MXPA00002074A - Method and system for automated detection of clustered microcalcifications from digital mammograms - Google Patents

Abstract

A method, a system for detecting, and displaying clustered micro-calcification in a digital mammogram wherein a single digital mammogram (100) is first automatically cropped (200) to a breast area sub-image which is then processed by means of an optimized difference of a Gaussian filter to enhance the appearance of potential micro-calcification in the sub-image. The potential micro-calcification is threshold, clusters are detected (300), features are computed for the detected clusters, and the clusters are classified (400) as either suspicious or not suspicious by means of a neural network. The locations in the original digital mammogram of the suspicious detected clustered micro-calcification are indicated. The results of the system are optimally combined with a radiologist's observation of the original mammogram by combining the observations with the results after the radiologist has first accepted or rejected individual detections reported by the system.

Description

METHOD AND SYSTEM FOR THE AUTOMATIC DETECTION OF MICRO-CALCIFICATIONS IN RACIMS OF DIGITAL MAMMOGRAMS BACKGROUND OF THE INVENTION 1. Field of the Invention This invention relates to a method and system for automatically detecting microcalcifications in clusters of digital images without detriment to the sensitivity of the radiologist. 2. Description of the Background Mammography, together with a physical examination, is the current screening procedure for breast cancer research. Mammography research has been responsible for an estimated 30 to 35 percent reduction in the mortality rate. However, in 1996 approximately 185,700 new cases of breast cancer were diagnosed and 44,300 women died of this disease. Women have about 1 chance in 8 to be diagnosed with breast cancer, and 1 in 30 will die from this disease. Even though mammography is a well-studied and standardized methodology, mammograms of 10 to 30 percent of women diagnosed with breast cancer were interpreted as negative. Additionally, only 10 to 20 percent of patients were ordered to perform a biopsy on the basis that the findings on mammography proved they had cancer. In addition, estimates indicate that malignancies omitted by radiologists are retrospectively evident in two-thirds of mammograms. The omitted detections can be attributed to many factors including: poor image quality, inappropriate patient placement, inaccurate interpretation, darkening of the fibroglandular tissue, faint nature of the radiographic findings, eyestrain, or inadvertence. To increase sensitivity, a double reading is suggested. However, the increasing increase in the number of mammogram research makes this option unlikely. Alternatively, a computer-aided diagnostic system (CAD or CADx) can act as a "second reader" to help the radiologist detect and diagnose injuries. Many researchers have tried to analyze mammographic abnormalities with digital computers. However, it is believed that recognized studies have achieved undesirably low rates of true-positive detections against false-positive detections. Microcalcifications represent an ideal goal for automatic detection because microcalcifications are often the first and sometimes the only early, curable findings of breast cancer, although individual microcalcifications in a suspect cluster have a limited range of radiographic appearances. Between 30 and 50 percent of breast carcinomas detected radiographically demonstrate microcalcifications on mammograms, and between 60 and 80 percent of breast carcinomas reveal microcalcifications when examined under a microscope. Any increase in the rate of detection of microcalcifications by mammography will lead to further improvements in their effectiveness of early detection of breast cancer. Although the promise of CAD systems is to increase the ability of doctors to diagnose cancer, the problem is that all CAD systems fail to detect some regions of interest that could be found by a human interpreter. However, human interpreters also fail to detect regions of interest that subsequently prove to be indicators of cancer. Omitting a region that is associated with a cancer is called a false negative error while a normal region is associated with a cancer with the term false positive error. It is not yet clear how practicing radiologists should incorporate the results of the CAD system into their mammographic analyzes. No CAD system can claim to have found all the suspicious regions detected by an average radiologist, and these tend to have highly unacceptable rates of false negative errors. However, CAD systems are able to find some suspicious regions that radiologists may have omitted.

SUMMARY OF THE INVENTION In accordance with the foregoing, one of the objectives of this invention is to provide a method and system for the automatic detection of cluster microcalcifications from digital mammograms. These and other objects according to the invention are achieved by means of supplying a novel method and system for the automatic detection of microcalcifications in cluster from digital mammograms in which a digital mammogram is obtained, the necessary parameters for the harvest are optimized of the digital mammogram image, the digital mammogram is harvested based on the optimized harvest parameters to select breast tissue for further analysis, the necessary parameters are optimized to detect cluster microcalcifications, and cluster microcalcifications are detected in the digital mammograms harvested based on optimized detection parameters of cluster microcalcifications. The detected cluster microcalcifications are then stored as detection images, the detection image is processed for visual display, and a computer-assisted detection image is produced for review by the radiologist.

First, the radiologist reviews the original mammograms and reports a series of suspicious regions of interest, Sl. A CAD system, or more particularly, the CAD system of this invention, works on the original mammogram and reports a second series of suspicious detections or regions of interest, S2. The radiologist then examines the S2 series, accepts or rejects the S2 members as suspects, and thus forms a third series of suspicious detections, S3, that is, a subset of the S2 series. The radiologist then creates a fourth series of suspicious detections, S4, that is, the union of the Sl and S2 series, for the subsequent survey of diagnoses. Thus, the results of the CAD system are incorporated into the mammographic analysis of the radiologist in such a way that the sensitivity in general is optimized to detect regions of true positive interest. Other objects and advantages of this invention will be apparent from the following description, the accompanying drawings and the appended claims.

BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a flow diagram illustrating the automated system for the detection of cluster microcalcifications on a digital mammogram. Figures 2 and 3 are flow diagrams illustrating the method and self-cutting system of this invention.

Figures 4 to 10 are flow diagrams illustrating in greater detail the method and self-cutting system of this invention. Figure 11 is a flow diagram illustrating in greater detail the cluster microcalcification detector of this invention. Figure 12 is a schematic diagram illustrating a cross-shaped 3 x 3 medium filter of this invention. Figure 13 is a three-dimensional plane of the core. of Gaussian units difference filter (Difference of Gaussians "DoG"). Figure 14 is a cross-sectional view through the core of the DoG filter of Figure 13. Figure 15 is a flow chart illustrating the overall threshold portion of the microcalcification detection system. Figure 16 is a flow diagram illustrating the double local threshold of this invention. Figure 17 is a flowchart illustrating the combination of double and global local threshold results. Figure 18 is a flow chart illustrating the slope of the local threshold of this invention. Figure 19 is a flow chart illustrating the grouping system of this invention. Figure 20 is a schematic diagram illustrating the grouping method of this invention. Figure 21 is a flow diagram illustrating the characteristic of the computing process of this invention; Figure 22 is a flow chart which illustrates a classifier that has a discriminant function by class. Figure 23 is a schematic diagram illustrating a neural network of a multilayer perceptron for a classifier of two classes. Figure 24 is a histogram of test results after detection and classification. Figure 25 is a flow diagram illustrating the parameter optimization method of this invention. Figure 26 is a free response receiving plane which operates a characteristic curve of this invention before making the classification of the detections. Figure 27 is a free response receiving plane that operates a characteristic curve of this invention after making the classification of the detections. Figure 28 is a probability plane of density functions that shows the relationship between the probabilities of false false and false positive detections. Figure 29 is a probability plane of density functions that shows the relationship between the probabilities of true negative and true positive detections. Figure 30 is a Venn diagram showing the relationship between radiologist and CAD system detections. Figure 31 is a flow diagram illustrating a computer-assisted method for incorporating diagnostic instructions with those of the human interpreter for optimum sensitivity. Figure 32 is a flow diagram illustrating an alternative embodiment of this invention that includes a density detector.

DETAILED DESCRIPTION OF THE PREFERRED MODE With reference now to the drawings, in which the numbers with the same reference designate identical or corresponding parts through many aspects, and more particularly to Figure 1 thereof, where a diagram of flow that illustrates a sequence of steps that are carried out for the purpose of detecting the locations of clusters of microcalcifications within a digital mammogram. In a first step 100, a digital mammogram is obtained by using hardware such as digital mammography systems, or by digitizing mammography films using laser beam digitizers or coupled charge devices. In an optimized harvest step 200, a rectangular region of analysis containing breast tissue is segmented from the digital mammogram image and a binary mask is created corresponding to the breast tissue to be used in later steps of the process to decrease the time required to process the mammogram image. The binary mask is also used to limit the detections to areas of the image that contain breast tissues. The microcalcifications in clusters are detected in a detection step 300 of a microcalcification in clusters. After filtering the harvested image first with a medium filter to reduce noise, the image is filtered using a Gaussian units optimized difference filter to improve the image of the microcalcifications. The image filtered by DoG is subject to optimized threshold tests to detect potential microcalcifications. The microcalcifications detected are reduced to representations of a single pixel (image element) and the detections are eliminated outside the area of the breast. The remaining microcalcifications are grouped in clusters. The characteristics for the clusters are then computed. The detected clusters are classified as either suspect or non-suspect in a classification step 400. The parameters used for auto-cutting, detection of cluster microcalcifications, and classification steps 200, 300, 400 are optimized in an optimization step of the parameters 500. The parameters are optimized by means of parameter optimization using a genetic algorithm (GA) in such a way that the true positive detection rate is maximized while the false positive detection rate is minimized. Of course, other optimization schemes can also be used. The detected cluster microcalcifications are stored in a list of image coordinates. The detection results are processed in a processing step 600 by simply adding an equivalent for each of the microcalcification coordinates to justify the transfer of the coordinates incurred as a result of the harvest procedure. The cluster microcalcifications detected on the digital mammogram are indicated by rectangles drawn around the cluster microcalcifications in a visual display step 700. Other indicators may be used such as, for example, arrows pointing to what is suspected as microcalcifications, or ellipses around what is suspected as microcalcifications.

ACQUISITION OF A DIGITAL REPRESENTATION OF A MAMMOGRAM A method for acquiring digital mammograms involves the digitization of radiological films by means of a laser beam or a coupled charge scanner (CCD) device. The digital images obtained in this manner typically have a sample space of approximately 100 μm per pixel, with a gray level resolution of 10 to 12 bits per pixel. In one embodiment of this invention, the radilagical films are scanned by using a CX812T digitizer manufactured by Radiographic Digital Imaging of Compton, California, to produce digital images having 50 μm of space per pixel and 12 bits of gray level resolution per pixel Another possible source of supply for digital imaging is a digital mammography unit from Tres Mexican Corporation of Danbury, Connecticut, which has a spatial resolution of approximately 45 μm per pixel and a gray level resolution of 14 bits per pixel. Digital images are stored as digital representations of the original images of the mammogram in the memory of media that can be read by computer. In a preferred embodiment, the digital representations or images are stored on a 2 GB hard drive of a general purpose computer such as a personal computer (PC) having dual Pentium II ® microprocessors operating at 200 MHZ, 512 MB of memory RAM, a ViewSonic PT813 ® monitor, a pointing device, and a Lexmark Optra S1625 ® printer. The system operates within a Windows NT ® operating system.

AUTOCORTING As can be seen in Figures 2 and 3, first a digital mammogram image 190 is cut to segment an analysis region 296 of the image and produce a binary mask 298 that corresponds to the breast tissue in the region of analysis. Preferably, the cutting is done automatically, although it can be cut manually. The image is cut as a preliminary step because the breast tissue does not cover the entire radiographic film. The image process is focused only on that portion of the image where the tissue of the breast reduces the time required to process the image. In addition, other things that appear in the film, such as labels and patient information, are excluded from consideration, and false positive indications that are outside the area of the breast tissue are eliminated. With reference to Figures 4 to 10, the cutting process will be described in detail. First, the 50 μm to 400 μm image is sub-sampled to reduce the amount of information to be processed in step 202. Of course, the sub-sampling of the image can be reduced to other resolutions as desired. You do not need all the original image information to reliably segment the breast tissue from the rest of the image. The sub sampling every eight pixels in both horizontal and vertical direction reduces 64 times the amount of information. For the purpose of segmenting the breast tissue from the rest of the image, the consequent loss of resolution does not matter. A white margin of twenty pixels wide is added around all sides of the sub-sample image in step 204. The objective corresponds to the maximum possible pixel value taking into account the number of bits used to represent each pixel. For images that have 12 bits of gray scale resolution, the maximum gray scale value is 4095. The framed image is passed to the threshold in step 206 with such a high threshold value that it is ensured that the majority of breast tissues is less than the threshold to produce a binary image. In one embodiment of the invention, the threshold is set equal to a percentage of the previously determined gray scale value of a pixel near the upper middle portion of the image. Then the image of the threshold is inverted, that is, the ones become zeros and the zeros become ones, in step 208. The inverted image is then dilated in step 210. The dilation is a morphological operation in which each pixel in a binary image is activated, that is, it is set to a value of one, if any of the nearby pixels are activated. If the pixel is already activated, it is left like this. In step 212 the dilated image is trimmed to the size of the largest drop. The drops are contiguous groups of pixels that have the value one. This step 212 eliminates the bright margins of the representation of the subsampled mammogram while ensuring that none of the areas of the breast is reduced. Other techniques that are located at the threshold to find the margin have serious difficulties in managing the shiny areas in areas of the breast adjacent to the margin, for example, when breast implants are visible in the image. The pixels of the original image, which result from step 202, which correspond to the locations of the pixels in the trimmed drop, are selected for subsequent processing. Note that this is a simple sub-series of pixels from the input image. The image of step 212 is the histogram matched in step 214. The average brightness of the image will vary widely from mammogram to mammogram. Likewise, the different digitizers that have characteristics of a different optical density are an additional source of variability in brightness levels in a digital representation of the mammogram. The breast mask that is the result of the self-cutter is mainly defined by the growth by region of algorithms that requires a single contrast background to work properly. However, it has been experimentally determined that a single contrast background will not work for a wide range of produced images. Therefore, each image is mapped to a normalized image space by using an automatic histogram improvement process, after which a single contrast background works well. First, you get a histogram of the image.

Typically, most of the information in the breast area will be in the lower bins (binary files) of the histogram (which correspond to gray scale values of approximately 0 - 1000), and the margins and labels are in the upper bins ( corresponding to gray scale values of approximately 4000 - 4095) for 12-bit information. The lower and upper bin values that contain the typical breast information are determined. The lowest value of the bin is the first highest peak found when going from the lowest value of the gray scale to the highest value of the gray scale. The upper bin is the last zero value bin that is found when the highest value of the gray scale is passed to the lowest value of the gray scale. The data is then reduced to an eight-bit representation and extended linearly over the range of the data type. For example, the values in the lower bins are set to zero. The values of the data in the upper bins are set to 255. Afterwards, the rest of the data is mapped linearly between the lower and upper bins. After the image has been matched by means of the histogram, the matched image can be considered to be like a matrix. The image matrix is divided into left and right halves, if possible of equal size, and the brighter side is selected in a step 216. The sums of all the pixels are calculated in the left and right halves.

The values of the sum are then compared and the side that has the highest sum is the brighter side. Before the region grows as the brighter side, the algorithm variables are started in step 218. Preliminarily in step 220 the mask size of the grown region is verified. If it is large enough, the mask is acceptable. Otherwise, the process to find the mask is continued. The side of the image that is to be the grown region is selected in step 22. In step 224 this region is searched to determine its maximum value in the gray scale. This maximum value is used to determine a pixel to start the algorithm of the growth region. Region growth is a process of grouping connected pixels that share some similar characteristics. The selection of characteristics influence the resulting region. The entry to a growth region function is a gray scale image and a starting point to start growth. The result is a binary image in which the digits indicate the pixels within the grown region, that is, drops. The growth of a region will create a single drop, but that drop can have internal holes, that is, pixels that are inactive. To grow a drop, each of the four adjacent pixels closest to a pixel of interest are contemplated. The proportion of the contrast is calculated for each of the closest neighboring pixels. If the ratio of the contrast is less than the ratio of the contrast threshold, then the neighboring pixel is set to one in a binary mask image. Otherwise, the adjacent pixel is set to zero. The growth region algorithm spirals outward from the start or seed pixel, and progressively sees the neighboring pixels closest to completion. For those skilled in the art, it is clear that the algorithms of other growing regions can also be applied. In step 226, the growth region begins with the pixel identified in the previous step 224 to produce a binary mask. The size resulting from the mask of step 226 is calculated in step 228 and verified in step 230. There may be three failure points for 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), then the search region approaches the side of the image and the search is made again. This is repeated three times, each time decreasing the threshold contrast value. This corresponds to the course taken through steps 232 and 234. Second, the side selection statement may be erroneous. Therefore, if a valid breast mask is not found on the first search side, then the other side of the breast is searched. This corresponds to the course taken through steps 236 and 238. Third, if a valid breast mask is not found on either side, then the full breast threshold is made and the larger object is taken as a breast mask in step 240. Since a constant contrast value is used in the algorithm of the growth region, some masks will be too large. Typically, there will be "tails" along the digital image of the mammogram where additional light was filtered while the original mammogram film is digitized. The tails are reduced by applying a series of wear and then a series of expansions to the image. Erosion is a morphological function in which each pixel in a binary image is deactivated unless all its neighbors are activated. If the pixel is already disabled, it remains deactivated. But first, the holes in the mask must be filled or multiple wear can break the mask into disjointed sections. Therefore, the holes in the mask are closed in step 242 by means of a majority operation. The majority operation is a morphological operation in which each pixel in a binary image is activated if most of the neighboring pixels are activated. If the pixel is already activated, it is left activated. However, another problem is that some smaller breast masks can not withstand as much wear as the larger breast masks. Therefore, as a safety measure against failures, the sum of the breast masks is taken before and after wear and tear. If the size is reduced too much (ie by more than 50%), the original mask is used before the morphological operation. Therefore, a duplicate of the mask is made in step 244 before the mask is worn out and dilated in steps 246 and 248, respectively. The size of the resulting mask is then calculated in step 250 and compared to the mask size of step 242 in step 252. If the new size is less than half the size above, then in step 254 it is selected the duplicate of the mask, from step 244, for its subsequent processing. Otherwise, the mask resulting from step 248 is used. The original image (from step 202) is cut to the size of the breast mask that was just found (either from step 242 or from step 248) in step 256 In case the resulting mask is too small for its subsequent processing, a cutting adjustment is always made in step 258. The adjustment comes in the form of an increase in the size of the box that confines the breast mask by means of additional pixels of the original image in the cropped image. The histogram of the cropped image is then automatically improved in step 260, as previously described in relation to step 214. This improved image is passed through the loose region growth of step 262 to produce a generous mask. This means that the image is subject to a lower threshold to yield more activated pixels. This mask is then subjected to the closure of holes, wear, and dilation in steps 264, 266 and 268, respectively, according to the above, but to a lesser degree. The same steps described above are repeated one more time in steps 270 through 276, but the cut-off settings are lower and the contrast value for growth is increased by tight region in step 276. This step 276 of the growth region tight can sustain the higher contrast value since the region will be growing just inside the cropped image. This results in a sparse estimate of the breast tissue. The resulting mask is segmented to find the largest object in step 278 and its confinement box shrinks just to the size to confine the object in step 280. Some holes may still be in the breast mask. Thus, after the cutting settings in step 282, the mask is inverted in step 284 and the largest object is found in step 286. This larger object is then extracted and after reversing in step 288 to obtain the penultimate mask. The final mask is obtained by closing holes in the penultimate mask with major operations and dilations in step 290. The image is then trimmed to the size of the resulting mask and the auto-cut is completed. An important result of the autocutter is the equivalent of the cropped image. This is the location of the pixel in the original image that corresponds to the pixel in the upper left pixel of the cropped image. The equivalent value is determined if a record of all cuttings and cutting settings is maintained. The result of the auto-cutting process is a rectangular arrangement of pixels that represent a binary mask in which the pixels corresponding to the breast tissue are assigned the value of one, while the rest of the pixels are assigned a value of zero. Put another way, the binary mask is a silhouette of the breast made of ones while the background is made of zeros. You can optimize the parameters of the auto cutter to obtain better breast masks. The procedure is described below in the optimization section.

DETECTION OF CLASS MICROCALCIFICATIONS Turning now to Figure 11, a flow diagram can be seen that illustrates in more detail the cluster microcalcification system of this invention. First, that portion of the digital representation of the mammogram corresponding to the analysis region 296, designated as a sub-image cut 302, produced in the cutting step 200, is processed to reduce the noise in a noise reduction step 310 to reduce the noise. Noise from digitization that contributes to false detections of microcalcifications. The image is then filtered with reduced noise by using an optimized spatial core of Gaussian Units Difference (DoG) dependent on the size of the target in step 320 to improve the differences between the targets and the background, which in this way creates maximums global and local in the filtered image. The optimized image filtered by DoG is placed on the threshold in step 340 to segment the maxima representing the potential detections of the microcalcifications. The maximums detected are converted into representations in coordinates of a single pixel in a conversion step 350. The representations in coordinates of the detected maximums are compared with the binary mask of the analysis area in a first step of false positive removal 360 to eliminate detections false out of the area of the breast mask. The remaining coordinate representations in the analysis area are put into clusters in a step 370 to make bunches. The characteristics for the rest of the clusters are calculated in a characteristic calculation step 380 and used to eliminate non-suspicious detections in a classification step 400 (Figure 1). The remaining detections are taken as cluster microcalcifications detected in an output step 600 in the form of cluster coordinates. Returning now to a more detailed description of the steps in the process of detection of cluster microcalcifications, first the digital image of the mammogram is filtered to reduce noise in the image. Even though the main limitation in the image quality must be the granulation of the emulsion of the film, the noise of the scanning process is introduced. This noise can later be detected as a pseudocalcification. In this system, a cross-shaped medium filter is used because it is well known that it is extremely effective at eliminating noise from a single pixel. The medium filter is a non-linear spatial filter that replaces each pixel value with the mean of the values of the pixel inside the nucleus of a chosen size and shape and centered on a pixel of interest. Referring to Figure 12, it can be seen that the cross shape is formed by the series of pixels that include the pixel of the center and its four closest neighbors. The cross shape preserves the lines and corners better than the typical medium filters in block form and limits the possible substitution to the four closest neighbors, in such a way that it reduces the margin displacement potential. After the noise has been reduced, the image is filtered with an optimized DoG core to improve the microcalcifications. The filtration is carried out by means of folding the reduced noise image with the DoG core. In an alternative mode, filtering is carried out first by obtaining fast Fourier transformations (FFTs) of the reduced noise image and the DoG core, then multiplying the FFTs together, and taking the inverse FFT of the result. The DoG core was chosen because neurophysiological experiments provide evidence that the visual path includes a series of "channels" that are selective for spatial frequency. Essentially, at each point of the visual field, there are filters or masks tuned to the size that analyze an image. The operation of these spatial receptor fields can be closely approximated by a DoG. The Gaussian mask 2-D is presented as: G (x, y) = ce 2s > OR) where c normalizes the sum of the elements of the mask to unity, x and are horizontal and vertical indexes, and d is the standard deviation. Using Equation 1, the difference of two Gaussian units with different d produces: DoG (x, y) = (2) It has been shown that when d2 = 1.66- ^ then the DoG filter response closely matches the response of human space receptor filters. Therefore, with the motivation of human psychology, the ratio of the standard deviation constants DoG is allowed to be 1: 1.6. Then, for the objective of the size (average width) of the pixels t, use d2 = t / 2 and, from the rule of the practical method, d1 = d2 / 1. 6 Since microcalcifications typically have a range of 100 to 300 μm in diameter, potential target sizes for 50 μm digital mammograms correspond to 2 to 6 pixels. It has been found that a DoG core constructed with the use of an optimization technique for the selection of the target size parameter, such as the GA that is detailed below, has an optimized target size of t = 6.01 pixels. The target t will vary according to factors such as the resolution and the scale of the image to be processed. The impulse response of a DoG filter having t = 6.01 pixels and d2 = l.dd-j ^ is shown in Figures 13 and 14. Once the clipped image with reduced noise has been filtered with DoG to improve the differences between the objectives and the fund, the filtered DoG sub-image contains differences in gray levels between the potential microcalcifications and the background. Although microcalcifications tend to be among the brightest objects in filtered DoG subpictures, gray levels can exist within regions of high average and thus prove difficult to segment reliably. The process for making the threshold that is used in an embodiment of the invention that generally exposes these subjects involves "ANDing" pixel pairs of the results of the global histogram and locally adaptive threshold. However, the preferred embodiment of the present invention uses the local oblique threshold. Since the targets tend to exist within the higher gray levels of the image, then you can approximate the global threshold by finding the level that segments a previously selected percentage of the highest pixel levels that correspond to the histogram of the image. One modality of the method to put in the global threshold is illustrated in Figure 15. It can be implemented to put the adaptive threshold locally by varying the mean and standard deviations of the high and low thresholds based on the local value of the pixel. Figure 16 illustrates one modality of a method to put on the dual local threshold. After calculating the histogram of the image, p (rk), it is found that the threshold of the gray level g of the histogram, used to segment a previously selected upper fraction, f, is being used: where rk is the gray level kth, 0 = g = gmax, and gmax is the maximum gray level in the image. It is found that the locally adaptive thresholds, tj0 and t r, use 'lo k' sN x * y + * H x'y) (4) ?, = **, < x ^ + 11 ^ (5) where klo and khí are used to previously select the multiple of dNN (x, y), the standard local deviation of the intensities of gray levels, and μNN (x, y) is the local gray level means of the adjoining N x N centered on the pixel in (x, y) of the filtered DoG image. Other abutting shapes may be used, such as rectangular, circular, and ellipsoidal. Pixels whose gray-level brightness falls within the threshold range, ie, tlo < bright > thi, they are set equal to one. The optimization of f, klo, khí, and N is discussed below in relation to the parameter optimization process. The results of the process of setting the global threshold can be combined with the results of the step of setting the local threshold by means of Boolean (AND) logic as shown in Figure 17. Alternatively, any of the methods can be used alone. to put on threshold. Figure 18 illustrates the preferred element to put on the threshold, where it can be seen that a window N x N is centered on a pixel x, and on the input image p (x, y). The means μ (x, y), and the standard deviation d (x, y), of the digital mammogram calculate the pixels of the image under the window. A local threshold value, T (x, y), is calculated as: T (x, y) = A + Bμ (x, y) + Cs (x, y) (6) wherein the values for N, A, B, and C are calculated during the parameter optimization step discussed below. The values for T (x, y) are calculated for each location x, and in the image. The digital mammogram has also been filtered by DoG, which produces an image d (x, y). Each pixel of the image filtered by DoG d (x, y) is compared to the threshold value T (x, y). The pixels in the image locally placed at the threshold l = (x, y) are set as one where the values of the image filtered by DoG are greater than the threshold, and set to zero elsewhere.

The advantage of this novel local method to put in the threshold inclined on previous methods to put in the threshold is that the threshold of the pixels in an image previously filtered by DoG is calculated instead of of an image later filtered by DoG. This eliminates the need for background trend correction. In a conventional threshold setting, the threshold is calculated as: T (x, y) = Bμ (x, y) + Cs (x, y) (7) of the mean and standard deviation of the image filtered by DoG. The problem of using the calculated local threshold of the image filtered by DoG is that the images filtered by DoG have typically low values close to zero and the standard deviations are significantly affected by the presence of targets. The local thresholds calculated from the statistics of the image filtered by DoG suffer the following adverse effects. First, since the low value is close to zero, a degree of freedom is lost in the calculation of the threshold, which essentially becomes a function of the standard deviation. Second, the absolute brightness of the input image is lost. To avoid the occurrence of many false detections, it is convenient to have high thresholds in the bright regions. However, information about the local means of image input in the image filtered by DoG is not available. Finally, the standard deviations of the images filtered by DoG are increased by means of objective detections. This is because when there are bright local points of appropriate size in the original image, the larger gray scale values result in the image filtered by DoG. Therefore, the presence of objectives in a region, increases the local standard deviation in this way increases the threshold of that region. The larger threshold reduces the likelihood of passing a bright spot to subsequent stages of the process. The novel method of putting in the local threshold that has just been described above solves the previous problems by means of calculating the thresholds of the input image, which are applied to the image filtered by DoG. Additionally, the threshold that is calculated here includes an equivalent term A, which is independent of the local image means. After putting in the threshold, the detections are converted to representations of a single pixel by means of calculating the centroid of the center of gravity of groups of conetual pixels found by the process of putting in the threshold. The detections are therefore represented as individual pixels having a value of a logical one while the remaining pixels have a value of logical zero.

False positive detections outside the area of the breast are eliminated by applying Boolean logic in the binary mask of the self-cutter with single-pixel representations. Calcifications associated with malignancies usually occur in clusters and may be extensive. The cluster detection module identifies the clusters based on the cluster algorithm as depicted in Figure 19. Specifically, a suspect cluster is declared when at least smj_n or more detected signals are separated by at least the nearest neighbor's distance, dnn. Next, the optimization of Csm¿n and dnn in relation to the parameter optimization process is discussed. Figure 20 illustrates the clustering process for the case in which μCsmin = 5 and dnn = 4. Additional false positive microcalcifications are eliminated by means of a classifier detailed below. The characteristics of each of the cluster microcalcifications are extracted as shown in Figure 21. In a preferred embodiment, the eight characteristics calculated for each of the potential microcalcifications in cluster are: 1. The largest eigenvalue (? x) of the covariant matrix of the points in a cluster; 2. The largest eigenvalue (?) Of the covariant matrix of the points in a cluster; 3. The proportion of the smallest eigenvalue of the covariant matrix to the largest eigenvalue of the covariant matrix of the points in a cluster. Equivalent to the ratio of the minor axis to the major axis of an ellipse made to measure to cover the points in a cluster; 4. The linear density calculated as the number of detected microcalcifications divided by the distance of a maximum inter point; 5. The standard deviation of the distances between the points in a cluster; 6. The intermediate lower than the distances between points in a cluster; 7. The range of points in a cluster calculated as the maximum intermediate distance minus the minimum intermediate distance; and 8. The density of a cluster calculated as the number of detections divided by the area of a box large enough to encompass the detections. Of course, other characteristics for cluster microcalcifications can be calculated, and the invention is not limited to the number or types of features enumerated herein.

CLASSIFICATION OF DETECTIONS Cluster characteristics are provided as inputs to the classifier, which classifies each microcalcification into a potential cluster either as suspect or non-suspect. In practice, the cluster microcalcification detector is only able to locate regions of interest in the digital representation of the original mammogram that could be associated with cancer. In any detector, there is a trade-off between locating as many potentially suspicious regions as possible against the reduction of the number of normal regions falsely detected as potentially suspicious. CAD systems are designed to provide as many detection ranges as possible at the cost of detecting potentially significant amounts of regions that are in fact normal. Many of these unwanted detections are left out of consideration by applying recognition pattern techniques. The pattern of recognition is the process of making decisions based on measurements. In this system, the regions of interest or detections are located by means of a detector, and are then accepted or rejected for visual display. The first step in the process is to characterize the detected regions. Towards this end, multiple measurements of each of the detected regions are calculated. Each measure is referred to as a characteristic. A collection of measurements for a detected region is referred to as a vector feature, while each element of the vector represents a value of the characteristic. The vector of the characteristic is an input to a discriminant function. With reference to Figure 22, in this one you can see a classifier that has a feature vector x applied to a series of discriminant functions g (x). The classifier shown in Figure 22 is designed with a discriminant function per class. A discriminant function calculates a single value as a function of a feature vector input. You can learn about the discriminating functions of training information and can be implemented in a variety of functional ways. At the output of a discriminant function it is referred to as a test statistic. The classification is to select a class according to the discriminant function with the largest output value. The test statistic is compared to a threshold value. For the values of the test statistic above the threshold, the region or detection associated with the feature vector is stopped and visually displayed as potentially suspicious. When the test statistic is less than the threshold, the region is not displayed visually. Many methods are available to design discriminant functions. One approach considered for this invention is a class of artificial neural networks.

Artificial neural networks require training, by means of which the discriminant function is formed with the assistance of qualified training information. In a preferred embodiment, the classification process is implemented by means of a neural network (NN) of the multi-layer perceptron (MLP). Of course, other means can be used to classify as, for example, a quadratic statistical classifier. Only potential cluster microcalcifications classified as suspicious are those that are retained for eventual appointment to the radiologist. Alternatively, it may be desirable to repetitively curve between the NN MLP analyzes of the individual microcalcification detections and the microcalcification in clusters. With reference to Figure 23, a schematic diagram of an NL MLP can be observed in this. The NN MLP includes a first layer of hidden protrusions or perceptrons 410, and a node or perceptron 420 for each class. The preferred embodiment of this invention utilizes two outlet protuberances, one for each of the classes of suspicious detections and another for the class, of non-suspicious detections. Of course, more or less classes can be used for the classification of the microcaiplication clusters. Each of the calculated characteristics x is first multiplied by the weight w j, where i is an index that represents the vector element of the characteristic ith, and j is an index that represents the first jth protrusion of the layer. The output y, -of each first perceptron 410 of the layer is a non-linear function of the heavy inputs and is provided by means of: where D represents the total amount of characteristics x and F (ß) is typically a non-linearity saturator. In this mode, f (•) = tanh (#). Then the first layer or the outlets y¿ of the protrusion of the layer hidden by the second weight layer Uj / k is multiplied and applied to the protuberances 420 of the output layer. The output of a layer extrusion 420 is a non-linear function of the heavy inputs and is provided by means of: where k is an index that represents the exit protrusion kth. The function of the hyperbolic tangent is used in a preferred embodiment of the system because it allows the NN MLP to be trained more quickly compared to other functions. However, other functions extraneous to the hyperbolic tangent can be used to provide outputs of the perceptrons. For example, linear functions can be used, as well as non-linear functions that vary slightly, such as the sigmoid function. Weight values are obtained through network training. The training consists of repeatedly presenting vectors characteristic of class memberships recognized as entries to the network. The values of the weight are adjusted with a background propagation algorithm to reduce the error of the average square between the current outputs of the network and those that are desired. The desired outputs of z and z2 of a suspicious input are +1 and -1, respectively. The desired outputs of z1 and z2 of a non-suspect entry are -1 and +1, respectively. Other metric errors and output values can also be used. In this modality of the system, the NN MLP is implemented by operating a software in a general purpose computer. Alternatively, the NN MLP could also be implemented in a hardware configuration by means that are readily apparent even to those not skilled in the art. After the training, each of the cluster microcalcifications detected is classified as either suspect or non-suspect by means that make up the difference z? ~ z2 'and is the difference compared to a threshold? For the values of z - z2 that are greater than or equal to the threshold?, Ie z - z2 =?, The classifier returns a value of +1 for the suspicious cluster microcalcifications, and for z? - z2 < ?, the classifier returns a value of -1 for non-suspect cluster microcalcifications. In order to arrive at the optimal values for the respective weights, and the number of protuberances of the first layer, the NN MLP has been trained with a training set of characteristic vectors derived from the database of 978 images of mammograms. To develop and test the CAD system of this invention, first true information was generated. True information provides a categorization of the tissue in digital images as a function of position. The truthful information was generated by certified radiologists who marked true cases on regions of images associated with cancer. In addition to the mammogram images, radiologists also had access to patients' medical history and pathology reports. By means of true cases, radiologists identified 57 regions of interest, with cancer confirmed by biopsy, associated with cluster microcalcifications. Then, all 978 images were processed by the microcalcification detector of this invention to produce a plurality of characteristic vectors, one of which sub-series was associated with the 57 true cases. Half of the characteristic vectors of the subseries were randomly chosen, along with almost three times more characteristic vectors not associated with cluster microcalcifications, to integrate the training set of characteristic vectors. The NN MLP, which has a number of hidden protuberances previously determined, was then trained using the training game. The remaining characteristic vectors were used as a test data base to evaluate the performance of the MLN NN after the training. The training of the NN MLP was carried out by means of the Levenberg-Marquardt background propagation algorithm. Alternatively, the NN MLP can be trained with other "learning algorithms that may not have non-hyperbolic linearity in any or both layers." In an alternative mode with sigmoidal output protrusions, Bayes' optimal solution to the problem may be obtained. classify detections of cluster microcalcifications either as suspect or non-suspect In the course of applying the preferred modality during the test, before applying the NN MLP classifier to eliminate false positive cluster microcalcifications, the detection procedure found about 93% of the false positive cluster microcalcifications in both training and testing databases while indicating approximately 10 false positive cluster microcalcifications per image was found after an NN MLP classifier having 25 bumps was used in the first layer, with the optimal weights Researchers found that during the training, 93% of true positive detections were retained, while 57% of false positive detections were successfully eliminated. With reference to Figure 24, in this one a histogram of the test results can be observed on the basis of test data after classification by means of the MLN NN. Of course, that the NN MLP of this invention can be operated as desired, with more or less protuberances of the first layer.

VISUAL DISPLAY OF THE DETECTIONS Once the positions of the cluster microcalcifications have been determined, they are indicated in the original digitized image of the mammogram, or in the original image copy, by means of drawing rectangular boxes around the microcalcifications. Other means may be used to indicate the positions of the microcalcifications, such as, for example, arrows placed on the image to indicate the detections or by means of ellipse-like patterns around the detections. The locations of the cluster microcalcifications are passed to the procedure of visual display of detections as a list of row and column coordinates of the upper left and lower right pixels that link each of the clusters. The minimum row and column coordinates and the maximum row and column coordinates are calculated for each cluster. Linked boxes that are defined by minimum and maximum row and column coordinates are added to the original digitized image, by means well known in the art. The resulting image is then stored as a computer readable file, displayed visually on the monitor, or printed as a hard copy image, as desired. In one system mode, the image that results in a hard disk of a general purpose computer having dual Pentium II ® processors and a Windows NT operating system is saved. The resulting image can be viewed on a VGA or SVGA monitor, such as a ViewSonic PT813 ®, or printed hard copy as a gray scale image of a laser printer, such as the Lexmark Optra S1625 ®. Of course, those not skilled in the art can use other hardware elements.

OPTIMIZATION OF PARAMETERS Genetic algorithms (GAs) have been successfully applied to different and difficult optimization problems. A preferred embodiment of this invention utilizes the implementation of a 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. reference, to find promisior adjustments of parameters. The parameter optimization process of this invention is shown in Figure 25. This is a novel application of optimization techniques compared to computer-aided diagnostic systems since these require manual tuning by experiment. GAs seek the solution space to maximize a goal function by using simulated evolutionary operators such as mutation and sexual recombinations. In this modality, the objective function to be maximized reflects the goals of maximizing the amount of true positive detections while minimizing the amount of false positive detections. The use of GA requires the determination of many issues: the design of the objective function, the representation of the parameter set, the initialization of the population, choice of the selection function, choice of genetic operators (reproduction mechanisms) for the simulated evolution, and identification of the conclusion criterion. The design of the objective function is a key factor in the realization of any optimization algorithm. The problem of the optimization function for detecting cluster microcalcifications can be described as follows: in some finite domain D, a particular set of cluster detection parameters, x =. { t, f, klo, khi, N, μCsmin, dnn} where ? e D, and an objective function f "sjbj: D - R, where R denotes the set of real numbers, find the x in D that maximizes or minimizes the ~ fobj When the slanting is used at the local threshold in The cluster detector optimizes the parameters?, A, B, and C. Radiological imaging systems can be optimized to maximize the TP proportion subject to the constraint of minimizing the FP rate. the functional form shown in the following equation: -FP (x), TP (x) = TPmm * / * > - (10) ^ penalty '^ e otherwise where maximization is the goal. For a particular set of cluster detection parameters, if the TP m, xn of the minimum acceptable TP rate is exceeded, the objective function returns the negative of the FP rate. Otherwise, if the TP rate falls below the TPmin, the objective function returns a constant value. The FPpenal ty = -10. Other functions of the lens can be used. Since the GA of real value is an order of magnitude more efficient in CPU time than the binary GA, and provides greater accuracy with more consistent results across the replicas, this embodiment of this invention uses a floating point representation of the GA . This modality also simulates the initial population with some members that are known in advance that they are in an interesting part of the search space so that they improve repetitively the solutions in existence. Likewise, the number of members is limited to twenty to reduce the computational costs caused by the evaluation of objective functions. In one embodiment of this invention, the standardized geometric classification is used for the process of probabilistic selection used to identify candidates for reproduction, as Houck et al., Supra, discusses in greater detail. The classification is less inclined to a premature convergence caused by individuals who are far above the average. The basic idea of classification is to select solutions for the common mating bag based on the relative objective between the solutions. This modality also uses the error operation schemes of arithmetic crossover and non-uniform mutation included in the GA of Houck et al. This mode continues the search for solutions until the goal function converges. Alternatively, the search may be terminated after a number of previously determined generations. Although the conclusion due to the loss of population diversity and / or lack of improvement is efficient when crossing is the primary source of variation in a population, homogeneous populations can occur with the best (highest) target when the mutation is used. Crossing refers to the generation of new members of a population by combining the elements of many of the most suitable members. This corresponds to saving solutions in the best part of the search space. Mutations refer to the random alteration of the elements of the fittest members. This allows the algorithm to exit the search space that could be just a local maximum. Since the utility of reinitialization of populations that may have converged is proven, many GA repetitions are run until a consistent lack of increase in average aptitude is recognized. Once the potentially optimal solutions are found through the use of the GA, the most suitable GA solution can be improved by means of the local search engines. An alternative embodiment of the invention uses the simplex method to further refine the optimized GA solution. The self-cutting system can also benefit from the optimization of its parameters including the contrast value, the number of protuberances and the number of dilatations. The method to optimize the self-cutter includes the steps to manually generate breast masks for some training information, selection and initial population, and the production of breast masks for training information. The method further includes the steps for measuring the overlap percentage of the manually and automatically generated masks as well as the fraction of self-cut breast tissue outside the manually generated masks. The method also includes the selection of winning numbers, the generation of new members, and repetition in a similar manner as previously described until a previously determined objective function converges. In Figures 26 and 27, the curves of the operating characteristics of the free answer answering machine for the system of the invention can be observed for the outputs of the detector and the optimized microcalcification classifier, respectively. Figure 26 represents the operation of the optimized detector before classifying the detections, while Figure 27 represents the operation of the system after classifying the detections. Although the GA has been described above in relation to the parameter optimization portion of the preferred embodiment, other optimization techniques are suitable, such as, for example, the response surface methodology. Of course, the processors systems other than those described in this one can also be optimized by means of the methods published in it, including the GA.

INCORPORATION OF CAD SYSTEM OUTPUTS FOR AN OPTIMAL SENSITIVITY Metric operation for the detection of suspicious regions associated with cancer are frequently reported in terms of sensitivity and specificity. Sensitivity measures how well a system finds suspicious regions and is defined as the percentage of suspicious regions detected from the total number of suspicious regions in the reviewed cases. The sensitivity is defined as: ... .... TP sensitivity TP + FN (11) where TP is the number of regions reported as suspicious by a CAD system that is associated with cancer, and FN is the number of regions known to be cancerous that are not reported as suspicious. The specificity is defined as: TN specificity = (12) FP + TN where TN represents regions correctly identified as non-suspect and FP represents regions reported as suspicious that are not cancerous. Current CAD systems increase specificity by reducing FP. However, the FP and the TP are paired amounts. That is, a reduction of FP leads to the reduction of TP. This implies that some of the suspicious regions that could have been detected were missed when the objective is to maintain a high degree of specificity. Figures 28 and 29 illustrate the relationships between the quantities TP, FP, TN and FN. A measurement of the projection of an image of mammography is presented by a statistical test, x. The probability of the density function of x is represented by p (x) and the decision threshold is represented by?. If x is greater than?, A suspicious region is reported. The areas under the probability of the density functions represent the probability of events. From Figure 28 it is observed that increasing the threshold reduces the probability of FP decisions. However, in Figure 29 it is observed that simultaneously increasing the threshold reduces the decision probability of TP. Another metric that exists for CAD systems is the positive prediction value (PPV), which is defined as the probability that cancer actually exists when a region of interest is labeled as suspicious. The PPV can be calculated from the following equation: PPV = TP TP + FP (13) Note that increasing the TP or reducing the FP increases the PPV. Radiologists and computers find different suspicious regions. Figure 30 is a Venn diagram representing a possible distribution of suspicious regions for human and machine detections. Some suspicious regions are found only through a human interpreter or radiologist, some only through a CAD system, some through both, and some do not find either one. With reference to Figure 31, one can find in this a preferred method for incorporating the outputs of a CAD system, and more particularly for the CAD system of the invention, with observations of a human interpreter of the projection of a mammography image. for optimal sensitivity, wherein the radiologist examines the projection of a mammography image 10 in a step 20 and reports a set of suspicious regions 30 designated as Sl. Then the CAD system operates on the image 10 in a step 40 and reports a series of suspicious regions 50 designated as S2. The radiologist then examines the set S2 and accepts or rejects members of the set S2 as suspects in a step 60, in such a way that a third set of suspicious regions 70 is formed, designated S3, which is a subset of S2. Then, in step 80 the radiologist creates a set of work regions 90 designated S4 which is the union of the sets Sl and S3. Then the working regions 90 are recommended for additional examinations such as the taking of additional mammograms with a higher resolution, examination of the breast tissue areas corresponding to the working regions by means of ultrasound, or performing biopsies of the breast tissue. With reference to Figure 32, an alternative embodiment of the invention can be observed which includes a density detector 800, a density classifier 900, and a combiner of detection results 1000 in addition to the elements previously described. The density detector 800 detects masses of lesions that appear in a digital representation of the projection of a mammogram. The density classifier 900 classifies the densities detected by means of an MLN NN as suspicious or non-suspicious in a manner similar to the NN MLP previously described with respect to the microcalcification classifier. The classified densities detected as suspicious are fused and combined with the suspicious microcalcifications detected in the combiner 100 of detection results. While the invention has been described in relation to the detection of cluster microcalcifications in mammograms, it should be understood that the methods and systems described herein may also be applied to other medical images such as chest radiographs. While the apparatus form described herein constitutes a preferred embodiment of this invention, it should be understood that the invention is not limited to this precise form of apparatus, and that changes can be made therein without departing from the scope of the invention, which is defined in the appended claims.

Claims (23)

1. A method for the automatic detection of cluster microcalcifications of a digital mammogram comprising the steps of: obtaining a digital mammogram; optimize the first parameters to cut said digital mammogram; cutting said digital mammogram, based on said optimization of the first parameters, to produce a cropped image representative of the breast tissue in said digital mammogram; optimizing the second parameters to detect the cluster microcalcifications in said cropped image; detecting cluster microcalcifications in said cropped image based on said optimization of second parameters; indicate the position of said cluster microcalcifications detected in said digital mammogram.

2. A method for segmenting an area of a digital mammogram image corresponding to breast tissue of the rest of the image, comprising the steps of: storing a digital representation of said digital image of the mammogram; improve the histogram of the digital representation to produce an improved image by means of which the contrast of the image area of the mammogram corresponding to the breast tissue increases; setting the enhanced image to the threshold to produce a binary image comprising a seed pixel; cutting in said region said seed pixel in said binary image to produce a mask; close holes in said mask; wear out said mask; dilate said mask; and cut that digital representation to the size of the largest object in that mask.

3. A method for detecting cluster microcalcifications in a digital mammogram image comprising the steps of: filtering said digital mammogram image for the first time to reduce the noise in said image to produce an image with reduced noise; filtering said image with reduced noise for a second time with a Gaussian unit difference filter to produce a filtered DoG image in which the appearance of potential microcalcifications has been improved; placing said filtered DoG image on the global threshold to segment the potential microcalcifications of said DoG filtered image; putting said filtered DoG image in the local threshold to segment the potential microcalcifications of said DoG filtered image; join with Boolean logic the potential global and local microcalcifications placed in the threshold; convert the potential microcalcifications of Boolean logic to coordinate representations of a single pixel; eliminate the representations of coordinates of a single pixel that are outside the area of the digital mammogram image that corresponds to the breast tissue; cluster the remaining representations of coordinates of a single pixel from the previous step to identify potential cluster microcalcifications; calculate the characteristics for each of the potential cluster microcalcifications; eliminate potential cluster microcalcifications based on the characteristics calculated for each of the potential cluster microcalcifications; and indicate in the digital mammogram image the positions of those potential cluster microcalcifications remaining after the previous step.

4. A method for automatic detection of cluster microcalcifications by processing a digital image on a mammogram comprising: storing a digital representation of a mammogram; refolding a filter core comprising a difference equation of Gaussian units with said digital representation, wherein the information in the image that does not conform to the size and characteristics of the figure of a microcalcification, is suppressed and the suspicious microcalcifications appear as points bright in a back image; and placing said posterior image on the local and global threshold, where a second image is obtained that essentially comprises only areas of suspicious microcalcifications.

5. A method for the automatic detection of cluster microcalcifications by means of processing a digital image on a projected mammogram comprising: storing a digital representation of a mammogram; use and optimize an algorithm and a database of training images to obtain optimized parameter values; and applying a filter algorithm using said optimized parameters to said digital representation to obtain a filtered image comprising essentially suspect microcalcifications; group such representations of a single pixel into clusters using said optimized parameters.

6. The method of compliance with the claim 5, where the step of using an optimizing algorithm, comprises: using a genetic algorithm.

7. The method of compliance with the claim 6, where: the step of using a genetic algorithm involves obtaining a value d- ^. and a μCsmin value where dnn represents the distance to the nearest neighbor and μCsmin represents a quantity of detected microcalcifications; and the step of making clusters of single pixel representations in clusters which represent microcalcifications that are within the distance d? m of other microcalcifications μCsmin.

8. The method of compliance with the claim 7, wherein the step of using a genetic algorithm further comprises the step of: repeatedly looking for a solution space containing possible values for cL ^ and μCsmin to identify sets of values in which at least one objective function is maximized.

9. The method of compliance with the claim 8, wherein the step of using a genetic algorithm comprises the step of: using a simplex method to identify at least one of the sets of said values in which a cost function is minimized.

10. A method for the automatic detection of cluster microcalcifications by means of the process of a digital image in the projected mammography, which comprises: storing the digital representation of a mammogram; locate potential clusters of microcalcifications in said digital representation; extract characteristics of said potential clusters of microcalcifications; use said extracted characteristics as inputs to a neural network of the artificial multi-layer perceptron; and using said neural network of the artificial multi-layer perceptron to classify said clusters of microcalcifications as suspect or non-suspect. The method according to claim 10, wherein the step of using said neural network of the artificial multi-layer perceptron to classify said clusters of microcalcifications as suspect or non-suspect comprises: using output values of a neural network of the artificial perceptron of multiple layers in a slightly variant output function to obtain a series of resulting values; and classifying a cluster associated with one of said resulting values as suspect if that value is greater than or equal to the threshold value, or as non-suspect if that value is less than or equal to the threshold value. The method according to claim 11, wherein said slightly varying output function comprises a hyperbolic tangent function. The method according to claim 11, wherein said slightly varying output function comprises a linear function. The method according to claim 11, wherein said slightly varying output function comprises a sigmoidal function. 15. The method of compliance with the claim 11, wherein the step of using said neural network of the artificial multi-layer perceptron to classify said clusters of microcalcifications as suspect or non-suspect comprises: multiplying at least one of said characteristics by a weight wj, where I is an index which represents the element of the characteristic vector i th of a characteristic vector x having elements N, and is an index representing the protrusion j of the first layer; and using protrusions of the first layer of said neural network of the artificial multi-layer perceptron to calculate the outputs fj that are calculated in accordance with the function: wherein x comprises a calculated characteristic vector element. 16. The method of compliance with the claim 15, wherein the step of using said neural network of the multi-layer artificial perceptron to classify said clusters of microcalcifications as suspect or non-suspect further comprises: multiplying at least one of said outputs fj of the first layer by a second weight Uj r k; and using a result of the multiplier step as an input to at least one output protrusion, calculating the output of at least said output protrusion in accordance with the function: eo where y, - = fj (x), k is an index that represents the output protrusion ktIl, and J is the number of outputs of the first layer that have to be multiplied. 17. A method for incorporating the output detections of a computer-aided detection system to detect cluster microcalcifications on a mammogram with detections on the same mammogram observed by a human interpreter without reducing sensitivity, comprising: obtaining said detections observed for form a first set of detections; obtain said observed detections to form a second set of detections; accept some output detections in the second set to form a third set of detections; combine the first set and the third set to form a fourth set of detections; and provide an output based on the fourth set of detections. 18. A method for automatically detecting cluster microcalcifications by processing a digital image on a projected mammogram, comprising: storing a digital representation of a mammogram; filtering said digital representation to obtain a filtered image comprising the suspicious microcalcifications; to put on the threshold said filtered image with a local threshold inclined to obtain an image that essentially comprises only areas of suspicious microcalcifications. The method according to claim 18, wherein the step of placing on the threshold comprises: centering a window on the pixel of interest having the coordinates (x, y) in said digital representation; calculate the mean (x, y) and standard deviations (x, y) of the pixels under said window, calculate a local threshold value T (x, y) for the pixel of interest in accordance with the function: T (x, y) = A + Bμ (x, y) + Cs (x.y) where A is a previously determined equivalent and B and C are previously determined coefficients; comparing the gray scale value of the corresponding pixel of interest in said filtered image to said local threshold value; and generating a binary image by placing a pixel corresponding to said binary image to one if the gray scale value is greater than or equal to the local threshold, and to zero if the gray scale value is less than the local threshold. 20. An apparatus for the automatic detection of cluster microcalcifications of a digital mammogram comprising: an element for obtaining a digital mammogram an element for optimizing the first parameters to cut the digital mammogram; an element to cut the digital mammogram. Said cutting means utilizing said first optimized parameters to produce a cropped image representative of the breast tissue in the digital mammogram; an element for optimizing the second parameters for detecting cluster microcalcifications in said cropped image; an element for detecting the cluster microcalcifications in said cropped image using said second optimized parameters; an element to indicate the positions of said cluster microcalcifications detected in the digital mammogram. 21. An apparatus for detecting cluster microcalcifications in a digital mammogram image comprising: a Gaussian unit difference filter to produce an image filtered by DoG in which the appearance of potential microcalcifications has been improved; an element for thresholding to segment said potential microcalcifications of said image filtered by DoG; extracting an element to generate coordinate representations of a single pixel for said potential microcalcifications; and an element for making clusters of groups of said representations of a single pixel. 22. An apparatus for the automatic detection of cluster microcalcifications by means of processing a digital image on a projected mammogram comprising: an element for storing a digital representation of a mammogram; an element for refolding the core of a filter comprising an equation of Gaussian units difference with said digital representation, where the information in the image that does not conform to the size and form of a microcalcification is suppressed and the suspicious microcalcifications appear as points bright in a resulting image; and an element for bringing said resulting image into the local and global threshold, from which a second resulting image is obtained which essentially comprises only areas of suspicious microcalcifications. 23. An apparatus for the automatic detection of cluster microcalcifications by means of processing a digital image on a projected mammogram comprising: an element for storing a digital representation of a mammogram; an element to optimize the values of the parameters, using an optimization algorithm and a database of training images; an element for applying a filtering algorithm, using said optimized parameters, to said digital representation to obtain a filtered image comprising essentially suspicious microcalcifications; an element for shrinking said filtered image to obtain an image that essentially comprises representations of a single pixel of said suspicious microcalcifications; and an element for making groups of said representations of a single pixel in clusters, using said optimized parameters. 2 . The apparatus in accordance with the claim 23 where the optimizing algorithm is a genetic algorithm. 25. The apparatus in accordance with the claim 24 wherein that element for optimizing parameter values comprises: an element for obtaining a value d ^ - and a value μCsmin where d ^ represents the distance to the nearest neighbor and μCsmin represents a quantity of microcalcifications detected; and where the element for making groups of representations of a single pixel in clusters which represent microcalcifications that are within the distance d ^ of other microcalcifications μCsmin. 26. The apparatus according to claim 25 wherein said element for optimizing parameters comprises: an element for repeatedly searching for a solution space containing possible values for dnn and μCsmin to identify sets of values in which at least one function is maximized of objective. 27. The apparatus in accordance with the claim 26 wherein that element for optimizing parameters comprises: an element for using a simplex method to identify at least one of said sets of values in which a function cost is minimized. 28. An apparatus for the automatic detection of cluster microcalcifications by means of the process of a digital image in the projected mammography, comprising: an element for storing the digital representation of a mammogram; an element for locating potential clusters of microcalcifications in said digital representation; an element to extract characteristics of said potential clusters of microcalcifications; an element to classify said potential clusters of microcalcifications as suspicious or non-suspicious using said characteristics. 29. The apparatus in accordance with the claim 28 wherein that element for classifying comprises a neural network of the artificial multi-layer perceptron. 30. The apparatus in accordance with the claim 29 wherein said neural network of the artificial multi-layer perceptron comprises: a slightly variant output element. 31. The apparatus in accordance with the claim 30 wherein the slightly varying output means comprises: an element for applying a hyperbolic tangent function to a sum of heavy inputs to provide an output indicating whether the microcalcification is suspicious or not suspicious. 32. The apparatus according to claim 30, wherein the slightly varying output means comprises: an element for applying a sigmoidal function to a sum of heavy inputs to provide an output indicating whether the microcalcification is suspicious or not suspicious. 33. The apparatus according to claim 30, wherein the slightly variant output element comprises: an element for applying a linear function to a sum of heavy inputs to provide an output indicating whether the microcalcification is suspicious or not suspicious. 34. The apparatus according to claim 30 wherein said neural network of the multi-layer artificial perceptron comprises: an element for multiplying at least one of. said characteristics by a weight w j, where i is an index representing the element of the characteristic vector ith of a characteristic vector x having elements N and is an index representing a protrusion jth of the first layer; and protuberances of the first layer to calculate the outputs /, which are calculated according to the function: where x ± comprises a calculated characteristic vector element. 35. The apparatus in accordance with the claim 34 wherein said neural network of the multi-layer artificial perceptron comprises: an element for multiplying at least one of said outputs of the first layer fj, by a second weight Uj ^ k; and at least one exit protrusion to calculate the outputs zk where k is an index representing that represents an exit protrusion kth, in accordance with the function: where y ^ = j (x) and J is the number of outputs of the first layer that multiply. 36. A method for automatically detecting cluster microcalcifications and density by processing an image on a projected mammogram, comprising: storing the digital representation of a mammogram; refolding a filter core comprising an equation of Gaussian units difference with said digital representation, by means of which the information in the image that does not conform to the size and characteristics of the figure of a microcalcification is suppressed, and the suspicious microcalcifications appear as bright spots in a subsequent image; placing said posterior image on the threshold, by means of which a second image is obtained essentially comprising only areas of suspicious microcalcifications; detect densities in said digital representation; And produce a third resulting image comprising only areas of suspicious densities. 37. A method for automatically cutting through the process of a digital image on a projected mammogram, comprising: storing a first digital representation of a mammogram; dilate said first digital representation to produce a dilated digital representation; cutting said dilated digital representation to the size of the largest contiguous group of pixels, by means of which a shortened digital representation is produced, and selecting for the subsequent process the pixels in said first digital representation which corresponds to the pixels in said representation digital dilated cropped. 38. A method for the automatic detection of microcalcifications by means of processing a digital image on a projected mammogram comprising: storing a digital representation of a mammogram; normalize the brightness values of said digital representation; using a contrast setting previously determined in an algorithm of the growth region to create a breast mask which defines an area of said digital representation containing breast tissue; and look for microcalcifications in said area that contains breast tissue. 39. The method of compliance with the claim 38, wherein the normalization step comprises a step of equalizing histogram. 40. A method for the automatic detection of microcalcifications by means of processing a digital image on a projected mammogram comprising: storing a digital representation of a mammogram; look for a region of that digital representation to find a seed pixel; use a growth algorithm per region to create a breast mask which defines an area of said digital representation that contains breast tissue, the step of using an algorithm in the growth region further comprises: using a function of growth region the which starts at the seed pixel and progressively makes groups of the closest neighboring pixels which share a characteristic, by means of which a contiguous group of pixels is created; using the contiguous group of pixels to determine a breast mask which defines an area of said digital representation containing breast tissue; and look for microcalcifications in that area that contains breast tissue. 41. The method according to claim 40, wherein the step of searching a region of said digital representation to find a seed pixel comprises finding a pixel within said region which has a maximum gray scale value. 42. A method according to claim 40, wherein the step of making groups of the closest neighboring pixels that share a characteristic comprises determining whether each adjacent pixel has a contrast ratio that is less than the threshold of the ratio of contrast. 43. A method for the automatic detection of microcalcifications by processing a digital image on a projected mammogram comprising: storing a digital representation of a mammogram; calculate the value of the local threshold of the pixels in said digital representation; filter said digital representation using a Gaussian unit difference filter to create a filtered image; creating an image set at the threshold by comparing the pixels in said filtered image for the local threshold value and adjusting the values of the pixels in accordance with a result of said comparison; and further processing said image placed on the threshold to identify the microcalcifications. 44. A method for the automatic detection of microcalcifications by means of processing a digital image on a projected mammogram comprising: storing a digital representation of a mammogram; use a genetic algorithm and a database of training images to obtain at least one parameter; and using the at least one said parameter in a function which processes said digital representation to identify the microcalcifications. 45. The method according to the claim 44, where the at least one parameter comprises a value cL ^. and a μCsmin value where d ^ represents the distance to the nearest neighbor and Csmin represents a quantity of detected microcalcifications; and wherein said function comprises making groups of single pixel representations in clusters which represent the microcalcifications that are within the distance dnn of other microcalcifications ri'-stan-46. The method according to claim 44, in where the at least one parameter comprises a target size used in a Gaussian unit difference filter. 47. The method according to claim 44, wherein the at least one parameter comprises at least one threshold value. 48. The method according to claim 44, wherein the at least one parameter comprises at least one neighboring value. 49. The method according to claim 44, wherein said function comprises a local inclined threshold function. 50. The method according to claim 44, wherein the at least one parameter comprises a percentage of a histogram. 51. The method according to claim 47, wherein the at least one threshold value comprises upper and lower threshold values.