Project Category
Medical Image Non-Rigid Registration (NRR)
Description
Image Guided Neurosurgery (IGNS) is different from conventional neurosurgery by the availability of imaging which helps surgeon to understand what is going on inside the patient's brain while the tissue resection is underway. The major advantage of this is that the surgeon can avoid the critical regions of the brain, as they shift due to changing conditions as the surgery goes on. Thus, there is a better possibility for good outcome of the operation and improved quality of his or her life afterwards.
Non-rigid registration is one of the enabling technologies for IGNS. As the surgery progresses, the brain changes its shape and shifts. Thus, multimodal imaging data acquired prior to the surgery becomes invalid. During IGNS brain deformation is tracked by acquiring low resolution MRI in an open MR scanner, which is subsequently used to deform, or register, the preoperative data.
 |
 |
Preoperative MRI |
Intraoperative MRI |
Results
Volumetric registration of MRI using intraoperative imaging has long been too computationally expensive to be practical. Our distributed implementation of non-rigid registration, for the first time, enabled in-time delivery of registration results to the surgeons. The implementation we have developed is able to span multiple computational sites distributed geographically, and communicate the results back to to operating room.
 |
 |
Deformation field is computed using intraoperative image and mesh-based biomechanical model. |
|
Real-Time NRR on Cooperative Architecture
Description
Multicores and GPU provide us a cooperative architecture, in which both Single Instruction Multiple Data (SIMD) and Single Program Multiple Data (SPMD) programming models can co-exist and complement each other. In [BIBM09] we presented a method to parallelize a NRR on this cooperative architecture. Our approach is first to separate the sequential algorithm into regular part (Block Matching) and irregular part (Incremental Finite Element Solver). We then map the regular part on GPU following SIMD paradigm and irregular part on multicores in a SPMD fashion. Unlike the approaches that use multicores or GPU alone, our approach leads to desirable speedup for the whole application by taking advantage of all components of the cooperative parallel architecture, for all individual parts of the application. This helps us to get closer to our goal: cheaper and faster NRR that leads to its more widespread use.
 |
Preoperative MRI |
Results
We compared the performance of the Block Matching on a typical modern workstation Dell Precision T3400 equipped with NVIDIA GeForce 8800 GT GPU with its MPI implementation running on a 8-node cluster (each node is Dell PowerEdge SC1435, 2 x dual-core Opteron 2218, 2.6 GHz CPU). The results were collected for computations on 6 retrospective brain tumor resection cases, with the imaging parameters similar to the ones used for acquisition of brain imaging in SPL Brain Tumor Resection dataset.
The following figure shows the comparison of performance for the considered implementations. Compared to the 4-node cluster (16 CPUs), the minimum speedup is 3.9 (case 1) and the maximum speedup is 7.7 (case 5). Compared to the 8-node cluster (32 CPUs), the minimum speedup is 1.9 (case 1) and the maximum speedup is 3.8 (case 5).
 |
Performance evaluation using six existing retrospective data from BWH. Image block: 9x9x9, Search window: 11x11x19, Thread block: 4x4x4 |
The solver runtime depends on the size of the mesh. Three meshes of increasing sizes were generated. Biconjugate gradient solver based on gmm is used for comparison with our parallel incremental solver. The execution times for assembling and iterative solution of the linear system are listed in the table.
 |
Performance comparison between sequential and parallel solver. |
Our experiments show that the difference of the accuracy between the parallel and the sequential implementations is below 0.006mm (large mesh)
 |
Precision comparison between sequential and parallel NRR on different size of meshes. |
A Novel Point Based Non-rigid Registration Method and Its Application for Brain Shift
Description
The purpose of point based NRR is to find mapping function, which generally requires to know correspondence. One kind of methods are to use some specific algorithm to find the correspondence and then solve mapping function. The other kind of methods do not rely on any specific algorithms to find correspondence, but solve them (correspondence and mapping function) simultaneously. The representative of this kind of methods is point matching method (RPM), which is an extension of well-known iterative closest point (ICP) algorithm. However, RPM employs thin-plate splines (TPS) as mapping function and therefore is incapable of estimating the deformation given sparse data because TPS is not compact support. Furthermore, RPM cannot deal with the outliers existing in both source and target point sets. To overcome these limitations, In [SPIE Medical Imaging 2010] we combine biomechanical model with RPM framework to deal with sparse point sets and employ robust regression technique to to deal with partially overlapping point sets.
This NRR method has a wide application because it has a very loose requirement for the input data: sparse and even partially overlapping point sets. One important application is surface based brain shift, which is characterized by 1. only sparse intraoperative information available(scanned surface). 2. Lots of outliers in both point sets. 3. entire brain deformation is required.
 |
||||
Scanned surface |
Point sets |
Deformed mesh |
Matching surface |
Deformation magnitude |
Results
Experiment results about surface based brain shift are shown as below.
 |
a: preMRI. b: deformed preMRI. c: iMRI used to evaluate the accuracy of this method. d: the superimposing of iMRI and the extracted boundary of preMRI. e: the superimposing of iMRI and the extracted boundary of deformed preMRI. |
Software
The following software is used in this project.
-
- CMake
-
- Paraview
Related Links