Fast Non-rigid Medical Image Registration Based On Advanced Similarity Metrics

Recent years, as a very important technology, medical image registration algorithm had gradually been employed into helping doctors to diagnose complicated diseases. In order to increase the success rate of operation and accelerate the procedure of surgery, we always fuse various kinds of medical data in different module by means of medical image registration to help doctors make operation plans and diagnose diseases more efficiently. In many medical applications, such as clinical surgery and computer assisted virtual surgery, medical image registration can enhance the reality and assist surgeons pose the surgical position precisely. In this way, it can reduce the risk of treatment and make the operations running successfully.To meet the requirements of highly-grown clinical or virtual surgery, multi-modality and automatic medical image registration algorithm with high accuracy, high performance is indispensable. In the thesis, we engaged into researching out a novel fast and automatic registration algorithm, and successfully improved the modules of traditional registration framework.We successfully proposed a novel registration algorithm based on support vector machine, which increases the performance of algorithm and also reaches highly accuracy. Experimental results reveal that comparing to traditional algorithm, our proposed novel registration algorithm is efficient and fast, which also fit for various kinds of applications including non-rigid registration. It has good robustness and expansibility.The main works and innovation points are as described as follows:1. The overview and the categories of the medical image registration, and kinds of popular algorithms are discussed, including pre-processing, transformation, interpolation, similarity metric and optimization. These above common modules of registration had been researched and analyzed in the thesis.2. Interpolation makes great effects on similarity metric in the registration framework. Based on which, we proposed an advanced interpolation algorithm to increase the smoothness of similarity metric distribution. It could increase the speed of optimization.3. The proposed novel similarity metric is divided into two stages in this thesis. Due to which, a novel fast registration algorithm framework had been formed. By means of support vector machine, estimated similarity metric distribution could be built up from the relationship between parameters of transform and prior sparse target metric values. Based on which, global optimal parameters of transform are finally searched out by an improved optimizer in order to guide moving image to match the referenced image.4. With the help of multi-resolution strategy, we could skillfully select the control points in the non-rigid registration application. The pre-processing step, including resample technology, followed by free-form deformation, which is responsible for adjusting the parameters of transformation to reach the registration target. Finally, we finish the local deformation to increase the accuracy of registration results.5. Our proposed novel algorithm is successfully applied for many automatic registration applications, including 3D/3D and 3D/2D rigid and non-rigid registration. After comparing to the traditional algorithm by reasonable evaluation criterion, clinician experimental results reveal that our proposed registration method could improve performance and also provide a precise registration result efficiently.