3D Keypoint-Based Neural Network for Rapid Needle Localization on Multislice 2D MRI

Motivation: Rapid, accurate needle localization on multislice images is essential for guiding needle-based MRI-guided interventions. Compared to segmentation-based methods, a keypoint-based approach can simplify reference creation and directly predict needle coordinates Goal(s): Develop a 3D keypoint-based method for needle localization on multislice 2D MRI. Approach: A 3D stacked hourglass neural network model with part affinity fields was developed for needle keypoint detection and compared to 3D UNet segmentation-based results and human-annotated references. Results: The keypoint model achieved median needle tip and axis errors of 2.1 mm and 0.9°, with ~30ms computational time. The accuracy outperformed 3D UNet and was comparable to human intra-reader variation. Impact: This work developed a 3D keypoint-based method for rapid, accurate needle localization on multislice 2D MRI, outperforming 3D UNet and matching human intra-reader variation. The keypoint-based method could be adapted to assist different needle-based MRI-guided interventions.