Motivation: MR-guided needle interventions require accurate and precise knowledge of the needle position and orientation, which is limited by image distortions. Goal(s): Provide differentiable image-based needle localization that inherently leverages image distortions that typically produce positioning errors. Approach: Needle susceptibility maps were modeled with a mixture of Gaussian functions that are related differentiably to a simulated distorted image. The simulated image was iteratively compared to a measured image to recover the true needle position and orientation. Simulations validated the method compared to a PCA approach. Results: The proposed approach achieved up to 50% lower position and angular errors than the PCA approach. Impact: Accurate needle localization is key to the safety and efficacy of interventional MR procedures. The proposed differentiable image-based method recovers needle position and pose accurately, regardless of needle-induced image distortions.
Garrow et al. (Tue,) studied this question.