Comparing two approaches to rigid registration of three-dimensional ultrasound and magnetic resonance images for neurosurgery

Abstract

We present a new technique for registering magnetic resonance (MR) and ultrasound images in the context of neurosurgery. It involves generating a pseudo-ultrasound (pseudo-US) from a segmented MR image and uses cross-correlation as the cost function to register the pseudo-US to the real ultrasound data. The algorithm's performance is compared with that of a state-of-the-art technique that uses a median-filtered MR image to register to a Gaussian-blurred ultrasound using a normalized mutual information (NMI) objective function. The two methods were tested on data from 15 patients with brain tumor, including low-and high-grade gliomas, in both first operations and reoperations. Two metrics were used to evaluate registration accuracy: (1) the mean distance between corresponding points, identified on both MR and ultrasound images by two experts, and (2) ratings based on visual comparison by one neurosurgeon. The mean residual distance of the pseudo-US technique, 2.97 mm, is significantly more accurate (p = .0011) than that of the NMI approach, 4.86 mm. The visual assessment shows that only 4 of the 15 cases had a satisfactory initial alignment based on homologous skin-point registration. There is a significant correlation between the quantitative distance measures and the qualitative ratings (rho = 0.785). The results show that the pseudo-US rigid registration technique robustly improves the MRI-ultrasound alignment when compared with the initial alignment, even when applied to highly distorted brains and a large range of tumor sizes and appearances.