Accurate Needle Localization in the Image Frames of Ultrasound Videos

Ultrasound imaging provides real-time guidance during needle interventions, but localizing the needle in ultrasound videos remains a challenging task. This paper introduces a novel machine learning-based method to localize the needle in ultrasound videos. The method comprises three phases for analyzing the image frames of the ultrasound video and localizing the needle in each image frame. The first phase aims to extract features that quantify the speckle variations associated with needle insertion, the edges that match the needle orientation, and the pixel intensity statistics of the ultrasound image. The features are analyzed using a machine learning classifier to generate a quantitative image that characterizes the pixels associated with the needle. In the second phase, the quantitative image is processed to identify the region of interest (ROI) that contains the needle. In the third phase, the ROI is processed using a custom-made Ranklet transform to accurately estimate the needle trajectory. Moreover, the needle tip is identified using a sliding window approach that analyzes the speckle variations along the needle trajectory. The performance of the proposed method was evaluated by localizing the needle in ex vivo and in vivo ultrasound videos. The results show that the proposed method was able to localize the needle with failure rates of 0%. The angular, axis, and tip errors computed for the ex vivo ultrasound videos are within the ranges of 0.3–0.7°, 0.2–0.7 mm, and 0.4–0.8 mm, respectively. Additionally, the angular, axis, and tip errors computed for the in vivo ultrasound videos are within the ranges of 0.2–1.0°, 0.3–1.0 mm, and 0.3–1.1 mm, respectively. A key advantage of the proposed method is the ability to achieve accurate localization of the needle without altering the clinical workflow of the intervention.