|Segmentation & Localization|
Novel Statistical Models and Segmentation Methods for
Fiber Bundles in DTI
Suyash Prakash Awate1, Hui Zhang1, James C. Gee1
1University of Pennsylvania, Philadelphia, USA
We present novel methods for statistically modeling and segmenting fiber bundles in diffusion-tensor images.
Temporal Dynamics 4D Level Set Method for
Segmentation of MR Renography Images
Ting Song1, Henry Rusinek2, Qun Chen2, Louisa Bokacheva2, Jeff L. Zhang2, Andrew F. Laine1, Vivian S. Lee2
1Columbia University, New York, New York, USA; 2NYU School of Medicine, New York, New York, USA
A novel 4D segmentation framework based on a temporal dynamics 4D level set requires less than one minute to automatically segment a 4D data set with more than 40 time points. The novelty of the method is that it combines both information from spatial anatomical structures and temporal dynamics from the time axis. The performance of the fully automatic 4D level set algorithm was found to be comparable to manual segmentation performed by two experts in renal anatomy.
Brain MR Image Segmentation by Minimizing
Scalable Neighborhood Intensity Fitting Energy: A Multiphase Level Set
Chunming Li1, Li Wang2, Chiu-Yen Kao3, Zhaohua Ding1, John C. Gore1
1Vanderbilt University, Nashville, Tennessee, USA; 2Nanjing University of Science and Technology, Nanjing, People's Republic of China; 3The Ohio State University, Columbus, Ohio, USA
Intensity inhomogeneity is often seen in MR images, and these cause considerable difficulties in applying existing image segmentation algorithms. In this work, we propose new multiphase level set method for brain MR images segmentation. The proposed method is able to segment images with intensity inhomogeneities, without the bias field correction. Our method has been applied to brain MR images at 3T with promising results.
MRI Measurement of Ischemic Brain Penumbra Using
Kohonen’s Multi-Parametric Self-Organizing Map (KMP-SOM) Technique
Hassan Bagher-Ebadian1, 2, Kourosh Jafari-Khouzani1, Panayiotis D. Mitsias1, Michael Chopp1, 3, James R. Ewing1, 3
1Henry Ford Hospital, Detroit, Michigan, USA; 2Amir-Kabir University of Technology, Tehran, Iran; 3Oakland University, Rochester, Michigan, USA
Experimental and clinical studies indicate that the likelihood for progression to infarction in the penumbra of physiologically impaired but potentially salvageable tissue surrounding in stroke ischemia is the most important factor in evaluating treatment efficacy. Thus, a multi-parametric analysis that increases the ability of investigators to characterize ischemic penumbra in the early stages of stroke may have a profound clinical significance. In this study, Kohonen’s Multi-Parametric Self Organizing Map (KMP-SOM) technique was used to detect the ischemic penumbra using MR acute information (T1 pre-contrast, T2, Diffusion-Weighted and proton density). Considering the DW and T2 lesions, the KMP-SOM maps have distinguished the penumbra and core of the lesions much better than ISODATA. We conclude that a KMP-SOM is capable of predicting the size and pattern of ischemic penumbra, from MR acute information and such modeling may play an important role in the assessment of subacute therapeutic interventions in the treatment of stroke.
Segmentation of Colorectal Cancer MR Images
Niranjan Bhaskar Joshi1, Sarah Bond2, Michael Brady1
1University of Oxford, Oxford, UK; 2Siemens Molecular Imaging, Oxford, UK
Relative distance of tumour from the mesorectal fascia (MF) in colorectal MR images provides important information about the stage of the cancer and likely success of the surgery. We present an image segmentation algorithm to estimate the position of the MF and size and location of the tumour using computer automated methods. Experiments are performed on oblique T2 weighted MRI dataset collected from 10 patients. Results show that the maximum of the average differences between an expert’s delineation of the MF and that segmented with our algorithm is just over 2 mm, which is equivalent to 3-5 pixels.
HippoQuant: Combining Geometrical and Intensity
Information for 3D Hippocampus Detection in 3D T1-Weighted MRI Images
Boubakeur Belaroussi1, Michael O'Sullivan2, Fabrice Vincent1, Chahin Pachai1
1Bio-Imaging Technologies S.A.S, Lyon, France; 2Neurologische Klinik, Ludwig-Maximilians-Universität, Munchen, Germany
In this work, we proposed HippoQuant, a new, fast, semi-automatic hippocampus segmentation procedure in 3D T1-weighted MRI images based on user-defined landmarks. From a set of user-defined landmarks, a 3D discrete hippocampus surface model is geometrically deformed using an Iterative Closest Point transform, supplemented by a Thin Plate Spline transform. This step leads to a “binary mask”. In parallel, the 3D MR image is segmented into 3 tissue classes (WM, GM and CSF) to identify CSF or CSF-like structures (uncal sulcus and “black holes”) located within the hippocampus. The output is a “presegmentation mask”. In the final step, the binary and presegmentation masks are merged to generate a hippocampal mask. HippoQuant combined both geometrical and statistical information, leading to potentially dramatic reductions in rater time and interaction, greater reliability, and a technique that is insensitive to variations in hippocampal size and shape, - all - important considerations for application in multi-centers studies.
Identification of Intratumour Low
Frequency Microvascular Components Via BOLD Signal Fractal Dimension
Graeme Wardlaw1, 2, Raimond Wong1, 3, Pierre Major1, 3, Michael D. Noseworthy1, 2
1McMaster University, Hamilton, Canada; 2Brain-Body Institute, Hamilton, Canada; 3Oncology, Hamilton, Canada
Typical clinical evaluation of tumour microvasculature utilises dynamic contrast enhanced MRI (dceMRI). However, this approach utilises numerous mathematical models with characteristic physiologic assumptions, often leading to inconclusive results. Alternatively, blood oxygen level dependent (BOLD) magnetic resonance imaging (MRI) is sensitive to the microvascular environment through fluctuation in the oxyhemoglobin to deoxyhemoglobin ratio. Consequently, quantification of the BOLD signal's temporal complexity using a fractal dimension index allows maps to be generated that are physiologically distinct and provide insight into tumour microvasculature without prior assumption as in dceMRI. This approach to tumour microvascular evaluation is presented using rectal cancer data.
An Approach to Prostate Segmentation on MR Images
Zhengyi Yang1, Aleksandra Zapotoczna2, Giuseppe Sasso2, John Simpson2, Gary Cowin1, Deming Wang1
1University of Queensland, Brisbane, Australia; 2The Townsville Hospital, Australia
While contouring the prostate manually, oncologists know the pelvic anatomy (not only the shape of prostate, but also its spatial relationship with other organs, such as bladder, rectum, and seminal versicles) and take the intensity distribution not only within but also surrounding the prostate into account. In order to incorporate this a prior knowledge of the neighbouring intensity distribution into model based segmentation method, Intensity Distribution Shell (IDS) model was proposed in this study. Given a prostate surface, by dilating and shrinking given numbers of voxels, we get two new surfaces enclosing the neighbouring tissues. The volume between these two surfaces is called a shell. The number of voxels is the shell thickness. The IDS model is the integration of a shape model obtained by principal component analysis (PCA) and an intensity distribution model represented by the histogram of the voxels within a shell with predefined shell thickness enclosing the shape surface.
Robustness of Morphologic Features for the
Characterization of Mass Lesions in Dynamic, Contrast-Enhanced Breast MR
Thomas Buelow1, Lina Arbash Meinel2, Axel Saalbach3, Rafael Wiemker1, Martin Bergtholdt1
1Philips Research Europe, Hamburg, Germany; 2Philips Research North America, Briarcliff Manor, New York, USA; 3Philips Research Europe, Aachen, Germany
Dynamic contrast enhanced breast MRI has been emerging as diagnostic tool. Due to its limited specificity there is a demand for computer aided diagnosis tools for this application. In order to build robust CAD applications yielding understandable and reproducible results, a carefully selected small set of features should be used. In this paper we compare three morphologic features with respect to their robustness against variations in the mass lesion segmentations that are the input to the feature computation step. The newly proposed feature "Normalized Mean Distance to Surface" proves more robust and more specific wrt. to lesion characterization than the common features "sphericality" and "compactness".
An Automated Assessment of White Matter Lesions Based
on Regional FLAIR Intensity Evaluation
Jacobus F.A. Jansen1, Paul A.M. Hofman1, Walter H. Backes1
1Maastricht University Hospital, Maastricht, Netherlands
White matter lesions (WML) are areas of bright, high signal intensity in the white matter depicted on T2-weighted magnetic resonance imaging. In a population with relatively mild WML, namely patients with localization-related, cryptogenic epilepsy and healthy volunteers, the performance of an automated WML detection algorithm, based on regional intensity evaluation, was assessed. The WML volumes from the automated segmentation method were found to be significantly correlated to the volumes obtained visually by neuroradiological assessment. The automated WML detection algorithm using a regional Z-score analysis can successfully segment and quantify the WML on FLAIR images.