Abdominal DR Image Processing Method and System

Abstract

The present invention discloses an abdominal DR image processing method and system. The method includes the following steps: 1) Image preprocessing: performing ROI annotation, binarization processing, resampling processing, and boundary generation on the abdominal DR image to obtain a boundary binary image; 2) Statistical shape analysis: extracting the centerline, calculating the target width, and calculating the statistical shape of the boundary binary image obtained in step 1) to obtain a number of statistical shape parameters; 3) Image classification: performing image classification based on the number of statistical shape parameters obtained in step 2). The present invention proposes a method and system for analyzing the intestinal morphology of DR images based on statistical shape. Compared with the existing geometric morphology analysis scheme, the present invention has better stability for intestinal DR images with characteristics such as intestinal morphological changes, intestinal overlap phenomena, and patient position changes. The statistical shape parameters obtained by the present invention have better classification accuracy for image classification.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to an abdominal DR image processing method and system. Background Art

[0002] The abdominal tissue structure is complex, especially the intestines are serpentine in three-dimensional space, which leads to the overlap and morphological changes of the intestines in two-dimensional DR images, which brings challenges to the analysis of intestinal shape in DR images. It is difficult to obtain satisfactory results by using conventional methods based on geometric morphological analysis to analyze DR images, so it is necessary to provide a more reliable solution. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide an abdominal DR image processing method and system in view of the deficiencies in the above-mentioned prior art.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: to provide an abdominal DR image processing method, comprising the following steps:

[0005] 1) Image preprocessing: perform ROI annotation, binarization, resampling and boundary generation on the abdominal DR image to obtain a boundary binary image;

[0006] 2) Statistical shape analysis: extracting the center line, calculating the target width, and calculating the statistical shape of the boundary binary image obtained in step 1) to obtain a number of statistical shape parameters;

[0007] 3) Image classification: Image classification is performed based on the statistical shape parameters obtained in step 2).

[0008] Preferably, the step 1) specifically comprises:

[0009] 1-1) ROI annotation: Obtain the region of interest (ROI) by performing manual region annotation, threshold segmentation annotation, connected domain analysis, hole filling, and curve delineation annotation based on image gradient changes on the abdominal DR images;

[0010] 1-2) Binarization: The marked area is regenerated into a binary image with the same size and spatial resolution as the original image through binarization, where the marked area is a non-zero value and the background area is a zero value; further preferably, the non-zero value is generally defined as a positive integer from 1 to 255

[0011] 1-3) Resampling: resampling the original image and the corresponding binary image to a specified size and spatial resolution according to the set parameters; the spatial resolution may be greater than the original image resolution, and preferably, the spatial resolution is set to be isotropic, with a resolution size of 0.5-1 mm;

[0012] 1 - 4) Boundary generation: Generate a binary image including the target boundary from the binary image obtained in steps 1 - 3), denoted as the boundary binary image;

[0013] Among them, the boundary of the boundary binary image is represented by a discrete single - pixel curve, that is, the boundary width is 1 pixel width; the pixel value corresponding to the boundary is non - zero, and the rest are 0; the boundary is a closed curve, and the interior of the closed curve is the originally labeled tissue target, and the exterior is the original background area.

[0014] Preferably, the center - line extraction in step 2) specifically includes:

[0015] 2 - 1 - 1) Input the boundary binary image: Input the boundary binary image obtained in steps 1 - 4), denoted as I E ; its boundary is denoted as E, and any pixel on its boundary is denoted as e i , where i represents the total number of pixels of the boundary;

[0016] 2 - 1 - 2) Extract the center - line using the skeleton algorithm:

[0017] Ⅰ. Extract the boundary binary image input in step 2 - 1 - 1);

[0018] Ⅱ. Output a center - line image with the same size and spatial resolution as the input boundary binary image generated by the skeleton algorithm;

[0019] Among them, the center - line image satisfies:

[0020] a. The center - line image is a binary discrete image, the center - line is represented by discrete pixel points, and the pixel value is non - zero, and the rest are 0;

[0021] b. The width of the center - line composed of non - zero value pixels is 1 pixel;

[0022] c. The center - line is continuous, that is, except for the first and last pixels, there are 2 pixels in the 8 - neighborhood range of any pixel representing the center - line;

[0023] d. The center - line image is denoted as I C ; its center - line is denoted as C, and any pixel on the center - line is denoted as c j , where j ≤ the total number of pixels of the center - line;

[0024] e. The center - line C is a directed curve, and the arrangement order of the pixels on the center - line is in order: that is, given any pixel position on the center - line, other any pixel can be traversed in order or in reverse order;

[0025] 2 - 1 - 3) Discretize the center - line.

[0026] Preferably, the calculation of the target width in step 2) specifically includes:

[0027] 2-2-1) Use Hilditch's algorithm to calculate the set of labeled tissue widths corresponding to the centerline. This width set is denoted as W, and each pixel c located on the centerline j corresponds to a width value, denoted as w j ;

[0028] 2-2-2) Extract the effective centerline:

[0029] Traverse the pixels forming the centerline from the top to the bottom of the intestine. Its maximum width is denoted as w max and the corresponding pixel is denoted as p max ; Starting from p max traverse towards the bottom of the intestine until the end of the ROI. All pixels located on the centerline form the effective centerline;

[0030] Among them, the effective centerline is a subset of the centerline, and the other characteristics are the same as those of the centerline;

[0031] 2-2-3) Extract the effective width set: Denote the ROI widths corresponding to the pixels in the effective centerline as the effective width set, denoted as W val , and each element in the W val set is denoted as w k .

[0032] Preferably, the calculated statistical shape parameters in step 2) include 19, specifically:

[0033] (1) Energy:

[0034] (2) Total energy:

[0035] (3) Entropy:

[0036] (4) Maximum value: maximum = max(W val );

[0037] (5) Minimum value: minimum = min(W val );

[0038] (6) 10% percentile value: The elements in W val are sorted from small to large, and the width value at the 10% position in the positive order;

[0039] (7) 90% percentile value: The elements in W val are sorted from small to large, and the width value at the 90% position in the positive order;

[0040] (8) Mean:

[0041] (9) Median: W val The width value at the 50% position in ascending order after sorting the elements in

[0042] (10) Interquartile range: P 75 -P 25 ;

[0043] (11) Maximum difference: maximum value - minimum value;

[0044] (12) Mean absolute deviation:

[0045] (13) Robust mean absolute deviation: rMAD;

[0046] (14) Mean square error:

[0047] (15) Standard deviation:

[0048] (16) Skewness:

[0049] (17) Kurtosis:

[0050] (18) Variance:

[0051] (19) Consistency:

[0052] where c is the offset, and c takes 0 or a positive value; N p is the maximum difference after discretization of the width values in the set; N w is the number of elements in the set W val ; P t is the width value at the t% position in ascending order after sorting the elements in W val ; p(w k ) is the probability of the corresponding value range of w in the set W val ; N k is the number of value ranges; ε is a preset constant used to ensure that the value of p(wk) in Equation (3) is greater than 0. Generally, ε is 0.001 - 0.00001. h Preferably, step 3) includes:

[0053] 3-1) Feature normalization: Using the 19 statistical shape parameters obtained in step 2) as features and performing feature normalization processing;

[0054] ​

[0055] 3-2) Feature screening: One or a combination of methods such as the t-test method, the Pearson correlation test method, and the maximum relevance minimum redundancy method are used to screen the 19 features processed in step 3-1). The feature set after screening is denoted as F sel ;

[0056] 3-3) Construct a machine learning classification model, and input F sel into the machine learning classification model. The machine learning classification model outputs the corresponding classification result and the corresponding confidence level.

[0057] Preferably, for the DR images of multiple imaging positions of the same patient, the machine learning classification model gives the classification result and the corresponding confidence level of each DR image of the imaging position, and then takes the classification result corresponding to the highest confidence level as the final classification result.

[0058] The present invention also provides an abdominal DR image processing system, which processes abdominal DR images by using the method described above. The system includes:

[0059] An image preprocessing module, which includes an ROI annotation sub-module, a binarization processing sub-module, a resampling processing sub-module, and a boundary generation sub-module. The image preprocessing module preprocesses the abdominal DR image according to the method of step 1);

[0060] A statistical shape analysis module, which includes a centerline extraction sub-module, a target width calculation sub-module, and a statistical shape calculation sub-module. The statistical shape analysis module performs statistical shape analysis according to the method of step 2);

[0061] And an image classification module, which includes a feature normalization sub-module, a feature screening sub-module, and a machine learning classification model sub-module. The image classification module performs image classification according to the method of step 3).

[0062] The present invention also provides a storage medium, on which a computer program is stored, and when the program is executed, it is used to implement the method described above.

[0063] The present invention also provides a computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method described above is implemented.

[0064] The beneficial effects of the present invention are: the present invention proposes a method and system for intestinal morphology analysis of DR images based on statistical shapes. Compared with the existing geometric morphology analysis scheme, the present invention has better stability of intestinal DR images with characteristics such as intestinal morphology changes, intestinal overlapping phenomena, and patient position changes. The statistical morphological parameters obtained by the present invention have better classification accuracy when used for image classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a principle block diagram of the abdominal DR image processing system of the present invention;

[0066] Figure 2 is a schematic diagram of a binary boundary image and its center line in the present invention;

[0067] Figure 3 It is a schematic diagram of the effective center line and its corresponding target radius in the present invention. DETAILED DESCRIPTION

[0068] The present invention is further described in detail below in conjunction with embodiments so that those skilled in the art can implement the invention with reference to the description.

[0069] It should be understood that the terms such as “having”, “including” and “comprising” used herein do not exclude the existence or addition of one or more other elements or combinations thereof.

[0070] Example 1

[0071] An abdominal DR image processing method of this embodiment includes the following steps:

[0072] 1) Image preprocessing: The abdominal DR image is subjected to ROI (region of interest) annotation, binarization, resampling and boundary generation to obtain a boundary binary image.

[0073] Specifically include:

[0074] 1-1) ROI annotation: Obtain the region of interest (ROI) by performing manual region annotation, threshold segmentation annotation, connected domain analysis, hole filling, and curve delineation annotation based on image gradient changes on the abdominal DR images;

[0075] 1-2) Binarization: Through binarization, the marked area is regenerated into a binary image with the same size and spatial resolution as the original image, where the marked area is a non-zero value and the background area is a zero value; non-zero values ​​are generally defined as positive integers from 1 to 255;

[0076] 1-3) Resampling process: Resample the original image and the corresponding binary image according to the set parameters into a specified size and spatial resolution; the spatial resolution can be greater than the original image resolution. In a preferred embodiment, the spatial resolution is set to be isotropic, and the resolution size is 0.5-1 mm;

[0077] 1-4) Boundary generation: Generate the binary image obtained in step 1-3) into a binary image including the target boundary, denoted as the boundary binary image; wherein, the boundary of the boundary binary image is represented by a discrete single-pixel curve, that is, the boundary width is 1 pixel width; the pixel value corresponding to the boundary is non-0, and the rest are 0 values; the boundary is a closed curve, and the inside of the closed curve is the originally marked tissue target, and the outside is the original background area.

[0078] 2) Statistical shape analysis: Perform centerline extraction, target width calculation, and statistical shape calculation on the boundary binary image obtained in step 1) to obtain a number of statistical shape parameters. The specific steps are as follows.

[0079] 2-1) Centerline extraction:

[0080] 2-1-1) Input the boundary binary image: Input the boundary binary image obtained in step 1-4), denoted as I E ; its boundary is denoted as E, and any pixel located on the boundary is denoted as e i , i ≤ the total number of pixels of the boundary;

[0081] 2-1-2) Extract the centerline using the skeleton line algorithm:

[0082] Ⅰ. Extract the boundary binary image input in step 2-1-1);

[0083] Ⅱ. Output a centerline image with the same size and spatial resolution as the input boundary binary image generated by the skeleton line algorithm;

[0084] Among them, the centerline image satisfies:

[0085] a. The centerline image is a binary discrete image, and the centerline is represented by discrete pixel points, and the pixel value is non-0, and the rest are 0;

[0086] b. The width of the centerline composed of non-0 value pixels is 1 pixel;

[0087] c. The centerline is continuous, that is, except for the head and tail pixels, there are 2 pixels in the 8-neighborhood range of any pixel representing the centerline;

[0088] d. The centerline image is denoted as I C ; its centerline is denoted as C, and any pixel located on the centerline is denoted as c j , j ≤ the total number of pixels of the centerline;

[0089] e. The centerline C is a directed curve, and the arrangement order of the pixels on the centerline is sequential: that is, given any pixel position on the centerline, any other pixel can be traversed sequentially or in reverse order;

[0090] 2-1-3) Discretize the centerline.

[0091] Refer to Figure 2 It is a schematic diagram of a binary boundary image (white line, label 1) and its centerline (yellow line, label 2), Figure 3 It is a schematic diagram of the effective centerline and its corresponding target radius.

[0092] 2-2) Target width calculation:

[0093] 2-2-1) Use Hilditch's algorithm to calculate the set of labeled tissue widths corresponding to the centerline. This width set is denoted as W, and each pixel c on the centerline j corresponds to a width value, which is denoted as w j ;

[0094] 2-2-2) Extract the effective centerline:

[0095] Traverse the pixels forming the centerline from the top to the bottom of the intestine. Its maximum width is denoted as w max and the corresponding pixel is denoted as p max ; Starting from p max traverse towards the bottom of the intestine until the end of the ROI. All the pixels on the centerline form the effective centerline;

[0096] Among them, the effective centerline is a subset of the centerline, and the other characteristics are the same as those of the centerline;

[0097] 2-2-3) Extract the effective width set: Denote the ROI widths corresponding to the pixels in the effective centerline as the effective width set, denoted as W val W val Each element in the set is denoted as w k .

[0098] 2-3) Statistical shape parameter calculation:

[0099] The 19 statistical shape parameters include:

[0100] (1) Energy:

[0101] (2) Total energy:

[0102] (3) Entropy:

[0103] (4) Maximum value: maximum = max(Wval )

[0104] (5) Minimum: minimum = min(W val )

[0105] (6) 10% percentile value: The width value at the 10% position in ascending order after sorting the elements in W val ;

[0106] (7) 90% percentile value: The width value at the 90% position in ascending order after sorting the elements in W val ;

[0107] (8) Mean:

[0108] (9) Median: The width value at the 50% position in ascending order after sorting the elements in W val ;

[0109] (10) Interquartile range: P 75 - P 25 ;

[0110] (11) Maximum difference: Maximum value - minimum value;

[0111] (12) Mean absolute deviation:

[0112] (13) Robust mean absolute deviation: rMAD;

[0113] (14) Mean square error:

[0114] (15) Standard deviation:

[0115] (16) Skewness:

[0116] (17) Kurtosis:

[0117] (18) Variance:

[0118] (19) Consistency:

[0119] Where c is the offset, c takes 0 or a positive value; N p is the maximum difference after discretization of the width values in the set; N w is the number of elements in the W val set; P t is the width value at the t% position in ascending order after sorting the elements in W val ; p(w k) is W val The probability of w in the set k corresponding to the value range, N h is the number of value ranges; ε is a preset constant used to ensure that the value of p(w k ) in Equation (3) is greater than 0. Generally, ε is 0.001 to 0.00001.

[0120] 3) Image classification: Based on the several statistical shape parameters obtained in step 2), image classification is performed.

[0121] Specifically, it includes:

[0122] 3-1) Feature normalization: Take the 19 statistical shape parameters obtained in step 2) as features and perform feature normalization processing;

[0123] 3-2) Feature screening: Use one or a combination of methods such as the t-test method, Pearson correlation test method, and maximum correlation minimum redundancy method to screen the 19 features processed in step 3-1). The screened feature set is denoted as F sel ;

[0124] 3-3) Construct a machine learning classification model, input F sel into the machine learning classification model, and the machine learning classification model outputs the corresponding classification result and the corresponding confidence level. Among them, for the DR images of multiple imaging positions of the same patient, the machine learning classification model gives the classification result and the corresponding confidence level of each DR image of the imaging position, and then takes the classification result corresponding to the highest confidence level as the final classification result.

[0125] In one embodiment, the method of Embodiment 1 can be used to classify abdominal DR images into Hirschsprung's disease and ordinary defecation disorders (such as indigestion). When applying, first construct the training data of these two types of images to train the machine learning classification model, and then the classification can be achieved according to the method of Embodiment 1. Of course, it can be understood that the method of the present invention can also be applied to other types of classification applications.

[0126] Embodiment 2

[0127] This embodiment provides an abdominal DR image processing system, which processes abdominal DR images using the method of Embodiment 1. Referring to Figure 1 , the system includes:

[0128] An image preprocessing module, which includes an ROI annotation sub-module, a binarization processing sub-module, a resampling processing sub-module, and a boundary generation sub-module. The image preprocessing module preprocesses abdominal DR images according to the method of step 1) in Embodiment 1;

[0129] A statistical shape analysis module, which includes a centerline extraction sub-module, a target width calculation sub-module, and a statistical shape calculation sub-module, and the statistical shape analysis module performs statistical shape analysis according to the method in step 2) of Embodiment 1;

[0130] And an image classification module, which includes a feature normalization sub-module, a feature screening sub-module, and a machine learning classification model sub-module, and the image classification module performs image classification according to the method in step 3) of Embodiment 1.

[0131] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed, it is used to implement the method of Embodiment 1.

[0132] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, it implements the method of Embodiment 1.

[0133] Although the embodiments of the present invention have been disclosed as above, they are not limited to only the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to specific details.

Claims

1. An abdominal DR image processing method, characterized in that, The following steps are involved: 1) Image preprocessing: perform ROI annotation, binarization, resampling and boundary generation on the abdominal DR image to obtain a boundary binary image; 2) Statistical shape analysis: extracting the center line, calculating the target width, and calculating the statistical shape of the boundary binary image obtained in step 1) to obtain a number of statistical shape parameters; 3) Image classification: image classification is performed based on the several statistical shape parameters obtained in step 2); The step 1) specifically includes: 1-1) ROI annotation: Obtain the region of interest (ROI) by performing manual region annotation, threshold segmentation annotation, connected domain analysis, hole filling, and curve delineation annotation based on image gradient changes on the abdominal DR images; 1-2) Binarization: The obtained region of interest ROI is regenerated into a binary image with the same size and spatial resolution as the original image through binarization, where the region of interest ROI is a non-zero value and the background area is a zero value; 1-3) Resampling: resample the original image and the corresponding binary image to the specified size and spatial resolution according to the set parameters; 1-4) Boundary generation: The binary image obtained in step 1-3) is generated into a binary image including the target boundary, which is recorded as a boundary binary image; The boundary of the binary image is represented by a discrete single-pixel curve, that is, the boundary width is 1 pixel; the pixel value corresponding to the boundary is non-zero, and the rest are zero; the boundary is a closed curve, the inside of the closed curve is the originally marked tissue target, and the outside is the original background area; The centerline extraction in step 2) specifically includes: 2-1-1) Input binary boundary image: Input the binary boundary image obtained in step 1-4), denoted as I E ; its boundary is denoted as E, and any pixel on its boundary is denoted as e i , i ≤ total number of pixels of the boundary; 2-1-2) Use skeleton line algorithm to extract the center line: Ⅰ. Extract the boundary binary image input in step 2-1-1); Ⅱ. Output a centerline image with the same size and spatial resolution as the input boundary binary image generated by the skeleton line algorithm; Among them, the centerline image satisfies: a. The centerline image is a binary discrete image. The centerline is represented by discrete pixels. The pixel values ​​are non-zero and the rest are zero. b. The width of the center line composed of non-zero value pixels is 1 pixel; c. The center line is continuous, that is, except for the first and last pixels, there are 2 pixels in the 8-neighborhood range of any pixel representing the center line; d. The centerline image is denoted as I C ; its centerline is denoted as C, and any pixel located on the centerline is denoted as c j , where j ≤ the total number of pixels of the centerline; e. The center line C is a directed curve, and the arrangement order of the pixels on the center line is sequential: that is, given any pixel position on the center line, any other pixel can be traversed sequentially or in reverse order; 2-1-3) Discretization center line; The target width calculation in step 2) specifically includes: 2-2-1) Calculate the set of labeled tissue widths corresponding to the centerline using Hilditch's algorithm. This set of widths is denoted as W, and each pixel c on the centerline j corresponds to a width value, which is denoted as w j ; 2-2-2) Extract the effective center line: Traverse the pixels forming the center line from the top to the bottom of the intestine, and record the maximum width as w max , and record the corresponding pixel as p max ; Traverse from p max to the bottom of the intestine until the end of the ROI, and all the pixels located on the center line form the effective center line; Among them, the effective centerline is a subset of the centerline, and the rest of the characteristics are consistent with the centerline; 2-2-3) Extract the set of effective widths: Denote the width of the ROI corresponding to the pixels in the effective center line as the set of effective widths, denoted as W val , W val Each element in the set is denoted as w k ; The statistical shape parameters calculated in step 2) include 19, specifically: (1) Energy: (2) Total energy: (3) Entropy: (4) Maximum value: maximum = max(W val ); (5) Minimum value: minimum = min(W val ); (6) 10% site value: W val The elements in are sorted from smallest to largest, and the width value at the 10% position in ascending order; (7)90th percentile value: W val The elements in are sorted from smallest to largest, and the width value at the 90th percentile in ascending order; (8) Mean: (9) Median value: W val The width value at the 50% position in ascending order after sorting the elements in (10) Interquartile range: P 75 -P 25 ; (11) Maximum difference: maximum value - minimum value; (12) Mean Absolute Deviation: (13) Robust mean absolute deviation: rMAD; (14) Mean square error: (15) Standard deviation: (16) Skew: (17) Kurtosis: (18) Variance: (19) Consistency: where c is the offset, and c takes 0 or a positive value; N p is the maximum difference after discretization of the width values in the set; N w is the number of elements in the set W val ; P t is the width value at the positive order t% position after sorting the elements in W val from small to large; p(w k ) is the probability of the corresponding value range in the set W val ; N k is the number of value ranges; ε is a preset constant; h ​ Step 3) includes: 3-1) Feature normalization: The 19 statistical shape parameters obtained in step 2) are used as features and feature normalization is performed; 3-2) Feature screening: One or a combination of methods among the t-test method, Pearson correlation test method, and maximum relevance minimum redundancy method are used to screen the 19 features processed in step 3-1), and the feature set after screening is denoted as F sel ; 3-3) Build a machine learning classification model and input F sel into the machine learning classification model, and the machine learning classification model outputs the corresponding classification result and the corresponding confidence level.

2. The abdominal DR image processing method according to claim 1, wherein, in, For the DR images of multiple imaging positions of the same patient, the machine learning classification model gives the classification results and corresponding confidence levels of the DR images of each imaging position, and then takes the classification result corresponding to the highest confidence level as the final classification result.

3. An abdominal DR image processing system, which processes abdominal DR images by using the method described in any one of claims 1-2. The system includes: An image preprocessing module, which includes an ROI annotation sub-module, a binarization processing sub-module, a resampling processing sub-module, and a boundary generation sub-module. The image preprocessing module preprocesses the abdominal DR images according to the method of step 1). A statistical shape analysis module, which includes a centerline extraction sub-module, a target width calculation sub-module, and a statistical shape calculation sub-module. The statistical shape analysis module performs statistical shape analysis according to the method of step 2). And an image classification module, which includes a feature normalization sub-module, a feature screening sub-module, and a machine learning classification model sub-module. The image classification module performs image classification according to the method of step 3).

4. A storage medium having a computer program stored thereon, characterized in that, When the program is executed, it is used to implement the method described in any one of claims 1-2.

5. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-2.

Images