2D/3D US-TO-MRI RIGID REGISTRATION BY DEEP LEARNING

Abstract

2D ultrasound (US) images to 3D magnetic resonance (MR) image registration is a crucial module in US-guided surgical navigation. US images need to be aligned with a preoperative image to provide good anatomy information guidance during interventions. However, the difference between the modality of US and MR makes the task challenging. To address this problem, we propose a learning-based rigid registration method between 2D US and 3D MR. The geodesic distance on the special Euclidean group SE(3) equipped with a left-invariant Riemannian metric is used as the loss function of a regression network. The registration result is optimized from the registration network by maximizing the similarity metric defined by a local structure orientation descriptor (LSOD). We achieve the angle and distance errors of 3.83 ± 0.39° and 0.017 ± 0.001 mm, outperforming the L2 norm loss function which results in 4.21 ± 0.19° angle error and 0.039 ± 0.001 mm distance error. Qualitative and quantitative evaluations confirm that the proposed method can achieve accurate 2DUS-3DMRI rigid registration.