Karsten Østergaard Noe

Accelerating the Nonequispaced Fast Fourier Transform on Commodity Graphics Hardware

We present a fast parallel algorithm to compute the nonequispaced fast Fourier transform on commodity graphics hardware (the GPU). We focus particularly on a novel implementation of the convolution step in the transform as it was previously its most time consuming part. We describe the performance for two common sample distributions in medical imaging (radial and spiral trajectories), and for different convolution kernels as these parameters all influence the speed of the algorithm. The GPU-accelerated convolution is up to 85 times faster as our reference, the open source NFFT library on a state-of-the-art 64 bit CPU. The accuracy of the proposed GPU implementation was quantitatively evaluated at the various settings. To illustrate the applicability of the transform in medical imaging, in which it is also known as gridding, we look specifically at non-Cartesian magnetic resonance imaging and reconstruct both a numerical phantom and an in vivo cardiac image.

Acceleration and validation of optical flow based deformable registration for image-guided radiotherapy

Materials and methods. Two registration methods based on optical flow estimation have been programmed to run on a graphics programming unit (GPU). One of these methods by Horn & Schunck is tested on a 4DCT thorax data set with 10 phases and 41 landmarks identified per phase. The other method by Cornelius & Kanade is tested on a series of six 3D cone beam CT (CBCT) data sets and a conventional planning CT data set from a head and neck cancer patient. In each of these data sets 6 landmark points have been identified on the cervical vertebrae and the base of skull. Both CBCT to CBCT and CBCT to CT registration is performed. Results. For the 4DCT registration average landmark error was reduced by deformable registration from 3.5±2.0mm to 1.1±0.6mm. For CBCT to CBCT registration the average bone landmark error was 1.8±1.0mm after rigid registration and 1.6±0.8mm after deformable registration. For CBCT to CT registration errors were 2.2±0.6mm and 1.8±0.6mm for rigid and deformable registration respectively. Using GPU hardware the Horn & Schunck method was accelerated by a factor of 48. The 4DCT registration can be performed in 37seconds. The head and neck cancer patient registration takes 64seconds. Discussion. Compared to image slice thickness, which limits accuracy of landmark point determination, we consider the landmark point accuracy of the registration acceptable. The points identified in the CBCT images do not give a full impression of the result of doing deformable registration as opposed to rigid registration. A larger validation study is being planned in which soft tissue landmarks will facilitate tracking the deformable registration. The acceleration obtained using GPU hardware means that registration can be done online for CBCT.

Quantitative Neuroimaging Software for Clinical Assessment of Hippocampal Volumes on MR Imaging

Background: Multiple neurological disorders including Alzheimer’s disease (AD), mesial temporal sclerosis, and mild traumatic brain injury manifest with volume loss on brain MRI. Subtle volume loss is particularly seen early in AD. While prior research has demonstrated the value of this additional information from quantitative neuroimaging, very few applications have been approved for clinical use. Here we describe a US FDA cleared software program, NeuroreaderTM, for assessment of clinical hippocampal volume on brain MRI. Objective: To present the validation of hippocampal volumetrics on a clinical software program. Method: Subjects were drawn (n = 99) from the Alzheimer Disease Neuroimaging Initiative study. Volumetric brain MR imaging was acquired in both 1.5 T (n = 59) and 3.0 T (n = 40) scanners in participants with manual hippocampal segmentation. Fully automated hippocampal segmentation and measurement was done using a multiple atlas approach. The Dice Similarity Coefficient (DSC) measured the level of spatial overlap between NeuroreaderTM and gold standard manual segmentation from 0 to 1 with 0 denoting no overlap and 1 representing complete agreement. DSC comparisons between 1.5 T and 3.0 T scanners were done using standard independent samples T-tests. Results: In the bilateral hippocampus, mean DSC was 0.87 with a range of 0.78–0.91 (right hippocampus) and 0.76–0.91 (left hippocampus). Automated segmentation agreement with manual segmentation was essentially equivalent at 1.5 T (DSC = 0.879) versus 3.0 T (DSC = 0.872). Conclusion: This work provides a description and validation of a software program that can be applied in measuring hippocampal volume, a biomarker that is frequently abnormal in AD and other neurological disorders.

Evaluation of an application for intensity-based deformable image registration and dose accumulation in radiotherapy

Abstract Background. Methods to accurately accumulate doses in radiotherapy (RT) are important for tumour and normal tissues being influenced by geometric uncertainties. The purpose of this study was to investigate a pre-release deformable image registration (DIR)-based dose accumulation application, in the setting of prostate RT. Material and methods. Initially accumulated bladder and prostate doses were assessed (based on 8–9 repeat CT scans/patient) for nine prostate cancer patients using an intensity-based DIR and dose accumulation algorithm as provided by the Dynamic Adaptive Radiation Therapy (DART) software. The accumulated bladder and prostate dose-volume histograms (DVHs) were compared on a range of parameters (paired Wilcoxon signed-rank test, 5% significance level) to DVHs derived using an in-house developed dose accumulation method based on biomechanical, contour-driven DIR (SurfaceRegistration). Finally, both these accumulated dose distributions were compared to the ‘static’ DVH, assessed from the planning CT. Results. Over the population, doses accumulated with DART were overall lower than those from SurfaceRegistration (p < 0.05: D2%, gEUD and NTCP (bladder); Dmin (prostate)). The magnitude of these differences peaked for the bladder gEUD with a population median of 47 Gy for DART versus 57 Gy for SurfaceRegistration. Across the ten bladder dose/volume parameters investigated, the most pronounced individual differences were observed between the ‘accumulated’ DVHs and the ‘static’ DVHs, with deviations in mean dose up to 22 Gy. Conclusion. Substantial and significant differences were observed in the dose distributions between the two investigated DIR-based dose accumulation applications. The most pronounced individual differences were seen for the bladder and relative to the planned dose distribution, encouraging the use of repeat imaging data in RT planning and evaluation for this organ.

Uncertainties of deformable image registration for dose accumulation of high-dose regions in bladder and rectum in locally advanced cervical cancer

PURPOSE To compare the dose accumulation for bladder and rectum by deformable image registration (DIR) and direct addition (DA) of dose volume histogram parameters in magnetic resonance image-guided adaptive brachytherapy (IGABT). Two DIR algorithms, contour- and intensity-based, also have been analyzed. METHODS AND MATERIALS Patients (n = 21) treated with IGABT for carcinoma cervix under the IntErnational study on MRI-guided BRachytherapy in locally Advanced CErvical cancer protocol were analyzed. Each patient underwent two HDR-BT applications, 1-week apart with two fractions of 7 Gy each delivered per application. For each application, magnetic resonance imaging, volume delineation, reconstruction, treatment planning (BT1 and BT2), and dose evaluation were carried out. BT1 and BT2 images were registered using an intensity-based DIR, followed by deformable dose accumulation (DDA), which was then compared with DA. To compare the intensity-based DIR to other DIR approaches, nine patients were further evaluated using an in-house contour-based DIR algorithm for bladder dose accumulation. RESULTS Mean (±standard deviation; range) percentage variation between DA and DDA was found to be 2.4% (±3.3;-1.8, 11.5) and 5.2% (±5.1;-1.7, 16.5) for the rectum and bladder, respectively. The differences between the DA and DDA were found to be statistically significant for both rectum (p = 0.008) and bladder (p = 0.0003). Intensity-based DIR algorithm resulted in a larger mean deviation between DDA and DA as compared with contour-based DIR, although statistically insignificant (p = 0.32). The difference between DDA and DA was 2.4 ± 2.0% and 1.3 ± 1.2%, for intensity- and contour-based DIR, respectively. CONCLUSIONS DA of dose volume histogram parameters provides a good estimate to the dose to the organs at risk; DIR based on image intensities may lead to systematic underestimation of dose due to implausible DIR.

Solid mesh registration for radiotherapy treatment planning

We present an algorithm for solid organ registration of pre-segmented data represented as tetrahedral meshes. Registration of the organ surface is driven by force terms based on a distance field representation of the source and reference shapes. Registration of internal morphology is achieved using a non-linear elastic finite element model. A key feature of the method is that the user does not need to specify boundary conditions (surface point correspondences) prior to the finite element analysis. Instead the boundary matches are found as an integrated part of the analysis. The method is evaluated on phantom data and prostate data obtained in vivo based on fiducial marker accuracy and inverse consistency of transformations. The parallel nature of the method allows an efficient implementation on a GPU and as a result the method is very fast. All validation registrations take less than 30 seconds to complete. The proposed method has many potential uses in image guided radiotherapy (IGRT) which relies on registration to account for organ deformation between treatment sessions.

Surface membrane based bladder registration for evaluation of accumulated dose during brachytherapy in cervical cancer

A new optimisation strategy for surface mesh registration is proposed based on energy minimisation of an elastic surface membrane. The energy term is evaluated in three-dimensional space, optimisation however is performed while constraining the moving surface to a parametrised spherical space of the fixed surface. Optional landmark based matches can be included in the suggested iterative solver. The technique is demonstrated for bladder registration in brachytherapy treatment evaluation of cervical cancer. It holds promise to better estimate the accumulated but unintentional dose delivered to organs at risk.

An optimised multi-baseline approach for on-line MR-temperature monitoring on commodity graphics hardware

Magnetic resonance Imaging (MRI) can be used for non invasive temperature mapping and is therefore a promising tool to monitor and control interventional therapies based on thermal ablation. The proton resonance frequency shift MRI technique gives an estimate of the temperature by comparing phase changes between dynamically acquired images. These temperature measurements are prone to motion induced errors however, particularly in abdominal organs due to breathing. Several computational approaches have been proposed previously to correct for these motion related errors on the measured temperature. They have required significant time to compute however, and have not been sufficiently fast for several real-time temperature mapping applications. This paper proposes to use modern graphics cards (GPUs) to assess on-line motion corrected thermal maps. The computation times obtained on the GPU are compared to an existing CPU reference implementation. An acceleration factor close to 7 was obtained for the processing of one slice (resolution 128 times 128 pixels), and higher than 21 for 12 slices, allowing a real-time implementation.

SSLPV: subsurface light propagation volumes

This paper presents the Subsurface Light Propagation Volume (SSLPV) method for real-time approximation of subsurface scattering effects in dynamic scenes with changing mesh topology and lighting. SSLPV extends the Light Propagation Volume (LPV) technique for indirect illumination in video games. We introduce a new consistent method for injecting flux from point light sources into an LPV grid, a new rendering method which consistently converts light intensity stored in an LPV grid into incident radiance, as well as a model for light scattering and absorption inside heterogeneous materials. Our scheme does not require any precomputation and handles arbitrarily deforming meshes. We show that SSLPV provides visually pleasing results in real-time at the expense of a few milliseconds of added rendering time.

Active contours in optical flow fields for image sequence segmentation

Using variational calculus we develop an active contour model to segment an object across a number of image frames in the presence of an optical flow field. We define an energy functional that is locally minimized when the object is tracked across the entire image stack. Unlike classical snakes, image forces and regularization terms are integrated over the full set of images in the proposed model. This results in a new formulation of active contours. The method is demonstrated by segmenting the ascending aorta in a phase-contrast cine MRI dataset. Techniques to compute the required optical flow field and a “one-click” contour initialization step are suggested for this particular modality.