WO1999028854A1

WO1999028854A1 - Method and system for automated multi-sampled detection of lesions in images

Info

Publication number: WO1999028854A1
Application number: PCT/US1998/024932
Authority: WO
Inventors: Robert M. Nishikawa; Kunio Doi
Original assignee: Arch Development Corporation
Priority date: 1997-11-28
Filing date: 1998-11-25
Publication date: 1999-06-10
Also published as: AU1597899A

Abstract

Method and system for automated detection of lesions in medical images using multi-sampling. An image is obtained and digitized. A multi-sampling (30) technique is used to generate a plurality of sampled images by sampling the digital image a corresponding plurality of times. The sampled images will have subtle differences from each other. All or a selected number of the sampled images are subjected to a lesion detection routine to detect regions suspected of being a lesion. The results of the detection are compared and regions that appear in less than a predetermined number of the sampled images are determined to be a false positive and eliminated. The remaining regions are determined to be a true positive. The detected lesions (13) could then undergo further analysis such as feature extraction (36) and feature analysis. Using the method and system according to the invention, the number of false positives can be reduced, improving the sensitivity of the automated detection.

Description

TITLE OF THE INVENTION

METHOD AND SYSTEM FOR AUTOMATED MULΗ-SAMPLED DETECTION OF LESIONS IN IMAGES

CROSS-REFERENCE TO RELATED PATENTS AND PATENT APPLICATIONS

This application is related to techniques for automated detection of abnormalities in

images, for example as disclosed in one or more of U.S. Patents 4.839,807; 4,841.555;

4,851.984; 4,875.165; 4,907,156; 4,918,534; 5,072,384; 5.133.020: 5,150.292; 5,224,177;

5,289.374; 5.319.549; 5.343.390: 5.359.513; 5,452,367; 5,463.548: 5.491.627; 5.537.485;

5.598.481 ; 5,622.171; 5.638.458; 5,657.362; 5,666.434: 5,668.888: and 5.673.332. as well as

U.S. Patent applications 08/158,388; 08,173,935; 08/220,917; 08/398,307; 08/428,867;

08/523.210; 08/536,149; 08/536,450; 08/515,798; 08/562,087; 08/757,61 1 ; 08/758,438;

08/900.188; 08/900,189; 08/900.191 ; 08/900,192; 08/900,361 ; 08/900.362; and 08/979 ,623

(Attorney Docket No. 730-018-20 filed on Nov. 28, 1997), each of which is incorporated

herein by reference.

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH

The present invention was made in part with support from U.S. Government grant

number NIH NCI CA 60187. The U.S. Government has certain rights in the invention.

BACKGROUND OF THE INVENTION

Field of the Invention

The present invention is related to detecting lesions in images and, more particularly. r to reducing the number of false positive detections in detecting lesions in medical images. Discussion of the Background

Digital images are commonly used in the detection of lesions. Digitized chest

radiographs have been used for the detection of lung nodules, such as in U.S. 5,289,374.

Digitized chest radiographs have also been used in the detection of pneumothorax (U.S.

5,668,888) and for texture analysis of infiltrates (U.S. 5,319,549). The techniques are

designed to maximize the detection of lesions (true positives) while minimizing false positive

detections.

Digital images are also commonly used in mammography. The images have been

used to detect many types, of lesions, including clustered microcalcifications. There are many

different approaches to the detection of clustered microcalcifications from mammograms. In

all previously reported approaches, a single version of a given mammogram is digitized and

analyzed. In all these methods both actual clusters of microcalcifications (true positives) and

false clusters (not an actual cluster of microcalcifications) will be detected. The goal of these

automated approaches is also to maximize the number of true positives while minimizing the

number of false clusters detected. It has been shown, however, that if the same mammogram

is digitized and subjected to an automated detection method multiple times, then in general,

the same true clusters are detected the majority of times, while many of the false clusters may

be detected on only a minority of the times. This is reported in R. M . Nishikawa, J.

Papaioannou and S. A. Collins, "Reproducibility Of an Automated Scheme for the Detection

of Clustered Microcalcifications on Digital Mammograms," Proc. SPIE 2710, 24-29 (1996),

which is hereby incorporated by reference.

For additional background for the invention, see R. M. Nishikawa, M. L. Giger, K.

Doi, C. J. Vyborny and R. A. Schmidt. "Computer-aided Detection and Diagnosis of Masses

.?. and Clustered Microcalcifications from Digital Mammograms." in State of the Art in Digital

Mammographic Image Analysis, edited by K. W. Bowyer and S. Astley (World Scientific

Publishing Co. Ltd., London 1994) pp. 82-102, B. Zheng, Y. H. Chang, M. Staiger. W. Good

and D. Gur, "Computer-aided Detection of Clustered Microcalcifications in Digitized

Mammograms." Academic Radiology, vol. 2, pp. 655-662 (1995), H. P. Chan. S.C. B. Lo. B.

Sahiner, K. L. Lam and M. A. Helvie, "Computer-aided Detection of Mammographic

Microcalcifications: Pattern Recognition with and Artificial Neural Network," Medical

Physics, vol. 22, pp. 1555-1567 (1995), and H. D. Li, M. Kallergi, L. P. Clarke, V. K. Jain

and R. A. Clark, "Markov Random Field for Tumor Detection in Digital Mammography,"

IEEE Trans. On Med. Imaging, vol. 14, pp. 565-576 (1995), each of which is also

incorporated herein by reference.

SUMMARY OF THE INVENTION

It is an object of the present invention to improve the automated detection of lesions

in medical images.

It is another object of the present invention to eliminate false positive detections in the

automated detection of lesions in medical images.

A further object of the present invention is to improve the detection of clustered

microcalcifications in mammograms.

These and other objects of the invention are obtained by the method and system

according to the invention. In the method, an image is obtained which may contain a region

suspected of being a lesion. The image is sampled a plurality of times to produce a

corresponding plurality of sampled images. Using the sampled images, lesions are detected

- and false positives are eliminated. The sampled images are sampled at different starting

points in the image or. in other words, using different groups of pixels. Thus, the sampled

images will be subtly different from each other. Since false positives may be due to noise in

the image or some other artifact, the false positive may not appear in each of the sampled

images.

The plurality of sampled images may be compared to determine whether the regions

suspected of being a lesion are present in each of the sampled images. An occurrence

threshold may be determined whereby if a region is present in only a selected number of the

sampled images (or less than a selected number of the sampled images), the region can be

determined to be a false positive thus eliminated from further consideration. Thus, the

accuracy of lesion detection is improved.

In the method according to the invention, the image may be digitized and the

sampling be carried out digitally. The image is digitized at a first pixel size and selected

number of the pixels are combined to form second pixels which make up the sampled images.

The sampling may be carried out by selecting pixel units, such as a sub-grid of the matrix of

digitized pixels, and combining the pixels in the pixel unit. The pixels may be combined by,

for example, pixel averaging, selecting the highest or lowest value of the pixels in the unit,

selecting the median value of the pixels in the unit or selecting an intermediate value of the

pixels in the pixel unit. A texture measure may also be used. The combined pixels make up

the pixels in the sampled images which are used to detect lesions and eliminate false

positives.

The method according to the invention can also include preprocessing of the image to

help eliminate image noise or artifacts. The method may also include feature extraction and feature analysis of the regions determined to be lesions after elimination of false positives by

the sampling technique. Also, feature extraction and analysis may be used on the sampled

images to perform the lesion detection and false positive elimination.

The sampling step may be implemented at a number of locations in the method. For

example, the sampling could take place along with the digitization or after the preprocessing

step. The sampling could also occur after identification of regions in an image which are

suspected of being a lesion. Using these locations, regions of interest may be selected

containing these locations and the sampling carried out on the regions of interest. Sampling

only the regions of interest should shorten the sampling time. The sampled regions of interest

would then be subjected to lesion detection and the occurrence thresholding to detect lesions

and eliminate false positives.

The system according to the invention includes a data acquisition device, such as an

x-ray machine. The image may be scanned using a digital scanning before sampling. The

sampling is carried out using sampling device and the sampled images are then fed to a lesion

detection apparatus for detection of lesions and eliminations of false positives. The system

can also include a feature extraction and analysis device to perform the lesion detection and

false positive reduction, or to further analyze lesions detected using the sampling technique.

The system may be advantageously implemented in software, allowing flexibility of how the

sampling is carried out and at what point in the method or system the sampling is performed.

BRIEF DESCRIPTION OF THE DRAWINGS

A more complete appreciation of the invention and many of the attendant advantages

thereof will be readilv obtained as the same becomes better understood bv reference to the following detailed description when considered in connection with the accompanying

drawings, wherein:

FIG. 1 is a block diagram of an embodiment of the system according to the invention;

FIG. 2 is a flow diagram of an embodiment of the method according to the invention;

FIGS. 3A-3D are diagrams illustrating multi-sampling according to the invention;

FIGS. 4A-4D are diagrams illustrating multi-sampled images according to the

invention;

FIG. 5 is a flow diagram of an embodiment of the method according to the invention;

and

FIG. 6 is a block diagram of an embodiment of the system according to the invention.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

With reference to the drawings, particularly FIG. 1 a first embodiment of the

invention will be described. FIG. 1 shows a block diagram of the system according to the

invention. An image, such as a chest radiograph or a mammogram is taken by image

acquisition device 10. In these two instances device 10 is an x-ray device, but other image

acquisition devices may be used. The output of device 10 is fed to a digitizing scanner 1 1.

which may be a laser scanner. Imaging device 10 and scanner 1 1 may be combined into a

single unit. Scanner 11 creates a digital image of the image obtained by device 10. The pixel

size of scanner 11 can be chosen given the desired level of resolution and taking into

consideration computational efficiency. A typical pixel size for a digital medical image is

about 50 μm.

The digitized image is sampled by sampling device 12. Device 12 performs multi- sampling on the digitized image. For example, the resolution of the image digitized at 50 μm

pixel size can be reduced to 100 μm. i.e., by combing four adjacent pixels. Larger numbers

of pixels of the digitized image can be combined. Also, the digitization pixel size can be

reduced, while keeping the multi-sampled image resolution the same to produce a larger

number of multi-sampled images. The multi-sampling is used to create multiple sampled

images from the single digitized image. This process is described in more detail below.

The multi-sampled images generated by device 12 are fed to a lesion detection

apparatus 13. Any type of detection apparatus may be used. Detection apparatus 13 could

be. for example, a trained neural network, a feature extraction and analysis device, wavelet

analysis, fuzzy network or a combination of the any of the above. Apparatus 13 can be

conveniently implemented in software and run by a general or specialized computer.

The system may also advantageously include a output device 14 to output an image

showing detected lesions, or regions suspected of being a lesion. Output device 14 is

typically a video display terminal or monitor, a printer, or both.

All of the components of the system are controlled by system controller 15.

Controller 15. for example, allows an operator to select the various imaging parameters used

in the image acquisition device 10 to obtaining the image, the parameters used in the scanner

11 to generate the digitized image, the parameters used in the apparatus 12 to detect lesions or

suspicious regions, and the output device 14 parameters. Controller manages the transfer of

the data between the components and storage of the data. The system controller can be a

personal computer or minicomputer which can also perform the various lesion and/or

suspicious region detection, and the lesion detection techniques such as feature extraction and

analysis. The operation of the system will now be described in more detail. When a film is

digitized, it is sampled by an aperture of a given size at a given sampling distance. Typically,

the sampling distance is equal to the sample aperture. The pixel size is equal to the sampling

distance. When a film is digitized, the values produced are averaged over the surface of the

film corresponding to the sample aperture. Because the point at which the digitizer starts to

scan the film is arbitrary, it is likely that if the same image is digitized more than once, each

digitized version will be slightly different from the others. The main reason for this is that

there will be small shifts in the starting point of the scan that will not be equal to an integral

multiple of the sampling distance. That is. the shift will correspond to a fraction of the

sampling distance and therefore the digitizer will average over slightly different areas of the

film. Therefore, even though each digitized version of an image will look identical, careful

examination of "corresponding" pixels will indicate subtle differences in the images.

The present invention takes advantage of this phenomenon. To illustrate this, the

method according to the invention will be described for the case of digitizing a medical

image. However, the invention is not limited to this one example and can be applied to

various types of images and detection. A flowchart illustrating the method is shown in FIG.

2. After the medical image is acquired using acquisition device 10. the image is digitized at

50 μm pixel resolution in step 20. The digitized image is reduced to a 100 μm pixel size by

multi-sampling (via sampling device 12) in step 21. As an example, four different starting

pixel locations were chosen to produce 4 unique sampled images. At a given pixel location in

the 50-μm acquired image, four pixels are combined to produce one pixel in the multi-

sampled (100-μm) image. The four pixels that are combined into one pixel are (x, y), (x+1,

y), (x. y+1). (x+1, y+1), where x is the column number and y is the row number. The offsets in this case are a multiple of two pixels, i.e., 1, 3, 5, .... or 0. 2, 4, ... . etc.

There are four different starting pixel locations, one for each of the four reduced

images. The four starting pixel locations, in terms of (x, y), are (0, 0), (1 , 0), (0, 1) and (1, 1)

assuming that the upper left most pixel in the image is located at (0, 0). That is, the first pixel

in the multi-sampled image is determined from the upper left four pixels of the 50-μm image,

which form a 2x2 pixel grid, specifically pixels (0, 0), (0, 1), (1, 0), and (1, 1), in standard

matrix notation. A second multi-sampled image, different from this first, can be produced if

the starting grid of four pixels is shifted to the right by one pixel. Now, for example, the

upper left pixel of the second multi-sampled image will be determined from pixels (0, 1),

(0,2), (1 , 1 ) and (1, 2). A third multi-sampled image is produced by shifting the grid down by

one pixel relative to the original 50-μm image, using pixels (1, 0), (1, 1), (2, 0), (2, 1), and a

fourth multi-sampled image is produced by shifting the grid down one pixel and to the right

one pixel, using pixels (1, 1), (1, 2), (2, 1), (2, 2).

FIGS. 3A-3D and 4A-4D illustrate the generation of the multi-sampled images.

FIGS. 3A-3D are a hypothetical 25 pixels segment of a 50-μm image, with pixel values

indicated. FIGS. 4A-4D are the corresponding 100-μm multi-sampled images. The pixels

values given in FIGS. 4A-4D are determined using pixel averaging of the corresponding four

pixels in the bold-line boxes in FIGS. 3A-3D. As is apparent from FIGS. 4A-4D. the four

images are each different from each other.

It should be pointed out that there are several different techniques for combining or

sampling the pixels, such as averaging the pixels, taking their median value, taking their

highest value, taking their second highest value, taking their third highest value, taking their

lowest value, and taking their average plus a multiple of the standard deviation of the pixels. Still other methods may be used in connection with the invention.

Furthermore, different numbers of pixels can be combined to produce correspondingly

more reduced images. The image may also be digitized at a smaller pixel size. For example,

if the image was digitized at 33 μm, nine pixels (in a 3x3 format) could be combined to

produce nine different reduced images. Because there are more reduced images produced, the

threshold value can range between one and nine. This could produce an even more effective

reduction in the false positive rate. The number of pixels combined to produce the reduced

images could be selected based upon diagnostic effectiveness for the type of lesion desired to

be detected. For example, if a reduced image pixel size of 150 μm is adequate, a 50-μm pixel

image could be used and nine pixels (3x3) could be combined to produce nine different

versions of a 150-μm pixel image.

It is also possible to use a selected number of the multi-sampled images for false

positive reduction. In the case where nine images are generated, any number of the images

can be selected for further processing, lesion identification or lesion detection. The choice

could be made based upon system parameters, such as storage capability or computational

capability, or upon the effectiveness of the false positive reduction or true positive detection.

Because the four 100-μm images are subtly different from one another, the detection

scheme used in detection apparatus 13 will not produce exactly the same output when

analyzing them. A lesion or a suspicious region can be determined by requiring a detection to

occur in a selected number of the 4 detection results, that is, a threshold test (step 22). It will

generally be more advantageous, that is, more false positives can be eliminated, by increasing

the number of required detection results. If a detection is required to appear in at least two of

the images in order to be reported, then some candidate detections (which appear in only one of the reduced images) will be eliminated. If a detection is required to appear in three of

reduced images in order to be reported, many more candidate detections will be eliminated,

and the largest number may be eliminated by requiring that a cluster appear in all of the

reduced images. The occurrence threshold (OT) is defined to be the number of reduced

images in which any candidate detections are required to appear in order to be reported.

In the case of FIGS. 4A-4D, the threshold is 1, 2, 3, or 4 images. The OT can thus

have values 1, 2, 3, or 4. In the general case, with M images, n detections are required

resulting in 1 < OT ≤ M. A cluster detected in less than the value OT images is eliminated

because it is considered to be a false detection. After eliminating some or all of the false

positives, the remaining candidates are considered to be lesions (step 23).

The detection routine used on the multi-sampled images may be the only detection

applied (or needed). If, however, this routine is rather complicated or requires significant

computations, a simpler detection or identification routine can be applied prior to the

thresholding, followed by a further, possible more complicated and more accurate detection

technique or techniques to improve the detection results.

It is also possible to combine the results of the multiple images at some other point in

the method or in the system, as shown in FIGS. 5 and 6. Instead of combining the results at

the output of the different reduced images, the results could be combined at some

intermediate point. Typically, there are three different steps in the detection of lesions,

processing (or preprocessing), signal identification, and detection (feature analysis). The

processing (or preprocessing) step typically consists of background-trend correction, signal or

image enhancement or some noise-reduction technique. The application of the multi-

sampling during or after preprocessing is rather straightforward. The image is multi-sampled either as part of the processing routine, or after the processing routine, to generate the multi-

sampled images. The sampled images are then subjected to lesion identification and

detection.

The second step is the identification of all possible lesions or signals. This step could

occur with or without the preprocessing routine. The identification routine produces location

information that is subsequently used in lesion detection. Multi-sampling would be applied

after the identification, so the detection would take place on the multi-sampled images using

the location information.

Since the locations of the suspicious regions are known, the multi-sampling could be

selectively applied to the suspicious region and a surrounding area. In other words, using the

location information, regions-of-interest (ROIs) of a specified size are selected in the image

and multi-sampling applied to generate a plurality of multi-sampled ROIs. The lesion

detection (feature analysis) then is applied to the multi-sampled ROIs. This is also illustrated

in FIGS. 3A-3D and 4A-4D where the 5 x 5 matrix of cells is the region-of interest and the

four 4 x 4 multi-sampled matrices are the multi-sampled ROIs.

In addition, this technique could be applied during the acquisition of the digital image.

If film digitization is used, then the digitizer could scan the film at 50 μm pixel size and the

image could be reduced to a 100-μm image, as outlined above. For a direct digital detector,

the data again could be acquired at high spatial resolution and then combined in either

hardware or software to obtain a reduced image at slightly lower spatial resolution.

The sampling circuit could also be available for use at all of these points, for

maximum flexibility in lesion identification and detection. To implement this, the sampling

could be performed as a subroutine available for use at any point in a software program designed to carry out all of the processing, identification, detection, feature extraction, etc.

operations.

FIG. 5 shows an example of the application of the multi-sampling at various points in

a lesion detection method. In step 31. the image, preferably a digital image, is obtained.

Next, an optional preprocessing step 32 is performed. The multi-sampling step 30 may be

performed as part of or after each of steps 31 and 32. The digital image or the preprocessed

image, either before or after multi-sampling is then subjected to the lesion detection or

identification routine in step 33. Step 33 identifies or detects regions in the input image

suspected of being a lesion. If step 33 is performed using a multi-sampled image,

thresholding (step 34) may be performed to eliminate false positives and to give detected

lesions (step 35). The result of step 35 may be output (on a device such as output device 14).

Step 33 may be performed on an image before multi-sampling. In this case, multi-

sampling is followed by feature extraction and analysis (step 36), or by thresholding (step

34). The feature analysis and extraction can use the suspect region identification information

obtained in step 33 to point to regions requiring the extraction and analysis. The suspect

region identification information may also be used to define ROIs around the suspect region,

followed by multi-sampling of the ROIs. The ROIs should be chosen large enough to contain

the suspected lesion and to take advantage of the multi-sampling. The size of the ROI is

based upon the lesion desired to be detected. For example, a 2 cm by 2 cm ROI is usually

adequate for clustered microcalcifications in a mammogram. The multi-sampling is then

performed only on the ROIs. Multi-sampling only the ROIs can reduce the time needed for

the sampling process and subsequent analysis.

Step 36 may be performed without performing step 33 on either the digital image after

-ι: multi-sampling or the preprocessed image after multi-sampling. In these two instances,

thresholding step 37 is performed to eliminate false positives and lesions are detected (step

35). The multi-sampling could be incorporated into step 36. Step 36 may also be performed

after thresholding and the lesions are detected (steps 34 and 35) for further analysis of the

detected lesions and/or false positives. The feature extraction and analysis may further reduce

the number of false positives.

FIG. 6 is a block diagram of a system where the multi-sampling is included at a

selected point in the processing of the image carried out by the system. In FIG. 6 the dashed

lines show the various connections that may be made with a sampling circuit 40. An image is

acquired with device 10 and digitized with scanner 11. As discussed above, these two

devices may be combined. The sampling circuit 40 may be included as a part of the scanner

11 or used to sample the digital image produced by scanner 11. In the later case the sampled

images may be then fed to (optional) preprocessing device 41.

Sampling circuit 40 may also be configured to receive the output image of

preprocessing circuit 41 and used the preprocessed image to produce the sampled images.

Sampling circuit 40 may also be included as a part of preprocessing circuit 41. Here, the

sampled images are then sent to signal identification or detection circuit 42 for identification

or detection of regions suspected of being a lesion. Circuit 42 can also perform the

occurrence thresholding processing for detection of lesions and elimination of false positives.

In this case, under the control of controller 15, the detected lesions may be output to output

device 14 or sent to feature extraction device 43 for further analysis.

Further, sampling circuit 40 could be configured to receive the output of circuit 42 (or

be included therein ) and use the location or identification information or regions suspected of being a lesion in the generation of the multi-sampled images. The detection or identification

information generated by circuit 42 may be used to position ROIs in the image containing the

regions suspected of being a lesion so that the multi-sampling may be implemented on only

the ROIs. Multi-sampling only the ROIS may reduce the time needed for the computations.

The ROIs should again be chosen large enough to contain the suspected lesion and to take

advantage of the multi-sampling. The occurrence thresholding may be performed at this

point on the sampled images to detect lesions, and the results may be output to output device

14 for viewing.

The resulting plurality of multi-sampled images using either the entire image or

regions-of-interest may be further analyzed by feature extraction and analysis device 43. The

occurrence thresholding may be performed prior to feature extraction and analysis, reducing

the amount of computation. Occurrence thresholding may also be performed after feature

extraction and analysis, by incorporating the function into device 43. Also, a single image

produced after occurrence thresholding of the multi-sampled images to eliminate false

positives may be subjected to feature extraction and analysis.

It is also possible to carry out the invention eliminating one or more of the operations

performed by the elements of the FIG. 6 system. For example, preprocessing may be omitted

and the digitized image, before or after multi-sampling, is sent to circuit 42 for detection or

identification. Also, the image (digitized or preprocessed) could be multi-sampled and then

feature extraction and analysis performed to detect lesions and eliminate false positives.

EXAMPLE

The method and system according to the invention was applied to detecting microcalcifications in mammograms. A database was generated of fifty-two mammograms

digitized at 50 μm pixel size using a Lumiscan 100 film digitizer. In twenty-six of the

mammograms. a single cluster of microcalcifications was identified by an experienced

radiologist. The remaining twenty-six mammograms were considered to be absent of

clustered microcalcifications.

From the 50-μm image, four 100-μm (reduced) images were generated, using the

multi-sampling methodology described above. The digital image was subjected to

preprocessing using linear filtering, signal extraction using gray-level thresholding, and

feature extraction and analysis of the size, contrast, shape, edge gradient, texture and spatial

distribution of the microcalcifications.

The results of analyzing each of the four reduced images are given in Table 1. Also

given in Table 1 are performance values for the mean and standard deviation in performance

averaged across the runs on the four different reduced images. Here, TP = true positive

(sensitivity) and FP/image = the average number of false positive clusters detected per image.

TABLE I

The results from the multi-sampling according to the invention are shown in Table II.

As the threshold is decreased from four (the cluster must be detected in all four reduced

images) to one (the cluster was detected in at least one of the four reduced images), both the

sensitivity and the false-positive (FP) rate increase. Table II shows four different sampling

techniques. The pixel value in the reduced image is obtained from the four pixel values in the

50-μm image by averaging, determining the median value of the pixels, taking the maximum

value of the four pixels, and taking the second highest value of the four pixels. There are

some differences in performance based on the type of reducing used. In general, using the

maximum pixel value in the reduced image, results in the reduced image being noisier than

using the average value. As a result, the FP rate is higher than with the other 3 reduction

methods. TABLE II

For any reduction method, the FP rate of the detection scheme is reduced compared to

what one would except if only a single 100-μm digitized image was analyzed. This is

because many of the FPs detected are caused by random noise in the image and this noise is

not reproducible in the different reduced images. If a threshold of one or two is used, the FP

rate and the sensitivity are actually higher than the expected value.

As can be seen from Table II, in this example the best performance was obtained at a

threshold of 3 using either averaging or the second highest pixel value as the method for

reducing the image, The sensitivity is not degraded and the false-positive rate is reduced by

35% (from 1.59 to 1.04 FP/image).

Although the method and system according to the invention has been illustrated using

the detection of microcalcifications as an example, the invention is clearly not limited to this

one example. The method can be extended to other detection schemes that either use

different techniques or search for different types of lesions (such as those present in chest

radiographs), or both. The invention is also not limited to radiographic images but can be applied to other types of images, such as the detection of lung nodules in CT images, or the

identification of buildings in an aerial photo. While the invention is well-suited to single-

image analysis, it could be applied to detection routines where multiple images are used.

Numerous further modifications and variations of the present invention are possible in light

of the above teachings. It is therefore to be understood that within the scope of the appended

claims, the invention may be practiced otherwise than as specifically described herein.

Claims

CLAIfrlS:

1. A method of detecting a lesion in an image; comprising:

obtaining an image of a subject;

sampling said image a plurality of times to produce a corresponding plurality of

sampled images; and

detecting a region suspected of being a lesion in said image using said plurality of

sampled images.

2. A method as recited in claim 1 , wherein said sampling step comprises:

sampling said image said plurality of times at a corresponding plurality of different

starting locations.

3. A method as recited in claim 1, wherein said sampling step comprises:

digitizing said image to produce a plurality of first pixels; and

selectively combining said first pixels to produce second pixels.

4. A method as recited in claim 3, comprising:

said first pixels having a first size; and

selectively combining said first pixels to produce second pixels having a second size

larger than said first size.

5. A method as recited in claim 3, comprising:

selecting said first and second pixel sizes based upon one of diagnostic effectiveness

and a size of a lesion desired to be detected.

6. A method as recited in claim 3, wherein said selectively combining step comprises:

selecting a first group of a predetermined number of first pixels; combining said first group using a selected sampling technique to produce a first one of said second pixels;

shifting said first group a selected distance in said image to form a second group of

said first pixels having said predetermined number of pixels;

combining said second group using said selected sampling technique to produce a

second one of said second pixels; and

repeating said shifting and combining steps to form said sampled images.

7. A method as recited in claim 6, comprising:

determining said region to be a false positive when detected in fewer than a

predetermined number of said sampled images.

8. A method as recited in claim 6, comprising:

selecting a plurality of said first groups in said image;

combining pixels in said plurality of first groups using said sampling technique to

form a first one of said sampled images;

shifting each of said plurality of said first groups said selected distance to form a

plurality of said second groups: and

combining pixels in said plurality of second groups using said sampling technique to

form a second one of said sampled images.

9. A method as recited in claim 1, comprising:

digitizing said image into a matrix of first pixels;

dividing said matrix into a plurality of first pixel units each having a selected number

of said first pixels;

sampling said selected number of first pixels in each of said first pixel units to produce a corresponding plurality of second pixels forming a first one of said sampled

images;

shifting said first pixel units at least one of at least one column and at least one row in

said matrix to obtain second pixel units each having said selected number of pixels: and

sampling said selected number of pixels in each of said second pixel units to produce

a corresponding plurality of second pixels forming a second one of said sampled images.

10. A method as recited in claim 9, comprising:

determining said region to be a false positive when detected in fewer than a

predetermined number of said sampled images.

1 1. A method as recited in claim 1 , comprising:

determining said region to be a false positive using results of said detecting step.

12. A method as recited in claim 11, comprising:

detecting whether said region is present in a corresponding location in each of said

sampled images; and

determining said region to be said false positive when said region is present in less

than a selected number of said sampled images.

13. A method as recited in claim 1 1. comprising:

determining said region to be said false positive if said region is not present in at least

one of sampled images.

14. A method as recited in claim 13, comprising:

two of sampled images.

15. A method as recited in claim 1. comprising: obtaining at least four sampled images;

detecting whether said region is present in each of said at least four sampled images;

and

determining said region to be a false positive when said region is not present in at

least two of sampled images.

16. A method as recited in claim 1, comprising:

preprocessing said digitized image to obtain a preprocessed image;

sampling said preprocessed image to obtain said plurality of sampled images;

identifying whether said region is present in each of said plurality of sampled images;

and

analyzing said region.

17. A method as recited in claim 16, comprising:

determining said region to be a false positive when detected in fewer than a

predetermined number of said sampled images.

18. A method as recited in claim 1, comprising:

preprocessing said digitized image to obtain a preprocessed image;

identifying said region in said preprocessed image;

sampling said preprocessed image to produce said sampled images; and

determining whether said region is a false positive using said sampled images.

19. A method as recited in claim 18, comprising:

determining said region to be a false positive when detected in fewer than a

predetermined number of said sampled images.

20. A method as recited in claim 1, comprising: identifying said region in said digital image suspected of being a lesion;

defining a region-of-interest in said image containing said region; and

sampling only said region-of interest a plurality of times to produce said plurality of

sampled images.

21. A method as recited in claim 20, comprising:

determining said region to be a false positive when detected in fewer than a

predetermined number of said sampled images.

22. A method as recited in claim 20, comprising:

identifying a plurality of said regions in said digital image suspected of being a lesion;

defining a corresponding plurality of regions-of-interest in said image containing said

regions; and

sampling only each of said regions-of interest a plurality of times to produce said

plurality of sampled images.

23. A method as recited in claim 1, comprising:

identifying regions in said digital image suspected of being said lesion;

determining at least one of said regions not to be said lesion using said plurality of

sampled images; and

performing analysis on remaining ones of said regions to detect said lesion.

24. A method as recited in claim 23, comprising:

determining said region to be a false positive when detected in fewer than a

predetermined number of said sampled images.

25. A method as recited in claim 1, wherein said sampling step comprises digitizing

said image a plurality of times.

26. An automated method of detecting a lesion, comprising:

obtaining an image containing a region suspected of being said lesion;

sampling said image using said location a plurality of times;

determining whether said region is said lesion using said plurality of samplings.

27. A method as recited in claim 26, comprising:

detecting whether said region is present in a selected number of said samplings;

determining said region to be a false positive when said region is detected in less than

a selected number of said plurality of samplings.

28. A method as recited in claim 26, comprising:

detecting whether said region is present in each of said samplings.

29. A method as recited in claim 26, comprising:

determining a location of said region in said image; and

sampling said image in a region-of-interest containing said location a plurality of

times to produce a plurality of sampled regions-of-interest.

30. A method as recited in claim 29, comprising:

sampling said image only in said region-of-interest containing said location said

plurality of times to produce said plurality of sampled regions-of-interest.

31. A method as recited in claim 26, wherein said determining step comprises:

detecting a plurality of regions suspected of being said lesion in said image;

determining whether said regions are said lesion using said plurality of samplings; and

determining a region to be a false positive if said region is present in less than a

predetermined number of said samplings.

32. A method as recited in claim 31, comprising: determining at least one of said regions to be said false positive; and

performing one of feature extraction and feature analysis on remaining ones of said

regions.

33. A method as recited in claim 26, wherein said determining step comprises:

performing one of feature extraction and feature analysis on said samplings; and

determining said region to be a false positive if said region is present in less than a

predetermined number of said samplings.

34. An automated lesion detection system, comprising:

an image acquisition device;

a multi-sampling circuit connected to said image acquisition device; and

a lesion detection circuit connected to said multi-sampling circuit.

35. A system as recited in claim 34, wherein:

said image acquisition device comprises a digital scanner producing a digital image;

said multi-sampling circuit digitally samples said digital image a plurality of times to

produce a corresponding plurality of sampled images.

36. A system as recited in claim 35, wherein said lesion detection circuit comprises:

means for detecting whether at least one of a true positive and a false positive is

present in each of said plurality of sampled images; and

means for eliminating at least one false positive using results of said means for

detecting from at least one of said plurality of sampled images.

37. A system as recited in claim 36, wherein said lesion detection circuit comprises:

detecting from at least two of said plurality of sampled images.

38. A system as recited in claim 34, wherein:

said system comprises an image preprocessing circuit connected to said image

acquisition device and adapted to produce a preprocessed image from said digital image; and

said multi-sampling circuit digitally samples said preprocessed image a plurality of

times to produce a corresponding plurality of sampled images.

39. A system as recited in claim 38, wherein said lesion detection circuit comprises:

present in each of said plurality of sampled images; and

detecting from at least two of said plurality of sampled images.

40. A system as recited in claim 34, wherein:

said system comprises a lesion identification circuit connected to receive said digital

image and produce locations of regions in said digital suspected of being a lesion: and

said multi-sampling circuit digitally samples said digital image after lesion

identification a plurality of times to produce a corresponding plurality of sampled images.

41. A system as recited in claim 40, wherein said lesion detection circuit comprises:

present in each of said plurality of sampled images; and

detecting from at least one of said plurality of sampled images.

42. A system as recited in claim 41 , wherein said lesion detection circuit comprises: means for eliminating at least one false positive using results of said means for

detecting from at least two of said plurality of sampled images.

43. A system as recited in claim 41, wherein:

said lesion identification circuit comprises means for determining regions-of-interest

containing corresponding ones of said regions; and

said multi-sampling circuit comprises means for sampling only said regions-of-

interest.

44. A system as recited in claim 34, wherein said lesion detection circuit comprises a

feature extraction and feature analysis circuit.

45. An automated lesion detection system, comprising:

a digital image acquisition device;

a digital image processing circuit connected to said image acquisition device;

a lesion identification circuit connected to said processing circuit;

a lesion detection circuit connected to said lesion identification circuit; and

a multi-sampling circuit adapted to be connected in at least one of said image

acquisition device, said image processing circuit and said lesion detection circuit.

46. A system as recited in claim 45. wherein:

said multi-sampling circuit is adapted to receive a digital image from said image

acquisition device and sample said digital image a plurality of times to produce a plurality of

sampled images; and

said lesion identification circuit identifies regions suspected of being a lesion in a

selected number of said plurality of sampled images.

47. A system as recited in claim 46. comprising: means for eliminating a false positive in said image using said selected number of sampled images.

48. A system as recited in claim 45, wherein:

said image processing circuit processes a digital image output from said image

acquisition device and produces a processed image;

said multi-sampling circuit is adapted to receive said processed image from said

image processing circuit and sample said processed image a plurality of times to produce a

plurality of sampled images; and

selected number of said plurality of sampled images.

49. A system as recited in claim 48, comprising:

means for eliminating a false positive in said image using said selected number of

sampled images.

50. A system as recited in claim 45, wherein:

said lesion identification circuit identifies regions suspected of being a lesion in an

image;

said multi-sampling circuit is adapted to receive said image after identification of said

regions by said lesion identification circuit and sample said image a plurality of times to

produce a plurality of sampled images; and.

said lesion detection circuit detects true positives and false positives in a selected

number of said plurality of sampled images.

51. A system as recited in claim 50, comprising means for eliminating a false positive

in said image using said selected number of sampled images.