Patent Grant
7187792
The present invention relates to an apparatus and method for determining the measure of similarity of two images.
Radiosurgery is useful for treating tumors and other lesions by delivering a prescribed high dose of high-energy radiation to the target area while minimizing radiation exposure to the surrounding tissue. In radiosurgery, precisely focused beams of radiation (e.g. very intense x-ray beams) are delivered to a target region in order to destroy tumors or to treat the tumor for other purposes. The goal is to apply a lethal or other desired amount of radiation to one or more tumors, without damaging the surrounding healthy tissue.
Conventional radiosurgery uses a rigid and invasive stereotactic frame to immobilize the patient prior to diagnostic CT or MRI scanning. The treatment planning is then conducted from the diagnostic images. The treatment planning software determines the number, intensity, and direction of the radiosurgical beams that should be cross-fired at the target, in order to ensure that a sufficient dose is administered throughout the tumor so as to destroy it, without damaging adjacent healthy tissue. Immobilization of patient is necessary in order to maintain the spatial relationship between the target and the radiation source that ensures accurate dose delivery. The frame is fixed on the patient during the whole treatment process, causing pain and inconvenience to the patient.
Image-guided radiosurgery allows the elimination of such invasive frame fixation, during treatment. In an image-guided radiosurgical process, the patient position and the relative alignment of the radiation beam with respect to the patient target is continuously adjusted. In order to ensure the delivery of the correct dose of radiation to the correct location, the patient (and target) position during treatment needs to be detected. This is accomplished by registering the x-ray image acquired at the treatment time with the diagnostic 3D scan data (e.g., CT, MRI, ultrasound, or PET scan data) obtained pre-operatively at the time of treatment planning. In the field of medical image registration, this problem is categorized as a 2D/3D registration.
In the 2D/3D registration process, similarity measures are useful for comparing the image intensities in the x-ray images and the DRR images, so that the change in patient position (and thus in target region position) that has occurred between the diagnostic scanning and the taking of real-time images can be accurately detected. Image-guided radiosurgery requires precise and fast positioning of the target at the treatment time. In practice, the accuracy should be below 1 mm, and the computation time should be on the order of a few seconds. Unfortunately, it is difficult to meet both requirements simultaneously. In order to optimize the 2D/3D registration process in image-guided radiosurgery, it is necessary to provide an accurate, robust, and efficient similarity measure method and system.
The present invention is directed to an improved method and system for comparing the measure of similarity of two digital images, in image-guided radiosurgery. The method is based on pattern intensity, and provides a more robust, accurate, and efficient solution to the 2D/3D medical image registration problem, as compared to other methods known and used in the art.
The present invention features a method in image-guided radiosurgery for determining the measure of similarity of a first image and a second image of an object. Preferably, the first image is a real-time 2D x-ray image, and the second image is an artificially synthesized DRR, constructed from previously generated 3D scans.
The method includes forming a difference image by subtracting the corresponding pixel values of the second (DRR) image from each pixel value of the first, “live,” x-ray image. The method further includes applying upon each pixel of the difference image a pattern intensity function, the pattern intensity function being an asymptotic function of the gradients of the difference image.
The mathematical formulation of pattern intensity function is given by:
where Idif(i,j) represents the array of pixel values for the difference image, where σ is a constant, and where R is a neighborhood that uses the pixel (i,j) as a center point.
In a preferred embodiment of the invention, the neighborhood R is defined so that the gradients of the difference image can be considered in at least four directions. In an exemplary embodiment, the four directions are: substantially horizontal; substantially vertical; about 45 degrees; and about −45 degrees.
The present invention features a system for determining the measure of similarity of a 2D x-ray image of an object and a 2D DRR of the object generated from previously obtained 3D scan data. Typically, the x-ray image and the DRR are discretized digital images, which are characterized by a first and second 2D array of pixel values. The system includes means for generating 3D scan data of the target, for example a CT scanner, an MRI system, or a PET system. A radiation source (typically an x-ray source) is provided for generating at least one imaging beam. Imaging means are provided for generating a 2D radiographic image of the target in near real time, by directing the imaging beam towards and through the target from a known location and angle and at a known intensity, and detecting the imaging beam after the beam has passed through the target.
The system includes a controller, which contains software for generating a set of DRR images of the target, using the known location, angle, and intensity of the imaging beam. The controller also includes software for determining the measure of similarity between the 2D x-ray image and the 2D DRR, by subtracting each pixel value of the second image from a corresponding pixel value of the first image to form a difference image, and then applying upon each pixel of the difference image to form a pattern intensity function. The pattern intensity function is an asymptotic function of the gradients of the difference image, and permits the pixel values within a neighborhood R defined about each pixel in the difference image to be considered. The gradients are defined over at least four directions, as discussed above.
The present invention is directed to a similarity measure, based on pattern intensity, for use in 2D/3D medical image registration. Similarity measures are used to compare two images, each of which have been generated using different modalities (e.g., CT versus x-rays), so that information from one image can be registered onto the other. Similarity measures are useful in procedures such as the 2D/3D medical image registration procedures in image-guided radiosurgery. The similarity measure method and system disclosed in the present invention allows for selected phases of the 2D/3D registration process in image-guided radiosurgery to be carried out in a more robust, efficient, and powerful manner, as compared to processes carried out using other similarity measures known and used in the art.
As explained in the background section, 2D/3D registration is necessary in order to correct patient position and properly align the radiosurgical beam relative to the target. The x-ray image acquired at treatment time is registered with the 3D scan obtained at the time of treatment planning. A CT scan is most frequently used for the 3D scan; however, other 3D scanning methods, such as MRI, ultrasound or PET scanning, may also be used. The 3D data is used as the reference to determine the patient position change during treatment. For this purpose, 2D reference images are reconstructed from the 3D scan data. Typically, digitally reconstructed radiographs (DRRs) need to be generated from 3D CT data and are used as the 2D reference images.
Typically, DRRs are generated by casting hypothetical beams or rays through the CT volumetric image of the target. Each ray goes through a number of voxels of the 3D CT image 60. By integrating the CT numbers for these voxels along each ray, and projecting onto an imaging plane (shown as 70A and 70B, respectively, in
Applications such as image-guided radiosurgery require that the comparison between the DRRs (that contain the 3D CT scan information) and the real-time x-ray images, and consequent adjustment of the position of the x-ray source, be made very rapidly and accurately. In practice, the accuracy should be below 1 mm, and the computation time should be on the order of a few seconds. Unfortunately, it is difficult to meet both requirements simultaneously, because of several reasons. First, the two different modality images, i.e. CT scan images and x-ray images, have different spatial resolution and image quality. Generally, x-ray image resolution and quality are superior to the resolution and quality of DRR images, which are only synthesized images. Typically, some structures in the DRR may appear more blurred (especially normal to the CT slice plane), compared to the x-ray image. Ideally, an optimal similarity measure for a 2D/3D registration process should allow for an accurate registration to be achieved, despite such differences.
Second, DRR generation relies on a proper attenuation model. Because attenuation is proportional to the mass intensity of the target volume through which the beam passes, the exact relationship between the traversed mass intensity and the CT image intensity needs to be known, in order to obtain an accurate modeling. Establishing this relationship is difficult, however, so the linear attenuation model is often used. As is known, the linear attenuation coefficient of a material is dependent on x-ray energy. CT machines and x-ray machines work at different effective energies, however. As a result, the attenuation coefficients measured by a CT scanner are different from the attenuation of a beam of x-rays passing through the target. The skeletal structures in DRR images cannot be reconstructed very well using the linear model, the DRRs being only synthetic x-ray projection images. At CT energies, the ratio of bone-to-soft-tissue attenuation is much lower than at x-ray radiographic energies. Thus, in a DRR produced from a 3D CT volume, the image contrast from soft tissue will be comparable with the image contrast from bone, reducing the clarity of bone details, for example.
Finally, x-ray images usually have a large image size (512×512). For better registration accuracy, it is desirable to use the full resolution image. Full resolution images are rarely used, in practice, however, because the resulting increase in computation time is excessive, and is incompatible with the requirements of image-guided radiosurgery.
Generally, similarity measure methods used in 2D/3D registration can be divided into two categories. The first method is based on image features. The image features could be anatomical edges or segmented objects. The registration accuracy depends on the accuracy of edge detection or object segmentation. The main advantage of this method is its fast computation. Feature-based similarity methods register on salient features that have been segmented from each image. They use a reduced amount of data, which makes the algorithms fast, once the segmentation has been undertaken. Because the full information content of the image is not used, however, the accuracy is sacrificed. Errors in the segmentation stage can lead to an error in the final registration.
The second method is based on image intensity content. Intensity-based methods compare the voxel and pixel values directly, using measures based on image statistics. The original images are used for registration. Usually, a good accuracy can be achieved. Although these methods require little or no segmentation, intensity-based methods are typically much slower. Because a long time computation is required, it is hard to apply intensity-based similarity measures to clinical practice.
In a co-pending patent application Ser. No. 10/652,786 a method and system for 2D/3D image registration is disclosed that uses a hierarchical and iterative framework for the registration algorithm, to achieve a higher accuracy with less computing time. The application Ser. No. 10/652,786 is owned by the assignee of the present application, and is hereby incorporated by reference in its entirety. In the application Ser. No. 10/652,786, a 2D/3D image registration procedure is disclosed that includes the steps of 1) performing a 3D multi-level matching to determine an initial estimate in-plane transformation parameters, 2) performing an initial 1-D search for each of a pair of out-of-plane rotation parameters, then 3) iteratively refining the in-plane parameters (x, y, θ) and the out-of-plane parameters (r, φ), until the parameters converge to a desired accuracy. The similarity measure methods disclosed in the present invention are designed to optimize the 2D/3D image registration procedure disclosed in the application Ser. No. 10/652,786.
The similarity measure method of the present invention is based on pattern intensity, and provides a powerful and efficient way to solve the 2D/3D image registration procedure, as disclosed in the application Ser. No. 10/652,786. In particular, the pattern intensity based method and system of the present invention is designed for the 1D search phase (for the out-of-plane parameters), and the iterative refining phases of the 2D/3D image registration procedure disclosed in the application Ser. No. 10/652,786.
For the 3D multi-level search phase, the “sum of absolute differences” (SAD) measure is used, which is a known, simple similarity measure. The SAD measure is widely used in medical image processing and video processing, in cases where the two images to be matched have high image quality. The main advantage of using SAD is its fast computation and its easy optimization in parallel computation. Its main disadvantage is that the solution is sensitive to image noise, artifacts and intensity difference between the live and DRR images. As a result, SAD is only used in the first search phase to get approximate results. SAD can be expressed as
where Ilive(i, j) represents the intensity of the “live” real-time x-ray image, and IDRR(i, j) represents the intensity of the reconstructed DRR image.
The pattern intensity similarity measure of the present invention is more accurate, and less sensitive to image noise, artifacts, and to the intensity difference between the images being compared. In the exemplary embodiment described in the following paragraphs, the first image is a 2D x-ray image of a radiosurgical treatment target, and the second image is a 2D DRR that is reconstructed from 3D CT scan data generated at the time of treatment planning. In a preferred embodiment, the two images are discretized, digital images, characterized by first and second 2D arrays of pixel values. The pixel arrays are equi-dimensional, i.e. the number of rows and columns of the first array is equal to the number of rows and columns of the second array. As well known, each pixel value of an image is a number representative of the intensity of the image at a unique corresponding 2D area element forming the image.
A difference image is formed from the real-time x-ray image and the DRR image, by subtracting the corresponding pixel values of the second image (the DRR image) from each pixel value of the first image (the real-time):
Idif(i, j)=ILive(i, j)−IDRR(i, j),
where Idif(ij) represents the intensity or pixel value of the ij-th pixel of the difference image,
Ilive(ij) represents the intensity or pixel value of the ij-th pixel of the live x-ray image; and
IDRR(i,j) represents the intensity or pixel value of the ij-th pixel of the artificial DRR image.
A pattern intensity function is defined, which operates on the difference image. The pattern intensity function is expressed as an asymptotic function of the gradients of the difference image:
where σ is a weighting constant and R is a neighborhood that is defined using the pixel (i, j) as the center point. The form of the mathematical formulation results in the similarity measure tending to a maximum value, as the number of structures tend to zero, and the similarity measure asymptotically tending to zero, as the number of structures increase. Because of the asymptotic nature of the pattern intensity measure, large differences in intensity have the same effect on the measure, regardless of their magnitude. This makes the measure robust to large differences in pixel intensity.
The function is weighted by the weighting constant σ. The constant σ is used to weight the function, so that small deviations in intensity (caused by noise, by way of example) results in the measure remaining proximate to its maximum value. The sensitivity of the solution to the variation of X-ray image can be minimized by careful selection of this constant. The larger the weighting constant, the more stable the results become. However, the choice of the weighting constant is a tradeoff between stability and accuracy. If the value of the weighting constant is too large, the smaller details in the images cannot be reflected in the similarity measure. Based on experimentation, the empirical value of σ is determined to be in the range from about 4 to about 16, although other values of σ are also within the scope of the present invention.
The pattern intensity function considers a selected neighborhood for each pixel. In a preferred embodiment of the invention, illustrated in
Based on the definition of the neighborhood R as shown in
The formulation of the pattern intensity function, given in equation (2) above, provides a number of advantages over other known similarity measures, as discussed in paragraph 41 below.
A radiation source 202 (typically an x-ray source) is provided for generating at least one imaging beam (typically an x-ray beam). An imaging system 206 is provided for generating a 2D radiographic image of the target in near real time, by directing the imaging beam towards and through the target from a known location and angle and at a known intensity, and detecting the imaging beam after the beam has passed through the target. The imaging system 206 is preferably an x-ray imaging system for generating a pair of orthogonal x-ray projection images of the target. The imaging system 206 preferably has a pair of x-ray sources for generating diagnostic imaging beams (having known positions, angles, and intensities), and a corresponding pair of x-ray image detectors which detect the beams after the beams have passed through the target.
The system includes a controller 208. The controller 208 includes software for generating a set of reconstructed 2D images (preferably DRRs) of the target, based on the 3D scan data from the 3D scanner 201, and the known intensity, location, and angle of the imaging beams. The controller also includes software for determining the measure of similarity between the 2D x-ray image and the 2D DRR, in a manner described above in paragraphs 28–36.
The pattern intensity similarity measure of the present invention provides a number of advantages over other methods known in the art. First, the difference image filters out the low frequency part that is basically the soft tissues and keeps the high frequency part that is mostly the skeletal structures. This feature makes the algorithm robust to some brightness intensity difference between live and DRR images. Second, because of the asymptotic nature of the pattern intensity function, the similarity measure is less affected by pixels whose intensity values deviate only slightly from its neighboring pixels. These kinds of pixels are thought to contain random noise, hence undesirable. Third, because the asymptotic function quickly approaches to zero when the variable increases, large intensity differences such as image artifacts have the same effects on the similarity measure, regardless of their magnitude. Accordingly, the pattern intensity is less sensitive to image artifacts.
While the invention has been particularly shown and described with reference to specific preferred embodiments, it should be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5117829 | Miller et al. | Jun 1992 | A |
| 5901199 | Murphy et al. | May 1999 | A |
| 6470207 | Simon et al. | Oct 2002 | B1 |
| 20040092815 | Schweikard et al. | May 2004 | A1 |
| 20040267113 | Thomson | Dec 2004 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20050049477 A1 | Mar 2005 | US |
