Surgery – Diagnostic testing – Detecting nuclear – electromagnetic – or ultrasonic radiation
Reexamination Certificate
2000-05-01
2002-08-06
Lateef, Marvin M. (Department: 3737)
Surgery
Diagnostic testing
Detecting nuclear, electromagnetic, or ultrasonic radiation
C128S920000, C128S923000, C128S925000, C382S293000, C382S294000
Reexamination Certificate
active
06430430
ABSTRACT:
2. FIELD OF THE INVENTION
This invention relates generally to the field of magnetic resonance imaging (“MRI”) and, in particular, to the automation of the interpretation of information present in MRI images to detect brain lesions. More specifically, the present invention relates to a method and/or system of detecting hyperintense regions in MRI images that are suspected of being related to various brain pathologies, such as Alzheimer's Disease, Multiple Sclerosis and like neurodegenerative conditions and/or determining volumetric measurements of cerebral anatomical regions.
3. BACKGROUND OF THE INVENTION
Present day computerized methods of hyperintensity identification in brain magnetic resonance images either rely heavily on human intervention, or on simple thresholding techniques. Consequently, these methods lead to considerable variation in the quantification of brain hyperintensities depending on image parameters such as contrast.
A review of a number of the published computerized segmentation techniques reveals the following. The manual or semi-automatic techniques, requiring informed judgment on the part of an experienced diagnostician, will be discussed first followed by a discussion of techniques considered fully automatic that generally require little or no operator interaction.
3.1. Manual or Semi-Automatic Techniques
Cline et al.
6
present a technique to produce a 3D segmentation of the head employing T1 and T2 images. This method requires an initial visual classification of a relatively small sample of the tissue types in the T1 and T2 images by an experienced radiologist to begin. The accuracy of this initial input directly affects the accuracy of the succeeding classification algorithm, which classifies the remaining tissue based on a bivariate distribution model for each tissue type. The possible classification outputs are; background, brain, CSF, WM, lesion, tumor, arteries or veins. Once the initial classification is made, a feature map is constructed employing clustering based on the bivariate normal probability distribution. With this feature map, segmentation is performed by replacing the voxel with the above tissue label assigned by the probabilistic calculation. Upon the completion of the segmentation, the resulting surfaces are filtered twice using a 3D diffusion filter to smooth discontinuities created by misclassifications and to improve rendering. A 3D connectivity algorithm then extracts surfaces from this smoothed, segmented data set. Finally, a dividing cubes algorithm is used to process the voxels marked in the connectivity algorithm for display. This method is rated with only preliminary results for three selected patients; one normal, one MS case and one tumor. It is important to reiterate that success of the classification depends upon the operator's level of expertise. This method may be most useful as a surgical planning tool, or perhaps as a visualization tool, which appears to be the original intent.
Hohol et al.
7
have adopted a similar technique. This method begins by manually isolating the intracranial cavity using the PD and T2 images. Regions of interest (ROI) are then generated for brain parenchyma and CSF to aid the Expectation Maximization (EM) tissue classifier. Each image for a patient is registered to a time reference image, and the EM classifier is used to classify each pixel of each image as either WM, GM, CSF, or lesion tissue. This classification includes an unspecified correction for partial voluming artifacts. Individual lesions are fixed by the use of a 4 dimensional connectivity algorithm that is assumed to similar to Cline
6
with the addition of a time dimension. With use of the time dimension element, lesion volumes are reported to change over time. No verification of the imaging technique is reported. However, the authors do report lesion burden correlation with the neuropsychological test scores of MS patients.
Wicks et al.
8
employ two manually selected thresholds to perform white matter lesion (WML) detection. The first is set to separate brain tissue from the skull, while the second is set to identify all areas definable as lesions having intensities greater than the brain tissue. In order to preserve accurate boundaries, the threshold is set somewhat lower in value than that is required to identify the lesions. The outlines of the identified lesions are superimposed on the original T2 image to allow manual correction of misclassification. It is reported that this requirement poses no additional burden since it prevents the operator from having to outline each lesion manually. It is also stated that the established threshold can be used for any successive serial study of one patient provided that the intensity histograms from later studies are scaled to match that of the initial study. Overall, this semi-automated thresholding technique generates approximately half the inter-observer variation than purely manual outlining technique, but it is stated that visual assessment of the delineation of the lesions between the manual technique and this semi-automatic one are “equally plausible”. The semi-automatic double threshold method is probably suitable for use with small patient sets in a non-critical research environment.
Wang et al.
9
report a similar global thresholding technique that employs intensity correction of interslice scans. In practice a reference histogram is generated from some arbitrary scan slice. From the next scan slice, another histogram is generated. The (low) background intensities are cut from the histograms using al sharp cut-off windowing function with a Hanning slope. A manually chosen parameter defines the number of bins required for this cut-off line to reach its maximum value of 1.0. The second histogram is matched to the first by minimizing the squared differences between them. This process permits a more satisfactory selection of global thresholds for a patient, as all subsequent scan slices are matched in sequence. Although the technique undoubtedly subdues the intensity variations for WM lesions across scans, the accuracy of the resulting lesion volumes cannot be judged as no independent measures are reported.
Zijenbos et al.
10
describe a semi-automatic method based on pattern recognition and built around a back-propagation artificial neural network (ANN). This is done to minimize the required input training points needed to achieve a successful tissue classification. The method begins with the use of an inter-cranial contour algorithm to remove the skull and incidental CSF, followed by an intensity correction algorithm to remove the shading, or intensity inhomogeneity artifact. This correction requires the operator to select 10 to 20 points in the WM that therefore strongly influences the detected lesion load. The intensity correction is followed by the same diffusion filter mentioned in Cline
6
to enhance the signal
oise ratio. Finally, the tissue classification occurs using the ANN with 3 input nodes (T1,T2, PD) and five output nodes; background, WM, GM, CSF, and WML. One 'sample of each class input tissue class is presented in sequence to the ANN until convergence (segmentation) occurs. Upon the completion of classification by the ANN, postprocessing is required to remove the WML that occur in close proximity to GM along the sulci and correct the classification errors caused by misregistration of the T1 image. In addition, all WML smaller than 10 pixels are eliminated as they are assumed to be the result of noise or misregistration in the T1 image. This semi-automated technique is compared with a completely manual method using two different observers. Their published results relating inter-rater and intra-rater variation using the kappa statistic indicate that the two techniques are well correlated, neither showing an obvious advantage over the other with respect to the observers used in the study. Although the technique is technically sophisticated, it is limited by its requirement of expert knowledge on the part of the operator. An experienced diagnostici
James Ialaya G
Katten Muchin Zavis & Rosenman
Lateef Marvin M.
University of South Florida
LandOfFree
Method and system for knowledge guided hyperintensity... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and system for knowledge guided hyperintensity..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and system for knowledge guided hyperintensity... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2906975