Image Processing & Analysis

Tuesday 13 May 2014

Refaat E Gabt¹, Xiaojun Sun¹, Amol Pednekar², and Ponnada A Narayana^1
1Diagnostic and Interventional Radiology, University of Texas Health Science Center at Houston, Houston, TX, United States, ²Philips Healthcare, Cleveland, OH, United States

The quantitative power of MRI in neuroimaging is largely limited to research purposes and hardly finds its way into the clinical world, primarily due to the computationally-intensive nature of quantitative neuroimaging and the large amount of MRI data. We developed an integrated image acquisition and analysis pipeline that seamlessly combines the data acquisition with the high performance computation. To demonstrate the power of this pipeline, in 16 multiple sclerosis patients, the brain was classified into white matter, grey matter, cerebrospinal fluid, and white matter lesions in under one minute. The tissue maps were imported back to the scanner for clinical review.

Esben Plenge¹ and Michael Elad^1
1Dept. of Computer Science, Technion - Israel Institute of Technology, Haifa, Israel

A new method for super-resolution reconstruction of single MRI slice-stacks is presented. It exploits the fact that both high-resolution and low-resolution image content is present in MRI slice-stacks: In-plane slices contain high-resolution content, while the orthogonal planes containing the slice-selection direction are of low anisotropic resolution. The method relies on the powerful assumption of self-similarity of local image structures across these two scales. By training a pair of dicitonaries to capture the relationship between the scales, the resolution LR image planes can subsequently be increased by application of the dictionaries.

James Martin O'Callaghan¹, Jack Wells¹, Simon Richardson¹, Holly Holmes¹, Yichao Yu¹, Bernard Siow^1,2, and Mark Lythgoe^1
1Centre for Advanced Biomedical Imaging, UCL, London, UK, United Kingdom, ²Centre for Medical Image Computing, UCL, UK, United Kingdom

In this study, we present a gradient calibration protocol, together with an open source phantom, for the correction of errors caused by imaging gradients that can be easily implemented in pre-clinical MRI environments. Mean gradient scaling errors were reduced from 2.7% to 0.3% through a system calibration which may be significant in studies of murine brain volumetrics. Displacements of greater than 100µm were detected in imaging regions affected by gradient non-linearity and corrected through a post processing correction.

Fan Lam^1,2, Ding Liu^1,2, Zhuang Song^3,4, Michael W Weiner^3,4, Norbert Schuff^3,4, and Zhi-Pei Liang^1,2
1Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Urbana, Illinois, United States, ²Beckman Institute, University of Illinois at Urbana-Champaign, Urbana, Illinois, United States, ³Department of Veteran Affairs Medical Center, San Francisco, California, United States,⁴Radiology and Biomedical Imaging, University of California, San Francisco, California, United States

We developed a new fast algorithm for joint rank and edge constrained denoising of 3D magnitude diffusion-weighted image sequences by extending a recently proposed majorize-minimize framework for statistical estimation with noncentral χ distributions. Specifically, we extended the framework to consider joint rank and edge constraints, deriving a new algorithm that decomposes the original noncentral χ denoising problem into a series of rank and edge constrained Gaussian denoising problems. We show that the proposed algorithm achieves similar or even better denoising performance compared to a previously proposed algorithm, both qualitatively and quantitatively, but in significantly less time.

Thomas Demarcy¹, Axel Saalbach², and Julien Sénégas^2
1Ecole des Mines, Saint-Etienne, France, ²Philips Research Laboratories, Hamburg, Germany

To cope with today’s healthcare challenges of providing cost-effective care with constant quality, advanced image analysis techniques that automatically extract anatomical information from complex, large image datasets are required. In this work, computer vision techniques based on trained classifiers and Haar-like features were extended to 3D and applied to automatically detect and localize a number of target organs in 3D water-fat separated, whole-body MR images. The benefit of using the joint information provided by water and fat separation was investigated.

Umit Yoruk^1,2, Manojkumar Saranathan², Andreas M Loening², Brian A Hargreaves², and Shreyas S Vasanawala^2
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Radiology, Stanford University, Stanford, CA, United States

Accurate aortic input function (AIF) measurement is essential for DCE-MRI applications. We present a method that estimates the AIF from groups of temporally neighboring k-space samples, which have better temporal resolution and temporal footprint than the full k-space data. Since intensity changes of a structure over time are reflected in intensity changes of its frequency components, a portion of k-space too small to form good images can still be used to quantitatively determine AIF dynamics. We demonstrated the feasibility of this new method on a pediatric MR Urography subject and validated the results using a high spatiotemporal resolution digital phantom.

Yaniv Gal¹, Sze-Liang Chan¹, Rosalind Lindy Jeffree², Michael Fay^1,2, Paul Thomas^1,2, Stuart Crozier¹, and Zhengyi Yang^1
1University of Queensland, Brisbane, Queensland, Australia, ²Royal Brisbane & Women’s Hospital, Brisbane, Queensland, Australia

Inadequate photon attenuation correction (AC) in Positron Emission Tomography (PET) has serious implications, such as inaccurate cancer staging or failure to detect tumours. In hybrid PET-MRI, however, attenuation has to be corrected in the absence of CT. MRI-based attenuation correction AC (MRAC) is challenging because there is no direct correspondence between photon attenuation coefficient and image intensity in conventional MRI. We present a regression based, subject specific, MRAC for PET-MRI using the raw PET image without AC and multiple MR images, including structural MRI, pre- and post-contrast dynamic contract enhanced MRI, all of which are clinically available for most oncology patients.

Venkata Veerendranadh Chebrolu¹, Dattesh D Shanbhag¹, Sheshadri Thiruvenkadam¹, Sandeep Kaushik¹, Uday Patil¹, Patrice Hervo², Sandeep N Gupta³, and Rakesh Mullick^4
1Medical Image Analysis Lab, GE Global Research, Bangalore, Karnataka, India, ²GE Healthcare, Buc, France, ³Biomedical Image Processing Lab, GE Global Research, Niskayuna, NY, United States, ⁴Diagnostics and Biomedical Technologies, GE Global Research, Bangalore, Karnataka, India

DCE-MRI and pharmacokinetic (pK) model parameters derived from the DCE data have been commonly used for characterizing tumor vascular properties quantitatively. The accuracy of pK parameters depends on the choice of arterial input function (AIF). Partial volume effects due to horizontal course of the axillary artery in the axial plane, motion from pulsation and breathing may reduce the accuracy of the AIF selected from the axillary arteries. In this work we demonstrate a completely automated method for detection of AIF in the ascending aorta for breast DCE MRI and compare the results with AIF manually selected by an experienced radiologist.

Andrew L Alexander¹ and Steven R Kecskemeti^1
1Waisman Center, University of Wisconsin, Madison, WI, United States

Double inversion recovery (DIR) methods are used to generate gray matter (GM) specific images by simultaneously nulling the signals from CSF and white matter (WM). By using a new, efficient MP-nRAGE method, it is possible to generate an entire set of whole-brain, high-resolution, 3D inversion recovery contrast images with tissues nulled in different frames. By taking the ratio of the minimum intensity projection to the maximum intensity projection images of the CSF and WM null frames, it is possible to generate high quality GM maps with minimal coil sensitivity effects.

Xeni Deligianni¹ and Oliver Bieri^1
1Department of Radiology, Division of Radiological Physics, University of Basel Hospital, Basel, NA, Switzerland

Assessing tissue fat fraction and composition is of increased clinical interest as a biomarker for various muscle and liver diseases. Here, we propose a new method to achieve an ultra-dense echo sampling in the tens of microseconds regime within clinically feasible scan times. This allows a model-free signal time course analysis based on a singular value decomposition matrix pencil method to resolve the spectral components of fat. The effect of the echo spacing and the pre-set of signal spectral components is analyzed. Four fat peaks are being resolved and the results are in agreement with literature. Reducing the preset signal components results in an underestimation of the fat percentage.