COMPUTER-AIDED METHOD AND SYSTEM FOR DETECTING SPICULATED LESIONS IN A MAMMOGRAM
Field of the Invention
The present invention relates to a method and system for digital mammography, and more particularly, to a computer-aided method and system for detecting spiculated lesions in a mammogram. Digital mammography, the computer analysis of mammo graphic images, is becoming an increasingly important tool for detecting breast cancer. The early detection and treatment of breast cancer can prolong or save the life of a patient. Until recently, the detection of mammographic abnormalities was the solely the domain of a trained radiologist who manually screened mammographic images. With advances in image acquisition and digital image processing, digital mammography has become a valuable clinical tool for identifying suspicious lesions in mammograms.
Many computer-aided diagnostic techniques have been developed to detect and classify potentially abnormal structures. Using density estimates, pattern recognition, or other techniques, abnormal structures can be extracted from a mammogram. However, the difficulty in the field of digital mammography is not only in identifying lesions or masses, but also in identifying lesions that a trained radiologist would classify as malignant. For example, a smooth, round or oval mass as shown in FIG. 1 A is usually classified as benign. Therefore, computer-aided techniques have been developed to distinguish between likely benign and likely malignant lesions according to calcification, spiculation, and the roughness and shape of a mass.
In particular, the margin of a mass, which is often characterized as circumscribed, lobulated, obscured, or spiculated, is considered as one of the most important indicators of a malignant lesion. A stellate structure having locally radiating spicules, as shown in FIG. IB, is considered a strong indication of malignancy. Thus, many techniques have been developed to detect and classify spiculated lesions. For example, in an article entitled "Recognition of Stellate Lesions in Digital
Mammograms (Digital Mammography: Proceedings of the 2nd International Workshop on Digital Mammography, York, England, 10-12 July, 1994, Elsevier Science 1994), which is hereby incorporated by reference, Karssemeijer described a mathematical algorithm for estimating the number of pixels along a line contained in a pre-defined pie-shaped bin that are oriented towards a central area.
The Karssemeijer feature extraction technique employs neighborhoods or bins to detect marginal spiculation. The area surrounding a selected pixel of a digital mammogram is divided into pie-shaped bins. Based on spiculation measurements calculated for each bin, the pixel is then classified as either benign or malignant. However, this approach and other feature extraction schemes have several disadvantages or shortcomings.
First, current feature extraction techniques yield a significant number of false- positive responses because the size and marginal spiculation characteristics of differently sized masses are not properly accommodated for. As shown in FIG. IB, a small lesion typically has a spiculated region closer to the central mass than a larger lesion. The typical spiculation characteristics of a large lesion are depicted in FIG. lC. Therefore, present techniques that simply measure the degree of spiculation according to the orientation of pixels along a line in a pre-defined, pie-shaped bin fail to accurately consider the depth or distance at which spicules are located from the central mass.
Second, present techniques fail to adequately adjust the size of a neighborhood or bin to accurately capture the spicules radiating from the margins of a mass. Because the choice of the neighborhood affects the detection and measurement of spiculation, it is important that the bin size be adjusted accordingly. Finally, the computational intensity of current detection schemes for measuring and analyzing spiculation information requires significant financial and computer resources.
In light of the foregoing, there is a need for a computer-aided method and system for detecting spiculated lesions having different sizes and different spiculation characteristics. Moreover, there is a need for a computer-aided method and system for computing and analyzing spiculation information for lesions of various sizes during one computational pass.
SUMMARY OF THE INVENTION
Accordingly, the present invention is directed to a computer-aided method and system for detecting spiculated lesions in a mammogram that substantially obviates one or more of the problems due to limitations and disadvantages of the related art.
One object of the present invention is to provide a computer-aided method and system for detecting and measuring spiculation information for lesions having different sizes.
Another object of the present invention is to provide a computer-aided method and system for identifying spiculated lesions by computing spiculation information for a combination or subset of neighborhoods.
A further object of the present invention is to provide an efficient computational method and system for detecting spiculated lesions.
Yet another object of the present invention is to provide a method and system for extracting spiculation features from a digital mammogram.
Yet another object of the present invention is to provide a method and system for distinguishing spiculated and non-spiculated lesions.
Yet another object of the present invention is to provide a method and system for adjusting the neighborhood or bin dimensions to better approximate the size of a lesion.
Yet another object of the present invention is to provide a method and system for decreasing the number false-positive responses by increasing the specificity in which bin depth is defined to capture the marginal spiculation of a mass.
Yet another object of the present invention is to provide a method and system for detecting and measuring the spiculation responses across a range of bin depths.
Yet another object of the present invention is to provide a method and system for detecting and measuring spiculation information that can inputted or integrated with other image processing techniques.
Additional objects and advantages of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the
invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory.
BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention. In the drawings:
FIG. 1 A illustrates a smooth, oval lesion;
FIG. IB illustrates a small lesion with marginal spiculation near the central mass; FIG. 1C illustrates a large lesion with marginal spiculation further from the central mass;
FIG. 2A is a block diagram showing a system for detecting spiculated lesions in a mammogram according to one exemplary embodiment of the present invention; FIG. 2B is a block diagram showing another exemplary embodiment of a system for detecting spiculated lesions in a mammogram;
FIG. 3A is a flowchart showing a method for detecting spiculation in a digital mammogram according to one exemplary embodiment of the present invention;
FIG. 3B is a flowchart showing a method for detecting spiculation in a digital mammogram according to an alternative exemplary embodiment of the present invention;
FIG. 4 is a photograph of an X-ray mammogram showing the outline of a breast mask according to one exemplary embodiment of the present invention; FIG. 5 graphically illustrates a number of depth bins according to one exemplary embodiment of the present invention; FIG. 6 is a flow diagram showing a method for detecting spiculation using depth bins according to one exemplary embodiment of the present invention;
FIGS. 7A-7C are flowcharts showing a detailed method for detecting spiculation using depth bins according to one exemplary embodiment of the present invention;
FIG. 8 is a flow diagram showing a method for detecting spiculated lesions in a mammogram according to an alternative exemplary embodiment of the present invention;
FIG. 9 is a flow diagram showing a method for detecting spiculation in a mammogram using a breast mask image according to an alternative exemplary embodiment f the present invention; and FIG. 10 is a photograph of an X-ray mammogram showing regions of interests highlighted.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
Reference will now be made in detail to the preferred embodiments of the present invention, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like elements.
FIG. 2A shows a block diagram of a computer-aided system 30 for detecting and measuring spiculation in a mammogram according to one embodiment of the present invention. The system 30 comprises a processor 38, such as a central processing unit (CPU) or parallel processor, memory (such as RAM, a hard drive, an optical media, and/or ROM) 32, a database 34, and various input/output devices 36, such as a keyboard, mouse or pointer, microphone, speaker, monitor, printer, and scanner. As one skilled in the art will appreciate, implementation of the present invention is not limited to the specific hardware configuration shown in FIG. 2A.
For instance, in another embodiment of the present invention shown in FIG. 2B, system 30 may be connected directly or indirectly to an imaging device 10, a digitizer 20, and one or more external storage devices 40. As one skilled in the art will appreciate, the imaging device 10 is used to produce X-ray mammograms that are converted into digital images by a digitizer 20. The digital mammogram is then processed by system 30 and stored in database 34 or in an external storage device 40,
such as a hard disk or optical disk. The imaging device 10 and digitizer 20 may be controlled by computer system 30 or by another computer system.
The software for implementing system 30 of the present invention is not limited to any one computer programming language. Because of the computational requirements associated with image processing, the software for implementing the present invention on system 30 was written in C++.
GENERAL OVERVIEW
FIG. 3 A is a flowchart showing a general overview of a method for detecting and measuring spiculation in a digitized mammogram according to one embodiment of the present invention. In the present embodiment, the X-ray mammogram has been previously produced and digitized according to methods and procedures well known in the art. The digitized mammogram 100 is input and processed to determine line orientation values for each pixel of the digital mammogram at step 200. Using the images output from step 200, spiculation information is computed at step 300 by examining the orientation of the pixels along one or more lines. The spiculation information is computed and recorded for a number of depth bins, which are located at varying distances from a central region or central pixel. The line detection step 200 and the spiculation detection step 300 will be discussed in more detail below.
INPUT
In an alternative embodiment shown in FIG. 3B, the present invention contemplates using a breast mask template 110 to identify which pixels represent the breast tissue and which pixels represent the film background. For instance, FIG. 4 shows a photographic image of a mammogram with the outline of a breast mask template overlaid. By restricting subsequent image processing to the breast tissue areas of the mammogram, a breast mask 110 reduces the computational requirements of the present embodiment. The breast mask 110 may be defined manually by a radiologist or by computer-aided techniques well known in the art. As will be recognized by one skilled in the art, the method and system of the present invention may be implemented independently or may be integrated with other
existing or future techniques. For example, the lesion extraction method and system of the present invention may alternatively be integrated with an imaging device 10 that produces an X-ray mammogram that is then converted into a suitable digital format. The digital mammogram may then be stored in system 30 or any type of compatible external storage device 40.
Moreover, one skilled in the art will appreciate that corrective post processing of the digital mammogram may be performed prior to feature extraction. For example, contrast or other properties of the digital mammogram may be enhanced and background noise may be identified and filtered. The digital mammogram 100 can then be processed in accordance with the line detection algorithm 200 and spiculation detection algorithm 300 contemplated by the present invention.
As discussed earlier in reference to FIG. 3 A, the present technique of detecting stellate patterns in a digital mammogram for purposes of identifying spiculated lesions involves two basic steps - line detection 200 and spiculation detection using depth bins 300. Each step will now be discussed in more detail.
LINE DETECTION
With respect to step 200 of FIG. 3 A, any number of the line detection algorithms known in the art may be utilized to estimate the orientation of pixels along lines detected in the mammogram. For example, in one embodiment line detection is performed on a digital mammogram at a spatial resolution of 200 microns per pixel to create three images: A line image, a line orientation image, and a line strength image.
The line image may be a template that contains line information for identifying bright lines against the dark background of the mammogram. The line detection technique may assign each pixel a positive value, e.g., "1," if it is located along a bright line and a negative value if the pixel is located along a dark line. Estimated values representing the orientation of each pixel in the line image may then mapped in a line orientation image. Finally, output from the preferred line detection technique may also include a line strength image that contains strength (i.e., magnitude or contrast) values for each pixel of the line image, and the distribution of the strength values of any bright lines that have been detected in the mammogram.
SPICULATION DETECTION
After the line-based orientation values have been estimated using a line detection algorithm 200, a spiculation detection technique may be performed at step 300 of FIG. 3 A. To identify spiculated lesions, which are a strong indication of malignancy, the current embodiment involves detecting stellate patterns (i.e., areas of locally radiating spicules) in the mammogram. The degree of marginal spiculation is quantified for a neighborhood of pixels according to both the orientation and distance of the pixels surrounding a central pixel. Therefore, more accurate spiculation information can be computed for a range of mass sizes that may have spiculation features located at varying distances from a central mass.
FIG. 5 shows a graphical representation of a number of neighborhoods or depth bins 440 defined according to one embodiment of the present invention. Namely, a circular neighborhood 400 centered on a candidate location 410 is divided into a number of sectors 420 or pie-shaped bins. The candidate location 410 may comprise a candidate pixel 410 or a group of pixels 410, however, the following description will only make reference to a candidate pixel 410 for simplicity and to avoid redundancy. The sectors 420 may measure between, for example, approximately 1° and 15°. Each sector 420 is further divided into a number of depth bins 440. As shown in FIG. 5, a depth bin 440 is located by the intersection of a sector 420 and the area between two concentric circles centered on pixel 410. The size of a concentric circle is defined by a radius Rx 454, which may range in value between a maximum allowable radius, Rmaχ 450, and a minimum allowable radius, Rmιn 452. For example, depth bin 442 shown in FIG. 5 is bound by an outer radius, Rmax 450 and inner radius, Rx 454. In the present embodiment, the depth bins 440 are located by concentric circles having radii Rx 454 set at fixed increments. One in the art will appreciate that the shape and position of the depth bins 440 may be defined using a number of techniques. For example, while concentric circles are shown in FIG. 5, other shapes or combinations of shapes may be used to bound a depth bin.
FIG. 6 shows a flow diagram of the general steps for accomplishing the spiculation detection step 300 of FIG. 3A. For a window of pixels that may correspond to a region of interest, the present embodiment calculates one or more metrics for each
candidate pixel 410. For example, the following two metrics may be calculated: (1) The percentage of pixels surrounding a candidate pixel 410 with line orientations directed toward a central region; and (2) the percentage of depth bins 440 which have more pixels oriented toward a central region than would be expected for a uniformly distributed random noise pattern.
Moving incrementally across an image or a region of an image, that may have been selected based on the results of a mass detection algorithm, the one or more metrics are calculated for each candidate pixel 410. A candidate pixel 410 may be chosen according to a predetermined sample spacing or sample rate. For example, a sample spacing may be defined such that every pixel, or every other or every third pixel is selected as the candidate pixel 410. As one skilled in the art will appreciate, the method for selecting the candidate pixel 410 or group of pixels 410 may rely on any number of techniques well known in the art.
After a candidate pixel 410 has been selected for analysis, a neighborhood of pixels 400 surrounding the central pixel 420 is defined by an outer radius Rmax 450 and an inner radius Rmιn 452. The neighborhood 400 is divided into sectors 420 and each sector is further subdivided into depth bins 440.
To determine whether any lines identified by the line detection algorithm 200 are radiating toward a candidate pixel 410, at step 310 of FIG. 6 the present technique first computes the number of pixels 460 (see FIG. 5) of neighborhood 400 that have line orientations directed towards a presumed or estimated central mass represented by region 430 (see FIG. 5) divided by the total number of pixels in the neighborhood. The central region 430, centered at the candidate pixel 410 and defined by radius R, is often referred to as the mass radius disk. In other words, the present technique determines the percentage of pixels 460 or subset of pixels 460 within neighborhood 400 that are contained along lines that point to a central mass region 430. The resultant metric for a pixel 460 is recorded according to both the orientation and depth of the pixel 460 with respect to the candidate pixel 410. Therefore, the number of pixels 460 of a particular depth bin 440 with the desired line orientation can be computed. Then at step 320 of FIG. 6, the present technique computes a second metric
(i.e., spiculation information) that quantifies the probability of finding a certain number
of pixels 460 of a particular depth bin 40 directed toward the mass radius disk 430. Since a random number of lines not associated with a spiculated lesion may point to mass radius disk 430, the spiculation information represents the degree in which the region of a depth bin 440 may be considered spiculated. Therefore, a depth bin 440 that has a large number of pixels 460 directed toward mass radius disk 430 above what is considered as random orientations or background noise would have an associated spiculation metric that quantifies the increase. The second metric is computed as the percentage of depth bins that have a spiculation metric greater than what is considered random. However, since different lesion sizes exhibit different marginal spiculation characteristics, the spiculation measurements computed for each depth bin 440 are subsequently analyzed.
As previously discussed, the spicules of small lesions are typically closer to the central mass than the spicules of a larger lesion. To better approximate the size of a lesion, spiculation information is combined from various subsets of depth bins 440 and compared at step 330 of FIG. 6. In one embodiment, the spiculation detection algorithm 300 computes and analyzes spiculation information for a number of bin configurations in one computational pass to produce a more accurate spiculation measurement for the candidate pixel 410. This technique indirectly approximates the size of a lesion. Therefore, the outer radius Rmax 450 and inner radius Rm,n 452 of a neighborhood 400 more likely to display characteristics of marginal spiculation for a particular mass is determined by analyzing the spiculation measurements from various combinations of depth bins 440. For example, a combination of three depth bins 440 close to the candidate pixel 410 may constitute the strongest spiculation measurement for a small mass. FIGS. 7A through 7C show in more detail the steps associated with the spiculation detection technique of step 300 of FIG. 3 A. FIG. 7 A shows the initial steps of defining operating parameters according to one embodiment of the present invention. Step 500 comprises the step of defining the spacing between candidate pixels 410 that are examined using the present technique. As one skilled in the art will appreciate, the precision of the present embodiment can be increased or decreased by adjusting the sample spacing rate. At step 510, the mass radius disk 430 may be computed according
to the estimated size of the lesion. Then at step 520 of FIG. 7 A, the depth bins 440 surrounding a candidate pixel 410 are defined.
FIG. 7B shows in more detail the step of computing the number of pixels 460 of a depth bin 440 that have a desired orientation. At step 540 the spiculation detection algorithm identifies the depth bin 440 that corresponds to the location 460 selected in step 530. The orientation of the pixel 460 towards the mass radius disk 430 is examined at step 550. A metric is recorded indicating whether the pixel 460 is located along a line that intersects mass radius disk 430. Then at step 560, the algorithm queries whether there are any remaining pixel locations 460 to examine. After examining the line orientations of the selected candidate locations 460, the technique computes the spiculation information for each depth bin 440, as shown in step 570 of FIG. 7C. To accommodate lesions of varying size, the spiculation information for a range of radial bins is computed and analyzed. In particular, at step 580 spiculation information is computed for a number of bin configurations that may include one depth bin 440 or multiple adjacent depth bins 440. The measurements are compared to determine better approximate the size of the mass, thereby resulting in a more accurate spiculation measurement. The outer and inner radii of the corresponding neighborhood 400 are retained at step 590. This process is repeated for each candidate pixel 410.
EXAMPLES OF ALTERNATIVE EMBODIMENTS
In an alternative embodiment shown in FIG. 8, the system computes line information at step 200 and spiculation information at step 300 for a digital mammogram 100. Then, the final spiculation measurement, which represents the best response across a range of bin sizes, is used in step 340 to classify the degree of spiculation. For example, a location may be classified as spiculated, moderately spiculated, or non-spiculated. As one skilled in the art will recognize, the classification will vary depending on the classification scheme. Furthermore, the operating parameters of many of the classification schemes can be adjusted to provide increased or decreased sensitivity.
In an alternative embodiment shown in FIG. 9, at step 150 regions of interest (ROI) or target areas are extracted from the digital mammogram 100 or from the breast region of a mammogram 100 using a breast mask 110. Therefore, only potentially interesting areas of the digital mammogram 100 are subsequently analyzed in the line detection step 200 and spiculation detection step 300. Alternatively, ROI may be extracted after computing line information in step 200 of FIG. 9. FIG. 10 shows a mammogram with exemplary ROIs highlighted. Restricting subsequent image processing to only a portion of the original mammogram increases the speed in which the present embodiment identifies spiculated lesions. The ROIs may be selected according to any number of reliable techniques known in the art - including manual and computer-aided detection techniques. In one embodiment, a multi-scale mass detection algorithm identifies ROIs that may contain a range of mass sizes, but not fatty tissue. First, contrast is computed by comparing the intensity of a pixel with a group of surrounding pixels. Then, by comparing the contrast measured at each pixel with a variable threshold value, the technique identifies roughly circular regions of relatively dense tissue. Depending on the sensitivity desired, the threshold value may be adjusted to create larger or smaller ROI.
It will be apparent to those skilled in the art that various modifications and variations can be made in the method and system for detecting spiculated lesions in a mammogram of the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention covers the modifications and variations of this invention provided they come within the scope of the appended claims and their equivalents.