Image analysis – Applications – Biomedical applications
Reexamination Certificate
1998-08-18
2001-06-12
Johns, Andrew W. (Department: 2621)
Image analysis
Applications
Biomedical applications
Reexamination Certificate
active
06246784
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to feature extraction and identification in medical imagery. More particularly, the invention relates to methods for generating efficient models of complex anatomical structures in the presence of “leakage,” and methods for employing such models as an aid to diagnosis.
BACKGROUND AND SUMMARY OF THE INVENTION
The science of medical image processing has taken tremendous strides in the past two decades, particularly in the field of three-dimensional visualization of internal anatomical structures. Such three dimensional models can be virtually rotated and viewed from any perspective, providing invaluable insights to surgeons, diagnosticians, researchers, and other scientists.
In its raw form, medical imagery typically consists of a large array of numbers representing the value of a physical property (e.g. radiological “density” or “intensity”) at each of a plurality of regularly-spaced locations within the patient. The methods for generating structural models from this data proceed by generally well known principles.
One familiar class of techniques is known as “volume growing” (sometimes termed “region growing”). In accordance with these techniques, a seed voxel (volume element) is first identified within the anatomical structure of interest. Other voxels are successively analyzed and identified as belonging to the same structure if (1) they adjoin a voxel already identified as belonging to the structure, and (2) they meet a specified physical attribute (typically a radiological density in a range characteristic of the structure of interest).
According to standard region growing methods, after the seed voxel is identified, the six voxels sharing a face with the seed voxel are analyzed to determine if their physical attribute is within the specified range. If so, such voxels are marked as belonging to the structure. These voxels form a first tier of volume growth.
Next, each voxel in the first tier of volume growth is processed like a seed voxel, with adjoining voxels analyzed to determine whether their physical attributes are within the specified range. Voxels so identified form a second tier of volume growth.
This process continues, each iteration adding a shell of further voxels within the structure of interest.
In the simple case, this march of voxel cubes proceeds until growth in each direction is stopped by an exhaustion of voxels meeting the specified physical criterion. Collectively, the set of voxels thus identified fills the volume of the anatomical structure being analyzed, permitting its three-dimensional modeling.
The below-cited General Electric patents more fully detail the foregoing volume growing techniques and improvements thereto, including techniques for particularly locating the structure's bounding surface with reference to the vertices of the outermost voxels, techniques for smoothing/shading the bounding surface to facilitate viewing, etc.
A problem with foregoing technique, and most other volume growing algorithms, is that of “leakage.” Leakage occurs when the march of cubes proceeds through the boundary of the structure being analyzed, rather than stopping as intended. Leakage causes the region growing to continue on the other side of the boundary, with a large number of voxels on the other side of the boundary spuriously identified as belonging to the structure of interest.
Leakage can occur for many reasons, including voxel dimensions larger than the boundary thickness, noise-induced imperfections in the bounding surface image data, etc.
As a general matter, leakage does not seriously impair the clinical usefulness of the extracted model. The leakage is an aesthetic distraction, but a reviewing physician can usually readily identify the leakage as a computer processing glitch. A more serious problem is the additional processing burden that leakage imposes on the three dimensional modeling software.
The computational complexity of three dimensional modeling is substantial. Such models typically include hundreds of thousands of data points, each of which must be processed every time the displayed model undergoes any change.
Much of the value of three dimensional modeling comes from the physician's ability to rotate the model, move his point of perspective, and zoom into and out from features of interest—all in real time. Each such operation requires that the display screen be “repainted” several times in quick succession to avoid the impression of jerky movement. Each such screen redraw, in turn, requires an enormous number of computations. The problem with leakage is that it vastly swells the dataset that must be processed, slowing the modeling software response time, and interfering with the physician's sense of real time interactivity.
The leakage problem in region/volume growing algorithms has been recognized for decades, and has been dealt with in various ways.
One way is simply to adopt a feature extraction technique relatively immune to leakage. One such class of techniques relies on deformable models. In “A Novel Volumetric Feature Extraction Technique, With Applications to MR Images,” Int'l. Conf. on Image Proc., pp. 564-67, IEEE (1995), for example, Ashton et al model an expanding bubble whose expansion continues until the bubble fills the structure of interest. The shape of the bubble is controlled by a constraining force imposed by surrounding tissue and by a penalty for deviation from the expected surface normal. More particularly, Ashton et al. expand a seed voxel outwardly in an ovoid shape until the expected volume is reached, or until no further expansion is possible due to constraining tissue. (The ovoid shape is tailored in accordance with a priori information about the expected shape and size of the structure of interest.)
Leakage is rarely an issue in such deformable model techniques because, like a balloon, the expanding outer surface will not generally tunnel through a small opening and spawn a large ballooned volume on the other side.
Feature extraction techniques offering immunity to leakage are rare, and suffer from various drawbacks that have prevented their widespread adoption. Accordingly, various other solutions to the leakage problem have been proposed.
One solution has keyed on the characteristic shape of leakage volumes (i.e. a growth volume linked to a more central volume by a single (or a few) voxels). Computerized feature recognition techniques can be applied to identify such characteristic shapes and automatically delete them from the dataset. However, such approaches are generally disfavored in medicine due to the possible inadvertent deletion of clinically significant features.
In cases where the boundary of concern is thin, leakage can sometimes be ameliorated by employing commensurately small voxels (or subvoxels). With this approach, a boundary won't be missed by a voxel simply spanning the physical boundary. However, halving the edge size of the voxel effects an eight-fold increase in the number of voxels to be processed, with a corresponding slow-down in manipulation of the resulting three-dimensional model. Moreover, thin boundaries are sometimes simply not manifested in the image data being analyzed, due to inherent resolution limitations of the data acquisition device (e.g. CT or MRI scanner). In such cases, small voxels offer no solution.
U.S. application Ser. No. 07/797,893, cited in U.S. Pat. No. 5,553,207 to Sekiguchi et al, proposes an interactive solution to the leakage problem in which an operator monitors progress of the volume growth on a plurality of display devices and interrupts the process if a leak extends to a volume outside the structure of interest. (Several display devices are required due to growth in three dimensions.) Sekiguchi'207 patent extends this technique by facilitating deletion of the spurious growth by reverse expansion from an operator-identified voxel within the leaked volume.
A drawback of Sekiguchi's technique is its requirement of human interaction and real-time vigilance, increasing th
Malley James D.
Pusanik Lynne M.
Selbie Scott
Summers Ronald M.
Johns Andrew W.
Klarquist Sparkman Campbell & Leigh & Whinston, LLP
Nakhjavan Shervin K.
The United States of America as represented by the Department of
LandOfFree
Method for segmenting medical images and detecting surface... 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 for segmenting medical images and detecting surface..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for segmenting medical images and detecting surface... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2510146