X-ray image collimation region detection and anatomical region decomposition method based on segmentation

Abstract

The invention relates to a segmentation-based X-ray image collimation region detection and anatomical region decomposition method. A cascaded k-means clustering block which is uniquely arranged in an X-ray protocol and a morphological operation block which is used for regularizing a segmentation result are included.

Description

Technical Field

[0001] The present invention relates to a segmentation-based X-ray image collimation region detection and anatomical region decomposition method, comprising a cascade k-means clustering block independent of an X-ray protocol and a morphological operation block for regularizing the segmentation result. Background Art

[0002] Radiography is the oldest imaging method and has been widely used since the discovery of X-rays. Imaging is achieved through the penetrating properties of X-rays. The basic principle is to allow X-rays to penetrate all layers of human tissue to form an image on a certain area.

[0003] The application of the collimation area is to limit the scattered X-rays used in X-ray imaging in order to reduce the radiation dose received by the imaged patient. The unit in the X-ray machine that performs the collimation process is called a collimator, which consists of two sets of lead blades, each set of blades moving in different directions on the same axis. The collimation area refers to the part that is not blocked by the lead blades and appears as a polygon of at least four sides in the X-ray image according to the position of the X-ray detector. However, the collimation area is not mandatory and the X-ray imaging technician can decide whether to keep the collimation area in the image based on his judgment.

[0004] Segmentation is the process of creating homogeneous image objects (segments) based on the defining features of the image. During segmentation, image objects should correspond to related real-world objects. Segmentation and feature detection are the basis of classification, and segmentation is the most important stage of object-based classification. Since out-of-collimation areas, anatomical areas, and directly exposed areas constitute fundamentally different image categories in X-ray images, detecting these areas using segmentation-based methods allows the development of algorithms that are independent of X-ray protocols.

[0005] With the digitization of direct X-ray imaging, various image processing algorithms can be used to achieve gains such as direct exposure on the acquired images through collimation and anatomical region segmentation. However, considering the radiology workflow, algorithms that can provide fast and accurate results are particularly important. The simplest and fastest method to generate different segmentation clusters is the threshold method, an adaptive version of which is applied in [1] and is independent of the X-ray protocol (i.e., the imaged anatomical region such as chest X-ray). However, its adaptive threshold update method assumes that the edge transitions between direct exposure-anatomical regions and between anatomical-collimation regions are sharp and have different speeds. The candidate segmentation edge transition detection algorithm needs to process each row and column of the X-ray image separately, which is very resource-intensive considering that a normal X-ray image contains about 6-7 million pixels. In addition, considering the scattering geometry of the X-rays and the thickness of the imaged patient, it can be observed that the edge transition speeds between different regions are very similar. On the other hand, although the method is documented as being independent of the X-ray protocol, paragraph

[0018] of the relevant literature [1] states that the segmentation performance improves when optimized according to the X-ray acquisition protocol, and that due to the shortcomings of the adaptive thresholding method, a region growing algorithm is used on the thresholding result to compensate for the defects of the segmentation mask.

[0006] Similar to [1], [2] assumes that the transition speed and length between different segmented regions in an X-ray image are different from each other. Unlike the adaptive threshold method, this method creates superpixels belonging to 10 different categories in the X-ray image and uses an algorithm based on the edge transition between superpixels. This method is not only faster but also better reflects the relationship between pixels. However, this method assumes that the transitions between different segmented regions have different characteristics, and the defects of the collimated regions on the superpixels need to be compensated by a region growing algorithm with higher computational cost.

[0007] Unlike references [1] and [2], the methods of references [3] and [4] do not use a threshold method that adapts to the transition between segmented regions. Instead, they use a line scoring method that uses an edge detection filter on the X-ray image and combines it with the Hough transform to consider the edge with the longest linearity as the collimated edge. Although the accuracy of the method using the Hough transform is considered to be higher than the threshold algorithm, this method assumes that there is a collimated region in the X-ray image. In the absence of a collimated region, the threshold-based method can detect the lack of collimation, while in references [3] and [4], additional solutions are required.

[0008] The results of the relevant research show that application EP742536B1 was discovered. The application relates to a method for identifying one or more irradiation areas, and mentions a method for automatically determining the position of the boundary between the signal and shadow areas between multiple exposures and within each exposure. However, the application does not mention the steps of image downsampling, adding virtual collimation, performing 3-center k-means segmentation, initial collimation mask detection, estimated collimation leaf detection, 2-center k-means segmentation of the area outside the collimation leaves, generating morphological structural elements through adaptive variance thresholding, completing 3-center segmented areas through morphological image closing operations, and achieving anatomical direct exposure area decomposition through 2-center k-means segmentation.

[0009] Therefore, due to the above-mentioned drawbacks and inadequacies of existing solutions, improvements in the related art become necessary. Summary of the invention

[0010] The present invention is inspired by the prior art and aims to solve the above-mentioned defects.

[0011] The main purpose of the present invention is to provide an additional collimation area detection independent of the X-ray acquisition protocol, to achieve segmentation-based X-ray image area detection and anatomical region decomposition through downsampling, and to output the results quickly.

[0012] To achieve the above-mentioned purpose, the present invention provides a segmentation-based X-ray image collimation region detection and anatomical region decomposition method, comprising the following steps: downsampling and image downsampling the X-ray image received by the detector to ensure that the distribution between pixels is preserved; virtually adding zero pixels in two dimensions of the downsampled X-ray image, and adding virtual collimation; for the downsampled X-ray image known to be composed of three image segments, creating image prior information for direct exposure, non-collimated area and anatomical area, and performing 3-center k-means segmentation; merging the direct exposure and anatomical region segments obtained by k-means into a single image mask, determining the four points in the mask with the smallest Euclidean distance to the four corners of the X-ray image, and linearly combining two consecutive points to generate Estimate the collimation blades, determine the initial collimation mask and detect the estimated collimation blades; perform 2-center k-means segmentation on the area outside the estimated collimation blades so that the initial center value is a specific ratio of the minimum and maximum pixel values ​​of the area involved; redetermine the 4 points closest to the 4 corners of the X-ray image based on the obtained pixel distribution of the anatomical region fragments and the direct exposure region fragments to generate new estimated collimation blades, and generate a morphological structure element through an adaptive variance threshold, that is, use the morphological structure element to perform a morphological image closing operation on the collimation mask, complete the 3-center segmented area through the morphological image closing operation, and finally complete the collimation fragment search process through the collimation mask of the original size, and separate the anatomical direct exposure area through 2-center k-means segmentation.

[0013] The structural and functional features of the present invention and all its advantages will be more clearly understood through the accompanying drawings and the detailed description made in conjunction with these drawings shown below. Therefore, the evaluation should be considered in conjunction with these drawings and the detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flow chart of the segmentation-based X-ray image alignment region detection and anatomical region decomposition method of the present invention is shown.

[0015] Part Number Description

[0016] 1000 Segmentation-based X-ray image alignment region detection and anatomical region decomposition method

[0017] 1001 Image Downsampling

[0018] 1002 Add virtual collimation

[0019] 1003 3-center k-means segmentation

[0020] 1004 Initial collimation mask detection and estimated collimation blade detection

[0021] 1005 Estimation of 2-center k-means segmentation of the outer area of ​​the collimator blade

[0022] 1006 Generating Morphological Structural Elements via Adaptive Variance Thresholding

[0023] 1007 Complete the 3-center segmentation region through morphological image closing operation

[0024] 1008 Anatomical Direct Exposure Region Decomposition via 2-center K-means Segmentation DETAILED DESCRIPTION

[0025] In this detailed description, preferred embodiments of the segmentation-based X-ray image collimation region detection and anatomical region decomposition method are described, which is only intended to provide a better understanding of the subject matter of the present invention.

[0026] The images acquired by digital X-ray imaging devices consist of three different segments: direct exposure, non-collimated area, and anatomical area. The commonly used digital image segmentation methods in existing literature are k-means and region growing algorithms. In theory, these two methods can segment the three segments in the X-ray image in a limited number of iterations under appropriate initial segment center estimation. However, in actual test scenarios, the accuracy of the segmentation algorithm is low due to factors such as the shooting position of the X-ray imaging and the medical equipment on the patient. In addition, with the development of technology, the resolution of detectors for acquiring X-ray images continues to increase, and the most common adult-sized detectors on the market contain about 6-7 million pixels. Due to the iterative structure of the segmentation algorithm, all pixels in the X-ray image need to be reprocessed in each iteration step, which will result in a loss of processing time, processor resources, and power consumption.

[0027] The incorrect results of the segmentation algorithm are mainly due to the confusion between anatomical regions and non-X-ray collimated regions. When trying to find 2 segment centers (collimated and non-collimated regions) instead of 3 segment centers, although the directly exposed areas in the X-ray image are correctly assigned to the corresponding segments, the anatomical regions and collimated regions are mixed with each other, which can form deep valleys in the segmented image depending on the patient's position and the anatomical regions.

[0028] In the segmented image, some holes, gaps or seams can be filled with appropriate structural elements through morphological image closing operations. However, the shapes formed by segmentation using 2 segment centers in X-ray imaging have no fixed geometric structure and vary greatly.

[0029]

[0030] The segmentation-based X-ray image alignment region detection and anatomical region decomposition method consists of a series of cascaded k-means blocks independent of the X-ray protocol and a morphological operation block for regularizing the segmentation results. The mathematical expression of the k-means algorithm is shown in (1), where k represents the number of segmentation centers and the subscript S (S i ) (if applicable) denotes the region regularized by the morphological operation and segmented by the k-means algorithm, and the subscript μ (“μ i ”) represents the average pixel value of the current segmentation center. The flowchart of the segmentation-based X-ray image collimation region detection and anatomical region decomposition method is as follows Figure 1 shown.

[0031] Downsampling is widely used in computer vision applications to optimize resource utilization. Although downsampling results in loss of original data, the relationship between pixels in the original image is largely preserved when downsampling is performed through appropriate interpolation methods. Since the size of X-ray images is about 6-7 million pixels, the images acquired by the detector are downsampled at a ratio of 1:16 using bicubic interpolation. In this way, each pixel in the downsampled image also contains the information of the 4-neighborhood in the original image.

[0032] The application of collimation area is to limit the scattered X-rays used in X-ray imaging to reduce the radiation dose received by the imaged patient. The unit in the X-ray machine that performs the collimation process is called a collimator, which consists of two sets of lead blades, each set of blades moving in different directions on the same axis. The collimation area refers to the part that is not blocked by the lead blades, and appears as a polygon of at least four sides in the X-ray image according to the position of the detector. However, the collimation area is not mandatory, and the X-ray technician can decide whether to keep the collimation area in the image based on his judgment. In order to provide flexibility for different scenarios, the method shown in the figure adds a virtual non-collimation area consisting of "0" pixels on both axes of the downsampled image.

[0033] By adding virtual collimation, it is known that the downsampled X-ray image consisting of 3 segments (direct exposure, non-collimated region and anatomical region) will be segmented by the 3-center k-means method, and its initial segmentation center value is adaptively determined according to the minimum and maximum values ​​of the corresponding X-ray image. The direct exposure region can be obtained from the segmentation result with high accuracy, but the accuracy of the anatomical region and non-collimated region is low. The direct exposure and anatomical regions are merged to form the initial collimated region, and the initial collimated region mask is obtained by determining the 4 points in the region with the smallest Euclidean distance to the 4 corners of the X-ray image. The initial collimated region mask is a quadrilateral whose two opposite sides correspond to the estimated positions of the collimator blades. The purpose of the 3-center k-means block is to obtain the direct exposure region and the anatomical region with different distributions of pixel brightness values. When the 2-center k-means method is used, the anatomical region and the non-collimated region may be obtained in the form of one segment. In addition, the direct exposure region alone carries limited information about the collimated region.

[0034] The 3-center k-means method and the resulting estimated collimation leaves are an underestimate of the actual collimation mask. Due to high X-ray exposure, the directly exposed area is easily separated, so the directly exposed-non-collimated area and directly exposed-anatomical boundaries obtained from 3-center k-means have high accuracy. On the other hand, the transition between the anatomical area with no or little X-ray exposure and the non-collimated area is highly uncertain. The result of 3-center k-means is that only part of the anatomical area can be correctly segmented. In order to add the missing anatomical area to the collimation mask, a 2-center k-means segmentation is performed on the area outside the initial collimation mask, and its initial segmentation center value is adaptively determined according to the minimum and maximum values ​​of the mask X-ray image. In this way, the anatomical area that is incorrectly classified as the non-collimated area segmentation in the 3-center k-means initial segmentation can be detected. However, due to X-ray scattering, a small amount of X-rays may fall outside the collimated area, and the pixel values ​​of some parts of the non-collimated area will be close to the anatomical area. Therefore, in the 2-center k-means method, it is also necessary to decompose the blocks segmented into anatomical areas. In this context, the X-ray image is divided into 4 non-intersecting regions using the equations of the lines that determine the four sides of the boundary region framed by the estimated X-ray leaves. In each of these 4 regions, the 2-center k-means method is checked for the presence of a pixel fragment that is identified as an anatomical region. If an anatomical region block exists and the number of pixels and the variance of this fragment exceed dynamically determined thresholds, this anatomical block is added to the collimation mask obtained by 3-center k-means.

[0035] Most of the collimation masks have possible additions. However, the anatomical regions added by the 2-center k-means method are not included in the anatomical regions classified as non-collimated within the area formed by the estimated collimation leaves of the collimation mask. Since these gaps are located within the estimated collimation leaves, it is incorrect to directly include them in the collimation area because the collimation area may not be a quadrilateral. In addition, the edges of the collimation mask must be linear and some areas may remain due to segmentation. In order to eliminate the segmentation gaps inside the estimated collimation leaves and impose linear constraints on the edges of the collimation mask, the quadrilateral region framed by the estimated collimation leaves is defined as a structural element, and a morphological image closing operation is performed on the collimation mask. The collimation mask of the downsampled image is thus obtained.

[0036] The downsampled collimation mask is adjusted to the original size of the X-ray image by bilinear interpolation at a ratio of 1:16. Unlike bicubic interpolation, bilinear interpolation is chosen instead of downsampling the original image because the collimation mask composed of 0s and 1s will produce a non-existent 0-1 transition band when enlarged, and bilinear interpolation can avoid this phenomenon. To segment the anatomical area and the direct exposure area in the original image with the non-collimation area (if any) removed, only the dual-center k-means method with adaptive determination of the initial segmentation center value is used - the initial segmentation center value is dynamically determined by the minimum and maximum values ​​of the masked X-ray image within the original size collimation mask.

[0037] The segmentation-based X-ray image collimation area detection and anatomical area decomposition method includes image downsampling, virtual collimation addition, 3-center k-means segmentation, initial collimation mask detection and estimated collimation leaf detection, 2-center k-means segmentation of the area outside the estimated collimation leaves, generation of morphological structure elements through adaptive variance threshold, completion of 3-center segmentation area through morphological image closing operation, and anatomical direct exposure area decomposition through 2-center k-means segmentation.

[0038] In the image downsampling step, the image acquired by the detector is downscaled by a ratio of 1:16 using bicubic interpolation to preserve the pixel distribution as much as possible. This step of the method is designed to obtain results quickly and use processor resources efficiently.

[0039] In the virtual collimation addition processing step, since it is uncertain whether the reduced X-ray image is collimated, zero pixels are virtually added in two dimensions of the image. This step of the method ensures that no matter whether the original X-ray image is collimated, the X-ray image always consists of three independent segments: direct exposure, non-collimated area, and anatomical area.

[0040] In the 3-center k-means segmentation processing step, the reduced X-ray image consisting of 3 image segments is input into the k-means algorithm, and its mathematical expression is shown in (1). The k in the mathematical expression (1) is taken as the number of segment centers, that is, 3. Each segment center represents the average pixel value of the pixels that are finally classified as direct exposure, non-collimated area or anatomical area. The initial average pixel value of each segment center is calculated as a multiple of the minimum and maximum pixel values ​​of the virtual collimated image (the coefficient is different for each segment). In each iteration of the k-means algorithm, the 1-distance (1-norm) of each pixel in the image to the 3 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 based on 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 unchanged in 2 consecutive iterations. The purpose of this algorithm step is to generate image priors for direct exposure, non-collimated area and anatomical area.

[0041] In the initial collimation mask detection and estimated collimation leaf detection processing steps, the direct exposure and anatomical region segments obtained by k-means are merged into a single image mask, and the four points in the mask with the smallest Euclidean distance to the four corners of the X-ray image are determined. Two consecutive points are linearly connected to form the estimated collimation leaves. This step of the method aims to obtain the position information of the segmented region to correct segmentation errors.

[0042] In the 2-center k-means segmentation of the area outside the estimated collimator blades, each estimated collimator blade drawn on the X-ray image in step 1004 is represented by a mathematical straight line equation. A total of 4 straight line equations corresponding to 4 edges divide the X-ray image into 4 regions. If there are pixels belonging to the collimation segment in these 4 regions, the 2-center k-means algorithm is run so that the initial center value in each region that meets this condition is a specific ratio of the minimum and maximum pixel values ​​of the region. The 2-center k-means algorithm in this step works in the same way as the 3-center k-means algorithm in step 1003, except that the number of centers is reduced to 2. This step of the method aims to complete the collimation area defects of the collimation segment (direct exposure + anatomical area) segmented in step 1003. The 2-center k-means algorithm is run up to 4 times to detect possible segmentation area defects.

[0043] In the processing step of generating morphological structural elements by adaptive variance thresholding, not all regions found in step 1005 correspond to missing collimation regions. Due to X-ray scattering, non-collimation regions may also contain parts with similar distribution to anatomical regions. In order to avoid adding these additional regions to the collimation region, variance values ​​are adaptively calculated based on the pixel distribution of the anatomical region fragments and the directly exposed region fragments obtained at the end of step 1003. Each possible collimation region (up to 4 regions) outside the estimated collimation leaves obtained at the end of step 1005 is thresholded according to the adaptively calculated variance value, and the region below the threshold is not included in the collimation mask. The region above the threshold is merged with the initial collimation mask to form a new collimation mask. At the end of step 1005, the boundary of the collimation mask has been basically confirmed after eliminating possible collimation regions by adaptive variance thresholding. However, due to the nature of the k-means algorithm used in step 1003, various gaps may exist in the collimation mask. In addition, the positional scatter of the possible collimation regions found by the 2-center k-means algorithm used in step 1005 may not physically conform to the linear structure of the collimation leaves. To eliminate these problems, new estimated collimation leaves are created by re-determining the 4 points in the collimation mask closest to the 4 corners of the X-ray image after adaptive variance thresholding. The area bounded by this rectangular structure is filled and converted into a binary structure element. Through this structure element, a morphological image closing operation is applied to the collimation mask. In this way, the edges of the collimation area are linearly constrained and the gaps within the collimation mask are filled.

[0044] In the step of completing the 3-center segmented regions by morphological image closing operation, the morphological image closing operation is applied to the alignment mask for the structure element created by step 1006. In this way, the edges of the alignment region are linearly constrained and the gaps in the alignment mask are filled. Thereafter, the binary image mask serving as the alignment mask is enlarged to the original image size at a ratio of 1:16. Unlike the downsampling of the original image, bilinear interpolation is used when enlarging the binary image to avoid values ​​between 0-1 pixels.

[0045] In the step of segmentation by 2-center k-means and decomposition of the anatomical direct exposure region, the detection of the collimated segments is completed using the original size collimation mask obtained in step 1007. As the last step of the method, the 2-center k-means algorithm is run only within the collimated region mask according to the k-means algorithm in (1). This method works in the same way as the 3-center k-means algorithm in step 1003, except that the number of centers is reduced to 2. In theory, the k-means algorithm in step 1003 can also find 3 segments: direct exposure, non-collimated region and anatomical region. However, in practical applications (real X-ray images), anatomical regions and non-collimated regions are mixed with each other. Therefore, the task of finding the collimated region is performed in stages and is regarded as an independent task. On the other hand, the decomposition of the direct exposure and anatomical regions by the k-means algorithm has a high accuracy.

[0046] References

[0047] [1] EP1501048B1, “Method for segmenting a radiological image into diagnostically relevant and diagnostically irrelevant regions”

[0048] [2] US5268967A, “Method for automatic foreground and background detection in digital radiological images”

[0049] [3] US5629989A, “Image line segment extraction device”

[0050] [4] US5901240A, “Method for detecting collimation area in digital radiography”

Claims

1. A segmentation-based X-ray image collimation region detection and anatomical region decomposition method, characterized in that The processing steps include: Downsampling the X-rays acquired by the detector and downsampling the image to preserve the distribution between pixels; Virtually adding zero pixels in both dimensions of the downsampled image and adding virtual collimation; Create image priors for direct exposure, non-collimated region, and anatomical region for downsampled X-ray images known to consist of 3 image segments, and perform 3-center k-means segmentation; Merge the direct exposure and anatomical region segments obtained by k-means into a single image mask, determine the four points in the image mask with the smallest Euclidean distances to the four corners of the X-ray image, linearly connect two consecutive points to generate estimated collimation leaves, and determine the initial collimation mask and the estimated collimation leaves; Perform 2-center k-means segmentation on the area outside the estimated collimation leaves, with the initial center value being a specific ratio of the minimum and maximum pixel values ​​of the area involved, and complete the collimation mask by filtering the fragments through a threshold; According to the pixel distribution of the obtained anatomical region fragment and the directly exposed region fragment, the four points closest to the four corners of the X-ray image are re-determined to generate new estimated collimation leaves, and the morphological structure elements are generated by an adaptive variance threshold; The morphological image closing operation is performed on the alignment mask using the morphological structure element, and the 3-center segmentation area is completed through the morphological image closing operation; The alignment segment detection process is completed through the original size alignment mask, and the anatomical direct exposure area decomposition is achieved through 2-center k-means segmentation.

Images