Multi-modality non-rigid medical image registration

This paper is concerned with the problem of multi-modality registration of images with complex geometrical relationship. We propose a high-dimensional approach based on point similarity measures and symmetric registration concept combined with a convolution-based geometric model. The high-dimensional registration approach follows the principles of elasticity as known from the physics. This is an iterative approach that consists of two steps (Fig. 1): estimation of forces that tend to deform the target image by making images more similar, and geometric model that realistically maps these forces into the actual transformation. The result of registration is transformation that maps the target image with the space of the reference image (Fig. 2). Measuring the image similarity is one of the main difficulties of non-rigid multi-modality registration approaches due to their limitations related to evaluation of local image differences. In our approach, point similarity measures are used to solve the problem. They enable a computationally effective measurement of the image similarity on arbitrary small image regions. We use a SUH point similarity measure with the point similarity function shown in Eq. 2 and measure the similarity of individual point pairs (Eq. 3). The similarities are employed to compute forces using the symmetric registration approach (Eq. 1) which estimates geometric differences in both registration directions. An increased amount of information obtained in the method improves registration consistency and correctness. The geometric model based on the convolution approach consists of four components (Fig. 3). The first component is a model of forces used to improve realisticity of forces. These forces are then mapped into transformation of independent voxels using the Hook's law. The last two components are a combined convolution model consisting of the incremental and absolute part. Such combined model reduces the systematic error of elasticity and error related to nonlinearity of forces. The overall registration approach is implemented in multi-resolution and is evaluated on images of the human head. The results indicate that the approach is suitable for solving practical clinical problems.