Segmentation-based collimation region detection and anatomical region decomposition method for X-ray images

Abstract

The present invention relates to a segmentation-based X-ray image collimation region detection and anatomical region decomposition method, which includes k-means blocks cascaded consecutively in an X-ray protocol independent manner, and morphological processing blocks used to regularize the segmentation results.

Description

[Technical Field]

[0001] The present invention relates to a segmentation-based X-ray image collimation region detection and anatomical region decomposition method, which includes a k-means block cascaded consecutively in an X-ray protocol independent manner, and a morphological processing block used to regularize the segmentation results. [Background technology]

[0002] Conventional technology Radiography is the first imaging method that began with the discovery of X-rays. Radiography has been widely used since the discovery of X-rays. Imaging is performed by taking advantage of the penetrating properties of X-rays. The basic principle of radiography is to generate an image of an entire area by passing X-rays through all layers of the body.

[0003] The application of a collimation zone is a control used to limit x-ray scatter during x-ray imaging to reduce the radiation dose received by the patient. The device that performs the collimation process within x-ray equipment is called a collimator, and consists of two pairs of lead blades, each pair moving in different directions on the same axis. The collimation zone refers to the unobstructed portion of the lead blades and appears in the x-ray image as a polygon with at least four sides, depending on the position of the x-ray detector. However, the collimation zone is not required and can be omitted from the x-ray image at the radiographer's discretion.

[0004] Segmentation is the process of generating homogeneous image objects (segments) based on features that are representative of the image. The segmentation process should ensure that the image objects correspond to real-world objects of interest. Segmentation and feature detection are the basis of classification, and segmentation is the most important stage of object-based classification. Because regions outside the collimation region, anatomical regions, and direct exposure regions constitute fundamentally different image classes in X-ray images, detecting these regions using segmentation-based methods allows the development of algorithms that are independent of the X-ray protocol.

[0005] With the digitalization of direct X-ray imaging, benefits such as direct exposure through collimation and segmentation of anatomical regions can be achieved on the resulting images using various image processing algorithms. However, considering the radiology workflow, algorithms that provide fast and accurate results become increasingly important. The easiest and fastest method for generating distinct segmentation clusters is thresholding, and Patent Document 1 provides an adaptive version of thresholding independent of the X-ray protocol (i.e., chest X-ray). However, the adaptive threshold update method assumes that the edge transitions between direct exposure and anatomical regions and between anatomical and collimated regions are sharp and have different speeds. Candidate segmentation and edge transition detection algorithms require processing each row and column of the X-ray image separately, which is very resource-intensive considering that an average X-ray image consists of approximately 6–7 million pixels. Furthermore, taking into account the geometry of X-ray scattering and the thickness of the patient being imaged, it is observed that the edge transition speeds between different regions are very close to each other. Meanwhile, although the method used is documented as being independent of the X-ray protocol, paragraph

[0018] of the related Patent Document 1 describes that optimizing according to the X-ray acquisition protocol improves the segmentation performance, and since the discrimination by adaptive thresholding is not sufficient, a region growing algorithm is used to cover defects in the segmentation mask in addition to the result of discrimination by thresholding.

[0006] Like Patent Document 1, Patent Document 2 assumes that the transition speed and length are different between different segmentation regions in an X-ray image. Unlike adaptive thresholding, it generates superpixels belonging to 10 different classes in the X-ray image and uses an algorithm based on edge transitions between superpixels, which is not only faster but also better reflects the relationship between pixels. However, this method assumes that the transitions between different segmentation regions have different characteristics, and that defects in the collimation region on the superpixels are compensated for by a region growing algorithm, which is computationally expensive.

[0007] Unlike Patent Documents 1 and 2, instead of a threshold-based discrimination method that adapts to the transition between segmentation regions, the methods of Patent Documents 3 and 4 use a line score method that uses an edge detection filter on the X-ray image, and when combined with a Hough transform, considers the edge with the longest straightness as the collimation edge. While the accuracy of the Hough transform method is known to be higher than that of threshold-based discrimination algorithms, the use of this method presupposes the presence of a collimation region in the X-ray image. Whereas threshold-based discrimination methods can detect the absence of collimation when a collimation region is absent, Patent Documents 3 and 4 require an additional solution.

[0008] As a result of related research, we found EP 742536 (Patent Document 5). This patent application relates to recording one or more exposure regions. This patent application mentions automatically determining the location of boundaries between multiple exposures and between signal and shadow regions within each exposure. However, this patent application does not mention the steps of image subsampling, adding virtual collimation, performing 3-center k-means segmentation, detecting an initial collimation mask, detecting estimated collimation blades, performing 2-center k-means segmentation of the region outside the collimation blades, generating morphological structuring elements by adaptive variance thresholding, completing the 3-center segmentation region by morphological image occlusion (inpainting, hole filling), and decomposing the anatomical direct exposure region by 2-center k-means segmentation.

[0009] Therefore, the shortcomings and inadequacies of the existing solutions discussed above have created a need for advances in the related art area. [Prior art documents] [Patent documents]

[0010] [Patent Document 1] European Patent No. 1501048 [Patent Document 2] U.S. Patent No. 5,268,967 [Patent Document 3] U.S. Patent No. 5,629,989 [Patent Document 4] U.S. Patent No. 5,901,240 [Patent Document 5] European Patent No. 742536 Summary of the Invention [Problem to be solved by the invention]

[0011] Object of the invention The present invention was conceived by the current state of the art and aims to overcome the above-mentioned drawbacks.

[0012] The main objective of the present invention is to provide extra collimation region detection that is independent of the X-ray acquisition protocol, segmentation-based X-ray image region detection, and anatomical region decomposition with fast results using sub-sampling. [Means for solving the problem]

[0013] To achieve the above-mentioned object, the present invention provides a segmentation-based method for detecting collimation regions and decomposing anatomical regions in X-ray images, which includes the steps of downsampling and image subsampling of an X-ray image received from a detector, ensuring that pixel-to-pixel distribution is preserved; adding virtual collimation by virtually adding zero pixels to both dimensions of the downsampled X-ray image; directly exposing the downsampled X-ray image, which is known to consist of three image segments; taking prior information about non-collimation regions and anatomical regions and performing 3-centered k-means segmentation; and merging the direct exposure segments and anatomical region segments by k-means to obtain a single image matrix. the steps of: determining an initial collimation mask; determining four points of the mask that have the smallest Euclidean distance to the four corners of the X-ray image; generating estimated collimation blades by connecting two consecutive points with a straight line; determining an initial collimation mask and detecting estimated collimation blades; performing 2-center k-means segmentation of the area outside the estimated collimation blades, with initial center values ​​that are a specific percentage of the minimum pixel value and a specific percentage of the maximum pixel value of the area of ​​interest; generating new estimated collimation blades by re-determining the four corners that are closest to the four corners of the X-ray image according to the pixel distribution of the acquired anatomical area segment and direct exposure area segment; and generating morphological structuring elements by adaptive variance threshold discrimination. These processing steps include performing morphological image occlusion processing on the collimation mask by structuring elements, completing the three-center segmentation region by morphological image occlusion, completing the collimation segment finding process by the original size collimation mask, and isolating the anatomical direct exposure region by two-center k-means segmentation.

[0014] The structural and functional features of the present invention, as well as all of the advantages thereof, will be more clearly understood from the following drawings and detailed description that follow, and should therefore be evaluated in light of the drawings and detailed description.

[0015] Drawings useful for understanding the present invention [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a flowchart of the segmentation-based X-ray image collimation detection and anatomical region decomposition method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] Reference Number Explanation 1000 Segmentation-based method for detecting collimation regions and decomposing anatomical regions in X-ray images 1001 Image Subsampling 1002 Add virtual collimation 1003 3-centered k-means segmentation 1004 Initial collimation mask detection and estimated collimation blade detection 1005 2-center k-means segmentation of the region outside the estimated collimation blades 1006 Morphological Structuring Element Generation by Adaptive Variance Threshold Discrimination 1007 Completing 3-Center Segmentation Regions by Morphological Image Closure (Hole Filling) 1008 Anatomical Direct Exposure Region Decomposition by 2-Centered k-Means Segmentation

[0018] Detailed Description of the Invention In this detailed description, a preferred embodiment of a novel segmentation-based X-ray image collimation region detection and anatomical region decomposition method is described solely for the purpose of better understanding the subject matter.

[0019] Images from digital radiography devices consist of three distinct segments: direct exposure, non-collimated regions, and anatomical regions. Commonly used methods in the literature for segmenting digital images are the k-means algorithm and region-growing algorithm. In theory, both methods can segment the three segments in an x-ray image with a finite number of iterations and a reasonable initial estimate of the segment centers. However, in practical test scenarios, segmentation algorithms often produce inaccurate results due to the location of the x-ray hit, medical equipment on the patient, etc. Additionally, the resolution of detectors that capture x-ray images has increased with technology, and the most common adult-sized detectors on the market now consist of approximately 6–7 million pixels. Because the iterative structure of the segmentation algorithm requires reprocessing previous pixels in the x-ray image at each iteration step, this situation results in losses in processing time, processor resources, and power consumption.

[0020] The inaccurate results of the segmentation algorithm are mainly due to the confusion of anatomical regions with non-collimated regions. When trying to find two segment centers (collimated and non-collimated regions) instead of three, the direct exposure regions in the X-ray image are assigned to the correct segments, but the anatomical and collimated regions are mixed together. This situation causes deep valleys in the segmented image depending on the patient position and anatomical regions.

[0021] In a segmented image, certain holes, gaps, or slits can be filled with appropriate structural elements by morphological image closing. However, the shape formed as a result of segmentation by two segment centers in an X-ray image does not have a fixed geometric shape but is highly variable.

number

[0022] The above-mentioned segmentation-based X-ray image collimation region detection and anatomical region decomposition method consists of a k-means block that is independent of the X-ray protocol and is continuously cascaded, and a morphology processing block that is used to regularize the segmentation results. The mathematical expression of the k-means algorithm is given in equation (1), where k represents the number of segmentation centers, the subscript S(Si) represents the region adjusted by morphology processing (if applied) and segmented by the k-means algorithm, and the subscript μ(μi) represents the average pixel value of the current segmentation center. A flowchart of the segmentation-based X-ray image collimation region detection and anatomical region decomposition method is shown in the figure.

[0023] Subsampling is widely used in computer vision applications to optimize resource utilization. Although subsampling introduces loss in the original data, when subsampling is performed with an appropriate interpolation method, the relationships between pixels in the original image are largely preserved. Since the size of an X-ray image is approximately 6-7 million pixels, the image from the X-ray detector is subsampled at a rate of 1 in 16 using bicubic (bicubic) interpolation. In this way, each pixel in the subsampled image also contains information about its four neighbors in the original image.

[0024] The application of a collimation region is a control used to limit x-ray scatter during x-ray imaging to reduce the radiation dose received by the patient. The device that performs the collimation process within x-ray equipment is called a collimator and consists of two pairs of lead blades, each pair moving in different directions along the same axis. The collimation region is the unobstructed portion of the lead blades in the x-ray image and appears as a polygon with at least four sides, depending on the position of the x-ray detector. However, the collimation region is not required and can be omitted from the x-ray image at the radiologist's discretion. To provide flexibility for different scenarios, a method identified in the figures adds a virtual non-collimation region consisting of "0" pixels on both axes of the subsampled image.

[0025] By adding virtual collimation, a subsampled X-ray image, known to consist of three segments (direct exposure, non-collimated region, and anatomical region), is segmented using 3-centered k-means. The initial segment center values ​​of the 3-centered k-means algorithm are adaptively determined according to the minimum and maximum values ​​of the corresponding X-ray image. The direct exposure region of the resulting image can be accurately captured, but the accuracy of the anatomical and non-collimated regions is very low. The direct exposure region and the anatomical region are combined to form an initial collimation region, and an initial collimation region mask is obtained by determining the points closest to the four corners of this region in the X-ray image. The initial collimation region mask is a rectangle, with two opposite sides corresponding to the estimated positions of the collimation blades. The purpose of the 3-centered k-means block is to capture the direct exposure region and the anatomical region in different segments, which differ from each other in terms of pixel intensity value distribution. Using 2-centered k-means, the anatomical region and the non-collimated region can be captured in a single segment. In addition, there is limited information that the direct exposure area is solely responsible for the collimation area.

[0026] The 3-center k-means algorithm and the resulting estimated collimation blades are an underestimation of the actual collimation mask. Directly exposed regions are easily separated due to their high X-ray exposure, and the boundaries between directly exposed regions and non-collimated regions and between directly exposed regions and anatomical regions obtained by 3-center k-means are highly accurate. On the other hand, the transition between anatomical regions with no or little X-ray exposure and non-collimated regions is very unclear. As a result of 3-center k-means, only a portion of the anatomical regions can be accurately segmented. To add the missing anatomical regions to the collimation region mask, the region outside the initial collimation mask is segmented using 2-center k-means, and the initial segmentation center values ​​of the 2-center k-means algorithm are adaptively determined according to the minimum and maximum values ​​of the masked X-ray image. In this way, anatomical regions that were mistakenly included in the segmentation of non-collimated regions in the 3-center k-means algorithm used for the initial segmentation are detected. However, due to X-ray scattering, a small amount of X-rays may enter regions outside the collimated region, causing some regions in the non-collimated region to approach the pixel values ​​of anatomical regions. Therefore, the 2-centered k-means algorithm also needs to resolve the blocks segmented as anatomical regions. In this context, the X-ray image is divided into four disjoint regions using a linear equation that determines the four edges of the region bounded by the estimated X-ray blade. Within each of the four regions, the 2-centered k-means algorithm checks for the presence of pixel segments that can be found as anatomical regions. If an anatomical region block exists and the number of pixels and variance of the segment exceed a dynamically determined threshold, the anatomical block is added to the collimated region mask obtained by the 3-centered k-means algorithm.

[0027] The majority of the collimation region mask is obtained through possible addition. However, the addition of anatomical regions performed using 2-centered k-means does not include anatomical regions classified as non-collimated within the region formed by the estimated collimation blades in the collimation region mask. Because these gaps remain within the estimated collimation blades, it is incorrect to include these gaps directly within the collimation region, as the collimation region may not be rectangular. Furthermore, the edges of the collimation mask must be straight, and there may be some residual regions due to segmentation. To confine the segmentation gaps within the estimated collimation blades and impose straightness constraints on the edges of the collimation mask, a rectangular region bounded by the estimated collimation blades is defined as a structuring element, and morphological image closing is applied to the collimation mask. In this way, the collimation mask for the subsampled image is obtained.

[0028] The collimation mask for the subsampled image is resized to the size of the x-ray image using bilinear interpolation at a factor of 1 in 16. The reason for using bilinear interpolation over bicubic, as opposed to subsampling the original image, is that a collimation mask of 0s and 1s creates a transition zone between 0 and 1 that does not exist when magnified, which is not the case with bilinear interpolation. To segment the anatomical and exposed regions of the original image, eliminating non-collimated regions (if possible), only 2-centered k-means is used, where the initial segmentation center values ​​are adaptively determined by the minimum and maximum values ​​of the x-ray image masked within the original size collimation mask.

[0029] A segmentation-based method for detecting collimation regions in X-ray images and decomposing anatomical regions; image subsampling, adding virtual collimation, 3-center k-means segmentation, detecting an initial collimation mask and detecting estimated collimation blades, 2-center k-means segmentation of the region outside the estimated collimation blades, generating morphological structural elements using adaptive variance threshold discrimination, completing 3-center segmentation regions using morphological image occlusion, and anatomical direct exposure regions using 2-center k-means segmentation.

[0030] In the image subsampling step, the image from the detector is reduced in number of pixels by bicubic interpolation to 1 in 16, so as to preserve as much of the pixel-to-pixel distribution as possible. This step of the method is performed for fast results and efficient use of processor resources.

[0031] In the virtual collimation addition process step, since it is not known whether the pixel-reduced X-ray image is collimated or not, zero pixels are virtually added to both dimensions of the image. This step of the method ensures that the X-ray image consists of three distinct segments, such as direct exposure, non-collimated region, and anatomical region, regardless of whether collimation was applied to the original X-ray image.

[0032] In the 3-center k-means segmentation process, a pixel-reduced X-ray image, known to consist of three segments, is fed to the k-means algorithm. The mathematical expression for the k-means algorithm is given in Equation (1). The number k in Equation (1) is the number of segment centers, i.e., 3. Each segment center represents the average pixel value of the pixels ultimately classified as direct exposure, non-collimated regions, or anatomical regions. The initial average pixel value for each segment center is calculated as a multiple of the minimum and maximum pixel values ​​of the virtual collimated image (with different coefficients for each segment). In each iteration of the k-means algorithm, the 1-norm distance from each pixel in the image to the three segment center values ​​calculated in the previous iteration is checked. Each pixel is temporarily assigned to the segment with the smallest 1-norm. The new segment center value for the next iteration is calculated from the average pixel value of the pixels temporarily assigned to each segment center in this iteration. The stopping criterion is that the segment center value remains constant for two consecutive iterations. The purpose of these algorithmic stops is to pre-image direct exposure, non-collimated areas, and anatomical areas.

[0033] In the initial collimation mask detection and estimated collimation blade detection processing steps, the direct exposure segment and the anatomical region segment found as a result of k-means are combined into a single image mask, and four points of this mask with the smallest Euclidean distance to the four corners of the X-ray image are determined. Two consecutive points are connected by a line to form an estimated collimation blade. This step of the method aims to obtain position information about the segmentation region and correct segmentation errors.

[0034] In the 2-center k-means segmentation of the region outside the estimated collimation blades, each estimated collimation blade drawn on the X-ray image in process step 1004 is represented by a mathematical line equation. In total, four line equations corresponding to the four edges divide the X-ray image into four regions. If pixels belonging to the collimation segment are present within these four regions, a 2-center k-means algorithm is run to ensure that the initial center value within each region that satisfies this condition is a specific percentage of the minimum pixel value and a specific percentage of the maximum pixel value within that region. The 2-center k-means algorithm in this step functions similarly to the 3-center k-means algorithm in process step 1003, only the number of centers is reduced to two. The purpose of this step of the method is to fill in defects (direct exposure + anatomical region) within the collimation region of the collimation segment obtained as a result of the segmentation in process step 1003. The 2-center k-means algorithm is used up to four times to detect possible defects in the segmentation region.

[0035] In the process step of generating morphological structuring elements using adaptive variance thresholding, not all of the regions found in process step 1005 correspond to missing collimation regions. Due to X-ray scattering, non-collimation regions may also contain portions with a distribution similar to that of anatomical regions. To avoid adding such extra regions to the collimation region, a variance value is adaptively calculated according to the pixel distribution of the anatomical region segment and the direct exposure region segment obtained at the end of process step 1003. Each of the possible collimation regions (up to four regions) outside the estimated collimation blades obtained at the end of process step 1005 is thresholded according to the adaptively calculated variance value, and regions below the threshold are not included in the collimation mask. Regions above the threshold are merged with the initial collimation mask to form a new collimation mask. At the end of process step 1005, after excluding possible collimation regions using adaptive variance thresholding, the majority of the boundary of the collimation mask is completed. However, due to the nature of the k-means algorithm used in process step 1003, various gaps may exist in the collimation mask. Furthermore, the positional scatter of possible collimation regions found by the 2-centered k-means algorithm used in process step 1005 may not be physically appropriate for the linear structure of the collimation blades. To address these issues, new estimated collimation blades are generated by re-determining the four corners of the collimation mask that are closest to the four corners of the X-ray image after adaptive thresholding. The region bounded by these rectangular structures is filled and converted into a binary structuring element. A morphological image closing process is applied to the collimation mask using these structuring elements. In this way, linearity constraints are applied to the edges of the collimation region and gaps within the collimation mask are filled.

[0036] In the processing step of completing the three-center segmentation region by morphological image closing, a morphological image closing process is applied to the collimation mask using the structuring element generated in processing step 1006. In this way, linearity constraints are applied to the edges of the collimation region and gaps inside the collimation mask are filled. After that, the binary image mask, which is the collimation mask, is expanded to the original image size at a ratio of 1:16. During the expansion of the binary image, bilinear interpolation is used in this part to avoid values ​​between 0 and 1.

[0037] In the processing steps of 2-center k-means segmentation and anatomical direct exposure region decomposition, collimation segment detection is completed using the original-sized collimation mask resulting from processing step 1007. As the final step of the above method, a 2-center k-means algorithm is performed only within the collimation region mask using the k-means algorithm in Equation (1). This method works similarly to the 3-center k-means algorithm in processing step 1003, only the number of centers is reduced to two. Theoretically, the k-means algorithm in processing step 1003 can find a total of three segments: direct exposure, non-collimation, and anatomical regions. However, in practical applications (real X-ray images), anatomical and non-collimation regions are intermixed. For this reason, the task of finding collimation regions is considered a separate step. Meanwhile, the decomposition of direct exposure and anatomical regions using the k-means algorithm is highly accurate.

[0038] References [1] EP 1501048, entitled “METHOD OF SEGMENTING A RADIOGRAPHIC IMAGE INTO DIAGNOSTICALLY RELEVANT AND DIAGNOSTICALLY IRRELEVANT REGIONS” [2] U.S. Patent No. 5,268,967, entitled "METHOD FOR AUTOMATIC FOREGROUND AND BACKGROUND DETECTION IN DIGITAL RADIOGRAPHIC IMAGES" [3] U.S. Patent No. 5,629,989, entitled "IMAGE LINE - SEGMENT EXTRACTING APPARATUS" [4] U.S. Patent No. 5,901,240, entitled "METHOD FOR DETECTING THE COLLIMATION FIELD IN A DIGITAL RADIOGRAPHY"

Claims

[Claim 1] A segmentation-based X-ray image collimation region detection and anatomical region decomposition method, comprising: a processing step of downsampling and image subsampling the X-ray image from the detector to ensure that pixel-to-pixel distribution is preserved; a processing step of virtually adding zero pixels to both dimensions of the X-ray image to add virtual collimation to the reduced pixel X-ray image; a processing step of pre-generating images for direct exposure, non-collimated regions, and anatomical regions of the reduced pixel X-ray image, known to consist of three image segments, and performing a 3-centered k-means segmentation; a processing step of determining an initial collimation mask and determining the estimated collimation blades by combining the direct exposure segment and the anatomical region segment into a single image mask using k-means, determining four points of the mask that have the smallest Euclidean distance to the four corners of the X-ray image, and connecting two consecutive points of the points with a line to generate estimated collimation blades; performing a 2-center k-means segmentation of the area outside the estimated collimation blades, with initial center values ​​being a specified percentage of the minimum pixel value and a specified percentage of the maximum pixel value in the area of ​​interest, and completing the collimation mask with segments that exceed a threshold; a processing step of generating new estimated collimation blades by re-determining the four corners closest to the four corners of the X-ray image according to pixel distributions of the acquired segment of the anatomical region and the segment of the direct exposure, and generating a morphological structuring element by adaptive variance threshold discrimination; a processing step of performing a morphological image closing operation on the collimation mask by the morphological structuring element to complete a three-center segmentation region by morphological image occlusion; a processing step of completing a collimation segment detection process using the collimation mask in its original size and decomposing the anatomical direct exposure region by bicentric k-means segmentation; A method comprising:

Images