[PAST EVENT] PhD Defense: Yixun Liu

July 15, 2011
McGlothlin-Street Hall, Room 002
251 Jamestown Rd
Williamsburg, VA 23185Map this location
Three critical issues about medical image non-rigid registration are performance, robustness and accuracy. A registration method, which is capable of responding timely with an accurate alignment, robust against the variation of the image intensity and the missing data, is desirable for its clinical use. This work addresses all three of these issues.

Unacceptable execution time of Non-rigid registration (NRR) often presents a major obstacle to its routine clinical use. We present a hybrid data partitioning method to parallelize a NRR method on a cooperative architecture, which enables us to get closer to the goal: accelerating using architecture rather than designing a parallel algorithm from scratch. To further accelerate the performance for the GPU part, a GPU optimization tool is provided to automatically optimize GPU execution configuration.

Missing data and variation of the intensity are two severe challenges for the robustness of the registration method. A novel point-based NRR method is presented to resolve mapping function (deformation field) with the point correspondence missing. The novelty of this method lies in incorporating a finite element biomechanical model into an Expectation and Maximization (EM) framework to resolve the correspondence and mapping function simultaneously. This method is extended to deal with the deformation induced by tumor resection, which imposes another challenge, i.e. incomplete intra-operative MRI. The registration is formulated as a three variable (Correspondence, Deformation Field, and Resection Region) functional minimization problem and resolved by a Nested Expectation and Maximization framework. The experimental results show the effectiveness of this method in correcting the deformation in the vicinity of the tumor. To deal with the variation of the intensity, two different methods are developed depending on the specific application. For the mono-modality registration on delayed enhanced cardiac MRI and cine MRI, a hybrid registration method is designed by unifying both intensity- and feature point-based metrics into one cost function. The experiment on the moving propagation of suspicious myocardial infarction shows effectiveness of this hybrid method. For the multi-modality registration on MRI and CT, a Mutual Information (MI)-based NRR is developed by modeling the underlying deformation as a Free-Form Deformation (FFD). MI is sensitive to the variation of the intensity due to equidistant bins. We overcome this disadvantage by designing a Top-to-Down K-means clustering method to naturally group similar intensities into one bin. The experiment shows this method can increase the accuracy of the MI-based registration.

In image registration, a finite element biomechanical model is usually employed to simulate the underlying movement of the soft tissue. We develop a multi-tissue mesh generation method to build a heterogeneous biomechanical model to realistically simulate the underlying movement of the brain. We focus on the following four critical mesh properties: tissue-dependent resolution, fidelity to tissue boundaries, smoothness of mesh surfaces, and element quality. Each mesh property can be controlled on a tissue level. The experiments on comparing the homogeneous model with the heterogeneous model demonstrate the effectiveness of the heterogeneous model in improving the registration accuracy.