JP2003502129A - Computer aided detection method for lumps and clustered microcalcifications - Google Patents

Abstract

SUMMARY The present invention is a computer assisted detection method and system for assisting a radiologist in reading a medical image (1000). The method and system are particularly applied to mammographic regions with clustered microcalcification (3000) and density detection. Microcalcification detection is provided, where the individual detections are ordered and classified (5000) by grade, and one of the classification features is obtained using a multilayer perceptron.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD OF THE INVENTION

The present invention relates to methods and systems for computer-aided analysis of detection on digital images, and more particularly to methods and systems for assisting radiologists in reading medical images.

[0002]

[Prior art]

Computer-aided detection for mammography
The ded detection (CAD) system has been advocated for decades. The purpose of any medical CAD system is to indicate areas of tissue that require medical attention from a physician without damaging normal tissue. Current systems use one, and possibly two, image inputs to the detector. If two image inputs are input to the detector, the method settles in one of two ways.

The first method is known as bilateral subtraction. In this method, the left and right chest images are subtracted to produce a difference image. Since cancer is usually unilateral and the human body is relatively symmetrical, difference images often highlight cancer areas. The second method is similar to two-sided subtraction. However, in this case a temporal subtraction is performed. The same image of the previously taken body part is subtracted from the current image. This temporal subtraction is intended to highlight areas that have changed during the time interval of the imaging visit.

Both methods can increase the sensitivity of the detector, but suffer from two problems. First, subtraction increases the number of false displays. Second, the images involved in the subtraction process must first be aligned. Because the tissue is elastic,
Alignment is not trivial. Therefore, the alignment method is
Rotation, scale, and space dependent stretching need to be compensated.

[0005]

[Problems to be Solved by the Invention]

The method in the present invention provides a significant improvement over one image-based detection method. Further, the CAD system described in this application leverages the information available from the entire set of images associated with the screening mammography practice of the patient without the need to subtract or align the images.

The present invention provides computer-aided detection methods and systems for assisting radiologists in reading medical images.

[0007]

[Means for Solving the Problems]

In the first step, a set of digital mammogram images, such as those from a mammogram obtained from a screening round of a patient, is obtained. A rectangular analysis region containing breast tissue is segmented from each digital mammogram image and a binary mask corresponding to the breast tissue is created.

[0008] Microcalcifications clusters are detected in the microcalcification detection stage. The point features from each microcalcification are applied to a classifier trained to pass the microcalcifications that show malignant clusters. These remaining microcalcifications after the classifier are indicated by their centroid location and are later grouped into clusters. Features are calculated from the clusters, and each is subsequently assigned a class label and a quality score.

Densities are detected in the density detection stage. At this stage, the sub-sampled image is applied to the DoG filter bank, which results in an image with locally bright and highlighted spots. This image is then evaluated by a threshold and the threshold evaluated image is input to the screening classifier. Detections that pass the screening classifier as suspicious are input to the region growing stage and concentration features are calculated from the region grown region.
The detections are input to the concentration classification stage, the output of which has the calculated features, class labels, and two quality scores for each detection.

Detections involving both microcalcifications and concentrations are further analyzed in the post-treatment stage. In the first post-processing stage, the detection across all images in the case is considered collectively by case-based grading. In a second post-processing stage, within each image, the set of detections across the detection categories is assigned to one of 29 image context categories based on the associated quality score. If a certain image context category for density is found on an image, then all detections on that image are eliminated. In the third post-processing stage, the remaining detections across all images in the case are considered to be within the category of microcalcification and concentration detection. Within each detection category, the "normality" characteristic is calculated using the quality score from all detections in cases from that category. These normality features are
Applies to classification stages designed to be assigned as “normal” or “non-normal.” When the classifier assigns a case as “normal,” all of the corresponding categories (microcalcifications or concentrations) Is detected from the image in the case, and the final output of the system is the set of displays that are overlaid and displayed on the set of digital images.

It should be noted that the method and system of the present invention is in contrast to the subtraction-based method proposed by the prior art. That is, while such prior art methods are essentially constrained to two input images, the present invention forms a patient case.
In that the analysis of detection extends over a set of images of the patient, such as a set of three images. By leveraging the information available from the entire set of images associated with patient screening mammography practice, the method of the present invention obtains both increased system sensitivity and reduced false positive detection. Bringing significant improvements in achieving the objectives.

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

[0013]

DETAILED DESCRIPTION OF THE INVENTION

Referring now to the drawings (and in particular to FIG. 1), the sequence of steps performed to detect the location of microcalcification clusters within a set of digital mammograms obtained from screen diagrams associated with screening mammography practice of a patient is described. A schematic flow chart is shown, in which like reference numerals refer to the same or corresponding parts throughout the several views.

In a first step 1000, a set of digital mammograms is digitized by using a hardware such as a digital mammography system or by using a laser or charge coupled device (CCD) digitizer. ,can get.

In the preprocessing stage 2000, a rectangular analysis region containing breast tissue is segmented from the digital mammogram image, which reduces the time required to process the mammogram image. A binary mask corresponding to breast tissue ("breast" mask) is created for use in later processing steps. Chest masks are used to limit detection to image areas that encompass breast tissue. The preprocessing stage 2000 further produces a subsampled input image. The sub-sampled image is input to subsequent processing stages. Due to the significantly smaller number of processed pixels in the subsampled image, the time required to analyze the subsampled image is much less than that required for the original image. The pre-treatment stage consists of a binary peculiar mask (binary pec
create a torral mask). The pectoral mask can be used to suppress detection in the pectoral region.

Clustered microcalcifications are detected in the microcalcification detection stage 3000. After filtering the cropped image with a DoG (difference of Gaussians) filter to enhance microcalcification, the resulting image is
It is evaluated by a threshold to detect microcalcifications. Besides the detection in the chest mask, the detection near the chest edges is eliminated and the "point" features are calculated from the individual microcalcification detections. Point features from each microcalcification are trained to pass through the classifier (clas
sifier) is applied. These microcalcifications remaining after the classifier are indicated by their centroid location and are clustered after this. The “cluster” features are calculated from the microcalcification clusters at step 4000. The detected clusters are classified into class labels '1' or '0' (suspicious,
Or any of the ones that are not in doubt). The output of the microcalcification classification stage includes calculated features, class labels, and quality scores for each detected region.

The concentration is detected in the concentration detection stage 6000. The input to this stage is the subsampled image from the preprocessing stage. After initially processing the image to reduce false detections associated with skin edge boundaries, locally bright densities are enhanced by a trend correction. The skew-corrected image is then applied to the DoG filterbank. The outputs of the filterbanks are non-linearly combined to produce an image with further enhanced densities. The image is threshold evaluated in the peak detection stage, the resulting pixel groups are grouped into concatenated objects, and a chest mask is applied. These object and image pixels from the original image and the filterbank output image are used to generate the "density screening" feature. These concentration screening features are input to a concentration screening classifier used to classify detected subjects as suspicious or non-suspicious. Subjects that pass the screening classifier as suspicious are input to the region growing stage which evaluates the shape of the detected local bright spots.

In step 7000, density features are calculated from the region growing region using the subsampled input image and the filterbank output. The expert filter stage removes certain detections by comparing the density features from the region growing region with a threshold. Detections with feature values below the threshold are excluded from further consideration. After the expert filter,
For the detection, in the concentration classification step 8000, the class label "1" or "0" (
Suspicious or non-suspicious) will be assigned. The output of the concentration classification stage is the calculated features for each detected region, the class label, and
And two quality scores corresponding to the categories "lumps" and "spiculated mass". The term "concentration" applies to two types of agglomerates.

The class labels and quality scores for the detected microcalcifications and concentration areas are further analyzed in the post-processing stage 9000. During this stage,
Detection is considered in two ways, radically different from the previous stages (across all images in the case, and across the microcalcification and concentration detection categories). Only detections with class label '1' at the end of the post-processing stage are displayed in the output of the system.

In the first post-processing stage, the detection across all images in the case is considered collectively by case-based grading. Case-based grading refers to the number of microcalcification and concentration detections with class label '1', using a quality score associated with each detection, and a predetermined maximum for each case and for each image. Limit to a number.

In a second post-processing stage, within each image, the set of detections across the microcalcification and concentration detection categories is based on an associated quality score of 2
Assigned to one of nine image situation combinations. When a combination of image contexts for detection is found on an image, all detections in that image set their class label to '0'.

In the third post-processing stage, all detections across all images in the case with a class label of '1' or '0' are considered within the microcalcification and concentration detection category. Within each detection category, "normality (
normalcy) "features are calculated using quality scores from all detections in cases from that category. These normality features are intended to assign the entire case as" normal "or" not normal ". If the classifier assigns a case as “normal”, all detections for the corresponding category (microcalcification or concentration) will have their class '0' over all images in the case. Set to.

In the final step 10000, the detection with class label '1' is displayed as "display" overlaid on the set of digital images.

Classifier Design The present invention uses a pattern recognition subsystem throughout the processing stages. This section gives an overview on the areas of pattern recognition and the design of classifiers when these areas are applied to the present invention.

Pattern recognition is the process of marking decisions based on measured values. In this system, suspicious areas or detections are located by the detector and finally accepted or rejected for display. For maximum sensitivity,
The detector is designed to locate many areas that may be malignant. This leads to the detection of many areas that are normal. The purpose of the pattern recognition subsystem is to eliminate as many false detections as possible while retaining as many truly malicious detections as possible at the same time. The detection corresponding to the cancer area is “true positive (true p
"Ositive)" detection, or simply TP. Similarly, the detection corresponding to normal regions is known as "false positive" detection, or FP.

The first step in the process is to characterize the detection. For this purpose, multiple measurements are calculated from each detected area. Each measurement is referred to as a feature. The set of measurements for this detected region is called its feature vector, where each element of the vector is
Indicates the feature value.

The second step in the process is to use the information in the feature vector to assign a quality score, or possibly a class label, to the detection. The feature vector from the detector is provided as an input to the classifier, which classifier
Compute a number for a given feature vector input. One way to identify a classifier is by a set of discriminant functions.

The discriminant function calculates one value as a function of the input feature vector. Referring to FIG. 2, a classifier having a feature vector x applied to a set of discriminant functions {g _i (x)} (i = 1, 2) is shown. The classifier shown in FIG. 2 uses one discriminant function for each class. The output value of the discriminant function is the discriminant
Is called. The discriminant is subtracted and compared with the threshold θ. If the discriminant difference is greater than the threshold, the class label is set to '1', otherwise the class label is set to '0'.

Many methods are valid for planning the discriminant function. In the preferred embodiment of the invention, the method for planning the discriminant function is a statistical method that yields a quadratic discriminant function. Under certain general conditions, a quadratic discriminant function can yield the minimum mean error probability. The design of the quadratic discriminant function begins with the classified training data.

The classified training data was generated in the following way. Mammographic images were examined by a licensed radiologist, which confirmed the cancerous area based on a pathological analysis of cells from the biopsied cancerous area. The true data was generated by the radiologist marking the truth boxes on the corresponding digital image areas associated with the cancer. Cancer categories were identified that corresponded to microcalcifications and "lumps" and "spiculated lumps". These images were then processed to yield a plurality of feature vectors, the subset of feature vectors associated with the truth boxes. The feature vector from the detection in the truth box receives a true label of '1', while the outer detection is classified as '0'.

In the present invention, these classes are ultimately cancer and not cancer. For each class to be identified, a mean and covariance matrix is calculated from the classified feature vector of the training data.
The discriminant function of the quadratic classifier is

[Equation 1] It is given by the formula. Where X is the feature vector, M _i is the mean of class i, and Σ _i is the covariance matrix of class i. The quantity β _i is a bias term that can be modified to fine-tune the decision process.

Since the discriminant function G _i (x) is a distance function, a small discriminant value indicates that the observation has a high probability of belonging to class i. The difference of the calculated discriminant,

[Equation 2] To consider as. Here, the discriminant difference is the delta discriminant
Is called. It is assumed that a class that is not cancer is class i = 0 and a class that is cancer is i ≠ 0. Next, if Delta _i is greater than zero, it indicates that the feature vector X is closer to the cancer category than the non-cancer category. Therefore, for any i ≠ 0, De
If lta _i > 0, the observation is considered cancerous. Furthermore, the larger the discriminant difference, the higher the quality of detection can be considered. For this reason, the discriminant difference can be used as the quality score.

In the quadratic classifier, the feature vector X has a Gaussian statistical distribution (Gaussian statistical dis).
tribution) is the optimal classification technique. If the distribution of individual measurements is in some other form of the gamma distribution, then the power transformation of the feature values (power
transformation) can change the distribution in order to more closely approximate the Gaussian distribution. This individual feature values is performed by exponentiating only 0.4 square (y = x ^{0. 4).}

Pre-Processing As can be seen in FIG. 3, in the pre-processing stage 2000 (FIG. 1), the digital mammogram image 201 is first low-pass filtered in stage 2010 and thereafter in horizontal and vertical direction in stage 2020. 1/1 in the direction
Decimated to 6. The above operation yields a low resolution image 2070 suitable for subsequent stages that do not require full spatial resolution. The low resolution image is
It is called a sub-sampled image. The sub-sampled image is input to the autocropper 2030.

The automatic cropping device identifies the location of breast tissue in the image by a chest mask (
Produces a binary mask 2040 known as a Breast Mask). Focusing the subsequent image processing on only the portion of the image that includes the breast tissue reduces the time required to process the image. 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.

Automatic clipping is described in US Pat. No. 5,995, which is hereby incorporated by reference.
It is described in detail in 9,639. In the preferred embodiment, the automatic cropper creates a Crop Rectangle 2050. The crop rectangle is the rectangular area that exactly contains the chest mask.

A further output of the pre-processing stage is the Pectoral Mask 2060.
The pectoral muscle mask is a binary mask that identifies the position of pectoral muscle tissue in the image, and methods for forming the pectoral muscle mask are known in the art. It is desirable to interfere with these detections in the pectoral muscle, because cancer occurs rarely in the pectoral muscle. In addition, this part of the digital mammogram is usually brighter than other areas, which can detect false positives. Therefore, interfering with detection in the pectoral muscle improves the performance of the system of the present invention.

<Detection of Clustered Microcalcifications> Turning now to FIGS. 4 and 5 (FIG. 5 is a continuation of FIG. 4), in these figures, the clustered microcalcifications of the present invention are shown. A flow chart is shown detailing the detection of calcification and the computational scheme 3000, 4000 for clustered microcalcification features. Digital image 201 and crop rectangle 2050 are used to highlight sub-image areas with localized bright spots in the microcalcification size range.
Input to the DoG filtering stage 3010, which produces a local maximum in the filtered image. Next, in step 3020,
The DoG filtered images are thresholded to segment the maxima that indicate the detection of microcalcifications. A coordinate representation of the detected maximum is calculated by the binary chest mask 2040 and pectoral muscle mask 2060, which results in step 3030.
At, false detections outside the breast tissue and within the breast tissue but near the chest edge are eliminated. The remaining coordinate representations in the analysis area are affected by the point feature calculation step 3040, where the measurements from the individual microcalcifications are calculated. These point features are input to a secondary classification stage 3050 intended to retain up to a predetermined number of detections that best represent the individual microcalcification features associated with cancer lesions. The retained detections are clustered at step 3060. The features of the microcalcification cluster are calculated for the cluster in the feature calculation stage 4000, which results in the feature vector of the microcalcification cluster and the position 4100 of the microcalcification cluster.

A more detailed description of the clustered microcalcification detection process steps follows. The digital image is first filtered by a DoG Kernel to enhance microcalcification. The DoG kernel is defined on the values of an N × N square array. In the preferred embodiment, the kernel is 77
It is x77. Filtering is done by convolving the input image with a DoG kernel. In another embodiment, the filtering is to first obtain a Fast Fourier Transform (FFT) of the input image and the DoG kernel, then multiply these FFTs together and, as a result, obtain an inverse FFT. Done by.

The DoG kernel is applied here as an optimal detector of individual microcalcifications, as described in US Pat. No. 5,999,639. DoG
The kernel is mathematically

[Equation 3] Expressed as Where σ ₁ and σ ₂ are the standard deviations of the individual Gaussian kernels, and c ₁ and c ₂ are the normalizing consta for the Gaussian kernels.
nt).

Optimal values for σ ₁ are detailed in US Pat. No. 5,999,639. When σ ₂ = 1.6σ ₁ , the DoG filter response closely matches the response of the human spatial receptive filter.
Therefore, the ratio of the DoG standard deviation constant is set to 1 to 1.6 by motivation from human physiological functions. Thus, for targets with an average diameter of t pixels, we use σ ₂ = t / 2. From the rule of thumb, σ ₁ = σ ₂ /1.6.

After DoG filtering to highlight the target against the background, the filtered image has a gray level difference between the microcalcifications and the background. Microcalcifications tend to be one of the brightest objects in a DoG filtered image, but microcalcifications may lie in regions of high average gray level, and this ensures segmentation. Can be difficult to do.
The threshold evaluation process used in one embodiment of the present invention to address these issues in general includes the locally graded thresholds detailed in US Pat. No. 5,999,639.

[0043]   This threshold evaluation is, in the neighborhood of the set of detections,

[Equation 4] Calculated as

Pixel values in the DoG filtered image to be displayed are defined as d (x, y)
To consider as. Each pixel is compared to a threshold T (x, y). The result of the threshold evaluation is the locally thresholded image l_s (x, y). l_s (x,
Pixels in y) are set to '1' (where the value of d (x, y) is T (
x, y)) and in other cases set to '0'. This gives rise to objects in the thresholded image. The predetermined number of detections is '1' in the target
Is held based on the difference between the brightest image pixel at the location and the threshold at that location.

After the threshold evaluation, the detection is converted to a one-pixel representation by calculating the centroid (or center of gravity) of the object found by the threshold evaluation process. Thereby, the detection is shown as one pixel with a logical value of "1", while the remaining pixels have a value of logical value "0". Subjects with a center of gravity less than a predetermined distance from the chest edge are rejected because they reduce the detection of false positives that occur due to discontinuities in brightness values between the chest and the background.

Calcification associated with malignancy has certain characteristics in common.
These properties can be exploited while distinguishing from regions of likely microcalcifications representing cancerous regions from insignificant microcalcifications or false detections. Separation of individual microcalcification detections is performed by first calculating the point features from each detection and then the subsequent point feature vector to a point classifier (
point classifier). A preferred embodiment for the classifier is a secondary classifier.

The number of false positive microcalcification detections is reduced by the point classifier, as described below. Point features are extracted for each microcalcification at location (x, y) as described above. In the preferred embodiment, the three point features calculated for each microcalcification are as follows: Response of DoG kernel (N × N) with target size 13 (t = 13) centered on (x, y), where N = 77 in the preferred embodiment.
. 2. Local average of a DoG filtered image of target size 7 (t = 7) around (x, y) (wherein in the preferred embodiment the local region is a square region of 77 × 77 pixels). .. 3. Non-linear combination of the following measurements: a) Response of a DoG kernel (N × N) with a target size of 7 (t = 7) centered on (x, y), where N = 77 in the preferred embodiment. B) a local average of the original image around (x, y) (wherein in the preferred embodiment the local area is a square area of 77 × 77 pixels); c) (x, local standard deviation of the original input image around y) (wherein in the preferred embodiment, the local area is a square area of 77 × 77 pixels).

The non-linear combination of the point feature 3 is performed by a multi-layer perceptron (MLP) artificial neural network having five input nodes and two output nodes. Nodes in the first layer use the hyperbolic tangent function, and output nodes use a linear function.

MLPs have been trained using backpropagation learning rules. The desired response of the first output node is the cluster truth box (
1 for detection of microcalcifications inside the cluster truth box), and zero otherwise. The desired response of the second output node is 1 for detection of microcalcifications outside the cluster truth box, and zero otherwise.

The third point feature is calculated as follows. First, the aforementioned measured values 3a, 3b, 3c are calculated and stored in vector form. Before applying to MLP, the input vector is normalized to zero mean and unit variance, which provides robustness over the range of input data variation. The normalized input vector is applied to the MLP and the value obtained at each output node. The value of the second output node is subtracted from the value of the first output node to yield the third point feature.

The microcalcification point classifier uses two classes of secondary classifiers. The two classes are normal and malignant. The key features of each detection are:
Tested by a classifier giving normal and malignant discriminant values. If the discriminant difference is negative, the detection is likely to be normal rather than malignant. Therefore, this point is rejected from further analysis. However, if the discriminant difference is positive, the detection is retained for further analysis and the associated point features and discriminant difference are stored for further processing.

The remaining detections are ranked ordered from largest to smallest by the difference in these discriminants. If these grades are greater than a predetermined number (50 in the preferred embodiment), they are rejected from further analysis. This operation keeps only the detection of microcalcifications that are most likely to be malignant. These remaining detections are applied to subsequent clustering steps.

To be considered a cluster, a set of microcalcification detections needs to meet two requirements. First, all microcalcifications within a cluster must be within a certain distance of other microcalcifications within the cluster. Second, there must be at least a minimum number of microcalcifications within the cluster. Therefore, some detections can be dropped if the set of detections from the point classifier is clustered. For example, isolated microcalcifications that are more than a predetermined maximum distance from other microcalcifications do not survive the clustering stage.

<Characteristics of Microcalcification Clusters> False-positive microcalcification clusters are reduced by a classifier described later. The symbol "PT" in the feature description indicates that the power transformation is applied to the measurement. The characteristics of the seven microcalcification clusters are calculated for each microcalcification cluster and in the preferred embodiment: Linear concentration-calculated as the number of microcalcifications detected divided by the maximum distance between points (PT). 2. Mean DoG value-the average value of DoG filtered image pixels with a target size of 7 (t = 7) calculated over the pixel positions of the detected cluster. 3. Average DoG Size-The average size (PT) of each target in the DoG filtered image of threshold-evaluated target size 7 (t = 7) with the centroid corresponding to the pixel position of the detected cluster. 4. Mean Gabor Value-The mean value that results from applying a set of pixels from the digital image corresponding to the pixel locations of the cluster detected through a bank of Gabor filters. The averaging is over the maximum for all N orientations of the filter in the Gabor filter bank. Gabor kernel is

[Equation 5] Where x is the coordinate of the pixel under the kernel, T is the period of the sinusoid, θ is the azimuth angle of the sinusoid, w is the width of the Gaussian function, and φ is the sinusoid. (0 = sine, π / 2 = cosine). In this embodiment, the Gabor filter is 0 ° to 1 in the M × M array.
It is defined at 24 azimuth angles between 80 °. The period T is found as the kernel size divided by twice the target size. In this embodiment, the kernel size corresponds to a 32 × 32 pixel neighborhood and the target size is 7.0 pixels. The width of the Gaussian function is equal to 1.0.
The phase of the sinusoid is equal to π / 2 (PT) to better respond to the local bright spot. 5. Three classifier classifier values-the points within the cluster are ordered by class based on the score of the discriminant difference from the point classification stage,
And the score of the M-th largest discriminant difference is selected as the feature value for the cluster, where in the preferred embodiment M = 3 (PT). 6. Average Intensity Ratio-This feature determines the position of a point as a seed pixel.
) Is used to compute the region growth of each detection in the cluster. The average intensity of image pixels inside each region growth region is divided by the average intensity of image pixels outside the region growth region within a rectangular neighborhood of a given size. The intensity ratios for all detections in the cluster are summed and then divided by the number of points in the cluster, which gives the average intensity ratio feature (PT).
. 7. Mean Second Central Moment-This feature is calculated by region growing each detection in the cluster using the point positions as seed pixels. A rectangular neighborhood area of a predetermined size centered on the point position of the input image is used after being normalized by the energy in the neighborhood area. Then, the second central moment of the energy-normalized region growing portion of the image is calculated. The second central moments are averaged, which gives this feature.

[0055]   The specific second central moment used in this embodiment is

[Equation 6] Is defined as Here, x-("x-" indicates a character with "-" on "x") is the x-component of the centroid of the region-grown detection.

These microcalcification cluster features are applied to the microcalcification cluster classifier in step 5000.

<Density Detection> Density detection is the process of locating the area of the mammography image associated with the malignant mass. Detecting concentrations is considered more difficult than detecting microcalcifications for several reasons. Significant microcalcifications, which are small, but range in size over relatively small intervals, are usually very bright and have high contrast within most breast tissue. However, important densities can have very large size variations (0-2.5 cm) and have low contrast. Furthermore, microcalcifications have a clear boundary with breast tissue, while the boundary between concentration and background tissue may be unclear.

Two major factors used by radiologists to identify suspect concentrations are lesion demarcation and lesion shape. Unclear boundaries indicate malignancy. Damage with spicules is more likely to be malignant than oval-shaped damage. Keggelmeyer, US Pat. No. 5,633.
, 948 and US Pat. No. 5,661,820, a detector was designed specifically for damage with spicules.

Turning now to FIGS. 6 and 7 (FIG. 7 is a continuation of FIG. 6), these drawings show the concentration detection of the present invention and the concentration feature calculation system 6000, 7000 in detail. A flow chart is shown. Subsampled digital image 2070
And the chest mask 2040 is input to the pre-processing stage 6010. At this stage, the border between the chest and the background is smoothed to reduce the number of false detections that occur only due to the large contrast between the chest tissue and the film background. In step 6020, the output image from the breast tissue processing step is deskewed so that local values of image intensity are evaluated in the neighborhood and then
Subtracted from this neighborhood. This leads to an increased contrast and thereby allows the detection of lesions with low contrast in the original image. The resulting image is then input to a MaxDoG filter bank at step 6030, the output of which has locally bright and generally circular (0.5-2.5 cm diameter) salient regions. It is an image. The peak in the MaxDoG image is detected at step 6040. The peak is MaxDo
Corresponds to a pixel location that is a locally bright area in the G image. These positions are grouped into connected objects in step 6050, and these connected objects are the first detections of the concentration detector. Screening features from these connected objects are calculated at step 6050 and are subsequently input to screening classification step 6070. The main purpose of the screening classifier is to reduce the number of initial detections before more detailed subsequent analysis, while preserving malignant areas for further consideration. The detections that pass the screening classifier are input to the region growing stage 6080. Region growth gives a more accurate assessment of the shape of the lesion. Region growth is a critical step in the present invention. Many methods have been described for region growth in mammography, implying a difficult problem. The region growing method described in the present invention is very general and adaptive weighting at each iteration of growth,
And state-of-the-art in two main areas: classifier-based selection of optimal region grown masks designed with a database of hand-truthed damage by humans. Make progress. The region grown object is analyzed in a density expert filter feature calculation step 6090.
Here, the features are calculated for the purpose of separating suspicious and non-suspect regions. These concentration features are analyzed in the expert filter stage 6100, which removes the subjects that are not most clearly associated with cancer. The detections remaining after the expert filter step are input to the density feature calculation step 7000. The output of the density detector comprises the density feature vector and the position of the detected density (step 7100).

A more detailed description of the steps in the density detection process is as follows. The subsampled digital image 2070 and chest mask 2040 are input to the preprocessing stage 6010. Here, the image is first prepared for subsequent processing to reduce the number of false detections due to discontinuities associated with the edges of the chest and mammography film boundaries. Large discontinuities cause ringing when the filter operates on the image. Swapping can cause spurious detections and obscure the damage, which makes these damage go undetected. The concentration pretreatment step 6010 reduces these undesirable effects.

Density pretreatment is used to reduce the hand-over on the chest edge due to the abrupt intensity discontinuity between background and tissue. Pretreatment is 2
Begins with the formation of two masks, an inflated chest mask, and an eroded chest mask. Tilt correction is then applied to the area between the eroded and dilated breast mask. This action preserves image detail and removes local background brightness. A local average is calculated at the edges of the dilated chest mask and this value is added to the tilt-corrected local area. This results in an image that retains detailed information on the transition between the chest and the background, while reducing the brightness discontinuity between the chest and the background.

After density pre-processing, a skew correction step 6020 is applied to the image. The tilt correction stage is a standard Bezier surface fitting algorithm.
algorithm)). Since the Bezier surface is a bi-cubic parametric patch, 16 control points are used to characterize the surface. These 16 control points are organized in a 4 × 4 grid, and the corresponding parametric surface is used as an approximation to the background image intensity in the local part of the image. Each pixel intensity within the boundaries of a 4 × 4 grid is subtracted from the approximation by an approximation to the background image intensity value that results in a pre-processed image for density for input to the next stage. There is.

The density pre-processed image is applied to the MaxDoG filterbank 6030. The MaxDoG filter bank shown in FIG.
Convolve the input image with the filter bank. The target size of the five DoGs is t = 12, 14, 16, 18, 20. These target sizes are related to the diameter of the lesion in the subsampled image. The relationship between the target size t of the DoG kernel and the standard deviation is σ ₁ = (t / 3.2) and σ ₂ = (t / 2.0). Each DoG kernel is individually convolved with the input image. The input image yields five DoG filtered images, each image being most sensitive to a particular target size. The maximum response at each pixel location is then selected across all DoG filtered images, which creates a MaxDoG output image.

Since the DoG kernel is an approximate approximation of how the cancer concentration appears on the mammogram, the kernel-like convolution response with the image is similar to this kernel and thus similar to the cancer concentration. The area of the image is shown. Do
The G kernel is given when the cardinality is roughly the same size as DoG (hence,
The strongest response (when 5 banks of different sized Dogs are used). In this way, concentrations of various sizes are detected over a diagnostically relevant size interval.

The MaxDoG image is input to the peak detector stage 6040. Here, the peaks in the MaxDoG image, which may be the concentration of cancer, are located for further processing. These stages are as follows. First, each MaxDoG
For a pixel location in the image, the peak detector examines every pixel 3 pixels away from that location. If the intensity value of the current pixel is greater than 90% of the intensity value of the pixel 3 pixels away, then the current pixel is set to a logical '1' on the local maximum image at the output of the peak selection stage. . Otherwise, the output image from peak selection is set to the logical value '0'.

The next step 6050 groups connected pixels having a value of '1' in the local maximal image into logical "objects". This is done by the following neighborhood algorithm. The local maximum image is searched in raster order (upper left of image to lower right of image row by row) for the pixels indicated to be the local maximum. For each local maximum image found, adjacent pixels above the image and adjacent pixels to the left of the image are examined. If neither adjacent pixel is part of the target,
A new object is created with its local maximum pixel as its first member.
If one of the neighboring pixels is already part of the object, the local maximum pixel is placed in this object. If both neighboring pixels are part of the object, then these objects are combined and the local maximum pixel is placed in the resulting object.
This continues until the entire image has been examined. At this point, the algorithm is generating an ordered list of N objects on the local maximum image. An object is identified by a set of pixel locations associated with each object index 1-N.

The next step in concentration detection is to remove many of the areas initially detected by the concentration detector and unlikely to be cancerous areas. In terms of quantity, the number of detections is often one
000 or more to a predetermined limit (here set to 20). Several steps contribute to this reduction. First, the pectoral mask 2060 is used to eliminate all detections located in the pectoral muscle. Then only the predetermined number of detections with the highest response in the MaxDoG image are retained. This is then graded above a predetermined number (3.5 in this embodiment) by performing a descending sort on the maximum value from the MaxDoG image under each detected object. By removing the subject from further consideration. The motivation for this step is that the larger the MaxDoG value, the better the damage matches the target model. The final step in this step involves calculating the features of the concentration screen and classifying the detections while keeping the number of detections passing through the secondary screening classifier below a second predetermined value.

The input to the screening classifier is the remaining detected subjects and their associated set of concentration screening features. The symbol "PT" in the feature description indicates that the power transformation is applied to the measurement. The concentration screening features in the preferred embodiment include: MaxDoG Maximum-The maximum value (PT) of the MaxDoG image from the set of pixels corresponding to the detected object. 2. MaxDoG standard deviation-The standard deviation of the MaxDoG image from the set of pixels corresponding to the detected object. 3. MaxDoG Normalized Mean-The average of the MaxDoG images below the detected object divided by the average of the sub-sampled images from the set of pixels corresponding to the detected objects. 4. Minimum MaxDoG value-Minimum MaxDoG value of a MaxDoG image in a 5 × 5 pixel block centered on the detection. 5. MaxDoG 5x5 Normalized Mean-5 divided by the mean of the subsampled input image around the central block of 5x5 pixels.
Average of MaxDoG images in the central block of × 5 pixels.

These feature vectors for concentration screening are input to the concentration screening classifier. The concentration screening classifier uses two discriminant functions, one for normal and one for abnormal. The difference is found after the discriminant is calculated. This difference is
It is compared with a predetermined threshold. If the discriminant difference is less than the threshold, the detected object is removed from further consideration. Regions having a density screening discriminant difference greater than a threshold are graded by the discriminant difference, whereby the largest discriminant difference has the smallest grade value (ie, 1). Any detected object with a magnitude greater than a further predetermined upper limit is excluded from further consideration. The remaining detected objects are passed to the next stage for further processing.

Each of the remaining detected objects is input to the region growing stage 6080 below. Here, the shape of each detected object is evaluated more strictly. The improved shape evaluation is used in the further density feature calculation 7000 for the next classification stage 8000. The shape evaluation by region growth will be described below.

Region Growth At this stage in the concentration detector, region growth is used to evaluate the boundaries of the detection. Ideally, all detections correspond to suspicious damage. However, in practice, the steps up to this point detect areas not associated with cancer. A useful way to reject unwanted detections is to compute features based on the shape of locally bright areas. However, due to the complexity of the chest tissue, the often subtle nature of the suspicious area, and the obscure edges of the suspicious area, shape evaluation is a difficult problem.

Area growth is still an active area of research in the medical imaging community. Due to the subtleties of the lesions, the complexity of the boundaries of these lesions, and the characteristics of the background image surrounding the lesions, shape evaluation is difficult to perform automatically in medical images. In many cases, the surrounding tissue has a bright structure that is not associated with injury. One drawback with conventional percentage-based floodfills is that they are prematurely terminated if bright and elongated structures are present in the image chip. This happens when the fill grows into the structure, because
This is because the tendency to fill is to follow bright and elongated structures to the boundaries of the image chips.

The region growing method described in the present invention has two unique features. First, iterative dynamic weighting for images is
Used to reduce the effect of background structures on the evaluated shape. Iterative dynamic weighting is the process of calculating a mixed image as the sum of an input image and a weighted input image. Weighting is a function of current shape evaluation. Secondly,
The classifier is used to select the shape that most represents the damage. The problem of selecting the optimal shape evaluation from the set of evaluation values created during the iterative process is solved by the use of a "region growth mask selection classifier". The classifier identifies the best shape estimate for a region based on the region growth discriminant function. The set of candidate masks is formed with a fill algorithm by iterative dynamic Gaussian weighting, which ensures that the grown regions are small and reasonably well formed. The output is a set of selected area growth masks and area growth discriminant values associated with each detection.

In the application of the invention, area growth is applied to the detected object. The detected objects that pass the concentration screening classifier are input to the region growing stage 6080. A block diagram outlining the operation is given in FIG. The inputs to the region growing stage are the subsampled image, the mask of the detected object, and a predetermined maximum allowed region growing neighborhood size. The output is a set of'optimal 'Boolean images associated with the damage. '1 in Boolean image
'Pixel indicates shape evaluation for damage. These Boolean images are known as region grown masks. The number of masks to be finally selected is predetermined as N, where the optimum selection of N masks is
It is based on the discriminant value associated with each mask.

The method for region growth described in the present invention uses two iterative loops. The outer iterations are used to establish a monotonically decreasing threshold. Within the outer loop, evaluation of multiple masks occurs through an iterative active weighting loop. The discriminant value associated with the region growth mask is further provided as an output for subsequent processing.

The region growing method will be described in detail below. Pixel locations, called seed pixels, are found as the brightest points in the subsampled image in the detection area. This is the beginning of the process. That is, the first current region growth mask is set to '1' at the seed pixel location. The image chip is then formed by selecting a square neighborhood twice the maximum region growth neighborhood size centered around the seed pixel.

Next, the percentage-based fills are successively applied to the image chips, thereby creating Boolean masks of the area. Filling is the process of creating the next mask by testing all the pixels connected to the current mask and then adding the pixels that pass the test to the current mask. In order to be added to the mask, the pixel under consideration must (1) have an intensity within x% of the intensity of the seed pixel, and (2) to other pixels already located in the region. Must be concatenated. The fill percentage x then increases at a given step size from a given lower limit to an upper limit. In a preferred embodiment, the percentage of fill varies from 1% to 40% in 1% steps. One criterion for ending a fill is when the fill percentage reaches an upper limit. The second criterion for ending the fill is when the area filled with a predetermined percentage exceeds the maximum area size. Here, exceeding the maximum size is defined as any part of the mask that touches the edge of the neighborhood of the square. In practice, the second termination criterion is more broad and general.

Increasing the fill percentage allows the area to grow. Control of growth is done by an iterative set of the following additional steps for each percentage. These further steps are based on the circularly sym
We need to dynamically create and apply a Gaussian weighting kernel to the image chip pixel values. The Gaussian weighting kernel is

[Equation 7] Found as. Where μ _x and μ _y are the x, y coordinates of the current mask centroid and σ _r is three times the maximum distance from the mask centroid to the current mask edge. This window is multiplied by the input image chip, which forms a weighted image. Next, a mixture image is created for the subsequent fill step as follows.

[Equation 8] Here, in the preferred embodiment, α = 0.5. FIG. 10 details dynamic Gaussian weighting and iteration within a particular fill percentage value.

For each iteration, a mixed image is formed. First, the Gaussian weighting kernel is created from the current mask. Weighting is then applied to the original image to form a weighted image, and a mixed image is created from the weighted image and the original image. Region growing is performed on the mixed image.

These steps create a set of masks, where each mask is for each percent fill value. The features are calculated from the mask and then input to a region growing mask selection classifier designed to select a predetermined number of'optimal 'density masks. The selected mask may be a mask resulting from any subset of the percentage fill values. These optimal masks, along with the discriminant values of these region growth mask selection classifiers, are stored for subsequent processing.

The subset of masks passed through the classifier is applied to a pod chopper to remove the outer tendrils from the core growth. First, a single erosion is performed on the area growth mask to break the connection between the pod and the core.
Second, the object is formed from the eroded mask and only the largest object is retained. Finally, the largest target mask expands once within the original mask, thereby forming a podchopped area growth mask.

Region Growing Mask Features The classifier is used to select the optimal set of region growing masks and thus the features are needed. These features calculated from the area growth mask are as follows, where the Bauer-Fisher ratio (Ba
uer-Fischer ratios) are described in detail in the section on concentration characteristics. The symbol "PT" in the feature description indicates that the power transformation is applied to the measurement. 1. Bower-Fisher ratio E5L5, first iteration ~ vs. ~ previous percent --- Rose texture using Rose kernel function E5L5 (Laws Tex
ture) Bauer-Fisher ratio of the measured value. The two regions used to produce this feature are the non-overlapping portion of the region of the first weighting iteration for the previous fill percentage and the region of the first weighting iteration for the current fill percentage. The non-overlapping part indicates a part of one area that is not included in the other area. This is the core area for the mask from the previous percentage, and
This results in a ring for non-overlapping parts from the current percent mask. The rose texture is calculated within the core and annulus, and then the Bauer-Fisher ratio is calculated. 2. Area Difference, Last Iteration ~ vs. ~ Previous Percent --- Area difference between the mask of the last weighted iteration for the current fill percentage and the mask of the last weighted iteration for the previous fill percent. (PT
). 3. Bauer-Fisher ratio L5E5, last iteration ~ vs.-previous percentage --- Bauer-Fisher ratio of rose texture measurements using the Rose kernel function L5E5. The two regions used to produce this feature are the non-overlapping parts of the region of the last weighted iteration for the previous fill percentage and the region of the last weighted iteration for the current fill percentage (PT). . 4. Bauer-Fisher ratio R5S5, first iteration vs. last iteration for previous percent --- Bauer-Fisher ratio of rose texture measurements using the Rose kernel function R5S5. The two regions used to produce this feature are the non-overlapping parts of the region of the last weighting iteration for the previous fill percentage and the region of the first weighting iteration for the current fill percentage. 5. Bauer-Fisher ratio E5L5, first iteration ˜vs.˜last iteration with respect to previous percent --- Bauer-Fisher ratio of rose texture measurements using the Rose kernel function E5L5. The two regions used to produce this feature are the non-overlapping parts of the region of the last weighting iteration for the previous fill percentage and the region of the first weighting iteration for the current fill percentage.

These features occur for all masks at the end of each fill percent repeat cycle. These features are tested by the region growing mask classifier to select the'optimal 'subset of masks.

Region Growth Mask Classifier The region growth mask classifier determines a set of'optimal 'shape evaluations from previous region growth stages. The region growth mask classifier uses a quadratic classifier to determine the optimal region growth mask for the detected object. Many region growth masks have been generated up to this point, and features from these grown regions are input to the classifier. For each mask, the classifier computes the discriminant associated with the'optimal 'category and the discriminant associated with the'non-optimal' category. These discriminants are subtracted to give the region growth mask quality score. These differences are then ordered by grade such that the largest difference receives a grade value of 1. The top N masks are selected based on the grading.

Density Features Density features are calculated for each detection from the region growth stage. These features are passed to a concentration classifier, where only the detections that are most likely cancer concentrations are retained. Many features are one form or another form of contrast, such as intensity contrast or texture contrast. The two measurements of contrast used in the density detector feature are:

[Equation 9] Bauer-Fischer ratio, and

[Equation 10] Is a mean ratio.

Subscripts 1 and 2 indicate two areas on the input image respectively corresponding to the area inside the detection and the second area outside the detection. Measurements are calculated in each area. In some cases, the measurements may simply be pixel values from the input image. The area within the detection is defined as all points that are part of the area growth mask. The undetected area is twice as large as the area growth mask (at least 1 cm in each direction).
Defined as the inverse mask of the region growth mask bounded by the (increased) box. Some features further erode the inner and outer masks so that they do not rely too much on the correct distinction between the inside and the outside. In the case of an eroded mask area, the inner mask is eroded once and the outer mask is eroded to half the original area. Neither mask is eroded to be smaller than 30 pixels.

The limitation of the Fisher ratio is that the relationship between the mean values is lost. That is,
There is no information about whether μ ₁ is greater than or less than μ ₂ . Bauer
The Fisher ratio incorporates a term with a sign (.) To add relevant information to the Fisher ratio.

Some of the features further use by-image normalization. Image-by-image normalization is a method that eliminates the need for global thresholds, which can be problematic in some situations. There are three different image-wise normalization schemes: ordering by magnitude with highest value first, ordering by lowest value first, and subtracting the mean value of each feature for the image from the feature vector. Used. Each of these methods has a rank ordered down:
ROD), rank ordered up (ROU), zero mean by image (ZMI). The feature value of ROD and the feature value of ROU are integer grade values, while the feature value of ZMI is a real number.

The final class of features generated in the concentration feature stage is the “delta feature (delta fe
deltas. The delta features result from obtaining the maximum, minimum, mean, or standard deviation of the feature values across all N masks through the region growing classifier.
Usually, false positives have less than a "good" mask of cancer growth than the concentration of cancer, and the delta feature takes advantage of this difference.

Before calculating the concentration features, the expert filter stage 618 is applied. The expert filter does a boundary check on the features of the expert filter from all remaining detections. If the features of the expert filter do not fit within the bounds, the region is removed from further consideration. What is "optimal"?
Note that we refer to the region growth mask graded as 1 from the region growth mask selection classifier and that the features of the expert filter are as follows: 1) the optimal region growth mask is Be circular. 2) Optimal area growth mask long axis (λ ₁ ) length. 3) Standard deviation of the Rose E5L5 texture image under the optimal region growing mask. 4) Bower-Fisher ratio of the Rose E5L5 texture image inside and outside the optimal area growing mask.

These features are formed from measurements on the image section. For example, expert filter 3 is found by first using the E5L5 texture image as a measurement. This feature is found as the standard deviation of the measurements.

These features are calculated for objects that pass the expert filter. The symbol "PT" in the feature description indicates that the power transformation is applied to the measurement. The following are the 18 density features calculated for subsequent processing that are input to the density classifier: 1) Mean (PT) of the MaxDoG image under the original detected object. 2) Optimum average ratio of image intensity inside and outside the area growth mask (PT
). 3) Optimal region growth mask perimeter, ZMI (PT). 4) Mean of MaxDoG in screener, ROD (PT), under original detected subject. 5) Optimal region growth mask region, ROD (PT). 6) Average ratio of image intensities inside and outside the optimum region growing mask, RO
D (PT). 7) Optimal region growth mask aspect ratio, ROU (PT). 8) Total pod area removed from the optimal area growth mask. 9) Average (PT) of the top four area growth masks of the circular area growth mask. 10) Average (PT) of the top four region growth masks of the average of the Rose R5S5 images inside the region growth mask. 11) Average of the top four area growth masks in the average ratio of the Rose R5S5 images inside and outside the area growth mask. 12) Strong bower inside and outside the eroded area growth mask
Average (PT) of the top four region growth masks by Fisher ratio. 13) Standard deviation (PT) of the top four area growth masks of the average ratio of the Rose L5E5 images inside and outside the eroded area growth mask. 14) Minimum value (PT) of the top four area growth masks in the length of the shorter axis (λ ₂ ) of the area growth mask. 15) Minimum perimeter (PT) to length ratio of the top four region growth masks. 16) Minimum value (PT) of the top four region growth masks of μ20 (instantaneous second). 17) Minimum of the top four area growth masks of the Bower-Fisher ratio of the Rose L5E5 image inside and outside the eroded area growth mask. 18) Maximum perimeter of the top four region growth masks.

At this point, it is important to distinguish the present invention from the prior art by noting that the optimal four masks do not necessarily result from consecutive fill percentage values. These density feature vectors associated with objects that pass the expert filter are applied to a density classifier 8000.

<Classification of Detection of Microcalcifications> The feature vector 4100 of the within-image microcalcification clusters is shown in FIGS. 11 and 12 (FIG. 12 is a continuation of FIG. 11). , As an input to the microcalcification classifier. The feature vectors of these microcalcification clusters are applied to two quadratic discriminant functions G _c (x), G _t (x) (steps 5010, 5020). G _c (x) is the discriminant function for the truthed microcalcification clusters, and G _t (x) is the discriminant function associated with the tissue region outside the truth box. Tissues outside the Truth Box are considered normal. The discriminant difference G _t (x) −G _c (x) is calculated at step 5030 to yield the “delta discriminant”. At step 5040, the delta discriminant is compared to a threshold (in this embodiment, the threshold is 0). If the delta discriminant is less than the threshold, then the detection is considered normal and step 5050
Receive class label '0' at. If not, the detection is considered suspicious and receives class label '1' in step 5060. The delta discriminant and classifier labels are stored at step 5070 for subsequent analysis at post-processing step 9000.

<Classification of Detecting Density> The density feature vector of (within-image) from the density detector is shown in FIG.
3 and FIG. 14 (FIG. 14 is a continuation of FIG. 13) are input to the concentration classifier. The feature vectors of these concentrations are three quadratic discriminant functions G _m (x
), G _s (x), G _n (x). G _m (x) (step 8010) is the discriminant function for an elliptical mass. G _n (x) (step 8010) is the discriminant function for tissues outside the truth box. Tissues outside the Truth Box are considered normal. G _s (x) (step 8020) is the discriminant function for agglomerates with spicules. Next, two delta discriminants are formed. Step 80
At 40, 8050, the discriminant of the mass and the discriminant of the mass with spicula are subtracted from the discriminant of the normal tissue. At step 8060, the two delta discriminants are compared to a threshold. In this embodiment, the threshold is 0. If both delta discriminants are less than the threshold, then the detection is considered normal and step 8
At 070, the class label “0” is received. However, if either discriminant is greater than the threshold value, then the detection is considered suspicious and thus receives the class label '1' in step 8080. The two delta discriminants and the label of the classifier are used in step 809 for subsequent analysis in post-processing step 9000.
Stored at 0.

Post-Processing The post-processing portion 9000 of the present invention performs three main functions. These stages are
Case-based grading (steps 9020, 9030), image context (step 9040)
, And normality classification (steps 9050, 9060).

FIG. 15 is a post-processing stage of the present invention. This process will be briefly described below. The inputs to the post-processing are the microcalcification delta discriminant in the case and the concentration delta discriminant in the case with the class label '1'. Within-case refers to a set of images input from the same patient by one mammography screening test. At each post-processing stage, the detection class label is' 1.
It may change from 'to' 0 '. In the output of the post-processing stage the class label '
Only detections with 1'are displayed to the end user.

The first stage of post-treatment is Radial Inhibition 9010. If more than one concentration detection in a local region has label 1 from the concentration classifier, then only the region with the highest likelihood of cancer is retained for consideration. The areas affected by this module are those that occur within a 1 cm radius, measured from center to center of each other. The criterion for considering that an area is more cancerous than another area is the sum of the discriminant differences from the feature classifier. Regions with a larger sum of discriminant differences are considered more likely to be cancer,
Therefore, it is retained as a potential cancer area. Regions with smaller discriminant sums are removed from further consideration.

<Case-Based Grading> In Case-Based Ranking (CBR) 910, only the top N detections from this case are within the M number held on any one image. Retained with detection. Retaining means keeping the class label as "1". Removal means changing the class label to "0". The number of detections that carry a classification label depends on how many images are present in this case, the distribution across the image, and the category of detection (microcalcification or concentration).

The CBR input is the delta discriminant for the concentration classifier and the microcalcification classifier,
And a class label for each detection in the case. CBR operates independently of both microcalcification and concentration detection. Within the injury category (microcalcification or concentration), the detections are ordered by quality score, from highest to lowest, and by grade. In a preferred embodiment, the quality score is the delta discriminant for microcalcifications or the sum of these delta discriminants for concentration. The microcalcification classifier has one delta discriminant (normal-
Recall that the concentration classifier gives two delta discriminants (normal-lump with spicule and normal-lump). In any case, the higher the quality score, the more likely the detection is cancerous.

The number of detections to be allowed for each case is determined based on a predetermined allowable number of detections N, assuming a case of 4 films. When cases other than 4 films make up the case, the allowable number of detections is

[Equation 11] Found as. Here, I represents the number of films forming the case, and ceil (.) Represents rounding up to the next integer. For example, the allowable number of detections for the case of 4 films is 7. The allowable number of detections for the case of 3 films is ceil (3 * 7/4), or 6.

Case-based grading proceeds as follows. First, the allowable detection number D is
Calculated as described above. The delta discriminant is then ordered by magnitude from largest to smallest. Detections with a magnitude greater than D will set their class label to '0.
Set to '.

Once the case detection is graded D, the second condition applies (ie no more than M detections can have the classification label '1' on any one image). ). If the image has more than M detections with a label equal to '1', the detections from this image are again ordered by maximum to minimum by these discriminants. Any detection on an image with a magnitude greater than M sets its classification label to '0'.

The values for M and N can be set independently based on the detection category. The maximum number of detected concentrations is equal to N_density and the maximum number of detected microcalcifications is equal to N_microcalc. Then the maximum number of detections per case is N_d
Equal to the sum of ensity and N_microcalc.

In a preferred embodiment, the detection of concentration and microcalcification is carried out first by M_dens
Limited to ity = 3 and M_microcalc = 2. That is, within each image, the density detections are subsequently ordered by grade from 1 to M_density. Density detections with grade values 1-M_density retain a class label of '1', while density detections above this limit receive a class label of '0'. Similarly, the detection of microcalcifications is ordered by grade from 1 to M_microcalc. Grade value 1-M_microcalc
Microcalcification detection retains a class label of '1', while microcalcification detection above this limit receives a class label of '0'.

This is a common and effective way to distribute a constrained number of CAD representations across an image in some cases. The only requirement is that the detections from all images in this case have an associated quality score to be ordered by grade. Moreover, in some cases, case-based grading has been found to reduce the total number of scars per case while simultaneously finding more cancers. The following section demonstrates the utility of case-based grading.

As an example of case-based grading, consider a set of concentration detections and the corresponding sum of delta discriminant scores, as shown in Table 1.

To better understand the advantages of case-based grading, first consider the strategy of image based ranking. An example of image-based grading is to pass the top two ranked detections from each image, the results of which are shown in Table 2. These are eight (two for each image) residual detections. In this example, some detections are true positives for a detection index of 9. All other detections are false positives.

In the following, consider the results of case-based grading for the same set of detections. Table 3 shows the advantages of CBR with the constraint of 6 detections per case. In this example, we classify the set of all detections in a case,
And, constraining to a maximum of 6, additional cancers are found with fewer false positives than the method for image-based grading. In this example, the cancer in image 2 is the third "optimal" detection in this case. Case-based grading makes it possible to preserve this detection, whereas image-based grading does not make it possible to preserve this detection.

[Table 1]

[Table 2]

While case-based grading checks for detection of all the same categories across all images for a case, image context processing involves detection over and within detection categories for images. Remove selected combinations. The motivation for this step is that it is observed that certain detection combinations occur preferentially on the normal images of the training database.

[Table 3]

An outline of the image situation processing will be given below. The detection of concentration and microcalcifications is assigned a credit rating of high or halo based on comparing the delta discriminant value to an appropriate threshold (where the threshold is derived from training data). Other embodiments may use more than two credit grades. The numbers, formats, and credits for the set of detections are recorded. Analysis of the training data showed that certain detection combinations occurred preferentially on normal images. Therefore, if these combinations occur, all detections from this image are always removed. This step greatly increases the number of flawless images without adversely reducing sensitivity.

The image situation processing will be described in detail below. An image context quality score is calculated for each detection as follows. For concentration detection, the delta discriminant is averaged. For microcalcification, the delta discriminant is used as is. The quality score is compared to a predetermined threshold to assign a credit rating to each detection. The predetermined threshold for each detection category is obtained from the analysis of the training database so that most of the true positive detections receive a high credit grade and most false positive detections the low credit grade. To receive.

In the preferred embodiment, up to three concentration detections and one microcalcification detection are considered. Each detection receives a high or halo credit rating. In this case, as shown in Table 4, there are 29 possible combinations of concentration detection and microcalcification detection.

Table 4 shows the results of assigning a combination of image contexts to about 3200 processed images, where about 600 images have malignant microcalcifications, and About 600 images have malignant density damage. The first column lists the combinations and the second column shows the category of combinations. The letters "C" and "D" indicate microcalcification detection and density detection, respectively. Uppercase letters indicate a high credit rating, while lowercase letters indicate a low credit rating. The last four columns show the number of occurrences of the corresponding combination on the image with the given true category. For example, for index 1, one confidence low density detection is performed 6 times on images with malignant mass, 0 on images with malignant microcalcification clusters, on images with both damage categories. , And 324 times on the normal image. It should be noted that the image is most often normal when the combination is one of the following: i) one credit low density detection (d); ii) three credit low density Detection (ddd); iii) One credibility low microcalcification display and one credibility low concentration display (cd). In such an example, all detections on the image can be dropped without unduly reducing sensitivity.

[Table 4]

<Classification of Normality> Classification of normality refers to a process of checking a discriminant from each detection category over all images in a case. If normality conditions are met, density detection and / or microcalcification detection is removed from all images in the case. The "normality" condition is assigned by the classifier using the quality score obtained as a feature of the input from all detections over a case. The quality score entered in the normality stage is the delta discriminant value of concentration and microcalcification.
A further input is the assigned class label for each detection. A separate health classifier is used for each injury category.

It is important to note that the above post-processing steps do not eliminate detections. The delta discriminant for all detections, whether identified by a '1' class label or a '0' class label, is used to generate the normality feature.

The microcalcification normality classifier requires two “microcalcification normality” characteristics. These are: Microcalcification Normality Features 1--The Delta discriminant values from the microcalcification classifier are sorted by image for all images in the case. The largest delta discriminant value for each image is sorted. The image-sorted delta discriminant is then searched across all images in the case. The maximum value from the image-wide search is used as the normality feature of the first microcalcification. Microcalcification Normality Feature 2 --- The maximum delta discriminant value for each image is stored. The maxima in the images are then averaged over all images in the case, which gives a second normality feature.

These two microcalcification normality features are used as inputs to the microcalcification normality classifier. One embodiment is a statistical quadratic classifier.
tic classifier). The classifier weights are planned using a training database of cases in two categories: normal cases and cases containing microcalcifications associated with being malignant. Two discriminant functions are obtained. One discriminant shows the case "normal" in the meaning of microcalcification, and the other discriminant shows the case "not normal" in the meaning of microcalcification. The “non-normal” discriminant is subtracted from the “normal” discriminant to yield the normal delta discriminant for microcalcifications, ie, the normality value. If this normality value is greater than a predetermined threshold, then all microcalcification detections are removed from this case.

First, the sum of the delta discriminants is calculated for all detections on each image. At this stage, even detections with a class label of '0' are considered. Since each detection is shown to have two delta discriminants (lumps and lumps with spicules), the sum of the delta discriminants is calculated for each detection and the maximum of the delta discriminants. The detection on each image with the sum is identified.

The normality characteristic of concentration is calculated from the delta discriminant. The symbol "PT" in the feature description indicates that the power transformation is applied to the measurement. The following normality features are calculated using the detections identified for each image as follows: Density normality feature 1--for the detections identified on each image,
Calculate the average of the mass and mass with spicula. Next, a search is performed over all images in the case, and the minimum value of the average delta discriminant is selected as the first feature (PT). Normality features of density 2 --- For the detections identified on each image,
Select the maximum of the two delta discriminants (lumps and chunks with spicules).
Next, the average of the selected delta discriminants is calculated as the second feature (PT). Normality features of density 3 --- For the detections identified on each image,
Select the minimum of the two delta discriminants (lumps and chunks with spicules).
Next, the average of the selected delta discriminants is calculated as the third feature.

These three density normality features are used as inputs to the density normality classifier. An embodiment may be a statistical quadratic classifier. The weights of this classifier are planned using a training database of cases in two categories (normal cases and cases containing malignant concentrations). Two discriminant functions are obtained. One of the discriminants indicates a “normal” case in the sense of density, and the other discriminant indicates a “not normal” case in the sense of concentration. The "non-normal" discriminant is subtracted from the "normal" discriminant to yield the normality delta discriminant of concentrations, or normality values. If this normality value is greater than the predetermined threshold,
Detection of all concentrations is removed from this case.

The output of the normality stage is a set of detections and associated class labels. Detections with a class label of '1' are displayed overlaid on the medical image as they require further judgment by the radiologist.

In view of the foregoing description, the present invention provides an improved method and system for identifying detections on digital images that meet criteria indicative of cancer damage, and
In particular, it is clear to provide methods and systems that can be conveniently used to identify cancer-induced damage on mammograms obtained from screening rounds of patients. As further noted from the above description, the present invention increases the sensitivity of the method and system while reducing the number of false positives, thereby
Presents a process for classifying microcalcifications and detections and for post-processing detections in such a way as to provide the additional benefit of increasing the accuracy with which cancer damage is identified. .

Although the present invention has been described in the context of analyzing detections found on mammograms, the methods and systems described herein are also applicable to the analysis of other medical images. Is further noted.

Although the methods and systems described herein form the preferred embodiments for the present invention, the present invention is not limited to the methods and systems as described herein. And, it is to be understood that changes can be made in the methods and systems without departing from the scope of the invention as defined in the appended claims.

[Brief description of drawings]

FIG. 1 is a flow chart showing an automated method for detection of agglomerates and clustered microcalcifications on a digital mammogram.

FIG. 2 is a flow chart showing a schematic classifier architecture.

FIG. 3 is a flowchart of a pretreatment process.

FIG. 4 is a flow chart of a clustered microcalcification detector.

FIG. 5 is a view similar to FIG.

FIG. 6 is a flow chart of a concentration detector.

FIG. 7 is a view similar to FIG.

FIG. 8 is a flow chart of a DoG filter bank.

FIG. 9 is a flowchart of a region growing method.

FIG. 10 is a flow chart detailing a fitness weighting unit of the region growing method.

FIG. 11 is a flow chart of a clustered microcalcification classifier.

FIG. 12 is a view similar to FIG. 11.

FIG. 13 is a flow chart showing a concentration classifier.

FIG. 14 is a view similar to FIG.

FIG. 15 is a flowchart of a post-treatment process.

[Explanation of symbols]

1000 Steps to obtain digital image 2000 Pretreatment stage 3000 Microcalcification detection stage 4000 Calculating microcalcification features 5000 Microcalcification classification stage 6000 Concentration detection stage 7000 Calculation of concentration characteristics 8000 concentration classification stage 9000 post-treatment stage

─────────────────────────────────────────────────── ─── Continued front page    (81) Designated countries EP (AT, BE, CH, CY, DE, DK, ES, FI, FR, GB, GR, IE, I T, LU, MC, NL, PT, SE), OA (BF, BJ , CF, CG, CI, CM, GA, GN, GW, ML, MR, NE, SN, TD, TG), AP (GH, GM, K E, LS, MW, MZ, SD, SL, SZ, TZ, UG , ZW), EA (AM, AZ, BY, KG, KZ, MD, RU, TJ, TM), AE, AL, AM, AT, AU, AZ, BA, BB, BG, BR, BY, CA, CH, C N, CR, CU, CZ, DE, DK, DM, EE, ES , FI, GB, GD, GE, GH, GM, HR, HU, ID, IL, IN, IS, JP, KE, KG, KP, K R, KZ, LC, LK, LR, LS, LT, LU, LV , MA, MD, MG, MK, MN, MW, MX, NO, NZ, PL, PT, RO, RU, SD, SE, SG, S I, SK, SL, TJ, TM, TR, TT, TZ, UA , UG, UZ, VN, YU, ZA, ZW (72) Inventor Randy P. Brossard             United States · Maryland · 20778 ·             West River Biltmore Aveni             1002 (72) Inventor Michael Jay Collins             United States Ohio 45431 Bee             Ver Creek Bandit Tray             Le-2225 (72) Inventor Martin P. Desimio             United States, Ohio, 45324, Fe             Aborn Aparosa Trail 4770 (72) Inventor Jeffrey W. Hoffmeister             United States, Ohio, 45324, Fe             Aborn Old Troon Dry             B-2945-H (72) Inventor Richard A. Mitchell             United States Ohio 45066 Sp             Ringborough McRae Boulevard             Do 125 (72) Inventor Thomas F. Rathvan             United States of America Colorado 80132 Moni             North Sherwood Grew             15 (72) Inventor John E. Rosen Stengel             United States Ohio 45431 Bee             Ver Creek Old Willow Do             Live3685 F term (reference) 4C093 DA06 FD01 FD03 FD05 FD09                       FD12 FD13 FF07 FF16 FF17                       FF18 FF19 FF28 FF29                 5B057 AA07 CA08 CA12 CA16 CE06                       DA01 DB02 DB09 DC01 DC22                 5L096 AA06 BA06 BA13 DA01 EA37                       GA10 GA51 GA55 HA11 JA11

Claims (105)

[Claims] 1. A method of detecting microcalcifications in a digital mammogram image, the method comprising the steps of: a) creating a filtered image with local bright spots corresponding to the size of the microcalcifications. Filtering, b) evaluating the filtered image with a threshold to segment bright spots showing microcalcifications formed at point locations on the digital image, and c) segmenting in b) Calculating point features for each microcalcification, and d) inputting the point features into a microcalcification point classifier to identify the microcalcifications that best characterize the damage caused by the cancer. And a method comprising: 2. The at least one of the point features comprises a DoG kernel response of a predetermined size and centered at each microcalcification point position. the method of. 3. The method of claim 1, wherein at least one of the point features comprises a local average value around each microcalcification point location in the DoG filtered image. Method. 4. The method of claim 1, wherein at least one of the point features comprises a non-linear combination of measurements obtained from each microcalcification. 5. The measurements obtained from each of the microcalcifications are obtained around i) the response of the DoG kernel centered on the microcalcification point position and ii) around the microcalcification point position. 5. The method of claim 4, further comprising: a local mean of the original image; and iii) a local standard deviation of the original image obtained around the microcalcification point locations. . 6. The method of claim 2, wherein the non-linear coupling is performed using a multilayer perceptron neural network. 7. The method of claim 1, wherein the microcalcification point classifier of step d) comprises a secondary classifier. 8. Three point features are calculated for input to a point classifier, which are: 1) Do of a predetermined size and centered on each microcalcification point location.
Comprising the response of the G kernel, 2) the local mean around each microcalcification point location in the DoG filtered image, and 3) the non-linear combination of the measurements obtained from each microcalcification. The method of claim 7, wherein the method comprises: 9. The measurements obtained from each of the microcalcifications are obtained around i) the response of the DoG kernel centered at the microcalcification point position and ii) around the microcalcification point position. 9. The method of claim 8, further comprising: a local average of the original image; and iii) a local standard deviation of the original image obtained around the microcalcification point locations. . 10. The step d) may include a step of ordering microcalcifications, which are likely to be cancers, by grade, based on an analysis of discriminant differences regarding each microcalcification, and indicating a cancer area. Removing microcalcifications having a grade above a predetermined grade limit value for retaining the most prominent microcalcifications. 11. A method of detecting microcalcification clusters in a digital mammogram image, the method comprising: a) identifying microcalcifications on the digital mammogram image; and b) forming microcalcification clusters to form microcalcification clusters. Clustering a group of calcifications, and c) i) calculating features of at least one cluster comprising features calculated from each microcalcification in each microcalcification cluster, ii) false positives Analyzing each microcalcification cluster by inputting features of the at least one cluster into a microcalcification cluster classifier to identify the clusters of. 12. The point features are calculated for each microcalcification in each cluster, and the point features are pointed to yield a discriminant score difference for each microcalcification. The method of claim 11, wherein the features of the at least one cluster that are input to the classifier are obtained from the difference of discriminant scores for microcalcifications. 13. The method of claim 11, wherein the at least one cluster feature is the difference in the Mth largest discriminant score for microcalcifications within the cluster. 14. The features of the at least one cluster are: i) by regionally growing each microcalcification within a cluster to form a region grown region, ii) for each microcalcification The average intensity of image pixels inside each region-grown region is divided by the average intensity of image pixels outside the region-grown region located within a region of a given shape and size to define the intensity ratio of Iii) that the sum of the intensity ratios for all microcalcifications in a cluster is calculated by dividing by the number of microcalcifications in the cluster to obtain the mean intensity ratio feature. The method of claim 11 characterized. 15. The feature of the at least one cluster is: i) centering the centroid of the cluster by locally growing each microcalcification in the cluster to form a region grown area. By normalizing the predetermined region with reference to the energy in the predetermined region, iii) the second of the region grown regions to define the characteristics of the average second moment.
The method of claim 11, wherein the method is calculated by calculating a moment. 16. A method of detecting microcalcifications in a digital mammogram image, the method comprising the steps of: a) creating a filtered image with localized bright spots corresponding to the size of the microcalcifications. Filtering, b) evaluating the filtered image with a threshold to segment bright spots showing microcalcifications formed at point locations on the digital image, and c) segmenting in b) Calculating point features for each microcalcification, and d) inputting the point features into a microcalcification point classifier to identify the microcalcifications that best characterize the damage caused by the cancer. And e) clustering groups of microcalcifications to form microcalcification clusters F) i) by calculating features of at least one cluster comprising features calculated from each microcalcification in each microcalcification cluster, ii) at least 1 to identify false positive clusters, Inputting the features of one cluster into a microcalcification cluster classifier, and analyzing each microcalcification cluster. 17. A method of iterative shape estimation of detections on a digital image, comprising the steps of: a) providing a pixel position corresponding to the detection on an input digital image; and b) at a given pixel position. Defining the current region-grown mask to be equal to 1 ′, c) calculating the current weighting kernel as a function of the current region-grown mask, and d) the input image and the current Calculating a current blended image from the weighting kernel, e) region growing the current blended image to obtain a new region grown mask, and f) satisfying a predetermined continuation condition. Test the new area grown mask,
And i) if the continuation condition is met: A) equal to the new region grown mask for the current iteration,
Assign the current region grown mask for the next iteration, B) repeat steps c) to f), ii) if the continuation condition is not satisfied, A) terminate the iteration, and B) continue condition. Selecting the final region-grown mask that satisfies as the evaluated shape for detection. 18. The area growth comprises: performing a percentage-based fill using a predetermined range of fill percentages; and g) initializing a fill percentage start value; and h) depending on a fill percentage start value. Performing steps a) to f) for each fill percentage over a range of starting fill percentages, and i) for each fill percentage to yield a set of masks, ii) B). Further comprising the steps of storing a final region-grown mask selected from, and j) selecting one of the masks from the set of masks as a final shape evaluation for detection. The method according to claim 17, wherein: 19. The method according to claim 18, wherein the step j) of selecting one of the masks is performed with a classifier. 20. The method of claim 17, wherein the continuation condition comprises a new area grown mask different from the current area grown mask. 21. The method of claim 17, wherein the continuation condition comprises a new region grown mask that has not reached a predetermined maximum mask size. 22. The method of claim 17, wherein the weighting kernel is a circularly symmetric Gaussian weighting kernel. 23. A method of calculating features from detections on an input image, comprising: a) 2 on the input image corresponding to a first region inside the detection and a second region outside the detection.
Giving two regions, b) calculating the measured values based on the values obtained from the two regions on the input image, c) calculating the average of the measured values of each region, and d) Calculating the standard deviation of the measured values of each region, and e) feature = SIGN (μ ₁ −μ ₂ ) * (μ ₁ −μ ₂ ) ² / (σ ₁ ² + σ ₂ ² ) (where SIGN (Μ ₁ −μ ₂ ) is +1 for positive declination and −1 for negative declination, μ _i is the average of the i-th region, and σ _i is the i-th region. A standard deviation.) According to the formula. 24. The method of claim 23, wherein the measurement comprises a pixel value. 25. The method of claim 23, wherein the measured value comprises a Rose E5L5 Bower Fisher ratio. 26. A method of obtaining a feature from a plurality of detections on an image, the method comprising: a) obtaining a measurement value from each detection of the plurality of detections on the image; and b) each detection on the image. Combining the measurements from all the detections to produce a unique feature for. 27. Combining the measurements, i) ordering all detections on the image by grade based on the value of the measurement for each detection, and ii) the grade of each detection. 27. The method of claim 26, comprising using a value as a characteristic of the detection. 28. Combining the measurements includes calculating an average of the measurements for detection on the image and combining the averages with the image to produce unique features for each detection. 27. Subtracting from the measurement value for each detection above. 29. A method of obtaining features from a plurality of masks corresponding to detections on an image, the method comprising: a) providing a plurality of masks corresponding to one detection on the image; and b) for each mask. C) ordering the masks by grade based on the quality score, d) selecting a subset of masks based on the grade value of the mask, and e) selecting the mask A method comprising: calculating a plurality of measurement values using a subset; and f) obtaining a feature from the plurality of measurement values. 30. The method of claim 29, wherein at least one of the features is an average of measurements. 31. The method of claim 29, wherein at least one of the features is a standard deviation of measurements. 32. The method of claim 29, wherein at least one of the features is a minimum of measurements. 33. The method of claim 29, wherein at least one of the features is a maximum of measurements. 34. A method for selecting a subset of detections on an image set in a process for analyzing detections on a digital image set comprising: a) for identified detections on the image set. , B) ordering the detections on the image set by grade based on the quality score assigned to the identified detections, and c) 1 to N, where N is , A predetermined upper bound on the number of detections retained on the image set), retaining only the detections on the image set. 35. The quality score indicates the probability of detection corresponding to a cancer area on an image set, and the probability of detection corresponding to a cancer area progressively decreases from 1 to N. 35. The method of claim 34, wherein: 36. The method of claim 34, wherein the image set comprises a plurality of images. 37. The method of claim 36, wherein the step of b) ordering by grade is performed with reference to all images in the image set. 38. The method of claim 37, wherein all images in the set comprise images from one case. 39. The method of claim 34, wherein the detection comprises two categories, microcalcification and concentration, and each category is separately ordered by grade. 40. The method of claim 39, wherein separate upper bounds N ₁ , N ₂ are defined for each detection category. 41. The method according to claim 40, wherein the upper limit N ₁ for microcalcification is different from the upper limit N ₂ for concentration. 42. The image set comprises a plurality of images, and 1-M (
Where M is an upper limit for a given image with respect to the number of detections retained on the image), only those detections graded for each detection category are retained on each image. 41. The method of claim 40, wherein: 43. The method of claim 42, wherein the per-image upper limit M ₁ for microcalcification is different than the per-image upper limit M ₂ for density. 44. A method for analyzing detections on a digital image, comprising the steps of: a) providing a quality score of the image context for the identified detections on the image; and b) quality for detection. Assigning a credit rating to the identified detections based on the score; c) comparing all identified detection credit ratings on the image to a predetermined credit rating combination; and d) a step. c), removing all detections from the image if the credit grades for all identified detections on the image match one of the predetermined credit grade combinations. Method. 45. The method of claim 44, wherein the quality score is obtained from at least one delta discriminant value for each detection. 46. The method of claim 44, wherein step b) comprises assigning one of two grades to each detection. 47. The method of claim 46, wherein the quality score comprises an average of delta discriminant values for each detection. 48. The quality score is compared with a predetermined threshold to assign a credit grade, and a first credit grade is assigned if the quality score is lower than the threshold, and a quality score is higher than the threshold. 48. The method of claim 47, wherein a second credit rating is assigned. 49. The detections on the image have at least two detection categories that are analyzed together, and step c) compares with reference to both the credit rating and the detection categories for each detection. 49. The method of claim 48, comprising: 50. The method of claim 49, wherein the at least two detection categories comprise microcalcification and concentration. 51. The steps a) to d) are performed on each image of the set of mammography images, and all detections on each image are false positives in the set of images. 51. The method of claim 50, which is either retained or removed to reduce the number of detections. 52. The steps a) to d) are performed on each image of the set of mammographic images, and all detections on each image are false positives in the set of images. 45. The method of claim 44, wherein the method is either retained or removed to reduce the number of detections. 53. A method for analyzing detections on a digital image, wherein at least two detection categories are analyzed together in the image, the method comprising: a) for the identified detections on the image. , Giving a quality score for the image situation, b) assigning a credit rating to the identified detections based on the quality score for the detection, and c) credit for all identified detections on the image. Comparing a grade to a combination of predetermined credit grades across said at least two detection categories, and d) in step c) the credit grades for all identified detections on the image are given credit grades. And removing all detections from the image if they match one of the combinations of. 54. The method of claim 53, wherein the quality score comprises an average of delta discriminant values for each detection. 55. The method of claim 53, wherein the at least two detection categories comprise microcalcification and concentration. 56. The steps a) to d) are performed on each image of the set of mammography images, and all detections on each image are false positives in the set of images. 54. The method of claim 53, which is either retained or removed to reduce the number of detections. 57. A method for selecting or rejecting a category of detections on an image set in a process for analyzing detections on a digital image set, the method comprising: a) selecting from categories of detections on the image set. Giving a calculated value corresponding to each detection, b) calculating a normality value using the calculated value corresponding to the detection, and c) comparing the normality value with a predetermined threshold value. And d) rejecting all detections of the corresponding category if the normality value does not meet a predetermined threshold condition. 58. The method of claim 57, wherein the image set comprises a plurality of digital images and the detection is spread through the image set. 59. The detection comprises at least two detection categories, and a separate normality value is calculated for each detection category, and the step of rejecting the detection is performed in each detection category. 58. The method of claim 57, wherein the method is performed separately. 60. The method of claim 59, wherein the detection category comprises microcalcification and concentration. 61. The method of claim 57, wherein the calculated value corresponding to each of the detections comprises a delta discriminant value. 62. The method of claim 61, wherein the features are derived from delta discriminant values for use in calculating a normality value. 63. The method of claim 62, wherein the features derived from the delta discriminant value are input to a classifier to obtain a normality value. 64. The method of claim 63, wherein the classifier comprises a quadratic classifier. 65. The detection comprises at least two detection categories, and the features derived from the delta discriminant values for each detection category are distinct normality values for each detection category. Used to calculate and
63. The method of claim 62, wherein the step of rejecting detection occurs separately in each detection category. 66. The method of claim 65, wherein the detection category comprises microcalcification and concentration. 67. The features derived from the delta discriminant values for the microcalcifications comprise two features, which are: i) for every image in the image set, image by image. , By sorting the delta discriminant values of microcalcifications, ii) for each image, by identifying the maximum delta discriminant value within the image, iii) the image to define the first feature By searching all the images in the set and selecting the maximum delta discriminant value, iv) the maximum delta discriminant in image from all the images in the image set to define the second feature Calculated by averaging the values, the first and second features of the microcalcification to obtain a microcalcification normality value for comparison with a predetermined microcalcification normality threshold. Characterized by having an input to a normality classifier 67. The method of claim 66, wherein 68. The first and second delta discriminant values are calculated for each concentration detection, and the features derived from the delta discriminant values for concentration comprise three features, which are: , I) by summing the first and second delta discriminant values for each density detection, ii) to define the identified detections for each image, of the summation operation of step i) Identifying the resulting density detection with the maximum value for each image, iii) calculating the average of the first and second delta discriminants for the identified detection in each image. Iv) by selecting the minimum of the mean values calculated in step iii) to define the first feature, v) to define the maximum delta discriminant value for each image , The maximum delta discriminant value, By selecting for each image from the first and second delta discriminants for separate detection, vi) the maximum delta discriminant for all images to define the second feature. Vii) determining the minimum delta discriminant value for each image by calculating the average of the values, the minimum delta discriminant value being the first and second delta discriminants for the identified detection. By selecting for each image, viii) by defining an average of the minimum delta discriminant values for all images to define a third feature, 67. The second, third, and third features comprise an input to a concentration classifier to obtain a concentration normality value for comparison with a predetermined concentration normality threshold. The method described in. 69. A method for selecting a subset of detections on an image set in a process for analyzing detections on a plurality of images forming a digital image set, comprising: a) each on the image set. Providing a graded order quality score for the detection of b), b) ordering the detections on the image set by grade based on the graded order quality score assigned to each detection, c) 1-N Retained only detections on the image set, where N is a predetermined upper bound on the number of detections retained on the image set, and d) remaining from step c) Assigning a credit rating to each detection based on the quality score of the image context for each detection, and e) each image. To compare all the credit grades for detection on the image with a predetermined combination of credit grades, and f) the credit grade for detection on the image is one of the combinations of predetermined credit grades. If there is a match, then removing all detections from every image, g) giving a calculated value corresponding to each detection on the category of detections on the image set, and h) calculating corresponding to the detections. The normality value is calculated using the calculated value, i) the normality value is compared with a predetermined threshold value, and j) if the normality value does not satisfy the predetermined threshold value condition, all detections are performed. And a step of rejecting. 70. A process for analyzing detections on a plurality of images forming a digital image set, the detections having at least two detection categories comprising microcalcifications and concentrations. A method for selecting a subset comprising: a) providing for each detection on an image set a graded order quality score derived from at least one delta discriminant value for each detection; b) Ordering the detections of each detection category on the image set by grade based on the grade order quality score assigned to the detections; c) 1 to N ₁ (where N ₁ is retained on the image set) Of microcalcifications on a set of images, which is graded to a given upper limit on the number of microcalcifications A step of holding the output only, d) 1 to N ₂ (where, N ₂ is the predetermined upper limit on the number of levels retained on the image set) is graded, the concentration on the image set Retaining only the detections of, e) ordering the remaining detections within each category by grade for each image, f) for each image, ₁ to M ₁ (where , M ₁ is an upper limit for each given image on the number of microcalcifications retained on the image set), retaining only detections of microcalcifications graded, and g) for each image To hold only the detection of densities graded from 1 to M ₂ (where M ₂ is an upper limit for a given image with respect to the number of densities held on the image set). h) For each detection remaining after step g), Assigning a credit grade determined with reference to the average of the delta discriminant values for: i) a combination of microcalcification and concentration detection and associated credit grades for each image , Together with a predetermined detection combination and its associated credit grade, j) on the image, the combination of the detection and its associated credit grade relates to the predetermined detection combination and its associated credit grade. Removing all detections from every image if it matches one of the credit grades, k) calculating features for each detection category on the image set, and l) normality for each detection category. Inputting features from the detection categories into the respective classifiers to give values, m) for each detection category, a normality value Comparing with a respective predetermined threshold, n) rejecting all detections for the category if the normality value for the category does not meet the predetermined threshold condition. And how to. 71. A system for analyzing detections on a digital image set and for selecting a subset of detections on the image set, comprising: a) for identified detections on the image set. A means for providing a graded order quality score, and b) based on the graded order quality score assigned to the identified detections,
Means for ordering the detections on the image set by grade, c) 1 to N, where N is a predetermined upper limit on the number of detections retained on the image set, Means for retaining only detections on the image set. 72. The quality score indicates the probability of detection corresponding to a cancer area on an image set, and the probability of detection corresponding to a cancer area decreases from 1 to N. 72. The system of claim 71, wherein: 73. The system of claim 71, wherein the image set comprises a plurality of images. 74. The means for ordering by grade refers to all images in an image set and orders by grade.
The system described in. 75. The system of claim 74, wherein all images in the set comprise images from one case. 76. The detection comprises two categories, microcalcifications and concentrations, and the means for ordering by grade determines the grade order for each category separately. 72. The system according to claim 71. 77. The system of claim 76, wherein a separate upper bound N ₁ , N ₂ is defined for each detection category. 78. The system of claim 77, wherein the upper limit N ₁ for microcalcification is different than the upper limit N ₂ for concentration. 79. The image set comprises a plurality of images, and 1-M (
Where M is an upper limit for a given image with respect to the number of detections retained on the image), only those detections graded for each detection category are retained on each image. 78. The system of claim 77, wherein the system is: 80. The system of claim 79, wherein the per-image upper limit M ₁ for microcalcifications is different than the per-image upper limit M ₂ for density. 81. A system for analyzing detections on a digital image, comprising: a) means for providing a quality score of the image context for the identified detections on the image; and b) for detection. Means for assigning a credit rating to the identified detections based on the quality score of c., C) for comparing all identified detection credit ratings on the image with a given combination of credit ratings And d) means for removing all detections from the image if the credit grades for all identified detections on the image match one of the predetermined credit grade combinations. System characterized by. 82. The system of claim 81, wherein the quality score is obtained from at least one delta discriminant value for each detection. 83. The system of claim 81, wherein the means for assigning a credit grade comprises means for assigning one of two grades to each detection. 84. The system of claim 83, wherein the means for providing the quality score provides an average of delta discriminant values for each detection. 85. The means for providing the quality score comprises comparing the quality score to a predetermined threshold to assign a credit grade, and assigning a first credit grade if the quality score is below the threshold; 85. A second credit grade is assigned if the quality score is higher than a threshold value.
The system described in. 86. The detection on the image has at least two detection categories, and the means for comparing credit ratings refers to both the credit rating and the detection category for each detection. 86. The system of claim 85, wherein the comparisons for the two detection categories are performed together. 87. The system of claim 86, wherein the at least two detection categories comprise microcalcification and concentration. 88. The means for providing the quality score provides a quality score for a set of mammography images, and all detections on each image are the number of false positive detections in the set of images. 88. The system of claim 87, which is either retained or removed to reduce the. 89. The means for providing the quality score provides a quality score for a set of mammographic images, and all detections on each image are the number of false positive detections in the set of images. 82. The system of claim 81, which is either retained or removed to reduce the. 90. A system for analyzing detection on a digital image, wherein at least two detection categories are analyzed together in the image, said system comprising: a) for identified detection on the image. , A means for giving a quality score for the image situation, b) a means for assigning a credit rating to the identified detections based on the quality score for the detections, and c) all the identified detections on the image Means for comparing a credit rating for a given credit rating to a given credit rating combination, and d) all identified detection credit ratings on the image are consistent with one of the given credit rating combinations. And a means for removing all detections from the image. 91. The system of claim 90, wherein the means for providing the quality score provides an average of delta discriminant values for each detection. 92. The system of claim 90, wherein the at least two detection categories comprise microcalcification and concentration. 93. The means for providing the quality score provides a quality score for a set of mammographic images, and all detections on each image are the number of false positive detections in the set of images. 91. The system of claim 90, which is either retained or removed to reduce the. 94. A system for selecting or rejecting a category of detections on an image set, comprising: a) providing a calculated value corresponding to each detection from the category of detections on the image set. Means, b) means for calculating a normality value using the calculated value corresponding to the detection, c) means for comparing the normality value with a predetermined threshold, and d) normality. Means for rejecting all detections of the corresponding category if the value does not meet a predetermined threshold condition. 95. The system of claim 94, wherein the image set comprises a plurality of digital images and the detection is spread through the image set. 96. The detection of claim 94, wherein the detection comprises at least two detection categories and has means for calculating a separate normality value for each detection category. System. 97. The system of claim 96, wherein the detection category comprises microcalcification and concentration. 98. The system of claim 94, wherein the means for providing a calculated value corresponding to each of the detections provides a delta discriminant value. 99. The system of claim 98, comprising means for deriving a feature from a delta discriminant value for use in calculating a normality value. 100. Means for inputting a feature obtained from the delta discriminant value to a classifier to obtain a normality value.
The system described in. 101. The system of claim 100, wherein the classifier comprises a secondary classifier. 102. The detection comprises at least two detection categories, and the means for characterization obtains each detection category to calculate a separate normality value for each detection category. 101. The system of claim 100, wherein the features are obtained from delta discriminant values for. 103. The system of claim 102, wherein the detection category comprises microcalcification and concentration. 104. The feature derived from the delta discriminant value for said microcalcification comprises two features, and the means for obtaining the feature are: i) for all images in the image set. A means for sorting the delta discriminant values of microcalcifications for each image; ii) a means for identifying the intra-image maximum delta discriminant values for each image; iii) a first Means for retrieving all the images in the image set and selecting the maximum delta discriminant value to define the feature of iv) and iv) to define the second feature Means for averaging the intra-image maximum delta discriminant values from all images; and v) a microcalcification normality classifier receiving said first and second features, said microcalcifications Normality classifier compares against a predetermined microcalcification normality threshold The system of claim 103, wherein the generating the health value of microcalcifications in order. 105. The first and second delta discriminant values are calculated for each concentration detection, and the features derived from the delta discriminant values for concentration comprise three features, which are: I) means for summing the first and second delta discriminant values for each density detection; and ii) summing to define the identified detections for each image The concentration detection with the maximum value resulting from the summation operation performed by the means,
Means for identifying for each image, iii) means for calculating an average of the first and second delta discriminants for the identified detections in each image, iv) first Means for selecting the minimum of the mean values calculated by the means for calculating the average of said first and second delta discriminants to define a feature; and v) a maximum for each image. Means for selecting for each image a maximum delta discriminant value from the first and second delta discriminants for identified detection to define a delta discriminant value; vi) A means for computing the average of the maximum delta discriminant values for all images to define the two features, and vii) a minimum to define the minimum delta discriminant value for each image. The delta discriminant value is used as the first value for the identified detection. And means for selecting for each image from the second delta discriminant, and viii) for computing the average of the minimum delta discriminant values for all images to define the third feature And ix) a concentration normality classifier that receives the first, second, and third features, the concentration normality classifier being compared to a predetermined concentration normality threshold. 104. The system of claim 103, wherein the system produces a normality value for the concentration.