KR20010023427A - Method and system for automated detection of clustered microcalcifications from digital mammograms - Google Patents

Abstract

One digital mammogram 100 is first automatically cropped into a breast region sub image (200) and then this sub image is processed by an optimized Gaussian fill to emphasize the appearance of latent microcalcifications in the sub image, Methods and systems for detecting and displaying clustered microcalcifications in digital mammograms. Potential microcalcification is thresholded and clusters are detected (300), features are computed for the detected clusters, and the clusters are classified as suspected or unsuspected by the neural network (400). The location within the original digital mammogram of the suspected detected clustered microcalcifications is indicated. The results of the system are optimally coupled to the radiologist's observations on the original mammogram by combining the observations with the results after the radiologist first accepts or rejects the individual detections reported by the system.

Description

Method and system for automated detection of clustered microcalcifications from digital mammograms

Mammography, along with physical examinations, is currently the process of choice for breast cancer screening. Screening mammography reduced breast cancer mortality rates by approximately 30 to 35 percent. However, in 1996 approximately 185,700 new breast cancer patients were diagnosed and 44,000 women died of the disease. One in eight women is likely to be diagnosed with breast cancer and one in 30 will die from the disease.

Although mammography is a well researched and standardized method, in 20-30 of women diagnosed with breast cancer, their mammograms were judged negative. Moreover, only 10 to 20 percent of patients referred for biopsy are found to have cancer based on findings on mammograms. In the past, it is also estimated that the malignancy missed by the radiologist is evident in two-thirds of the mammograms. Detection may be missed due to several factors, including poor picture quality, incorrect patient positioning, inaccurate judgment, obscuring fibroglandular tissue, the indistinguishable nature of radiographic findings, eye fatigue, or overlook have.

In order to increase the sensitivity, double reading has been proposed. However, increasing the number of screening mammograms makes this option unlikely. Alternatively, a computer aided diagnostic (CAD or CADx) system can act as a "second reader" to assist the radiologist in detecting and diagnosing a region. Some researchers have attempted to interpret the abnormalities of mammograms using digital computers. However, known studies suggest that the true-positive detection rate for error-positive detection is undesirably low.

Microcalcification is an ideal target for automatic detection because microcalcification, which is often difficult to distinguish, is often the first and only radiographic finding in early, treatable breast cancers, but individual microcalcifications of suspect clusters are quite sensitive to radiographs. Appear in a limited range. Thirty to fifty percent of carcinomas detected by radiographs show microcalcifications on the mammograms and 60 to 80 percent of breast carcinomas show microcalcifications upon microscopic examination. Increasing the detection rate of microcalcification by mammography will lead to improved photographic efficacy in detecting early breast cancer.

Although the prospect of CAD systems is to enhance the physician's ability to diagnose cancer, the problem is that not all CAD systems can detect some of those areas that may be found at human judgment. However, a person's judgment misses those areas that later manifest as signs of cancer. Missing a cancer-related site is called a false negative error, and relating a normal site to a cancer is called a false positive error.

It is not yet clear how the radiologist will combine CAD system output with mammography. No existing CAD system has found all suspected sites detected by the average radiologist and they tend to have unacceptably high false positive error rates. However, CAD systems can find some suspicious areas that radiologists may miss.

The present invention relates to a method and system for automatically detecting clustered microcalcification from a digital image without reducing the radiosensitivity.

1 is a flow diagram illustrating an automated system for detection of clustered microcalcifications in digital mammograms.

2 and 3 are flowcharts illustrating the automatic cropping method and system of the present invention.

4-10 are flowcharts illustrating in more detail the automatic cropping method and system of the present invention.

11 is a flow chart illustrating in more detail the clustered microcalcification detector of the present invention.

12 is a schematic diagram illustrating a 3 x 3 cross-shaped median filter of the present invention.

13 is a three dimensional view of a Gaussian difference (DoG) filter kernel.

14 is a cross-sectional view through the center of the DoG filter kernel of FIG. 13.

15 is a flow diagram illustrating the global thresholding portion of a microcalcification detection system.

Figure 16 is a flow diagram illustrating dual region thresholding of the present invention.

17 is a flow chart illustrating the combination of results of global and dual-region imposition.

18 is a flow chart illustrating a sloped area thresholding of the present invention.

19 is a flowchart illustrating a clustering method of the present invention.

20 is a schematic diagram illustrating a clustering method of the present invention.

Fig. 21 is a flowchart showing a feature calculation process of the present invention.

22 is a flowchart illustrating a classifier having one discrimination function per class.

23 is a schematic diagram illustrating a multilayer perceptron neural network for a class 2 classifier.

24 is a histogram testing the results after detection and classification.

25 is a flowchart illustrating a parameter optimization method of the present invention.

Figure 26 is a curve diagram of a free response receiver operating characteristic of the present invention prior to classifying detection.

27 is a curve diagram of a free response receiver operating characteristic of the present invention after classifying detection.

28 is a probability density function diagram showing a relationship between an error negative detection probability and an error positive detection probability.

29 is a probability density function diagram showing the relationship between true negative detection probability and true positive detection probability.

30 is a Venn diagram showing the relationship between the detection of the radiologist and the detection of the CAD system.

FIG. 31 is a flow diagram illustrating a method for combining computer aided diagnostic detection with human interpreter detection for optimal sensitivity.

32 is a flow diagram illustrating an alternative embodiment of the present invention including a density detector.

<Summary of invention>

It is therefore an object of the present invention to provide a method and system for automatically detecting clustered microcalcifications from digital mammograms.

These and other aims are to obtain digital mammograms, optimize the parameters needed to crop (crop) digital mammogram images, and cut out the digital mammograms based on the optimized cropping parameters to obtain breast tissue for subsequent analysis. Cluster-phase microcalcification from digital mammography, selecting and optimizing the necessary parameters for cluster-phase microcalcification and detecting cluster-phase microcalcification in the cropped digital breast radiograph based on optimized cluster-phase microcalcification detection parameters. It is achieved according to the invention by providing a novel method and system for the automatic detection of.

The detected clustered microcalcifications are then stored as a detection image, which is provided for display and a computer aided detection image is generated for viewing by the radiologist.

The radiologist first reviews the original mammogram and reports the set S1 of suspected areas involved. The CAD system, in particular the CAD system of the present invention, operates on the original mammogram to report a first set of suspected detections or sites involved. The radiologist examines the set S2 and accepts or rejects the members of the S2 as suspect, thereby forming a suspect third set which is a subset of the S2 set. The radiologist then generates a suspected fourth detection set S4 which is the union of sets S1 and S2 for later series of overhaul diagnosis. As such, the CAD system output is coupled to the radiologist's mammogram to optimize the overall sensitivity of the true true site detection.

Other objects and advantages of the invention will be apparent from the following description, the accompanying drawings and the claims.

In the drawings, which refer to the same or corresponding parts, in particular FIG. 1 shows a flow chart illustrating a series of steps performed to detect the location of microcalcification clusters in a digital mammogram.

In a first step 100, a digital mammogram is obtained using hardware such as a digital mammography system or by digitizing a mammogram using a laser or charge-coupled device (CCD) digitizer. In an optimized cropping step 200, a rectangular analysis site comprising breast tissue is partitioned from the digital mammogram image for use in subsequent processing steps to reduce the time required to process the mammogram image. Create a binary mask corresponding to the breast tissue. Binary masks are also used to limit detection for image areas containing breast tissue.

Clustered microcalcification is detected in clustered microcalcification detection step 300. The filtered image is first filtered with a median filter to reduce noise, and then the image is filtered using a Gaussian difference (DoG) filter optimized to enhance microcalcification. DoG-filtered images are then subjected to threshold tests optimized to detect latent microcalcifications. The detected microcalcifications are reduced to a single pixel representation and detection outside the breast region is eliminated. The remaining microcalcifications are grouped into clusters. The feature for the clusters is then calculated. The detected clusters are classified as suspicious or unsuspecting in the classification step 400.

The parameters used by the automatic cropping step, the cluster phase microcalcification detection step, and the classification step 200, 300, 400 are optimized in the parameter optimization step 500. The parameters are optimized by parameter optimization means using genetic algorithm (GA) to maximize true-positive detection rate while minimizing false positive detection rate. Of course, other optimization methods can also be used.

The detected microcalcifications on the cluster are stored in the image coordinate list. The detection result is processed in processing step 600 by simply adding an offset to each of the microcalcified coordinates to handle the coordinate transformation resulting from the cropping process. The detected clustered microcalcifications are displayed on the digital mammogram as a rectangle drawn around the clustered microcalcifications in display step 700. For example, other marks may be used, such as an arrow pointing to a suspected microcalcification, or an oval around the suspected microcalcification.

Acquisition of digital representation of mammography

One method of obtaining digital mammograms includes digitizing a radiographic film with a laser or a charge-coupled device (CCD) scanner. The digital image obtained in this way has a sample interval of about 100 μm per pixel and a gray level resolution of 10 to 12 bits per pixel. In one embodiment of the present invention, the radiographic film was scanned using a Model CX812T digitizer manufactured by Radiographic Digital Imaging Compton, California, to have a digital image with a 50-micron spacing per pixel and a gray-level resolution of 12 bits per pixel. Create Another possible source for digital images is a digital breast radiography device from Trex Medical of Danbury, Connecticut, which has a spatial resolution of about 45 μm per pixel and a gray-level resolution of 14 bits per pixel.

The digital image is stored as a digital representation of the original mammogram image on a computer readable storage medium. In a preferred embodiment, the digital representation or image is a dual Pentium II® microprocessor operating at 200 MHz, 512 MB of RAM memory, a ViewSonic PT813® monitor, pointing device, and a Lexmark Oprah S1625® printer. It is stored on a 2 GB hard drive of a general purpose computer such as a PC. The system runs on the Windows NT® operating system.

Automatic cropping

As shown in FIGS. 2 and 3, the digital mammography image 190 is first defined to partition the analysis region 296 from the digital mammogram image to generate a binary mask 298 corresponding to the breast tissue in the analysis region. ) Cut out (cropping). Preferably, cropping may be cut manually but is performed automatically. Because breast tissue does not cover the entire radiographic film, the image is cropped as a preliminary step. By focusing the image processing on only that portion of the image of the breast tissue, the time required for image processing is reduced. In addition, other items appearing on the film, such as labels and patient information, are excluded from consideration and false positive indications outside the breast tissue area are removed.

4 to 10, the automatic cropping process will be described in detail. First, the image is subsampled from 50 μm to 400 μm to reduce the amount of data to be processed in step 202. Of course, the image may be down sampled to another desired resolution. Not all of the original image data is needed to reliably partition breast tissue from the rest of the image. By sampling every eighth pixel in the horizontal and vertical directions, the data amount is reduced by 64 times. The resulting resolution loss is not important for the purpose of partitioning breast tissue from the rest of the image.

In step 204 a white border, 20 pixels wide, is added around all sides of the subsampled image. White corresponds to a given maximum pixel value, where possible, as the number of bits used to represent each pixel. For an image with grey-level resolution of 12 bits, the maximum gray-scale value is 4095. The image bounded at step 206 is thresholded to a relatively high threshold so that most breast tissue is below the threshold to produce a binary image. In one embodiment of the invention, the threshold is set equal to a predetermined percentage of the gray-scale value of the pixel near the top middle portion of the image. The thresholded image is then inverted. That is, in step 208 ones become zeros and zeros become ones. The inverted image is expanded in step 210. Expansion is a morphological operation in which each pixel in the binary image is on, i.e., set to a value of 1 if any of the neighboring pixels is on. If the pixel is already on, leave it on.

In step 212, the expanded image is cropped to the largest blob size. The blob is a contiguous group of pixels with a value of one. This step 212 removes bright borders in the subsampled breast radiographic representation while none of the breast regions are reduced. Other techniques of thresholding to find a border are sometimes very difficult to process bright areas within the breast adjacent to the border, such as when a breast implant is visible in an image. Pixels from the original image from step 202, corresponding to pixel positions in the cropped blob, are selected for subsequent processing. Here it is a simple subset of the pixels from the input image.

The image from step 212 is the histogram equalized in step 214. The average brightness of the image will vary widely from breast radiograph. Moreover, other digitizers with different optical density characteristics are another cause of the change in brightness level in the digital representation of the mammogram. Breast masks, which are the output of an automatic cropper, are mainly formed by region-growing algorithms that require a single contrast setting to work properly. However, it has been experimentally determined that one contrast setting will not work for a wide range of image inputs. Therefore, each image is mapped to normalized image space using automatic histogram enhancement processing, after which a single contrast setting works well.

First, get a histogram of the image. Typically, most of the data in the breast region will be in low histogram bins (corresponding to grayscale values of about 0-1000) and borders and labels will be in high bins (gray-scale of about 4000-4095 for 12-bit data). Corresponding to a value). The upper and lower limit values for obtaining typical breast data are determined. The lower bin value is the first largest peak encountered when going from the lowest gray-scale value to the largest gray-scale value. The upper bound bin is the last zero value bin encountered when going from the largest gray-scale level to the lowest gray-scale value. The data is then reduced to an 8-bit representation and stretched linearly over a range of data types. For example, the value in the lower limit bin is set to zero. The data value in the upper bound bin is set to 255. The remaining data is linearly mapped between the lower and upper bins.

After equalizing the histogram of the image, the equalized image can be considered a matrix. The image matrix is divided into left and right halves of the same size if possible, and the brighter side is selected in step 216. The sum of all pixels in the left half and right half is calculated. The side with the larger sum by comparing the sum values is the brighter side.

Prior to area-expanding the bright side, the algorithm variable is initialized in step 218. The size of the area extension mask is checked in advance at step 220. If large enough, a mask can be used. If not, processing continues to find the mask. The area of the image-the side to be expanded is selected in step 222. In step 224 this area is searched to find its maximum gray-scale value. This maximum value is used to find the pixel to start the area-expansion algorithm. Region-expansion is the process of grouping connected pixels that share some of its features. The choice of features affects the resulting area. Input to the area-extension function is the starting point for starting the gray-scale image and expansion. The output is a binary image with an expanded area, ie those representing pixels in the blob. The area-extension will create a single blob, but the blob will have an inner hole in it, i.e., an off state pixel. To extend the blob, consider each of the four nearest neighbors of that pixel. The contrast ratio is calculated for each nearest neighboring pixel. If the contrast ratio is below the contrast ratio threshold, the neighboring pixels are set to 1 in the binary mask image. Otherwise, neighboring pixels are set to zero. The region expansion algorithm sequentially considers the nearest neighboring pixels until it completes spiraling outward from the starting or seed pixel. For those skilled in the art, it is clear that other area-extension algorithms can also be applied.

In step 226, the region expansion starts with the pixel identified in the previous step 224 to generate a binary mask. The size of the mask from step 226 is calculated in step 228 and checked in step 230. There can be three points to this approach. First, the brightest point in the search region may be an artifact outside the breast. Therefore, if the resulting mask is not large enough (50 pixels), the search area is moved closer to the side of the image and searched again. This is repeated three times, each time lowering the contrast value threshold. This corresponds to the path taken through steps 232 and 234. Second, side selection can be error prone. Therefore, if no valid breast mask is found on the searched first side, the other side of the breast is searched. This corresponds to the path taken through steps 236 and 238. Third, if no effective breast mask is found on either side, the entire breast is thresholded to take the largest object as the breast mask in step 240.

Some masks may be too large because the area-expansion algorithm uses constant contrast. Typically, there will be a "tail" along the edge of the digitized mammogram image when extra light leaks in while the original mammogram film is digitizing. This tail can be reduced by applying a series of erodes and then applying a series of expansions to the image. Erosion is a morphology operation in which each pixel in a binary image is turned off unless all of its neighbors are on. If the pixel is already off, leave it off. However, firstly the holes in the mask must be filled or a plurality of erosions can divide the mask into disassembled parts. Thus, the holes in the mask are filled by a majority operation in step 242. The majority operation is a shape processing operation in which each pixel in the binary image is turned on if the majority of adjacent pixels are in the on state. If the pixel is already on, leave it on.

However, another problem is that some small breast masks cannot receive as much erosion as a large breast mask can. Therefore, the safety mask measures take the sum of the breast mask before and after erosion and inflation. If the size is reduced too much (i.e. more than 50%) then the original mask before the shape processing operation is used. Thus, the mask is duplicated in step 244 before eroding the mask in steps 246 and 248, respectively. The size of the resulting mask is calculated in step 250 and compared with the size of the mask from step 242 in step 252. If the new size is less than half the previous size, then the duplicated mask from step 244 is selected in step 254 for the next processing. Otherwise, the resulting mask from step 248 is used.

The original mask (from step 202) is cut out (from step 242 or step 248) to the size of the breast mask found in step 256. If the resulting mask is too small for subsequent processing, always adjust the cut in step 258. The adjustment is in the form of increasing the size of the breast mask bounding box by including additional pixels from the original image in the cropped image.

The cropped image is automatically histogram highlighted in step 260 as described above with respect to step 214. This highlighted image goes through a loose region expansion step 262 to create a generous mask. This means lower thresholding the image resulting in more "on" pixels. This mask is hole-filled, eroded, expanded in steps 264, 266, and 268, respectively, as described above, but to a lesser extent.

The same step described above is repeated one final time in steps 270 to 276. However, the cropping adjustment is less and the contrast value is increased in the dense region expansion step 276. This compact region expansion step 276 may provide a higher contrast value because the region will be extended only to the cropped image. This is a stingy estimate of breast tissue. In order to find the largest object in step 278, the resulting mask is split and in step 280 its bounding box is reduced to surround only the object. There may still be some holes in the breast mask. Therefore, after the cropping adjustment in step 282, the mask is reversed in step 284 to find the largest object in step 286. This largest object is extracted and inverted in step 288 to obtain a penultimate mask.

In step 290 the final mask is obtained by filling the holes in the secondary mask by a plurality of majority operations and expansions. The image is then cut out to the size of the resulting mask, and automatic cropping is complete. An important result from the automatic cropper is the offset of the cropped image. This is the pixel position in the original image corresponding to the pixel in the pixel on the upper left side of the cropped image. This offset value is determined by keeping track of all cropping and cropping adjustments.

The output of the automatic cropping process is a rectangular pixel array representing a binary mask where a value of 1 is assigned to pixels corresponding to the breast tissue and a zero value is assigned to the remaining pixels. Alternatively, the binary mask is a silhouette of the breast consisting of ones with a background of zero.

A better breast mask can be obtained by optimizing the parameters of the automatic cropper. This process is described in the next optimization section.

Detecting Clustered Microcalcifications

11 is a flow chart illustrating in more detail the clustered microcalcification detection system 300 of the present invention.

The portion of the digital representation of the mammogram that corresponds to the analysis region 296 referred to as the cropped sub-image 302 generated in the cropping step 200 is first processed to reduce noise in the noise reduction step 310. Microcalcification to reduce digital noise contributing to false detection. The noise reduced image is then filtered using an optimized target size dependent Gaussian difference (GoD) spatial kernel in step 320 to emphasize the difference between the target and the background, thereby increasing the global maximum and local maximum within the filtered image. Create a value. The optimized DoG-filtered image is thresholded at step 340 to define a maximum value indicating that microcalcification is detectable.

The conversion step 350 converts the detected maximum value into a single pixel coordinate representation. In the first error positive elimination step 360, error detection outside the breast mask region is removed by comparing the detected coordinates of the maximum value with the binary mask of the analysis region. The remaining coordinate representations in the analysis region are clustered in cluster step 370. Feature calculation step 380 calculates features for the remaining clusters and these features are used to eliminate unsuspected detection in classification step 400 (FIG. 1). The remaining detection is output in the form of cluster coordinates as clustered microcalcifications detected in the output step 600.

To discuss the steps in the clustered microcalcification detection process in more detail, first filter this image to reduce noise in the digital mammogram image. Although the main limitation of image quality is the particle size of the film emulsion, noise is introduced in the digitization process. This noise can later be detected as pseudo-calcification. In this system, the cruciform median filter is used because it is known to be extremely effective at removing single pixel noise. The median filter is a nonlinear spatial filter that replaces each pixel value with the median of the pixel values in the kernel of the selected size and shape centered on that pixel. In FIG. 12, a cross is formed with a pixel set including a pixel and its four nearest neighbors in the center. The criss-cross maintains better lines and edges than typical block-shaped median filters and limits possible substitutions to four nearest neighbors, thus reducing the possibility of edge displacement.

After reducing the noise, the image is filtered with an optimized DoG kernel to enhance microcalcification. Filtering is accomplished by convoluting the noise-reduced image with the DoG kernel. In an alternative embodiment, filtering is achieved by first obtaining fast Fourier transforms (FFTs) for the noise reduced image and the DoG kernel and multiplying these FFTs and then taking the resulting inverse FFT.

Neurophysiology experiments were chosen because the DoG kernel provided evidence that the human visual pathway contained a set of "channels" that are spatial frequency selective. Essentially, there is a mask that analyzes an image or a scaled filter at each point in the field of view. The behavior of these spatial receptive fields can be closely approximated by DoG.

The 2-D Gaussian mask is given by

(One)

Where c normalizes the sum of mask elements to unity, x and y are horizontal and vertical indices, and σ is the standard deviation. Using equation (1), two Gaussian differences with deviation σ provide:

(2)

When σ ₂ = 1.6σ ₁ , the response of the DoG filter is shown to be in close agreement with the response of the human spatial acceptance filter. Therefore, derived from human physiology, if the ratio of DoG standard deviation constants is 1: 1.6, then use σ ₂ = t / 2 for a target of size (average width) t pixels and approximately σ ₁ = σ ₂ /1.6 Becomes

Since microcalcifications are typically 100 to 300 μm in diameter, possible target sizes for 50 μm digitized mammograms correspond to 2 to 6 pixels. The DoG kernel constructed using optimization techniques for selecting target size parameters, such as GA in detail below, was found to have an optimized target size t = 6.01 pixels. The target size t will vary depending on factors such as the resolution and scale of the image to be processed. The impulse response of a DoG filter with t = 6.01 pixels and σ ₂ = 1.6σ ₁ is shown in FIGS. 13 and 14.

If the cropped image with reduced noise is DoG filtered to emphasize the difference between the target and the background, the DoG filtered subimage contains the potential microcalcification and the difference in gray level between the background. Although microcalcifications tend to be among the brightest objects in DoG filtered subimages, they are present in the region of high average gray levels and are therefore difficult to partition reliably. The thresholding process used in one embodiment of the present invention to address these issues involves "AND" global histogram results in pairs and locally adaptive thresholding. However, a preferred embodiment of the present invention uses slope area thresholding.

Since the target tends to be within the higher gray level of the image, the global threshold can be approximated by finding a level that partitions a preselected percentage of the corresponding high pixel level in the image histogram. An embodiment of a global thresholding method is shown in FIG. 15. The adaptive thresholding can be implemented by changing the high threshold and the low threshold based on the local pixel value mean and standard deviation. An embodiment of a dual zone thresholding method is shown in FIG. 16.

After calculating the image histogram p (r _k ), the gray level threshold g used to partition the preselected upper portion f of the histogram is found using:

(3)

Where r _k is a k-th gray level, 0≤g≤g _max, g _max is the maximum gray level in the image.

The locally adaptive thresholds t _lo and t _hi are found using

(4)

(5)

Where k _lo and k _hi are used to predetermine multiples of the local standard deviation σ _NN (x, y) of gray-level intensity, and μ _NN (x, y) is the (x, in y) is the local gray-level mean near N x N centered on the pixel. Other shapes of neighborhoods such as squares, circles, and ovals can be used. A pixel whose brightness or gray level value falls within a threshold interval, that is, t _lo &lt; brightness <t _hi is equal to one. Regarding the parameter optimization process, the f, k _lo , k _hi and N optimizations are discussed next. The result of the global thresholding process and the result of the local thresholding step can be logically ANDed and combined as shown in FIG. 17. Alternatively, any thresholding method may be used alone.

A preferred thresholding means is shown in FIG. 18, where the N x N window can be seen centered on the pixels x, y in the input image p (x, y). The mean μ (x, y) and standard deviation σ (x, y) of the digital mammogram images of the window are calculated. The local threshold T (x, y) is calculated as follows.

T (x, y) = A + Bμ (x, y) + Cσ (x, y) (6)

Here the values for N, A, B, C are calculated in the parameter optimization step described below. The values for T (x, y) are calculated for all x, y positions in the image.

Digital mammograms were DoG filtered to generate images d (x, y). Each pixel of the DoG filtered image d (x, y) is compared with a threshold value T (x, y). The pixels in the locally thresholded image I _s (x, y) are set to 1 where the values of the DoG filtered image are greater than the threshold and 0 otherwise.

The advantage of this novel regional slope thresholding method over the prior art thresholding method is that the threshold is calculated from the pixels in the pre-DoG-filtered image, not from the post-DoG-filtered image. This eliminates the need to correct the background slope. In conventional local thresholding, the threshold is calculated from the mean and standard deviation of the DoG filtered image as follows.

T (x, y) = Bμ (x, y) + Cσ (x, y) (7)

The problem with local thresholds calculated from DoG-filtered images is that DoG-filtered images typically have a mean value close to zero and a standard deviation that is significantly adversely affected by the presence of the target.

Regional thresholds calculated from statistics of DoG-filtered images are adversely affected. First, since the mean is close to zero, there is no degree of freedom in the threshold calculation, which is essentially a function of the standard deviation. Second, the absolute brightness of the input image is lost. It is desirable to have a high threshold in bright areas so as not to generate much spurious detection. However, information about the local mean of the input image cannot be obtained from the DoG filtered image. Finally, the standard deviation of the DoG filtered image is increased by target detection. This is because a large gray scale value becomes a DoG filtered image when there are locally bright spots of a size suitable for the original image. Thus, the presence of a target in a region increases the regional standard deviation, thereby raising the threshold of that region. If the threshold is high, the probability of sending bright spots to the next processing step is reduced.

The novel regional thresholding method described now solves this problem by calculating thresholds from the input image and applying them to the DoG-filtered image. The threshold calculated here also includes an offset term A, which is independent of the regional image mean.

After thresholding, detection is converted into a single pixel representation by calculating the center of gravity of adjacent groups of pixels found by the thresholding process. The detection is thus represented as single pixels having a value of logic one and the remaining pixels have a value of logic zero.

Error positive detection outside the breast region is eliminated by logically ANDing the binary mask from the auto cropper and the single pixel representation for detection.

Malignant calcification usually occurs in clusters and can be extensive. The cluster detection module identifies the cluster based on the clustering algorithm as shown in FIG. Specifically, assert as a cluster in question when at least μCs _min or more of detected signals are less than the nearest adjacent distance d _nn . Regarding the parameter optimization process, the optimization of μCs _min and d _nn is discussed next. 20 shows the clustering process in the case of μCs _min = 5 and d _nn = 4.

Additional error positive clustered microcalcifications are eliminated by the classifier described in detail below. As shown in FIG. 21, the characteristics of each of the latent cluster phase microcalcifications are extracted. The eight features calculated for each of the latent clustered microcalcifications in the preferred embodiment are:

1. a large eigenvalue (λ ₁ ) of the covariance matrix of points in a cluster,

2. small eigenvalues (λ ₂ ) of the covariance matrix of points in the cluster,

3. The ratio of small eigenvalues to large eigenvalues of the covariance matrix for points in a cluster. Equivalent to the ratio of elliptical to long axis that covers points in the cluster.

4. Linear density calculated by dividing the number of detected microcalcifications by the maximum point-to-point distance,

5. The standard deviation of the distance between the points in the cluster,

6. The median minus mean of the distances between the points in the cluster,

7. The range of points in the cluster calculated as the maximum point-to-point minus the minimum point-to-point distance,

8. The density of clusters calculated by dividing the number of detections by the area of the box that is large enough to encompass the detection.

Of course, other features for potential microcalcification clusters may be calculated, and the present invention is not limited to the number or type of features listed herein.

Detection classification

Cluster features are provided as input to the classifier, which classifies each potential clustered microcalcification as suspect or not. Indeed, clustered microcalcification detectors can only find that area in the digital representation of the original mammogram, which may be related to cancer. In any detector, there is a tradeoff between finding as many potentially suspicious areas as possible and reducing the number of normal areas that were falsely detected as suspicious. CAD systems are designed to provide the largest possible detection rate in exchange for detecting a potentially significant number of areas that are actually normal. The majority of these unwanted detections are removed from consideration by applying pattern recognition techniques.

Pattern recognition is a process of determining based on measurement. In this system, the relevant area or detection is found by the detector and accepted or rejected for display. The first step of this process is to characterize the detected area. For this purpose, a plurality of features are calculated from the detected areas. Each measurement is called a feature. The entire measurement for the detected area is called a feature vector, where each element of the vector represents a feature value. This feature vector is input to the discrimination function.

Fig. 22 shows a classifier having a feature vector x applied to a set of discrimination functions g (x). The classifier shown in Fig. 22 is designed with one discrimination function per class. The discriminant function computes a single value as a function of the input feature vector. Discriminant functions can be learned from the training data and implemented in various functional forms. The output of the discriminant function is called test statistics. Classification is to choose a class based on the discriminant function as the largest output value. Compare test statistics with thresholds. For values of test statistics above the threshold, the area or detection associated with the feature vector is retained and displayed as potentially suspicious. When the test statistic is below the threshold, the area is not displayed.

There are many ways to design discriminant functions. One method considered for this invention is the artificial neural network class. Artificial neural networks require training, whereby a discriminant function is formed using training data classified by labels.

In a preferred embodiment, the classification process is implemented by a multilayer perceptron (MLP) neural network (NN). Of course, other classifier means may be used, for example a statistical quadratic classifier. Only latent clustered microcalcifications classified as suspicious are retained for final indication to the radiologist. Alternatively, it may be desirable to repeat the MLP NM analysis of individual microcalcifications and loops between microcalcifications.

A schematic of MLP NN is shown in FIG. 23. The MLP NN includes a J hidden layer node or perceptron 410 of the first layer, and one output node or perceptron 420 for each class. The preferred embodiment of the present invention uses two output nodes, one for each class of suspicious detection and one for non-suspicious detection. Of course, more or fewer classes can be used to classify clusters of microcalcifications. First, each calculated feature x _i is multiplied by the weight w _ij . Where i is an index indicating an i-th feature vector element and j is an index indicating a j-th first layer node. The output y _i of each first layer perceptron 410 is a nonlinear function of the weighted input and is given by

(8)

Where d represents the total number of features x _i and f (·) is saturated nonlinearity. In this embodiment, f (·) = tanh (·). The first layer or hidden layer node output y _i is multiplied by the weight u _{j, k} of the second layer and applied to the output layer node 420. The output of the output layer node 420 is a nonlinear function of the weighted input and is given next.

(9)

Where k is the exponent indicating the kth output node.

The hyperbolic tangent function is used in a preferred embodiment of the system because it allows for training relatively quickly compared to other functions. However, other hyperbolic tangents can be used to provide output from the perceptron. For example, a linear function may be used, or a smoothly varying nonlinear function such as an S-shaped function may be used.

Weights are obtained by training the network. Training consists of providing inputs to the network, iterating over the feature vectors of known class members. The weight is adjusted by the backpropagation algorithm to reduce the mean square error between the actual network power and the desired network power. The desired outputs z ₁ and z ₂ for the suspicious input are +1 and -1, respectively. Other error metrics and outputs may be used.

In this embodiment of the system, the MLP NN is implemented in software running on a general purpose computer. Alternatively, MLP NN may be implemented in a hardware configuration by means apparent to those of ordinary skill in the art.

After training, each detected clustered microcalcification is classified as suspect or unsuspecting by means of forming the difference z ₁ -z ₂ , and then compares the difference with the threshold θ. For z ₁ -z ₂ values above the threshold θ, i.e. z ₁ -z ₂ ≥θ, the classifier returns a +1 value for suspected clustered microcalcifications, and for a value of z ₁ -z ₂ <θ. Classifiers are used for microcalcifications on unsuspected clusters. Returns a value of -1.

MLP NN was trained with a set of training feature vectors derived from a database of 978 mammogram images to arrive at an optimal value for each weight and an optimal value for the number of first layer nodes.

In order to develop and test the CAD system of the present invention, truth data was first generated. Authenticity data provides for the differentiation of tissue in the digital image as a function of location. Authenticity data was generated by the radiologist who indicated the authenticity box for the image area related to cancer. In addition to mammogram images, the radiologist has access to patient history and disease progress reports.

The radiologist has identified 57 relevant areas, including cancers identified by biopsy involving clustered microcalcifications, by authenticity boxes. All 987 images were processed with the microcalcification detector of the present invention to generate a plurality of feature vectors, and a subset of this feature bettor was associated with 57 authenticity boxes. Half of the subset feature vectors were randomly chosen to include a set of training feature vectors, with about three times the number of feature vectors not related to clustered microcalcifications. MLP NNs with a certain number of hidden nodes were trained using a training set. The remaining feature vectors were used as a test database to evaluate the performance of the MLP NN after training. Training of the MLP NN was performed by the Revverter-Marquart backpropagation algorithm.

Alternatively, the MLP NN may be trained with other learning algorithms and may have nonlinear other than hyperbolic tangents on either or both layers. In an alternative embodiment with an S-shaped output node, Bayes's optimal solution to the problem of classifying clustered microcalcification detection as suspect or unsuspecting can be obtained.

Prior to applying the MLP NN classifier to remove the microcalcifications on error positive clusters, in one performance of the preferred embodiment in the test, the detection process found about 93% of the true positivity clusters on true-positive clusters in the training database and the test database. Microcalcifications on about 10 error positive clusters per image were shown. After using the MLP NN classifier with 25 first layer nodes with each optimal weight found in training, 93% true-positive detection was maintained and 57% false positive detection was successfully removed. FIG. 24 shows histogram results of testing with a test database after sorting by MLP NN. FIG. Of course, the MLP NN of the present invention can be operated with more or less first layer nodes as desired.

Detection display

After the location of the microcalcifications on the clusters are determined, they are displayed on the original digitized mammogram or a copy of the original image by drawing a rectangular box around the microcalcifications. Other means of indicating the location of microcalcification may be used, for example, by placing an arrow indicating detection on the image or by drawing an ellipse around the detection.

The location of microcalcifications on the clusters is sent to display detection as a list of row and column coordinates of the upper left and lower right pixels bounding each cluster. For each cluster, the minimum row and column coordinates and the maximum row and column coordinates are calculated. The bounding box defined by the minimum and maximum row and column coordinates is added to the original digitized image by means known in the art. The resulting image can be stored as a computer readable file, displayed on a monitor, or printed as a hard-copy image, as desired.

In one embodiment of the system, the resulting image is stored on a hard disk of a general purpose computer equipped with a dual Pentium II® processor and operating on a Windows NT® operating system. The resulting image can be viewed with VGA or SVGA, such as a ViewSonic PT813® monitor, or printed as a hard-copy gray-scale image using a laser printer such as Lexmark Optra S1625®. . Of course, other hardware elements may be used by those skilled in the art.

Parameter optimization

Genetic algorithms (GAs) have been successfully applied to many diverse and difficult optimization problems. Preferred embodiments of the present invention employ an implementation of GA developed by Houck, et al. ("A Genetic Algorithm for Function Optimization," Tech. Rep., NCSU-IE TR 95-09, 1995), which is incorporated herein by reference. To find promising parameter settings. The parameter optimization process of the present invention is shown in FIG. This is an application of new optimization techniques over current computer-aided diagnostic systems that require manual adjustment by experiment.

GA uses simulated evolutionary operators, such as mutation and male recombination, to find a solution space that maximizes the fitness function. In this embodiment, the fitness function to be maximized reflects the purpose of maximizing the number of true-positive detections and the purpose of minimizing the number of false positive detections. In using GA, several issues must be determined: evaluation function design, parameter set representation, population initialization, selection of selection functions, selection of genetic operators (regeneration mechanisms) for simulated evolution, and identification of termination criteria.

The design of the evaluation function is a major factor in the performance of any optimization algorithm. Function optimization problems for detecting microcalcifications on clusters can be described as follows: any finite domain D, a specific set of cluster detection parameters, x = {t, f, k _lo , k _hi , N, μCs _min , d _nn }, where x D, and the evaluation function f _obj : D → R, R is given by a set of real numbers, where x finds x that maximizes or minimizes f _obj in D. When using slope area thresholding for the cluster detector, the parameters N, A, B, C are optimized. The radiation imaging system can be optimized to maximize the TP rate with constraints that minimize the FP rate. This evaluation can be rewritten in the form of a function shown in the following equation.

10

Maximization is the purpose here. For a particular set of cluster detection parameters, the evaluation function returns a negative TP rate if the minimum acceptable TP rate TP _min is exceeded. Otherwise, if the TP rate falls below TP _min , the evaluation function returns a constant value, FP _penalty = -10. Other evaluation functions may be used.

This embodiment of the present invention uses a floating point representation of GA because real GA is more useful in CPU time than binary GA, and provides higher precision for more consistent results over replication. .

This embodiment puts some previously known members into the initial population so that they are in relevant parts of the search space to iteratively refine existing solutions. In addition, the number of members is limited to 20 to reduce the calculation cost of evaluating the evaluation function.

In one embodiment of the invention, for the promising selection process used to identify playback candidates, normalized geometric rankings are used as detailed in Houck, et al. The ranking results in less premature convergence caused by individuals much higher than average. The basic idea of the ranking is to choose a solution for the mating pool based on the relative goodness of the solutions. This example uses the internal genetic algorithm of arithmetic crossover and nonuniform mutations included in GA of Houck, et al.

This embodiment continues to look for solutions until the evaluation function converges. Alternatively, the search may end after a certain number of generations. Although termination is due to loss of population changes and / or unimproved when mating is the main cause of variation in the population, a homogeneous population can be followed with better (higher) fitness when using mutations. Crossing refers to creating a new member of the population by combining elements from some of the most suitable members. This corresponds to keeping the solution in the best part of the search space. Mutation refers to the random change of elements from the most suitable member. This only allows the algorithm to exit an area of search space that may be a local maximum. Since it has proven useful to restart a population that might converge, GA is repeated several times until it is recognized that the average goodness of fit is not consistently raised.

Once a potentially optimal solution is found using GA, the most suitable GA solution can be further optimized by local search. Alternative embodiments of the present invention use a simple method to further refine the optimized GA solution.

Automatic cropping systems benefit from the optimization of the parameters of the system, including the contrast value, erosion number, and expansion number. Methods for optimizing automatic croppers include generating a breast mask by hand for some training data, selecting an initial population, and generating a breast mask for training data. The method further includes measuring the percentage of overlap between the automatically generated mask and the automatically generated mask as well as the portion of breast tissue that is automatically cropped out of the manually generated mask. The method includes selecting a dominant member, creating a new member, and repeating in the same manner as described above until the predetermined evaluation function converges.

26 and 27 show the free response receiver operating characteristic curves for the system of the present invention for the output of each of the optimized microcalcification detector and classifier. FIG. 26 shows the performance of the optimized detector before classifying the detection, and FIG. 27 shows the performance of the system after classifying the detection.

Although GA has been described above in relation to the parameter optimization portion of the preferred embodiment, other optimization techniques, such as, for example, response representation methods, are suitable. Of course, processing systems other than those described herein can be optimized by the methods described herein, including GA.

Combination of CAD system outputs for optimum sensitivity

Performance metrics for the detection of suspicious sites related to cancer are often reported in terms of sensitivity and specificity. Sensitivity is a measure of how well the system finds suspect sites and is defined as the percentage of suspect sites detected from the total number of suspect sites in the observed patient. Sensitivity is defined as follows.

Sensitivity = TP / (TP + FN) (11)

Where TP is the number of sites reported as suspected by the CAD system involved in cancer, and FN is the number of sites known to have cancer not reported as suspected. The characteristic measures how well the system reports the normal as normal. The property is defined as follows:

Attribute = TN / (FP + TN) (12)

Where TN represents a region correctly identified as not suspect and FP represents a site reported as suspect without cancer.

Current CAD systems increase properties by creating FPs. However, FP and TP are combined amounts. In other words, the reduction of FP is the reduction of TP. This means that when the goal is to maintain high characteristics, it has missed some suspect areas that might have been detected.

28 and 29 show the relationship between the two TP, FP, TN, FN. The measurements from the screening breast radiation images are expressed as test statistics x. The probability density function of x is represented by p (x) and the decision threshold is represented by θ. If x is greater than θ, the thinner area of doubt is reported. The area under the probability density function represents the probability of the event. 28, it can be seen that increasing the threshold reduces the probability of FP determination. However, it can be seen from FIG. 29 that increasing the threshold reduces the probability of TP determination at the same time.

Another metric that exists for CAD systems is a positive predictive value (PPV), which is defined as the probability that cancer actually exists when the area is marked as suspect. PPV can be calculated from the following equation.

PPV = TP / (TP + FP) (13)

Increasing TP or decreasing FP increases PPV.

The radiologist and the computer find different areas of doubt. FIG. 30 is a Venn diagram showing the possible distribution of the site in question for human and machine detection. Some suspect thin spots were found alone by human interpreters or radiologists, some only by CAD systems, some by all, and some by neither.

FIG. 31 illustrates a preferred method including the CAD of the present invention, in particular the output of the CAD system, with the human interpreter's observation of the screening mammogram image 10 for optimal sensitivity, the radiologist comprising the steps The screening breast radiograph image 10 is examined at 20 to report a suspected set of sites 30 as S1. The CAD system operates on the image 10 in step 40, reporting a suspect set of regions 50, represented as S2. The radiologist examines the set of S2s and accepts or rejects the members of the set of S2s as suspect in step 60, thereby allowing a third set of suspected sites 70, which is a subset of S2, represented by S3. To form. The radiologist generates in step 80 a set of overhaul sites 90 represented by S4, which is the union of the S1 and S3 sets. A series of overhaul sites 90 may be used to take additional breast radiographs with greater resolution, examine areas of breast tissue corresponding to a series of overhaul sites by ultrasound, or perform biopsy of breast tissue. It is recommended to perform further tests.

FIG. 32 illustrates an alternative embodiment of the present invention that includes a density detector 800, a density classifier 900, and a detection result combiner 1000 in addition to the components described above. The density detector 800 detects mass and regions that appear in the digital representation of the screening mammogram. The density classifier 900 classifies the density detected by the MLP NN as suspect or unsuspecting, in a manner similar to the MLP NN described above with respect to the microcalcification classifier. The detected densities classified as suspicious are combined together and bound to suspicious detected microcalcification in detection result combiner 1000.

Although the present invention has been described in connection with detecting clustered microcalcifications in breast radiographs, it will be appreciated that the methods and systems described herein may be applied to other medical images such as chest x-rays.

Although the form of the device described herein constitutes a preferred embodiment of the invention, the invention is not limited to this precise form of the device, and changes can be made without departing from the scope of the invention as defined in the appended claims. You should know

The present invention is applicable to methods and systems for automatically detecting clustered microcalcifications from digital images without reducing the radiosensitivity.

Claims (51)

A method for automatically detecting clustered microcalcification from a digital mammogram, Obtaining a digital mammogram; Optimizing first parameters for cropping the digital mammogram; Cropping the digital mammogram based on the optimized first parameters to produce a cropped image representing breast tissue in the digital mammogram; Optimizing second parameters for detecting clustered microcalcifications in the cropped image; Detecting clustered microcalcifications in the cropped image based on the optimized second parameters; Indicating the location of the detected clustered microcalcifications in the digital mammogram Automatic detection of clustered microcalcifications in digital mammography comprising a. A method of partitioning an area of an image corresponding to breast tissue from a remainder of a digital mammogram image, the method comprising: Storing a digital representation of the digital mammogram image; Emphasizing the histogram of the digital representation to produce a highlighted image with increased contrast of the area of the breast radiographic image corresponding to breast tissue; Thresholding the highlighted image to generate a binary image comprising a seed pixel; Region growing the seed pixel in the binary image to generate a mask; Filling the holes in the mask; Eroding the mask; Dilating the mask; And Cropping the digital representation to the size of the largest object in the mask Breast tissue region segmentation method in a digital radiographic image comprising a. A method for detecting clustered microcalcifications in digital mammography images, First filtering the digital mammogram image to reduce noise in the image to produce a noise reduced image; Second filtering the noise reduced image with a Gaussian difference filter to produce a DoG-filtered image with emphasis on the appearance of latent microcalcification; Global thresholding the DoG-filtered image to partition latent microcalcification from the DoG-filtered image; Local thresholding the DoG-filtered image to partition latent microcalcification from the DoG-filtered image; Logically ANDing the globally and locally thresholded latent microcalcifications together; Converting the logically ANDed latent microcalcifications into a single pixel coordinate representation; Removing single pixel coordinate representations that lie outside the area of the digital mammogram image corresponding to breast tissue; Clustering the single pixel coordinate representations remaining from the previous step together to identify potential clustered microcalcifications; Calculating features for each of the latent clustered microcalcifications; Removing the latent clustered microcalcifications based on the calculated features for each of the latent clustered microcalcifications; And Marking on the digital mammogram the position of the latent clustered microcalcification remaining after the previous step. Clustered microcalcification detection method in a digital mammography image comprising a. In a method for automatically detecting clustered microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Convolving a filter kernel containing Gaussian differences into the digital representation, thereby suppressing information in the image that does not conform to the size and shape characteristics of the microcalcifications, and suspicious microcalcification is bright in the resulting image. Making it appear as a dot; And Thresholding the resulting image globally and locally to obtain a second resulting image that includes only the areas of fundamentally suspect microcalcification; Automatic detection of cluster phase microcalcification in screening mammography comprising a. In a method for automatically detecting clustered microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Obtaining optimized parameter values using a database that trains an optimization algorithm and images; And Applying a filtering algorithm to the digital representation using the optimized parameters to obtain a filtered image comprising essentially suspicious microcalcifications; Reducing the filtered image to obtain an image comprising a single pixel representation of the suspected microcalcification; Grouping the single pixel representations into clusters using the optimized parameter Automatic detection of cluster phase microcalcification in screening mammography comprising a. 6. The method of claim 5, wherein using the optimization algorithm comprises using a genetic algorithm. 7. The method of claim 6, wherein using the genetic algorithm comprises obtaining a nearest neighbor distance d _nn value and a detected microcalcification number μCs _min value, And the grouping step groups the single pixel representation into clusters representing microcalcifications that are different microcalcifications within a distance d _nn of μCs _min . 8. The method of claim 7, wherein using the genetic algorithm comprises repeatedly searching for a solution space containing possible values for d _nn and μCs _min to identify sets of values for which at least one fitness function is maximized. Automatic detection of cluster phase microcalcification in screening mammography further comprising. 10. The method of claim 8, wherein using the genetic algorithm comprises using a simple method of identifying at least one of the sets of values for which a cost function is minimized. Detection method. In a method for automatically detecting clustered microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Finding latent microcalcification clusters in the digital representation; Extracting features of the latent microcalcification cluster; Using the extracted features as input to a multilayer perceptron artificial neural network; And Classifying the microcalcification cluster as suspect or unsuspecting using the multilayer perceptron ligament neural network Automatic detection of cluster phase microcalcification in screening mammography comprising a. The method of claim 10, wherein the step of classifying the microcalcification cluster as suspect or unsuspecting using the multilayer perceptron diameter neural network, Obtaining a series of resulting values using the outputs of the multilayer perceptron artificial neural network at a smoothly varying output function; And Classifying the cluster associated with this value as suspect if one of the resulting values is above a threshold and classifying as suspect if the value is less than the threshold, and the clustered microcalcification in screening mammography. Automatic detection method. The method of claim 11, wherein the smoothly varying output function comprises a hyperbolic tangent function. 12. The method of claim 11, wherein the smoothly varying output function comprises a linear function. 12. The method of claim 11 wherein the smoothly varying output function comprises a sigmoid function. 12. The method of claim 11, wherein the step of classifying the microcalcification cluster as suspect or unsuspecting using the multilayer perceptron diameter neural network, Multiplying at least one of the features by a weight w _{i, j} , where I is an index representing the i th feature vector element of the feature vector x having N elements and j is an index representing the j th first layer node; And Using the first layer nodes of the multilayer perceptron artificial neural network, calculating a first layer output f _j , where x _i is calculated according to the following function comprising the calculated feature vector element: Automatic detection of cluster phase microcalcification in screening mammography comprising a. 16. The method of claim 15, wherein classifying the microcalcification cluster as suspect or unsuspecting using the multilayer perceptron diameter neural network, Multiplying at least one of the first layer outputs f _j by a second weight u _{j, k} ; And Using the result of the multiplying step as input to at least one output node, yj = f _j (x), k is the exponent representing the kth output node, J is the number of first-layer outputs to be multiplied And calculating the output of the at least one output node according to the clustered microcalcifications in screening mammography. A method of combining output detection of a computer-aided detection system for detecting clustered microcalcifications in a mammogram with detection by a human interpreter of the photograph without diminished degradation, Obtaining the observed detection to form a first detection set; Obtaining the output detection to form a second detection set; Accepting some output detection from the second set to form a third detection set; Combining the first set and the third set to form a third detection set; And Providing an output based on the second detection set Combination method of output detection of the computer-aided detection system comprising a human analyst observation detection. In a method for automatically detecting clustered microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Filtering the digital representation to obtain a filtered image that includes suspected microcalcifications; Thresholding the filtered image to a sloped area threshold to obtain an image that includes only an area of suspicious microcalcification Automatic detection of cluster phase microcalcification in screening mammography comprising a. The method of claim 18, wherein the thresholding step, Positioning a relevant pixel with coordinates (x, y) in the digital representation at the center of the window; Calculating an average (x, y) and standard deviation (x, y) of the pixels under the window; A is a predetermined offset and B and C are predetermined coefficients T (x, y) = A + Bμ (x, y) + Cσ (x, y) Calculating a local threshold T (x, y) for the relevant pixel as a function of; Comparing the gray-scale value and the local threshold value of the corresponding pixel in the filtered image; And Generating a binary image by setting a corresponding pixel in the binary image to 1 if the gray-scale value is greater than or equal to the local threshold and setting to 0 if the gray-scale value is less than the local threshold. Automatic Detection of Clustered Microcalcifications in Mammography. An apparatus for automatically detecting clustered microcalcification from a digital mammogram, Means for obtaining a digital mammogram; Means for optimizing first parameters for cropping the digital mammogram; Means for cropping the digital mammogram based on the optimized first parameters to produce a cropped image representing breast tissue in the digital mammogram; Means for optimizing second parameters for detecting clustered microcalcifications in the cropped image; Means for detecting clustered microcalcifications in the cropped image based on the optimized second parameters; Means for indicating the location of the detected clustered microcalcifications in the digital mammogram Automatic detection device for clustered microcalcifications in digital mammography comprising a. An apparatus for detecting clustered microcalcifications in digital mammogram images, Differences in Gaussian filters for producing DoG-filtered images with emphasis on the appearance of latent microcalcifications; Thresholding means for partitioning latent microcalcifications from the DoG-filtered image; Extraction means for generating a single pixel coordinate representation for said latent microcalcification; And Clustering means for grouping the single pixel representation into clusters Clustered microcalcification detection device in a digital mammography image comprising a. An apparatus for automatically detecting clustered microcalcification by digital image processing in screening mammography, Means for storing a digital representation of a mammogram; Convolving a filter kernel containing Gaussian differences into the digital representation, thereby suppressing information in the image that does not depend on the size and shape of the microcalcifications and causing suspicious microcalcifications to appear as bright spots in the resulting image. Way; And Means for globally and regionally thresholding the resulting image to obtain a second resulting image that includes only the area of suspicious microcalcification Automatic detection of cluster phase microcalcification in screening mammography comprising a. An apparatus for automatically detecting clustered microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Means for optimizing parameter values using a database that trains an optimization algorithm and images; Means for applying a filtering algorithm to the digital representation using the optimized parameters to obtain a filtered image comprising essentially suspicious microcalcifications; Means for reducing the filtered image to obtain an image including a single pixel representation of the suspected microcalcification; And Means for grouping the monopixel representations into clusters using the optimized parameter Automatic detection of cluster phase microcalcification in screening mammography comprising a. 24. The apparatus of claim 23, wherein the step of using the optimization algorithm is a genetic algorithm. The method of claim 24, wherein the means for optimizing the parameter values comprises: Means for obtaining a nearest adjacency distance d _nn value and a detected microcalcification number μCs _min value, Said grouping means grouping said single pixel representation into clusters representing microcalcifications, said microcalcifications being different microcalcifications within a distance d _nn of μCs _min . 26. The apparatus of claim 25, wherein the means for optimizing the parameter comprises means for iteratively searching the solution space including possible values for d _nn and μCs _min to identify sets of values for which at least one fitness function is maximized. Automatic detection of cluster phase microcalcification in screening mammography. 27. The automatic detection of clustered microcalcifications in screening mammography according to claim 26, wherein the means for optimizing the parameter comprises means for using a simple method of identifying at least one of the sets of values for which a cost function is minimized. Device. An apparatus for automatically detecting clustered microcalcification by digital image processing in screening mammography, Means for storing a digital representation of a mammogram; Means for finding latent microcalcification clusters in the digital representation; Means for extracting features of the latent microcalcification cluster; Means for classifying the microcalcification cluster as suspect or unsuspecting using the features. Automatic detection of cluster phase microcalcification in screening mammography comprising a. 29. The apparatus of claim 28, wherein the sorting means is a screening mammography comprising a multilayer perceptron ligament neural network. 30. The apparatus of claim 29, wherein the multilayer perceptron neural network comprises smoothly varying output means. 31. The apparatus of claim 30, wherein the smoothly varying output means comprises means for applying a hyperbolic tangent function to the sum of the weighted inputs to provide an output indicating whether or not the microcalcification on the cluster is suspicious or unsuspecting. Automatic detection of cluster phase microcalcification in screening mammography comprising. 31. The apparatus of claim 30, wherein the smoothly varying output means comprises means for applying an S-shaped function to the sum of the weighted inputs to provide an output indicating whether or not the microcalcification on the cluster is suspect or not. Automatic detection of cluster phase microcalcification in screening mammography. 31. The apparatus of claim 30, wherein the smoothly varying output means comprises means for applying a linear function to the sum of the weighted inputs to provide an output indicating whether or not the microcalcification on the cluster is suspect or not. Automatic detection of clustered microcalcifications in screening mammography. The neural network of claim 30, wherein Means for multiplying at least one of the features with a weight w _{i, j} , where _i is an exponent representing the i th feature vector element of the feature vector x having N elements and j is an exponent representing the j th first layer node; And x _i contains the computed feature vector element First-layer nodes for calculating first-layer output f _j as a function of Automatic detection of cluster phase microcalcification in screening mammography comprising a. The neural network of claim 34, wherein Means for multiplying at least one of the first layer outputs f _j by a second weight u _{j, k} ; And y _j = f _j (x) and J is the number of first-layer outputs to be multiplied According to claim, k _k , k is an automatic detection device for clustered microcalcifications in screening mammography further comprises at least one output node for calculating the index, k-th output node. In the method of automatically detecting clustered microcalcification and density by digital image processing in screening mammography, Storing a digital representation of the mammogram; Convolving a filter kernel containing Gaussian differences into the digital representation, thereby suppressing information in the image that does not conform to the size and shape characteristics of the microcalcifications, and suspicious microcalcification is bright in the resulting image. Making it appear as a dot; And Thresholding the resulting image to obtain a second resulting image that includes only the areas of suspicious microcalcification; Detecting a density in the digital representation; And Generating a third resulting image that includes only the areas of suspicious density Clustered microcalcification and density automatic detection method in screening mammography comprising a. A method for automatically cropping an image by digital image processing in screening mammography, Storing a digital representation of the mammogram; Dilatating the first digital representation to produce an expanded digital representation; Cutting the expanded digital surface to the size of the largest group of adjacent pixels to produce the cropped expanded digital representation; And Selecting for subsequent processing pixels in the first digital representation that correspond to pixels in the cropped expanded digital representation Automatic cropping method of the image in the screening mammography comprising a. In the method for automatically detecting microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Normalizing bright values of the digital representation; Generating a breast mask defining an area of the digital representation comprising breast tissue using a predetermined contrast setting in an area growth algorithm; And Searching for said area containing breast tissue for microcalcification Automatic detection of cluster phase microcalcification in screening mammography comprising a. 39. The method according to claim 38, wherein the normalization step comprises a histogram equalization step. In the method for automatically detecting microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Searching an area of the digital representation to find a seed pixel; Using a region-expansion algorithm, generating a breast mask that defines a region of the digital representation comprising breast tissue, using the region-expansion algorithm, Causing adjacent groups of pixels to be created using a region-expansion function that gradually groups together the nearest adjacent pixels starting at the seed pixel and sharing characteristic; Using a contiguous group of pixels to determine a breast mask defining an area of the digital surface containing breast tissue; And Searching for an area comprising the breast tissue for microcalcification Automatic detection of cluster phase microcalcification in screening mammography further comprising. 41. The method of claim 40, wherein searching the region of the digital representation to find the seed pixel comprises finding a pixel within the region having a maximum gray-level value. . 41. The method of claim 40, wherein grouping the nearest neighbor pixels that share the feature together comprises determining whether each nearest neighbor pixel has a contrast ratio that is less than the contrast ratio threshold. Automatic microcalcification detection method. In the method for automatically detecting microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Calculating an area threshold from pixels in the digital representation; Generating a filtered image by filtering the digital representation using a Gaussian difference filter; Generating a thresholded image by comparing pixels in the filtered image with the local threshold and adjusting pixel values according to the comparison result; And Further processing the thresholded pixels to identify microcalcifications Automatic microcalcification detection method in screening mammography comprising a. In the method for automatically detecting microcalcification by digital image processing in screening mammography, Storing a digital representation of the mammogram; Obtaining values of at least one parameter using a genetic algorithm and a database to train the imagers; And Using the at least one parameter in a function to process the digital representation to identify microcalcifications Automatic microcalcification detection method in screening mammography comprising a. 45. The method of claim 44, wherein the at least one parameter comprises a nearest neighbor distance d _nn value and a detected number of microcalcifications μCs _min value, the function converting the single pixel representation to another microcalcification within a distance d _nn of μCs _min . A method for automatically detecting microcalcifications in screening mammography comprising grouping into clusters representing phosphorus microcalcifications. 45. The method of claim 44, wherein the at least one parameter comprises a target size used in a Gaussian difference filter. 45. The method of claim 44, wherein the at least one parameter comprises at least one threshold. 45. The method of claim 44, wherein the at least one parameter comprises an adjacent size. 45. The method of claim 44, wherein the function comprises a slope area threshold function. 45. The method of claim 44, wherein said at least one parameter comprises a percentage of the histogram. 48. The method of claim 47, wherein said at least one threshold comprises an upper and lower threshold.