A Boundary-Guided Needle Localization Approach for MRI-Guided Robotic Interventions

Accurate and rapid needle localization in three-dimensional (3D) magnetic resonance images is critical for MRI-guided robotic interventions. However, the slim and slender shape of the needle, combined with the influence of the surrounding tissue interference, poses significant challenges to accurate needle tip localization. In this paper, we develop a real-time boundary-guided needle localization approach to improve the accuracy and efficiency of MRI-guided robotic interventions. To acquire the 2D MR slice images containing the entire needle, the Brown-Roberts-Wells frame is first utilized to automatically and quickly locate the two-dimensional (2D) needle plane in the 3D space. A novel boundary-guided segmentation network named BFine-Mask is then designed to segment the needle mask accurately with the guidance of the shape prior. Specifically, we design a balanced fine-grained feature pyramid network (BF-FPN) to better integrate multi-scale features and improve the representation of fine-grained boundary details. Additionally, a boundary feature branch is proposed to leverage shape information at the boundary to enhance segmentation accuracy in the tip region. Finally, a post-processing algorithm based on centerline detection is implemented to accurately locate the needle tip from the segmented mask. The proposed approach is thoroughly evaluated using a 15-gauge ceramic needle and a 14-gauge medical titanium needle in the pig brain and heart tissue phantoms with the guidance of a 5T MRI machine. In three different scenarios, our approach achieves average needle tip localization errors of 0.597 mm (1.01 pixels), 0.856 mm (1.45 pixels), and 0.590 mm (1.00 pixels). This approach achieves sub-millimeter positioning accuracy within one second, demonstrating its potential for future routine clinical applications.