Image registration is a key step in computer-assisted analysis of medical images and image-guided procedures. For instance, it is used to detect and quantify normal and pathological changes over time, such as monitoring the course of neurodegeneration in head examinations, monitoring of tumor development, contouring of critical structures in images for radiotherapy planning through registration of topological atlases, etc. Furthermore, by registering pre- and inter-operative images in real time, we can localize the patient's anatomy in the operating room during the procedure, transfer the associated pre-operative plan of the procedure and thus enable minimally invasive surgical procedures. In a particular clinical context, the applicability of medical image registration method is critically determined by the appropriate balance and/or trade-offs between accuracy, reliability, and computational efficiency.
Image registration is performed by searching for spatial transform parameters via a numerical optimization algorithm, which aims to find the (global) optimum of a measure of similarity between the reference and moving image. Non-rigid registration of images is poorly conditioned because the space of solutions is generally infinite, and at the same time, due to the large number of free transformation parameters, it is also computationally demanding. A widely used computationally efficient optimization approach is the iterative gradient descent, however, the achievable alignment accuracy and reliability may be limited due to local minima in the similarity measure, especially in applications where the initial transformation is far from the global optimum. The latter problem is addressed by global optimization procedures, among which a typical representative is the genetic algorithm. Due to the large space of possible solutions, finding the global optimum remains computationally very demanding.
With the increasing capacity of graphic processing units, both in terms of computing and memory, the field of parallel programming is entering the domain of medical image analysis and, in certain cases, promises high acceleration compared to classical serial algorithms. The genetic algorithm in particular allows an efficient parallel implementation of the optimization process. Therefore, the aim of this thesis was to develop and evaluate a method for non-rigid alignment of medical images, which uses a genetic optimization algorithm that is independent of image acquisition technique, can be performed on one or more graphical processing units, and has a flexible software interface to achieve the desired relationship between accuracy, reliability and time efficiency of the registration.
The procedure was based on non-rigid transformation using B-splines. As a measure of image similarity we compared the normalized gradient fields between the reference and the moving images. To eliminate implausible solutions, we included two regularization terms in the criterion function; first, to penalize large shifts and, second, to prevent folding of the moving image. Due to the memory limitations on the graphics processing unit, as well as to take advantage of within-cycle dual 16-bit floating point math operations capability, we successfully tested a 16-bit floating point number format of grayscale images as a possible solution. Due to the large number of spatial transformation parameters we adjusted the mutation process and thereby improved convergence and solved the problem of completely random, so called blind mutation.
The developed non-rigid registration method was objectively and quantitatively evaluated and compared with the established freely available software packages for the non-rigid registration of medical images. We evaluated the methods on a private image database and showed comparable results of the developed method compared to the established methods, but achieving a shorter registration time. This features creates new accuracy-reliability-efficiency trade-off opportunities and thus increase the potential for practical application. Due to the high speed of alignment, the developed method is suitable for time-critical applications, for example in the analysis and adjustment of irradiation plans and in image-guided interventions. Flexibility and scalability to multiple graphics processing units thus offers a competitive alternative to the existing established methods.
|
