Deformable registration of ultrasound and magnetic resonance imaging using a new self-similarity based neighborhood descriptor

Abstract

Purpose. Computer-aided intervention often depends on multi-modal deformable registration to make full use of images taken at different times to provide complementary information. However, multi-modal deformable registration remains a challenging task and is an active research area in medical image analysis, especially in the deformable registration between ultrasound (US) and magnetic resonance imaging (MRI). Methods. This paper addresses this problem by proposing a more robust and representative structural image descriptor that combines the merits of modality independent neighborhood descriptor (MIND) and self-similarity context (SSC) using a distance weight (MINDWSSC) and transforms the complex multi-modal registration problem into a mono-modal registration. Gauss-Newton optimization with the sum of the square differences as a metric is used to build the registration framework. Results. The BrainWeb database and a set of 13 pairs of pre-operative MRI and pre-resection three-dimensional US images from the Brain Images of Tumors for Evaluation (BITE) database were used to validate the advantage of MINDWSSC compared with the MIND and SSC descriptors. The MINDWSSC achieved the best overall registration accuracy of 2.29±1.2 mm, significantly outperforming the MIND and the SSC. Conclusions. The MINDWSSC is new descriptor that is more robust to noise, is more representative, and combines the merits of the MIND and the SSC. It has the potential to deal with the highly challenging MRI-US registration.