A method and device for automatic segmentation of artificial valve calcification area

Abstract

The present invention discloses a method and device for automatic segmentation of artificial valve calcification areas. Through the implementation of the present invention, before processing a target image based on a deep learning model, image pre-processing and feature extraction operations are set, thereby improving the consistency, reliability and accuracy of image data input into the deep learning model; then, the image pre-processing results and feature extraction results are learned by the deep learning model to output model output results including a segmentation result mask tensor, thereby achieving high-precision segmentation of the calcification area, which is beneficial to improving the segmentation efficiency and accuracy of the target image; finally, by setting up post-segmentation processing, the position and shape of the calcification area are intuitively displayed, and the area information of the calcification area is also provided, thereby improving the convenience of consulting relevant information of the calcification area.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and a device for automatically segmenting a calcified region of an artificial valve. Background Art

[0002] In the relevant research field of cardiac technology, artificial valves are increasingly used due to their durability, stability and other advantages. However, artificial valves also have corresponding usage defects. Specifically, over time, the valve material may fail due to calcification, resulting in decreased valve function and even serious complications. Therefore, accurate assessment of the calcification of artificial valves is crucial to the formulation of their use plans. Traditional calcification assessment methods for artificial valves rely on technicians to manually analyze CT images, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, making it difficult to guarantee the accuracy and consistency of the evaluation results.

[0003] Although the intelligent analysis of images has made significant progress in recent years with the development of computer vision and deep learning technology, the automatic segmentation technology for artificial valve calcification areas still faces many challenges. On the one hand, the artificial valve structure is complex, the calcification area has various morphologies, and is easily disturbed by surrounding tissues, making accurate segmentation particularly difficult; on the other hand, existing methods often lack the means to extract specific features of artificial valves, resulting in limited prediction capabilities of deep learning models, and the accuracy and robustness of segmentation results need to be improved. Summary of the invention

[0004] The present invention provides a method and device for automatically segmenting a calcified region of an artificial valve, which can improve the feature extraction accuracy of the artificial valve, and simultaneously improve the segmentation efficiency and segmentation accuracy of the calcified region of the artificial valve.

[0005] In order to solve the above technical problems, the present invention discloses, in a first aspect, a method for automatically segmenting a calcified region of an artificial valve, the method comprising:

[0006] Acquire a target image to be segmented, and perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image; the target image includes a plurality of CT images of cross sections of artificial valves; the image pre-processing result includes at least an image tensor corresponding to the target image;

[0007] According to a preset section feature extraction module, a feature extraction operation is performed on the image pre-processing result to obtain a feature extraction result corresponding to the image pre-processing result, wherein the feature extraction result at least includes an artificial valve intersection feature;

[0008] Inputting the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result; the model output result at least includes a segmentation result mask tensor;

[0009] Performing post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and a calcification region area.

[0010] As an optional implementation manner, in the first aspect of the present invention, performing a preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image includes:

[0011] Performing image stacking processing on the target image to obtain a stacked image corresponding to the target image; and determining a first section to be analyzed from the stacked image based on pixel distance; the first section includes a coronal plane, a sagittal plane, and a plurality of cross sections;

[0012] Selecting a second section to be analyzed from the first section, and determining section parameters adapted to the second section, the section parameters including window width and window level, and adjusting the second section according to the section parameters to obtain a target section corresponding to the second section;

[0013] The target section is subjected to tensor transformation to obtain an image pre-processing result corresponding to the target section.

[0014] As an optional implementation, in the first aspect of the present invention, the feature extraction operation is performed on the image pre-processing result according to the preset section feature extraction module to obtain a feature extraction result corresponding to the image pre-processing result, including:

[0015] According to a preset global threshold, threshold segmentation is performed on the image pre-processing result to obtain a threshold segmentation result corresponding to the image pre-processing result; the threshold segmentation result includes a binary image;

[0016] The pixel points in the binary image whose brightness is greater than the global threshold are grouped into a target point set;

[0017] According to a preset point set processing algorithm, point set processing is performed on the target point set to obtain a point set processing result corresponding to the target point set; the point set processing sequentially includes median statistics for pixel point coordinates, outlier elimination based on a box plot method, and median statistics for pixel point coordinates; the point set processing result includes a new target point set and its corresponding target center point;

[0018] Performing point set traversal on the new target point set to obtain multiple coordinate extreme value points corresponding to the new target point set and their corresponding bounding boxes, the multiple coordinate extreme value points including the maximum value point and the minimum value point of the horizontal coordinate and the maximum value point and the minimum value point of the vertical coordinate;

[0019] Embedding transformation is performed on the target center point and its corresponding center point coordinates and the bounding box to obtain a feature vector corresponding to a specified dimension as a feature extraction result corresponding to the image pre-processing result.

[0020] As an optional implementation manner, in the first aspect of the present invention, performing point set processing on the target point set according to a preset point set processing algorithm to obtain a point set processing result corresponding to the target point set includes:

[0021] Performing median statistics on all first pixel points in the target point set to obtain first median coordinates corresponding to all first pixel points and their corresponding first center points;

[0022] According to the box plot method, combined with the first median coordinate, outlier elimination is performed on all the target pixel points to obtain a new target point set;

[0023] Perform the median statistics on all second pixel points in the new target point set to obtain second median coordinates corresponding to all second pixel points and their corresponding second center points as the target center point;

[0024] The new target point set, the target center point and its corresponding center point coordinates are determined as point set processing results.

[0025] As an optional implementation, in the first aspect of the present invention, the inputting the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result includes:

[0026] Performing scaling processing on the image pre-processing result according to a bilinear interpolation method to obtain a target image tensor of a target image size;

[0027] Performing normalization and standardization processing on the elements in the target image tensor in sequence to obtain a standardized tensor corresponding to the target image tensor;

[0028] The feature extraction result is embedded in a preset deep learning model, and the standardized tensor is input into the deep learning model to obtain a model output result corresponding to the standardized tensor.

[0029] As an optional implementation manner, in the first aspect of the present invention, performing post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result includes:

[0030] Normalizing the first dimension of the segmentation result mask tensor to obtain a normalized result corresponding to the segmentation result mask tensor;

[0031] Calculate the maximum value corresponding to the first dimension in the normalized result, and determine the category subscript with the highest probability from the maximum value to obtain a first output tensor of a preset size;

[0032] Performing bilinear interpolation processing on the first output tensor to obtain a second output tensor corresponding to the first output tensor; the image size corresponding to the second output tensor is consistent with the image size of the target image;

[0033] Overlaying the second output tensor on the target image to obtain a segmentation result image;

[0034] The segmentation result map is added to the post-segmentation processing result corresponding to the model output result.

[0035] As an optional implementation, in the first aspect of the present invention, performing post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result further includes:

[0036] For the second output tensor, counting the number of pixel points in the second output tensor that are classified as calcified areas;

[0037] Calculating the product of the number of pixels and a predetermined pixel interval length to obtain a calcified area corresponding to the calcified area;

[0038] The calcification area is added to the post-segmentation processing result.

[0039] The second aspect of the present invention discloses an automatic segmentation device for artificial valve calcification area, the device comprising:

[0040] An acquisition module, used for acquiring a target image to be segmented; the target image includes a plurality of CT images of cross sections of the artificial valve;

[0041] An image pre-processing module, used to perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image; the image pre-processing result at least includes an image tensor corresponding to the target image;

[0042] A section feature extraction module, used to perform a feature extraction operation on the image pre-processing result according to a preset section feature extraction module, to obtain a feature extraction result corresponding to the image pre-processing result, wherein the feature extraction result at least includes an artificial valve intersection feature;

[0043] A model processing module, used for inputting the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result; the model output result at least includes a segmentation result mask tensor;

[0044] The post-segmentation processing module is used to perform post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and a calcification region area.

[0045] As an optional implementation, in the second aspect of the present invention, the image pre-processing module performs preset image pre-processing on the target image, and a method of obtaining an image pre-processing result corresponding to the target image specifically includes:

[0046] Performing image stacking processing on the target image to obtain a stacked image corresponding to the target image; and determining a first section to be analyzed from the stacked image based on pixel distance; the first section includes a coronal plane, a sagittal plane, and a plurality of cross sections;

[0047] Selecting a second section to be analyzed from the first section, and determining section parameters adapted to the second section, the section parameters including window width and window level, and adjusting the second section according to the section parameters to obtain a target section corresponding to the second section;

[0048] The target section is subjected to tensor transformation to obtain an image pre-processing result corresponding to the target section.

[0049] As an optional implementation, in the second aspect of the present invention, the section feature extraction module performs a feature extraction operation on the image pre-processing result according to a preset section feature extraction module, and a method of obtaining a feature extraction result corresponding to the image pre-processing result specifically includes:

[0050] According to a preset global threshold, threshold segmentation is performed on the image pre-processing result to obtain a threshold segmentation result corresponding to the image pre-processing result; the threshold segmentation result includes a binary image;

[0051] The pixel points in the binary image whose brightness is greater than the global threshold are grouped into a target point set;

[0052] According to a preset point set processing algorithm, point set processing is performed on the target point set to obtain a point set processing result corresponding to the target point set; the point set processing sequentially includes median statistics for pixel point coordinates, outlier elimination based on a box plot method, and median statistics for pixel point coordinates; the point set processing result includes a new target point set and its corresponding target center point;

[0053] Performing point set traversal on the new target point set to obtain multiple coordinate extreme value points corresponding to the new target point set and their corresponding bounding boxes, the multiple coordinate extreme value points including the maximum value point and the minimum value point of the horizontal coordinate and the maximum value point and the minimum value point of the vertical coordinate;

[0054] Embedding transformation is performed on the target center point and its corresponding center point coordinates and the bounding box to obtain a feature vector corresponding to a specified dimension as a feature extraction result corresponding to the image pre-processing result.

[0055] As an optional implementation, in the second aspect of the present invention, the section feature extraction module performs point set processing on the target point set according to a preset point set processing algorithm, and a method of obtaining a point set processing result corresponding to the target point set specifically includes:

[0056] Performing median statistics on all first pixel points in the target point set to obtain first median coordinates corresponding to all first pixel points and their corresponding first center points;

[0057] According to the box plot method, combined with the first median coordinate, outlier elimination is performed on all the target pixel points to obtain a new target point set;

[0058] Perform the median statistics on all second pixel points in the new target point set to obtain second median coordinates corresponding to all second pixel points and their corresponding second center points as the target center point;

[0059] The new target point set, the target center point and its corresponding center point coordinates are determined as point set processing results.

[0060] As an optional implementation, in the second aspect of the present invention, the model processing module inputs the image pre-processing result and the feature extraction result into a preset deep learning model, and the manner in which the model output result corresponding to the image pre-processing result and the feature extraction result is obtained specifically includes:

[0061] Performing scaling processing on the image pre-processing result according to a bilinear interpolation method to obtain a target image tensor of a target image size;

[0062] Performing normalization and standardization processing on the elements in the target image tensor in sequence to obtain a standardized tensor corresponding to the target image tensor;

[0063] The feature extraction result is embedded in a preset deep learning model, and the standardized tensor is input into the deep learning model to obtain a model output result corresponding to the standardized tensor.

[0064] As an optional implementation, in the second aspect of the present invention, the post-segmentation processing module performs post-segmentation processing on the model output result, and the manner in which the post-segmentation processing result corresponding to the model output result is obtained specifically includes:

[0065] Normalizing the first dimension of the segmentation result mask tensor to obtain a normalized result corresponding to the segmentation result mask tensor;

[0066] Calculate the maximum value corresponding to the first dimension in the normalized result, and determine the category subscript with the highest probability from the maximum value to obtain a first output tensor of a preset size;

[0067] Performing bilinear interpolation processing on the first output tensor to obtain a second output tensor corresponding to the first output tensor; the image size corresponding to the second output tensor is consistent with the image size of the target image;

[0068] Overlaying the second output tensor on the target image to obtain a segmentation result image;

[0069] The segmentation result map is added to the post-segmentation processing result corresponding to the model output result.

[0070] As an optional implementation, in the second aspect of the present invention, the post-segmentation processing module performs post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result, and further includes:

[0071] For the second output tensor, counting the number of pixel points in the second output tensor that are classified as calcified areas;

[0072] Calculating the product of the number of pixels and a predetermined pixel interval length to obtain a calcified area corresponding to the calcified area;

[0073] The calcification area is added to the post-segmentation processing result.

[0074] The third aspect of the present invention discloses another device for automatically segmenting calcified areas of artificial valves, the device comprising:

[0075] A memory storing executable program code;

[0076] a processor coupled to the memory;

[0077] The processor calls the executable program code stored in the memory to execute the method for automatic segmentation of artificial valve calcification area disclosed in the first aspect of the present invention.

[0078] The fourth aspect of the present invention discloses a computer storage medium, wherein the computer storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the method for automatic segmentation of artificial valve calcification areas disclosed in the first aspect of the present invention.

[0079] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0080] In an embodiment of the present invention, a method for automatic segmentation of calcified areas of artificial valves is provided. By implementing the present invention, image pre-processing and feature extraction operations are set before processing the target image based on the deep learning model, thereby improving the consistency, reliability and accuracy of the image data input into the deep learning model; then, the image pre-processing results and the feature extraction results are learned by the deep learning model to output the model output results including the segmentation result mask tensor, thereby achieving high-precision segmentation of the calcified area, which is beneficial to improving the segmentation efficiency and accuracy of the target image; finally, by setting the segmentation post-processing, the position and shape of the calcified area are intuitively displayed, and the area information of the calcified area is also provided, thereby improving the convenience of consulting the relevant information of the calcified area. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0082] Figure 1 It is a flow chart of a method for automatic segmentation of artificial valve calcification area disclosed in an embodiment of the present invention;

[0083] Figure 2 It is a flow chart of another method for automatic segmentation of artificial valve calcification area disclosed in an embodiment of the present invention;

[0084] Figure 3 It is a structural schematic diagram of an automatic segmentation device for artificial valve calcification area disclosed in an embodiment of the present invention;

[0085] Figure 4 It is a structural schematic diagram of another device for automatically segmenting calcified areas of artificial valves disclosed in an embodiment of the present invention;

[0086] Figure 5 It is a model structure diagram corresponding to a deep learning model adopted by a method for automatic segmentation of artificial valve calcification areas disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0088] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or ends.

[0089] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0090] The present invention discloses a method and device for automatic segmentation of artificial valve calcification area. Before processing the target image based on the deep learning model, image pre-processing and feature extraction operations are set to improve the consistency, reliability and accuracy of the image data input into the deep learning model; the image pre-processing results and feature extraction results are then learned by the deep learning model to output the model output results including the segmentation result mask tensor, thereby realizing high-precision segmentation of the calcified area, which is beneficial to improving the segmentation efficiency and accuracy of the target image; finally, by setting the segmentation post-processing, the position and shape of the calcified area are intuitively displayed, and the area information of the calcified area is also provided, thereby improving the convenience of consulting the relevant information of the calcified area. The following are detailed descriptions.

[0091] Embodiment 1

[0092] See also Figure 1 , Figure 11 is a flow chart of an automatic segmentation method for artificial valve calcification area disclosed in an embodiment of the present invention. Figure 1 The method for automatic segmentation of artificial valve calcification region described above can be applied to an automatic segmentation device for artificial valve calcification region, and the embodiment of the present invention does not limit this. Figure 1 As shown, the method for automatic segmentation of artificial valve calcification area may include the following operations:

[0093] 101. Obtain a target image to be segmented, and perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image.

[0094] In an embodiment of the present invention, the target image includes a plurality of CT images of cross sections of artificial valves; the image pre-processing result includes at least an image tensor corresponding to the target image;

[0095] 102. According to a preset section feature extraction module, a feature extraction operation is performed on the image pre-processing result to obtain a feature extraction result corresponding to the image pre-processing result.

[0096] In the embodiment of the present invention, the feature extraction result at least includes the artificial valve intersection feature.

[0097] 103. Input the image pre-processing results and the feature extraction results into a preset deep learning model to obtain a model output result corresponding to the image pre-processing results and the feature extraction results.

[0098] In the embodiment of the present invention, the model output result at least includes a segmentation result mask tensor.

[0099] 104. Perform post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and the calcification area.

[0100] In the embodiment of the present invention, by acquiring the target image to be segmented, i.e., multiple CT images of the cross section of the artificial valve, basic data is provided for subsequent analysis and processing. Subsequently, the preset image pre-processing is performed on the target image to improve the image quality and reduce interference factors, thereby obtaining an image pre-processing result corresponding to the target image. This step ensures that the image data input into the deep learning model is consistent and reliable, which helps to improve the accuracy of the segmentation results.

[0101] In an embodiment of the present invention, in the feature extraction stage, a preset section feature extraction module is used to perform feature extraction operations on the image pre-processing results to accurately identify and extract key features such as artificial valve intersections; these features are crucial for distinguishing calcified areas from surrounding tissues, and provide a strong basis for the subsequent segmentation of deep learning models.

[0102] In the embodiment of the present invention, the image pre-processing results and feature extraction results are input into a preset deep learning model. After extensive training and optimization, the model can accurately understand the complex structures and features in the image, thereby generating model output results including the segmentation result mask tensor. This step achieves high-precision segmentation of the calcified area and significantly improves segmentation efficiency and accuracy.

[0103] In an embodiment of the present invention, post-segmentation processing is finally performed on the model output result to obtain a post-segmentation processing result corresponding to the model output result; this result not only includes a segmentation result map corresponding to the segmentation result mask tensor, which intuitively displays the position and shape of the calcified area, but also provides area information of the calcified area, providing relevant personnel with more detailed image data.

[0104] It can be seen that the implementation Figure 1 The described method for automatic segmentation of artificial valve calcification areas sets up image preprocessing and feature extraction operations before processing the target image based on the deep learning model, thereby improving the consistency, reliability and accuracy of the image data input into the deep learning model; the image preprocessing results and feature extraction results are then learned through the deep learning model to output model output results including the segmentation result mask tensor, thereby achieving high-precision segmentation of the calcification area, which is beneficial to improving the segmentation efficiency and accuracy of the target image; finally, by setting up post-segmentation processing, the position and shape of the calcification area are intuitively displayed, and the area information of the calcification area is also provided, thereby improving the convenience of consulting relevant information of the calcification area.

[0105] In an optional embodiment, the above step 101 performs preset image pre-processing on the target image, and the method of obtaining the image pre-processing result corresponding to the target image specifically includes:

[0106] Performing image stacking processing on the target image to obtain a stacked image corresponding to the target image; and determining a first section to be analyzed from the stacked image based on the pixel distance; the first section includes a coronal plane, a sagittal plane, and a plurality of cross sections;

[0107] Selecting a second section to be analyzed from the first section, and determining section parameters adapted to the second section, the section parameters including window width and window level, and adjusting the second section according to the section parameters to obtain a target section corresponding to the second section;

[0108] The target section is subjected to tensor transformation to obtain an image pre-processing result corresponding to the target section.

[0109] In this optional embodiment, in the CT image pre-processing stage, the initial input is a CT image of the cross section of the artificial valve, which is output as an image tensor after processing. Specifically, firstly, multi-planar reconstruction is performed on the target image, that is, the above-mentioned image stacking processing is performed to integrate multiple cross-sectional CT images into a complete stacked image. This step ensures the integrity and continuity of all relevant image information; then, based on the pixel distance, the first section to be analyzed is determined from the stacked image. The first section includes the coronal plane, the sagittal plane and multiple cross sections. This step ensures that the selected section can fully reflect the structural characteristics of the artificial valve and its calcified area through precise spatial positioning, and provides key information for subsequent analysis and processing.

[0110] In this optional embodiment, after the above steps, the second section to be analyzed is selected from the first section, and the section parameters adapted to the second section are determined, including the window width and window level. These parameters are crucial for adjusting the image brightness and contrast, and can ensure that the key structural features in the image are clearly visible, and reduce interference factors. The second section is then adjusted according to the section parameters to obtain the target section, providing high-quality image data for subsequent tensor conversion and deep learning model input; finally, tensor conversion is performed on the target section to convert it into a format suitable for deep learning model processing. This step ensures the compatibility and consistency of the input data, and provides a solid foundation for subsequent model training and segmentation.

[0111] In this optional embodiment, it should be noted that in the CT image, the Hu value reflects the degree of absorption of X-rays at that location. Due to the differences in the absorption capacity of different human tissues to X-rays, the Hu value becomes an effective indicator for distinguishing different human tissues and organs. For example, the Hu value of bone tissue in human tissue is generally greater than 400; the Hu value of calcification foci is generally 100 to 2000; it can be clearly seen that the Hu value of calcification foci is very close to the Hu value of bone tissue, and the Hu value range of calcification foci is relatively wide, usually ranging from 100 to 2000. Therefore, in order to effectively identify all calcified areas, the width and window position of the above-mentioned section parameters can be set to a window width of 1500 and a window position of 1250; without adjustment, the default window width and window position of the bone window can be a window width of 1500 and a window position of 300.

[0112] It can be seen that in this optional embodiment, through a series of image pre-processing operations such as image stacking processing, section determination, section parameter adjustment and tensor conversion, the quality and usability of the target image are significantly improved, which provides key technical support for the subsequent automatic segmentation of artificial valve calcification areas. Through this image pre-processing operation, not only the accuracy of the segmentation results is improved, but also the analysis time is shortened and manual intervention is reduced.

[0113] Embodiment 2

[0114] See also Figure 2 , Figure 2 FIG. 1 is a flow chart of another method for automatically segmenting the calcified region of an artificial valve disclosed in an embodiment of the present invention. Figure 2 The method for automatic segmentation of artificial valve calcification region described above can be applied to an automatic segmentation device for artificial valve calcification region, and the embodiment of the present invention does not limit this. Figure 2 As shown, the method for automatic segmentation of artificial valve calcification area may include the following operations:

[0115] 201. Obtain a target image to be segmented, and perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image.

[0116] 202. Perform threshold segmentation on the image pre-processing result according to a preset global threshold value to obtain a threshold segmentation result corresponding to the image pre-processing result; the threshold segmentation result includes a binary image.

[0117] 203. Pixel points in the binary image whose brightness is greater than the global threshold are grouped into the target point set.

[0118] In the embodiment of the present invention, the purpose of step 202 and step 203 is that the intersection of the artificial valve and the calcified area usually have a higher brightness value. Through global threshold segmentation, the coordinate information of these points can be effectively obtained. However, this step cannot remove human tissues with higher Hu values ​​such as bone tissue in the CT image, so subsequent processing is required.

[0119] 204. Perform point set processing on the target point set according to a preset point set processing algorithm to obtain a point set processing result corresponding to the target point set.

[0120] In an embodiment of the present invention, point set processing includes median statistics for pixel point coordinates, outlier elimination based on box plot method, and median statistics for pixel point coordinates in sequence; the point set processing result includes a new target point set and its corresponding target center point.

[0121] 205. Perform point set traversal on the new target point set to obtain multiple coordinate maximum points and their corresponding bounding boxes corresponding to the new target point set.

[0122] In the embodiment of the present invention, the multiple coordinate maximum value points include the maximum value point and the minimum value point of the horizontal coordinate and the maximum value point and the minimum value point of the vertical coordinate.

[0123] In the embodiment of the present invention, the purpose of setting step 205 is to obtain a bounding box that includes the intersection of the artificial valve and the calcified area. Specifically, by traversing the coordinates in the new target point set, the minimum and maximum values ​​of the x coordinates and the minimum and maximum values ​​of the y coordinates are obtained. In order to prevent information loss caused by threshold segmentation and box plot method, 10 pixels can be filled in the upper, lower, left and right sides of the statistically obtained bounding box to ensure its effectiveness.

[0124] 206. Perform embedding transformation on the target center point and its corresponding center point coordinates and bounding box to obtain a feature vector corresponding to the specified dimension as a feature extraction result corresponding to the image pre-processing result.

[0125] In the embodiment of the present invention, first, the coordinate point, distance and other information are encoded into a tensor f of size (1, n_patch), where n_patch represents the number of patches in each dimension of the input Transformer encoder in the subsequent network model, which is generally 14. Then, a tensor e of size (pic_size, n_patch) is generated as an embedding module, and a learnable weight weight is used, where pic_size represents the maximum value of each dimension of the input image coordinate + 1. Then, the tensor f is mapped to the tensor space of e. Specifically, for each element f of the tensor f i , f i ∈[0,pic_size), all f i As a subscript, get the tensor e(f i ), and finally concatenate to obtain a feature vector of size (n_patch, n_patch).

[0126] 207. Input the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result.

[0127] 208. Perform post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and the calcification area.

[0128] In the embodiment of the present invention, for other descriptions of step 201 and step 207-step 208, please refer to other specific descriptions of step 101 and step 103-step 104 in the first embodiment, and the embodiment of the present invention will not be repeated here.

[0129] In the embodiment of the present invention, a preset global threshold is used to perform threshold segmentation on the image pre-processing result to obtain a binary image. By setting a reasonable threshold, the calcified area and the surrounding tissue are effectively distinguished, providing a clear image basis for subsequent feature extraction.

[0130] In the embodiment of the present invention, a preset point set processing algorithm is used to process the target point set, including median statistics, outlier elimination based on the box plot method, and median statistics again. This step further refines the target point set through statistical analysis and data cleaning, eliminates noise and outliers, and ensures the accuracy and stability of subsequent analysis. At the same time, by calculating the target center point, the present invention provides a key reference point for subsequent embedding transformation.

[0131] In the embodiment of the present invention, by performing point set traversal on the new target point set, multiple coordinate maximum points and their corresponding bounding boxes are obtained, thereby realizing the boundary determination of the calcified area and providing necessary spatial information for subsequent embedding transformation and feature vector extraction.

[0132] In an embodiment of the present invention, by performing embedding transformation on the target center point and its corresponding center point coordinates and bounding box, a feature vector corresponding to the specified dimension is obtained, which can convert complex image features into concise and efficient numerical representations, thereby providing easy-to-process and understand input data for subsequent deep learning models.

[0133] It can be seen that the implementation Figure 2 The described method for automatic segmentation of artificial valve calcification areas can accurately extract key features related to artificial valve calcification areas from image preprocessing results through refined feature extraction steps. These features not only accurately reflect the structure and position information of the calcification areas, but also provide high-quality input data for subsequent deep learning models, further improving the accuracy and efficiency of automatic segmentation of artificial valve calcification areas.

[0134] In an optional embodiment, the above-mentioned method of performing point set processing on the target point set according to the preset point set processing algorithm to obtain the point set processing result corresponding to the target point set specifically includes:

[0135] Perform median statistics on all first pixel points in the target point set to obtain first median coordinates corresponding to all first pixel points and their corresponding first center points;

[0136] According to the box plot method, combined with the first median coordinate, outlier elimination is performed on all target pixels to obtain a new target point set;

[0137] Perform median statistics on all second pixel points in the new target point set to obtain second median coordinates corresponding to all second pixel points and their corresponding second center points as the target center point;

[0138] The new target point set, the target center point and its corresponding center point coordinates are determined as the point set processing result.

[0139] In this optional embodiment, by performing median statistics on all first pixel points in the target point set, the first median coordinates and the first center point corresponding thereto are obtained. The coordinates of the first center point are marked as (C x , C y ), the significance of this step is to obtain a rough artificial valve center point through a fast method, which can calculate the central trend of the target point set and provide a benchmark for subsequent outlier judgment. Compared with the mean statistics, the median statistics are more resistant to the influence of outliers, thus ensuring the stability and accuracy of the center point.

[0140] In this optional embodiment, outlier elimination is performed on all target pixels according to the box plot method in combination with the first median coordinate. The box plot method is an effective data visualization tool that can intuitively display the distribution of data and outliers. Specifically, the box plot method first calculates Q1, Q2, and Q3 based on the values ​​in the target sequence. Respectively represent the lower quartile, the middle quartile, and the upper quartile. The calculation formulas of Q1, Q2, and Q3 are shown in Formula 1.1.

[0141]

[0142] Where n means there are n numbers in the sequence. The calculation results here all represent the subscript of a number in the sequence. The box plot method stipulates that the distance between Q3 and Q1 is the interquartile range (IQR), and the value that is more than 1.5 times the IQR from the quartile is an outlier.

[0143] In order to express the degree of outlier of the point in the point set, this method calculates the distance from each point to the rough center point (C x , C y )’s Manhattan distance, the calculation formula is shown in Formula 1.2.

[0144] MD(P, C)=|P x -C x |+|P y -C y |(1.2)

[0145] Among them, MD(P, C) represents the point (P x , P y ) and point (C x , C y ) is the Manhattan distance of .

[0146] In order to remove outliers in the point set, this paper uses the box plot method to set the value greater than 1.5 times the IQR of the upper quartile as the outlier limit. Exceeding this limit means that the distance between the point and the rough center point is greater than that of the vast majority of points.

[0147] In this optional embodiment, the method can successfully identify and eliminate outliers in the target point set, which may be caused by noise, errors or abnormal phenomena, and their existence will interfere with subsequent analysis and processing. After the outliers are eliminated, a new target point set is obtained, which is purer and more accurate, providing a better data basis for subsequent median statistics and center point calculations.

[0148] In this optional embodiment, median statistics are performed again on all second pixel points in the new target point set to obtain the second median coordinates and the corresponding second center point as the final target center point. This step further refines the target point set and ensures the accuracy and representativeness of the center point. At the same time, by comparing the two median statistics, the effect of outlier elimination and the stability and consistency of the target point set can be evaluated.

[0149] It can be seen that in this optional embodiment, an efficient point set processing algorithm is adopted, which can perform detailed median statistics and outlier elimination on the target point set, further improving the accuracy and stability of feature information extraction.

[0150] In another optional embodiment, the above step 207 inputs the image pre-processing result and the feature extraction result into a preset deep learning model, and the method of obtaining the model output result corresponding to the image pre-processing result and the feature extraction result specifically includes:

[0151] Performing scaling processing on the image pre-processing result according to the bilinear interpolation method to obtain a target image tensor of the target image size;

[0152] Perform normalization and standardization processing on the elements in the target image tensor in sequence to obtain a standardized tensor corresponding to the target image tensor;

[0153] The feature extraction results are embedded into the preset deep learning model, and the standardized tensor is input into the deep learning model to obtain the model output result corresponding to the standardized tensor.

[0154] In this optional embodiment, in the deep learning model processing stage, the input is an image tensor and the output is a segmentation mask tensor. Assume that the size of the input image tensor is C×H×W, where C represents the number of image channels, generally 3, and H and W represent the image height and width, respectively. First, it is scaled to the specified size C×H'×W' by bilinear interpolation, where H' and W' represent the height and width of the scaled image, respectively. In order to ensure that the input image can retain feature information and to control the complexity of the network, H' and W' in this article are both 224.

[0155] Then, the elements in the tensor are normalized and standardized. Specifically, all the elements in the image tensor are converted from integers distributed in the range of [0,255] to floating-point types distributed in the range of [0,1]. Standardization needs to be calculated based on the mean and variance of the elements in the channel. The specific calculation formula is shown in Equation 1.3.

[0156]

[0157] Among them, px′ c,i,j Represents the element value of the normalized tensor at position (c, i, j), px c,i,j Represents the element value of the tensor at position (c, i, j) before normalization, μ c represents the mean of all elements of the channel, σ c Represents the standard deviation of all elements of channel c.

[0158] The normalized and standardized tensor is fed into the deep learning model. The specific processing of the tensor in the model depends on the network model structure. The output segmentation mask tensor has the size of n×H'×W'. Among them, n represents the number of segmentation categories. In the above-mentioned calcification area segmentation task, n is 2, that is, the calcification area and the background are two categories.

[0159] It can be seen that in this optional embodiment, by adopting a technical solution combining bilinear interpolation, normalization processing, standardization processing and deep learning model, automatic segmentation and accurate identification of artificial valve calcification areas are achieved, further improving the prediction performance, stability and accuracy of the deep learning model.

[0160] In this alternative embodiment, see Figure 5 , Figure 5 is a model structure diagram corresponding to a deep learning model used in a method for automatic segmentation of artificial valve calcification area disclosed in an embodiment of the present invention, such as Figure 5As shown in the figure, compared with the traditional TransUNet model, this solution makes two improvements to the model: 1) adding an artificial valve section feature extractor, fusing the section features with the features extracted by the convolutional neural network, and sending them together to the Transformer module for learning using the self-attention mechanism; 2) changing the third layer of the convolutional neural network module to a maximum pooling layer, and reducing the number of SkipConnections in the network model from 3 to 2. The following will elaborate on these two improvements.

[0161] In this optional embodiment, the first improvement to the traditional TransUNet model is the introduction of the artificial valve section feature extractor mentioned above. Specifically, the number of output channels of the last fully connected layer of the convolutional neural network module is modified from D layers to D-1 layers, that is, the output tensor is (D-1, n_patch, n_patch), where D is 768, representing the dimension of the input tensor of the Transformer. Subsequently, the artificial valve section features of size (1, n_patch, n_patch) are feature concatenated with the output tensor of the fully connected layer in the first dimension, and finally a tensor of size (D, n_patch, n_patch) is obtained as the output of the convolutional neural network module, which is reshaped as the input of the Transformer module. The motivation for proposing this change in this scheme is to better learn how to segment calcified areas by converting prior knowledge in related technical fields into feature information and utilizing the characteristics of the self-attention mechanism of the Transformer.

[0162] In this optional embodiment, the second improvement to the traditional TransUNet model is to streamline the network model. The feature extraction method in the traditional TransUNet model is maintained, that is, ResNet-50 proposed by He et al. is used for feature extraction. However, through experiments, it is found that the traditional TransUNet model performs well on the training set, but performs poorly on the validation set and the test set, and the segmentation accuracy is quite different from the training set. This paper speculates that the reason is that the model is seriously overfitting. To solve this problem, this study modified the model to achieve its simplification. Specifically, this paper modifies the last layer of ResNet residual unit to a maximum pooling layer. The reason behind this change is that as the depth of the convolutional neural network increases, the feature size gradually decreases, resulting in a larger receptive field for each pixel in the deeper feature tensor. Considering that in the artificial valve calcification area segmentation task proposed in this paper, the calcification area is smaller than the entire image, and the maximum pooling method can better retain the edge features of the image than the average pooling, which is conducive to maintaining the accuracy of segmentation. Therefore, by replacing the deepest layer of the ResNet residual unit with the maximum pooling layer, this paper reduces the complexity of the model and reduces the model parameters, thereby effectively alleviating the overfitting phenomenon.

[0163] In this optional embodiment, the traditional TransUNet model uses a mixed loss function of the cross entropy loss function (Cross Entropy Loss) as shown in Formula 1.4 and the dice loss function (Dice Loss) as shown in Formula 1.5. The mixed loss function is shown in Formula 1.6.

[0164]

[0165] Loss total =αLoss ce +(1-α)Loss dice (1.6)

[0166] Among them, C represents all segmentation categories, G represents all pixels in the image, Indicates the true category to which the pixel p belongs, represents the probability distribution of the model's prediction for pixel p, ε represents a very small real number used to avoid division by 0, and α represents the weight. Chen et al. set α to 0.5 in the traditional TransUNet.

[0167] As can be seen from formula 1.4, the cross entropy loss function is used to comprehensively consider the categories of all pixels and assume that all pixels contribute equally to the total loss function. However, in the calcified area segmentation task proposed in this paper, due to the imbalance in the pixel distribution of the background class and the calcified area class, the number of background class pixels far exceeds the number of calcified area pixels, which causes the network model to be more inclined to capture the characteristics of the background during the learning process, and may classify all pixels as background pixels to reduce the overall loss function. Some past studies have shown that the use of weighted cross entropy loss can alleviate this phenomenon well. For example, Long et al. and Ronneberger et al. used the idea of ​​weighting when training FCN and U-Net to alleviate the problem of unbalanced pixel segmentation categories. The formula for weighted cross entropy is shown in formula 1.7.

[0168]

[0169] Among them, w i Represents the weight given to the i-th category pixel. In particular, since the artificial valve calcification area segmentation task in this paper is a binary classification task, Formula 1.7 can be simplified to Formula 1.8.

[0170]

[0171] Where β represents the weight given to the pixels in the calcified area category, β∈[0,1]. The higher the β value, the fewer false negative results; the lower the β value, the fewer false positive results. In this paper, β is uniformly set to 0.01.

[0172] The focal loss proposed by Lin et al. focuses on samples that are more difficult to learn. It improves on the cross entropy loss function and introduces a regulation factor γ. By penalizing easy-to-learn samples, the model is more focused on difficult-to-learn samples. The formula for focal loss in the binary classification scenario is shown in Equation 1.9.

[0173]

[0174] In the calcification area segmentation task proposed in this paper, the background area accounts for a large proportion, and the number of negative samples is much higher than the positive samples. It is difficult to effectively train using the conventional cross entropy function. In order to improve the generalization ability of the model, this paper uses the loss function Focal Dice Loss, which is a mixture of focal loss and dice loss. The focal loss is used to evaluate the classification loss of pixel points, and the dice loss is used to evaluate the overlap between the overall segmentation result and the true segmentation result, taking into account both the local and the overall. The Focal DiceLoss loss function formula is shown in Equation 1.10.

[0175] Loss FocalDice =αLoss focal +(1-α)Loss dice (1.10)

[0176] Among them, represents the weight of the two loss functions. In order to ensure that the two loss functions work at the same time, α in this paper is taken as 0.5.

[0177] In another optional embodiment, the above step 208 performs post-segmentation processing on the model output result, and the method of obtaining the post-segmentation processing result corresponding to the model output result specifically includes:

[0178] Normalize the first dimension in the segmentation result mask tensor to obtain a normalized result corresponding to the segmentation result mask tensor;

[0179] Calculate the maximum value corresponding to the first dimension in the normalized result, and determine the category subscript with the highest probability from the maximum value to obtain the first output tensor of a preset size;

[0180] Performing bilinear interpolation processing on the first output tensor to obtain a second output tensor corresponding to the first output tensor; the image size corresponding to the second output tensor is consistent with the image size of the target image;

[0181] Overlay the second output tensor on the target image to obtain the segmentation result image;

[0182] Add the segmentation result map to the segmentation post-processing result corresponding to the model output result.

[0183] In this optional embodiment, the above step 208 performs post-segmentation processing on the model output result, and the method of obtaining the post-segmentation processing result corresponding to the model output result specifically includes:

[0184] For the second output tensor, counting the number of pixel points in the second output tensor that are classified as calcified areas;

[0185] Calculate the product of the number of pixels and the predetermined pixel interval length to obtain the calcification area corresponding to the calcification area;

[0186] Add the calcification area to the segmentation post-processing results.

[0187] In this optional embodiment, a tensor of size n×H'×W' is obtained through a deep learning model, and the elements in the tensor are the outputs of the fully connected layer, that is, the unnormalized classification probability of each pixel. Through the above steps, the first step is to normalize the first dimension of the tensor. The significance of normalizing the first dimension is that in an image of size H'×W', the probability of the classification to which each element belongs is normalized so that the sum of the probabilities is 1; the second step is to calculate the maximum value in the first dimension of the tensor, and take the category subscript with the highest probability as the output. This step will obtain a tensor of size 1×H'×W'; the third step is to use bilinear interpolation to scale the image from 1×H×W, that is, to restore it to the input size; finally, the segmentation result tensor is overlaid on the original image and displayed to relevant personnel as a segmentation result diagram. At the same time, by counting the number of pixels classified as calcified areas and multiplying them by the pixel interval length in the Dicom tag, the calcified area of ​​the plane is obtained.

[0188] It can be seen that in this optional embodiment, by introducing a sophisticated post-segmentation processing step, not only the accuracy and readability of the segmentation results are improved, but also quantitative information of the calcification area is provided, the precision and reliability of the target image information are improved, and the efficiency and accuracy of medical image analysis are further improved.

[0189] Embodiment 3

[0190] See also Figure 3 , Figure 3: is a schematic diagram of the structure of an automatic segmentation device for artificial valve calcification area disclosed in an embodiment of the present invention. The automatic segmentation device for artificial valve calcification area can be an automatic segmentation terminal, device, system or server for artificial valve calcification area. The server can be a local server, a remote server, or a cloud server (also called a cloud server). When the server is a non-cloud server, the non-cloud server can communicate with the cloud server, which is not limited in the embodiment of the present invention. Figure 3 As shown, the artificial valve calcification area automatic segmentation device may include an acquisition module 301, an image pre-processing module 302, a section feature extraction module 303, a model processing module 304 and a segmentation post-processing module 305, wherein:

[0191] The acquisition module 301 is used to acquire the target image to be segmented; the target image includes a plurality of CT images of cross sections of the artificial valve.

[0192] The image pre-processing module 302 is used to perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image; the image pre-processing result at least includes an image tensor corresponding to the target image.

[0193] The section feature extraction module 303 is used to perform feature extraction operations on the image preprocessing results according to the preset section feature extraction module 303 to obtain feature extraction results corresponding to the image preprocessing results, and the feature extraction results at least include artificial valve intersection features.

[0194] The model processing module 304 is used to input the image pre-processing results and the feature extraction results into a preset deep learning model to obtain a model output result corresponding to the image pre-processing results and the feature extraction results; the model output result at least includes a segmentation result mask tensor.

[0195] The post-segmentation processing module 305 is used to perform post-segmentation processing on the model output result to obtain the post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and the calcification area.

[0196] It can be seen that implementation Figure 3The described automatic segmentation device for artificial valve calcification areas sets up image pre-processing and feature extraction operations before processing the target image based on the deep learning model, thereby improving the consistency, reliability and accuracy of the image data input into the deep learning model; the image pre-processing results and feature extraction results are then learned through the deep learning model to output model output results including the segmentation result mask tensor, thereby achieving high-precision segmentation of the calcification area, which is beneficial to improving the segmentation efficiency and accuracy of the target image; finally, by setting up post-segmentation processing, the position and shape of the calcification area are intuitively displayed, and the area information of the calcification area is also provided, thereby improving the convenience of consulting relevant information of the calcification area.

[0197] In an optional embodiment, the image pre-processing module 302 performs preset image pre-processing on the target image, and the manner in which the image pre-processing result corresponding to the target image is obtained specifically includes:

[0198] Performing image stacking processing on the target image to obtain a stacked image corresponding to the target image; and determining a first section to be analyzed from the stacked image based on the pixel distance; the first section includes a coronal plane, a sagittal plane, and a plurality of cross sections;

[0199] Selecting a second section to be analyzed from the first section, and determining section parameters adapted to the second section, the section parameters including window width and window level, and adjusting the second section according to the section parameters to obtain a target section corresponding to the second section;

[0200] The target section is subjected to tensor transformation to obtain an image pre-processing result corresponding to the target section.

[0201] It can be seen that in this optional embodiment, through a series of image pre-processing operations such as image stacking processing, section determination, section parameter adjustment and tensor conversion, the quality and usability of the target image are significantly improved, which provides key technical support for the subsequent automatic segmentation of artificial valve calcification areas. Through this image pre-processing operation, not only the accuracy of the segmentation results is improved, but also the analysis time is shortened and manual intervention is reduced.

[0202] In another optional embodiment, the section feature extraction module 303 performs a feature extraction operation on the image pre-processing result according to the preset section feature extraction module 303, and the manner of obtaining the feature extraction result corresponding to the image pre-processing result specifically includes:

[0203] According to a preset global threshold, threshold segmentation is performed on the image pre-processing result to obtain a threshold segmentation result corresponding to the image pre-processing result; the threshold segmentation result includes a binary image;

[0204] The pixels in the binary image whose brightness is greater than the global threshold are grouped into the target point set;

[0205] According to the preset point set processing algorithm, point set processing is performed on the target point set to obtain a point set processing result corresponding to the target point set; the point set processing includes median statistics for pixel point coordinates, outlier elimination based on the box plot method, and median statistics for pixel point coordinates; the point set processing result includes a new target point set and its corresponding target center point;

[0206] Performing point set traversal on the new target point set to obtain multiple coordinate maximum value points and their corresponding bounding boxes corresponding to the new target point set, wherein the multiple coordinate maximum value points include the maximum value point and the minimum value point of the horizontal coordinate and the maximum value point and the minimum value point of the vertical coordinate;

[0207] Embedding transformation is performed on the target center point and its corresponding center point coordinates and bounding box to obtain a feature vector corresponding to the specified dimension as the feature extraction result corresponding to the image pre-processing result.

[0208] It can be seen that in this optional embodiment, through the refined feature extraction step, the key features related to the artificial valve calcification area can be accurately extracted from the image pre-processing results. These features not only accurately reflect the structure and position information of the calcified area, but also provide high-quality input data for the subsequent deep learning model, further improving the accuracy and efficiency of the automatic segmentation of the artificial valve calcification area.

[0209] In another optional embodiment, the section feature extraction module 303 performs point set processing on the target point set according to a preset point set processing algorithm, and obtains a point set processing result corresponding to the target point set in a manner specifically including:

[0210] Perform median statistics on all first pixel points in the target point set to obtain first median coordinates corresponding to all first pixel points and their corresponding first center points;

[0211] According to the box plot method, combined with the first median coordinate, outlier elimination is performed on all target pixels to obtain a new target point set;

[0212] Perform median statistics on all second pixel points in the new target point set to obtain second median coordinates corresponding to all second pixel points and their corresponding second center points as the target center point;

[0213] The new target point set, the target center point and its corresponding center point coordinates are determined as the point set processing result.

[0214] It can be seen that in this optional embodiment, an efficient point set processing algorithm is adopted, which can perform detailed median statistics and outlier elimination on the target point set, further improving the accuracy and stability of feature information extraction.

[0215] In another optional embodiment, the model processing module 304 inputs the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result, specifically including:

[0216] Performing scaling processing on the image pre-processing result according to the bilinear interpolation method to obtain a target image tensor of the target image size;

[0217] Perform normalization and standardization processing on the elements in the target image tensor in sequence to obtain a standardized tensor corresponding to the target image tensor;

[0218] The feature extraction results are embedded into the preset deep learning model, and the standardized tensor is input into the deep learning model to obtain the model output result corresponding to the standardized tensor.

[0219] It can be seen that in this optional embodiment, by adopting a technical solution combining bilinear interpolation, normalization processing, standardization processing and deep learning model, automatic segmentation and accurate identification of artificial valve calcification areas are achieved, further improving the prediction performance, stability and accuracy of the deep learning model.

[0220] In yet another optional embodiment, the post-segmentation processing module 305 performs post-segmentation processing on the model output result, and the manner in which the post-segmentation processing result corresponding to the model output result is obtained specifically includes:

[0221] Normalize the first dimension in the segmentation result mask tensor to obtain a normalized result corresponding to the segmentation result mask tensor;

[0222] Calculate the maximum value corresponding to the first dimension in the normalized result, and determine the category subscript with the highest probability from the maximum value to obtain the first output tensor of a preset size;

[0223] Performing bilinear interpolation processing on the first output tensor to obtain a second output tensor corresponding to the first output tensor; the image size corresponding to the second output tensor is consistent with the image size of the target image;

[0224] Overlay the second output tensor on the target image to obtain the segmentation result image;

[0225] Add the segmentation result map to the segmentation post-processing result corresponding to the model output result.

[0226] In this optional embodiment, the post-segmentation processing module 305 performs post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result, and further includes:

[0227] For the second output tensor, counting the number of pixel points in the second output tensor that are classified as calcified areas;

[0228] Calculate the product of the number of pixels and the predetermined pixel interval length to obtain the calcification area corresponding to the calcification area;

[0229] Add the calcification area to the segmentation post-processing results.

[0230] It can be seen that in this optional embodiment, by introducing a sophisticated post-segmentation processing step, not only the accuracy and readability of the segmentation results are improved, but also quantitative information of the calcification area is provided, the precision and reliability of the target image information are improved, and the efficiency and accuracy of medical image analysis are further improved.

[0231] Embodiment 4

[0232] See also Figure 4 , Figure 4 FIG. 1 is a schematic diagram of the structure of another device for automatically segmenting the calcified region of an artificial valve disclosed in an embodiment of the present invention. Figure 4 As shown, the device for automatically segmenting the artificial valve calcification region may include:

[0233] A memory 401 storing executable program codes;

[0234] a processor 402 coupled to the memory 401;

[0235] The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the method for automatic segmentation of artificial valve calcification region described in the first embodiment of the present invention or the second embodiment of the present invention.

[0236] Embodiment 5

[0237] An embodiment of the present invention discloses a computer storage medium storing computer instructions. When the computer instructions are called, they are used to execute the steps in the method for automatic segmentation of artificial valve calcification areas described in Embodiment 1 or Embodiment 2 of the present invention.

[0238] Embodiment 6

[0239] An embodiment of the present invention discloses a computer program product, which includes a non-transitory computer storage medium storing a computer program, and the computer program is operable to enable a computer to execute the steps in the method for automatic segmentation of artificial valve calcification areas described in Example 1 or Example 2.

[0240] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0241] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or partly contributed to the prior art in the form of a software product, and the computer software product can be stored in a computer storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0242] Finally, it should be noted that the method and device for automatic segmentation of artificial valve calcification areas disclosed in the embodiments of the present invention only disclose the preferred embodiments of the present invention, which are only used to illustrate the technical scheme of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical schemes described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical schemes from the spirit and scope of the technical schemes of the embodiments of the present invention.

Claims

1. A method for automatic segmentation of artificial valve calcification area, characterized in that: The method comprises: Acquire a target image to be segmented, and perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image; the target image includes a plurality of CT images of cross sections of artificial valves; the image pre-processing result includes at least an image tensor corresponding to the target image; According to a preset section feature extraction module, a feature extraction operation is performed on the image pre-processing result to obtain a feature extraction result corresponding to the image pre-processing result, wherein the feature extraction result at least includes an artificial valve intersection feature; Inputting the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result; the model output result at least includes a segmentation result mask tensor; Performing post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and a calcification region area; The step of inputting the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result includes: Performing scaling processing on the image pre-processing result according to a bilinear interpolation method to obtain a target image tensor of a target image size; Performing normalization and standardization processing on the elements in the target image tensor in sequence to obtain a standardized tensor corresponding to the target image tensor; Embedding the feature extraction result into a preset deep learning model, and inputting the standardized tensor into the deep learning model to obtain a model output result corresponding to the standardized tensor; The performing post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result includes: Normalizing the first dimension of the segmentation result mask tensor to obtain a normalized result corresponding to the segmentation result mask tensor; Calculate the maximum value corresponding to the first dimension in the normalized result, and determine the category subscript with the highest probability from the maximum value to obtain a first output tensor of a preset size; Performing bilinear interpolation processing on the first output tensor to obtain a second output tensor corresponding to the first output tensor; the image size corresponding to the second output tensor is consistent with the image size of the target image; Overlaying the second output tensor on the target image to obtain a segmentation result image; The segmentation result map is added to the post-segmentation processing result corresponding to the model output result.

2. The method for automatic segmentation of artificial valve calcification area according to claim 1, characterized in that: The performing of a preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image includes: Performing image stacking processing on the target image to obtain a stacked image corresponding to the target image; and determining a first section to be analyzed from the stacked image based on pixel distance; the first section includes a coronal plane, a sagittal plane, and a plurality of cross sections; Selecting a second section to be analyzed from the first section, and determining section parameters adapted to the second section, the section parameters including window width and window level, and adjusting the second section according to the section parameters to obtain a target section corresponding to the second section; The target section is subjected to tensor transformation to obtain an image pre-processing result corresponding to the target section.

3. The method for automatic segmentation of artificial valve calcification area according to claim 1 or 2, characterized in that: The step of performing a feature extraction operation on the image pre-processing result according to a preset section feature extraction module to obtain a feature extraction result corresponding to the image pre-processing result includes: According to a preset global threshold, threshold segmentation is performed on the image pre-processing result to obtain a threshold segmentation result corresponding to the image pre-processing result; the threshold segmentation result includes a binary image; The pixel points in the binary image whose brightness is greater than the global threshold are grouped into a target point set; According to a preset point set processing algorithm, point set processing is performed on the target point set to obtain a point set processing result corresponding to the target point set; the point set processing sequentially includes median statistics for pixel point coordinates, outlier elimination based on a box plot method, and median statistics for pixel point coordinates; the point set processing result includes a new target point set and its corresponding target center point; Performing point set traversal on the new target point set to obtain multiple coordinate extreme value points corresponding to the new target point set and their corresponding bounding boxes, the multiple coordinate extreme value points including the maximum value point and the minimum value point of the horizontal coordinate and the maximum value point and the minimum value point of the vertical coordinate; Embedding transformation is performed on the target center point and its corresponding center point coordinates and the bounding box to obtain a feature vector corresponding to a specified dimension as a feature extraction result corresponding to the image pre-processing result.

4. The method for automatic segmentation of artificial valve calcification area according to claim 3, characterized in that: The performing point set processing on the target point set according to a preset point set processing algorithm to obtain a point set processing result corresponding to the target point set includes: Performing median statistics on all first pixel points in the target point set to obtain first median coordinates corresponding to all first pixel points and their corresponding first center points; According to the box plot method, combined with the first median coordinate, outlier elimination is performed on all the target pixel points to obtain a new target point set; Perform the median statistics on all second pixel points in the new target point set to obtain second median coordinates corresponding to all second pixel points and their corresponding second center points as the target center point; The new target point set, the target center point and its corresponding center point coordinates are determined as point set processing results.

5. The method for automatic segmentation of artificial valve calcification area according to claim 1, characterized in that: The performing post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result also includes: For the second output tensor, counting the number of pixel points in the second output tensor that are classified as calcified areas; Calculating the product of the number of pixels and a predetermined pixel interval length to obtain a calcified area corresponding to the calcified area; The calcification area is added to the post-segmentation processing result.

6. An automatic segmentation device for artificial valve calcification area, characterized in that: The device is used to perform the method for automatic segmentation of artificial valve calcification area according to any one of claims 1 to 5, and the device comprises: An acquisition module, used for acquiring a target image to be segmented; the target image includes a plurality of CT images of cross sections of the artificial valve; An image pre-processing module, used to perform preset image pre-processing on the target image to obtain an image pre-processing result corresponding to the target image; the image pre-processing result at least includes an image tensor corresponding to the target image; A section feature extraction module, used to perform a feature extraction operation on the image pre-processing result according to a preset section feature extraction module, to obtain a feature extraction result corresponding to the image pre-processing result, wherein the feature extraction result at least includes an artificial valve intersection feature; A model processing module, used for inputting the image pre-processing result and the feature extraction result into a preset deep learning model to obtain a model output result corresponding to the image pre-processing result and the feature extraction result; the model output result at least includes a segmentation result mask tensor; The post-segmentation processing module is used to perform post-segmentation processing on the model output result to obtain a post-segmentation processing result corresponding to the model output result; the post-segmentation processing result includes a segmentation result map corresponding to the segmentation result mask tensor and a calcification region area.

7. An automatic segmentation device for artificial valve calcification area, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the method for automatic segmentation of artificial valve calcification area according to any one of claims 1-5.

8. A computer storage medium, characterized in that The computer storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the method for automatic segmentation of artificial valve calcification areas as described in any one of claims 1-5.

Images