efficiency of needle tip tracking during image-guided prostate biopsy. Significance: The proposed method shows the potential to improve both clinical diagnosis and therapy for prostate cancer.