Yixun Liu

Modeling class cohesion as mixtures of latent topics

The paper proposes a new measure for the cohesion of classes in Object-Oriented software systems. It is based on the analysis of latent topics embedded in comments and identifiers in source code. The measure, named as Maximal Weighted Entropy, utilizes the Latent Dirichlet Allocation technique and information entropy measures to quantitatively evaluate the cohesion of classes in software. This paper presents the principles and the technology that stand behind the proposed measure. Two case studies on a large open source software system are presented. They compare the new measure with an extensive set of existing metrics and use them to construct models that predict software faults. The case studies indicate that the novel measure captures different aspects of class cohesion compared to the existing cohesion measures and improves fault prediction for most metrics, which are combined with Maximal Weighted Entropy.

Real-Time Non-rigid Registration of Medical Images on a Cooperative Parallel Architecture

Unacceptable execution time of Non-rigid registration (NRR) often presents amajor obstacle to its routine clinical use. Parallel computing is aneffective way to accelerate NRR. However, development of efficient parallelNRR codes is a very challenging task. One desirable approach is to map theexisting sequential algorithm to the parallel architecture to gain speedupinstead of designing a new parallel algorithm. Multicores and GPU provide usa cooperative architecture, in which both Single Instruction Multiple Data(SIMD) and Single Program Multiple Data (SPMD) programming models can co-existand complement each other. We present a method to parallelize aNRR on this cooperative architecture. Our approach is first to separate thesequential algorithm into regular and irregular parts. We then map the regularpart on GPU following SIMD paradigm and irregular part on multicores in a SPMD fashion. Unlike the approaches that use multicores orGPU alone, our approach leads to desirable speedup for the whole applicationby taking advantage of all components of the cooperative parallelarchitecture, for all individual parts of the application. This helps us toget closer to our goal: cheaper and faster NRR that leads to its morewidespread use. The results on clinical brain MRI data showthat the GPU-based Block Matching (regular part) can run at least 1.9 timesfaster than on a typical cluster of workstations with eighthigh-performance nodes. The multicores-based implementation of theincremental finite element solver (irregular part) achieves speedup of up to7 times compared to its sequential version. As a result, the total run timeof the NRR code can be reduced to less than 1 minute therefore satisfyingthe real time requirement for its clinical application.

Multi-tissue Mesh Generation for Brain Images

We develop a multi-tissue mesh generation method that is suitable for finite element simulation involved in non-rigid registration and surgery simulation of brain images. 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. This method consists of two steps. First, a coarse multi-tissue mesh with tissue-dependent resolution is generated according to a predefined subdivision criterion. Then, a tissue-aware point-based registration method is used to find an optimal trade-off among fidelity, smoothness, and quality. We evaluated our method on a number of images ranging from MRI, visible human, to brain atlas. The experimental results verify the features of this method.

Mesh deformation-based multi-tissue mesh generation for brain images

Multi-tissue meshing is necessary for the realistic building of a biomechanical model of the brain, which has been widely used in brain surgery simulation, brain shift, and non-rigid registration. A two step multi-tissue mesher is developed. First, a coarse multi-tissue mesh is generated by redistributing labels of a body-centered cubic (BCC) mesh. Second, all the surfaces of the submeshes are deformed to their corresponding tissue boundaries. To deform the mesh, two strategies are developed. One is based on a point-based registration (PBR) and the other is based on a robust point matching (RPM). The PBR method explicitly calculates the correspondence, which takes both smoothing and quality into account, then resolves the displacement vector by minimizing an energy function. Unlike PBR method, RPM does not require the correspondence between the source points and the target points to be known in advance. To simultaneously resolve the displacement vector and the correspondence, the Expectation and Maximization optimization is employed to alternately estimate the correspondence and the displacement vector. To effectively cope with outliers, least trimmed square, a robust regression technique, is employed to correct the regression bias induced by outliers. Both methods are effective in deforming the multi-tissue mesh. However, the PBR method favors quality and smoothing, and the RPM method favors fidelity. The resulting mesh is characterized by its flexible control of four mesh properties: (1) tissue-dependent resolution, (2) fidelity to tissue boundaries, (3) smoothness of mesh surfaces, and (4) element quality. Each mesh property can be controlled on a tissue level. Our experiments conducted on synthetic data, clinic MRI, visible human data, and brain atlas effectively demonstrate these features of this multi-tissue mesher.

A point based non-rigid registration for tumor resection using iMRI

This paper presents a novel feature point based non-rigid registration of preoperative MRI with resected intra-operative MRI (iMRI) to compensate for brain shift during tumor resection.

An ITK implementation of a physics-based non-rigid registration method for brain deformation in image-guided neurosurgery

As part of the ITK v4 project efforts, we have developed ITK filters for physics-based non-rigid registration (PBNRR), which satisfies the following requirements: account for tissue properties in the registration, improve accuracy compared to rigid registration, and reduce execution time using GPU and multi-core accelerators. The implementation has three main components: (1) Feature Point Selection, (2) Block Matching (mapped to both multi-core and GPU processors), and (3) a Robust Finite Element Solver. The use of multi-core and GPU accelerators in ITK v4 provides substantial performance improvements. For example, for the non-rigid registration of brain MRIs, the performance of the block matching filter on average is about 10 times faster when 12 hyperthreaded multi-cores are used and about 83 times faster when the NVIDIA Tesla GPU is used in Dell Workstation.

An Evaluation of Tetrahedral Mesh Generation for Nonrigid Registration of Brain MRI

In this chapter, we assess the impact of mesh generation on nonrigid registration of brain MR images. The solution accuracy and the speed of finite element solvers depend on how well the underlying mesh approximates the surface of the biological object (fidelity), and how well the elements of this mesh are shaped (quality). Fidelity and quality, however, are two contradicting requirements, as increased fidelity usually implies poor quality and vice versa. In this chapter, we evaluate three public mesh generators and examine how this quality-fidelity trade-off affects the accuracy and the speed of nonrigid registration solvers for brain images.

Toward a real time multi-tissue Adaptive Physics-Based Non-Rigid Registration framework for brain tumor resection

This paper presents an adaptive non-rigid registration method for aligning pre-operative MRI with intra-operative MRI (iMRI) to compensate for brain deformation during brain tumor resection. This method extends a successful existing Physics-Based Non-Rigid Registration (PBNRR) technique implemented in ITKv4.5. The new method relies on a parallel adaptive heterogeneous biomechanical Finite Element (FE) model for tissue/tumor removal depicted in the iMRI. In contrast the existing PBNRR in ITK relies on homogeneous static FE model designed for brain shift only (i.e., it is not designed to handle brain tumor resection). As a result, the new method (1) accurately captures the intra-operative deformations associated with the tissue removal due to tumor resection and (2) reduces the end-to-end execution time to within the time constraints imposed by the neurosurgical procedure. The evaluation of the new method is based on 14 clinical cases with: (i) brain shift only (seven cases), (ii) partial tumor resection (two cases), and (iii) complete tumor resection (five cases). The new adaptive method can reduce the alignment error up to seven and five times compared to a rigid and ITKs PBNRR registration methods, respectively. On average, the alignment error of the new method is reduced by 9.23 and 5.63 mm compared to the alignment error from the rigid and PBNRR method implemented in ITK. Moreover, the total execution time for all the case studies is about 1 min or less in a Linux Dell workstation with 12 Intel Xeon 3.47 GHz CPU cores and 96 GB of RAM.

Tetrahedral Image-to-Mesh Conversion Approaches for Surgery Simulation and Navigation

In this paper we evaluate three different mesh generation approaches with respect to their fitness for use in a surgery simulation and navigation system. The behavior of such a system can be thought of as a trade-off between material fidelity and computation time. We focus on one critical component of this system, namely non-rigid registration, and conduct an experimental study of the selected mesh generation approaches with respect to material fidelity of the resulting meshes, shape of mesh elements, condition number of the resulting stiffness matrix, and the registration error. We concluded that meshes with very bad fidelity do not affect the accuracy drastically. On the contrary, meshes with very good fidelity hurt the speed of the mesher due to the poor quality they exhibit. We also observed that the speed of the solver is very sensitive to mesh quality rather than to fidelity. For these reasons, we think that mesh generation should first try to produce high quality meshes, possibly sacrificing fidelity.

Moving propagation of suspicious myocardial infarction from delayed enhanced cardiac imaging to CINE MRI using hybrid image registration

Cardiac magnetic resonance imaging has proved its effectiveness to determine the patient-specific myocardial motion/functional information via the cine imaging and to detect myocardial infarction in the delayed enhanced MRI (DE-MRI). Standard cardiac MR protocols usually acquire these two sets of images across multiple acquisitions with varying imaging slice geometry, pixel spacing and different breath-holdings, which could make the joint inspection of myocardial motion and infarction difficult. The purpose of this work is therefore to develop dedicated post-processing algorithms to register DE-MRI to corresponding cine image and propagate suspicious infarction to all cardiac phases. Suspicious infarction regions delineated in the DE-MRI can be used to define the region-of-interest for the quantification of regional wall motion abnormality. The proposed approaches are applied to 6 patients and the evaluation shows the feasibility of a joint DE-MRI and cine assessment which can yield clinically valuable outcomes.