Image analysis – Applications – Biomedical applications
Reexamination Certificate
1998-04-29
2001-10-30
Johns, Andrew W. (Department: 2621)
Image analysis
Applications
Biomedical applications
Reexamination Certificate
active
06310967
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to systems and methods for analyzing medical images, and, more particularly, to systems and methods for analyzing digital mammograms.
2. Description of Related Art
Many computer-aided diagnosis (CAD) schemes have been devised for mammographic image analysis [1-27]. A general review of digital radiography has been given by Doi et al. [1]. Many of these methods are based on multiresolution techniques.
Work related to the use of various multiresolution methods for investigating mammograms includes Refs. 3, 11, 12, 19, 23, and 26. Dengler et al. [11] use a difference of two Gaussians for the detection filter, and the final detection is based on a global threshold. Valatx et al. [12] generate a smooth approximation of the image with a &bgr;-spline expansion and apply a mixed distribution based local thresholding technique to both the raw and approximated image; the output image is formed by subtracting the two thresholded images. A calcification segmentation method is developed by Qian et al. [3] using two-channel and multichannel wavelet transforms [19], based on subband selection and a rescaling (thresholding) technique for feature detection [24]. Strickland and Hann [23] apply the wavelet transform at full resolution (no downsampling) and detect independently in two sets (HH and LH+HL) of three full resolution subband images. The detection results are combined, further processed, and the inverse wavelet transform is implemented. De Vore et al. [26] implement the standard wavelet transform, select the important subbands, and invert the transform after wavelet coefficient suppression. The resulting image is empirically thresholded in order to remove the remaining background information.
Various statistical approaches have been used to study mammograms [12-14, 18,21,23,27]. Wavelet domain coefficient probability modeling has also been utilized in other areas of research: selecting optimized coding methods [28, 29], Gauss-Markov field representation [30-32], and texture identification [32].
It is known that film grain noise in mammograms is signal dependent [33, 34]. Typically, the accepted noise field for radiographs results from three independent components: (1) spatial fluctuations in the number of x-ray quanta absorbed in the screen; (2) spatial fluctuations in the screen absorption associated with random structural inhomogeneities in the phosphor coating; and (3) spatial fluctuations in film sensitivity due to the silver halide random distribution per unit area in the emulsion [35]. Many CAD methods have found it essential to carefully treat the image noise with a preprocessing step [3, 15, 22, 27, 36].
SUMMARY OF THE INVENTION
It is therefore an object of the present invention to provide a system and method for identifying normal and abnormal tissue in medical images such as mammograms.
It is an additional object to provide such a system and method that permit significant time savings in reading clinically normal mammograms.
It is a further object to provide such a system and method for providing a second opinion strategy.
It is another object to provide such a system and method having sufficient performance to detect a predetermined portion of the normal images with a low probability of false negatives.
It is yet an additional object to provide such a system and method for detecting calcifications.
These objects and others are attained by the present invention, a system and method for identifying normal tissue in medical images. Here the term normal is intended to define an image that does not contain a suspicious area, an image aberration, or small image medium defects. As radiologists spends an enormous amount of time investigating images lacking abnormalities, the invention can save a great deal of valuable time. This system and method may also be considered a second-opinion strategy, since the image can be declared “normal” by the detection system and method, and then reviewed by a radiologist, and thus the image has been analyzed twice.
In a preferred embodiment, the invention addresses the detection of microcalcifications in mammograms, with a performance of detecting 40-50% of the normal images with a low probability of false negatives.
The invention comprises the use of a multiresolution statistical model for normal tissue. This model is then used to make comparisons with local image regions. If a small region deviates significantly from the global model, it is flagged as potentially suspicious; if a region is in agreement, it can be discarded. The systematic identification of abnormal regions can be regarded as a detection algorithm that can be tested and evaluated using a standard database. If no suspicious regions are located, an image lacking any pathology can be identified by the detection process.
A fundamental distinction exists between the techniques of the present invention and the prior art. Herein a multiresolution approach is used as a simplifying device for statistical modeling, in order to show that a multiresolution statistical analysis has the potential for simplifying what has previously been considered an intractable statistical problem. Specifically, the statistical interpretation of a raw image is very difficult, but is reasonably simple when applied separately to various resolutions of the image, after the decomposition into independent subspace images.
Specifically, the method of the present invention comprises the step of applying a wavelet expansion to a raw image. The raw image, which is typically in electronic form, comprises an array of sectors (e.g., pixels), wherein each sector has an intensity level. The wavelet expansion is for obtaining a plurality of subspace images of varying resolution.
The next step comprises selecting at least one subspace image that has a resolution commensurate with a desired predetermined detection resolution range. For example, if it is desired to examine for the presence of a neoplasm or calcification having dimensions in a particular size range, one or more subspace images are selected that encompass that size range.
Next is determined a functional form of a probability distribution function (pdf) for each selected subspace image and an optimal statistical normal image region test for each selected subspace image. There is a test statistic associated with the normal image region test that has some pdf. From this pdf a threshold level is established for the probability distribution function from the optimal statistical normal image region test for each selected subspace image. Preferably this step is accomplished with the use of a “test statistic, ” which will be described in the following.
Finally, an output image is created, such as in electronic and/or visualizable form. A region is defined as comprising at least one sector, typically a plurality of sectors, against which the threshold level is compared. The output image has a first value (e.g., “1”) for each region when the region intensity level is above the threshold and a second value (e.g., “0”) when the region intensity level is below the threshold. This image then permits the localization of a potential abnormality within the image.
Preferably, this method is taken a step further, although this is not intended as a limitation. The further step comprises determining for the presence of a plurality of above-threshold regions within a predetermined larger area. Such a plurality of above-threshold regions can be indicative of a likelihood of abnormality.
There are two distinctions between the present invention and previously reported statistical approaches, including the noise processing approach. First, the statistical analysis is applied to independent subspace images. The accepted noise components and the signal are lumped together; the aggregate is considered as a random field. Second, the focus is on the
Clarke Laurence P.
Cullers David Kent
Deans Stanley R.
Heine John J.
Stauduhar Richard Paul
Allen Dyer Doppelt Milbrath & Gilchrist, P.A.
Johns Andrew W.
Nakhjavan Shervin
University of South Florida
LandOfFree
Normal and abnormal tissue identification system and method... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Normal and abnormal tissue identification system and method..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Normal and abnormal tissue identification system and method... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2615047