Complex aortic lesion intelligent analysis method and system and storage medium

Abstract

The invention relates to the technical field of medical image analysis and processing, and discloses a complex aortic lesion intelligent analysis method and system and a storage medium. The method comprises the following steps: collecting a plurality of aorta images of multiple image types containing different lesion types, collecting lesion information as label information, and labeling the lesion aorta images based on the label information; extracting a lesion area and a key area, generating first training data and training a first recognition model; acquiring to-be-analyzed aortic images of multiple image types of a patient, performing enhancement processing, using a trained first identification model to identify each key region in the to-be-analyzed aortic images, acquiring a lesion score of each lesion type, calculating a first lesion score of each lesion type of the to-be-analyzed aortic images, and performing image processing on the to-be-analyzed aortic images according to the first lesion score; and determining a corresponding lesion type based on the first lesion score. According to the method, various image types can be effectively analyzed, and the recognition accuracy of aortic lesion types is improved.

Description

Technical Field

[0001] The present application relates to the technical field of medical image analysis and processing, and in particular to a method, system and storage medium for intelligent analysis of complex aortic lesions. Background Art

[0002] Complex aortic lesions, such as Stanford type B aortic dissection and complex abdominal aortic aneurysm, are serious vascular diseases that require precise vascular reconstruction techniques for their treatment. Currently, in vitro pre-fenestration combined with beam diameter technology has become an effective means of treating such lesions. However, this technology has extremely high requirements for the accuracy of the fenestration position and the accuracy of stent release. In clinical applications, the analysis of lesions and the formulation of surgical plans mainly rely on the doctor's experience and subjective judgment, and lack of systematic and intelligent analysis methods, which to some extent affects the success rate of the operation and the patient's prognosis.

[0003] A similar prior art Chinese patent application with publication number CN110675375A provides a method for automatically distinguishing aortic dissection in thoracic and abdominal aorta images, including: projecting and counting the CTA image of the medical aorta region contour through Radon transform, and combining the area and circumference characteristics of the aorta contour, to quickly and accurately determine whether the aorta region contains a dissection. However, this method only relies on CTA images and does not involve the fusion of other modalities (such as MRI, echocardiography, etc.). A single modality image may not be able to fully reflect the complexity of the lesion, and the discrimination ability may be limited for some images with unclear dissection features.

[0004] Similar prior art also has a Chinese patent application with publication number CN118735797A, which provides an aortic lesion assessment method based on multi-scale feature fusion, including step S1, obtaining chest and abdominal plain scan CT images to be evaluated, extracting and reconstructing the aortic three-dimensional structure image; step S2, obtaining multi-scale image features based on the aortic three-dimensional structure image, including multi-scale fusion image features and aortic three-dimensional image features; step S3, obtaining an aortic lesion assessment model, and inputting the multi-scale image features into the aortic lesion assessment model; step S4, the aortic lesion assessment model evaluates the aortic lesion assessment result according to the multi-scale image features. However, this method is mainly based on chest and abdominal plain scan CT images. For some lesion types (such as early dissection or microlesions), the resolution and contrast of plain scan CT images may not be sufficient to provide sufficient diagnostic information.

[0005] Therefore, the present invention provides a complex aortic lesion intelligent analysis method, system and storage medium. Summary of the invention

[0006] The present application provides a complex aortic lesion intelligent analysis method, system and storage medium for improving the accuracy of lesion recognition.

[0007] In a first aspect, the present application provides a method for intelligent analysis of complex aortic lesions, the method comprising:

[0008] Step S1, collecting a number of aortic images of multiple image types including different lesion types, wherein each image type of aortic image includes a normal aortic image and a lesion aortic image, collecting corresponding lesion information as label information, wherein the label information includes a lesion type and a lesion range, and annotating the lesion aortic image based on the label information;

[0009] Step S2, extracting a lesion area from the aorta image using a first extraction model, extracting a key area of ​​a preset first size from the lesion area, combining the key area and corresponding lesion information based on the annotation information to generate first training data, training a first recognition model based on the first training data, wherein the first recognition model can output a corresponding lesion type and a lesion score according to the key area of ​​the aorta image, wherein the lesion score represents a lesion probability of the corresponding lesion type;

[0010] Step S3, obtaining aortic images to be analyzed of multiple image types of the patient, enhancing all the aortic images to be analyzed, inputting the aortic images to be analyzed of each image type after the enhancement into the first extraction model to obtain the corresponding lesion area, then extracting a key area of ​​a predetermined size from each lesion area, using the trained first recognition model to identify each key area in the aortic image to be analyzed, obtaining a lesion score for each lesion type, calculating a first lesion score for each lesion type of the aortic image to be analyzed, and determining the corresponding lesion type based on the first lesion score.

[0011] In combination with the first aspect, in a first implementation of the first aspect of the present application, extracting a lesion area from an aorta image using a first extraction model includes:

[0012] The brightness value of each pixel in the aortic image is obtained, the frequency distribution of each brightness value is counted, the maximum frequency and the minimum frequency are obtained, and a first value is preset. The frequency of each brightness value is subtracted from the minimum frequency to obtain a first result value, the maximum frequency is subtracted from the minimum frequency to obtain a second result value, the first result value is divided by the second result value to obtain a third result value, the third result value is multiplied by the first value to obtain a fourth result value, the fourth result value is used as the normalized frequency of the corresponding brightness value, and a histogram of the brightness value of the aortic image is generated based on the normalized frequency. A plurality of different mathematical models are used to fit the normalized frequency distribution, and the parameters of each of the mathematical models are adjusted. The plurality of mathematical models are combined to fit the distribution of the normalized frequency. Based on the mathematical model after adjusting the parameters, the pixels in the aortic image are divided into different groups, and the pixels divided into the same group are regarded as belonging to the same area. The multiple different areas are analyzed to identify the diseased area.

[0013] In combination with the first aspect, in a second implementation of the first aspect of the present application, analyzing multiple different regions to identify lesion regions includes:

[0014] Extract the brightness value of the pixel from each area, calculate the average brightness value and the first standard deviation of each area, and also calculate the second value of each area, use the brightness average, the first standard deviation and the second value as eigenvalues, normalize the eigenvalues, and draw the characteristic polygon of each area based on the normalized eigenvalues. First, set an origin, draw multiple line segments along a first number of directions with the origin, each line segment represents an eigenvalue, mark the normalized eigenvalue on the corresponding line segment for each area, connect all the eigenvalues ​​corresponding to each area in sequence to form a polygon, calculate the area of ​​each polygon, obtain the first polygon whose area is greater than the preset first threshold, calculate the second standard deviation of the eigenvalues ​​corresponding to each of the first polygons, obtain the first polygon whose second standard deviation is greater than the second threshold, and call it the second polygon, and use the area corresponding to the second polygon as the lesion area.

[0015] In combination with the first aspect, in a third implementation of the first aspect of the present application, calculating the second value of each region includes:

[0016] A small area containing a preset number of pixels is obtained from each area, and the central pixel of each of the small areas is obtained. For other pixels in the small areas, the grayscale value is compared with that of the central pixel. If the grayscale value of other pixels is greater than or equal to the grayscale value of the central pixel, it is marked as 1, otherwise it is marked as 0. The marked values ​​corresponding to all other pixels contained in each of the small areas are combined into a binary number, and the binary numbers of all the small areas are counted to obtain the total number of all non-repeating binary numbers, and the total number is used as the second value.

[0017] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, calculating a first lesion score for each lesion type of the aorta image to be analyzed includes:

[0018] For each type of lesion, the lesion score of each key area is obtained, the average score of the lesion scores is calculated, and the standard deviation of the average score is calculated as the third standard deviation. If the third standard deviation is greater than or equal to a preset third threshold, the area ratio of the lesion area corresponding to each key area to the aorta image to be analyzed is also calculated, all the area ratios are normalized, and the normalized area ratio is used as the first weight corresponding to the lesion score. Based on the first weight and the lesion score, a weighted average score of the lesion score is calculated, and the weighted average score is used as the first lesion score of the lesion type. If the third standard deviation is less than the third threshold, the average score is used as the first lesion score of the lesion type.

[0019] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, determining the corresponding lesion type based on the first lesion score includes:

[0020] Calculate the first lesion score of the lesion type of the aortic image to be analyzed for each image type, set a corresponding second weight for each image type according to historical data, calculate the comprehensive lesion score of each lesion type of multiple image types based on the second weight and the first lesion score, obtain the maximum comprehensive lesion score among all the comprehensive lesion scores, and determine whether the maximum comprehensive lesion score is greater than or equal to a preset fourth threshold. If so, use the lesion type corresponding to the maximum comprehensive lesion score as the lesion type of the corresponding patient. If not, have professionals further review the results of each step.

[0021] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, all aortic images to be analyzed are enhanced, including:

[0022] Prepare multiple first-category images and second-category images in advance, associate the first-category images and the second-category images with each other to generate an image combination, and generate a first model based on the image combination training. The first model can generate the second-category images based on the first-category images. After the first model generates the second-category images, use the second-category images generated by the first model as third-category images, compare the third-category images with the corresponding second-category images, and calculate the difference between the third-category images and the second-category images. If the difference is greater than or equal to a preset fifth threshold, the first model adjusts the model parameters based on the difference, and repeats this step until the difference is less than the fifth threshold.

[0023] In combination with the first aspect, in a seventh implementation manner of the first aspect of the present application, generating a first model based on the image combination training includes:

[0024] In the process of training and generating the first model, the feature maps of the first category of images are divided into several groups, each group uses multiple convolution kernels of different sizes to perform convolution operations to obtain intermediate feature maps, all the intermediate feature maps are spliced ​​together to generate the third category of images, and the weight of each group is generated based on the difference. The intermediate feature map of each group is multiplied by the corresponding weight to obtain a new intermediate feature map, and the new intermediate feature maps are combined to generate the third category of images.

[0025] In a second aspect, the present application provides a complex aortic lesion intelligent analysis system, the system comprising:

[0026] An image collection module is used to collect a number of aortic images of multiple image types including different lesion types, wherein the aortic images of each image type include normal aortic images and lesion aortic images, collect corresponding lesion information as label information, wherein the label information includes lesion type and lesion range, and annotate the lesion aortic images based on the label information;

[0027] A model training module, configured to extract a lesion area from an aortic image using a first extraction model, extract a key area of ​​a preset first size from the lesion area, combine the key area and corresponding lesion information based on the annotation information to generate first training data, and train a first recognition model based on the first training data, wherein the first recognition model can output a corresponding lesion type and a lesion score according to the key area of ​​the aortic image, wherein the lesion score represents a lesion probability of the corresponding lesion type;

[0028] A lesion analysis module is used to obtain aortic images to be analyzed of a patient of multiple image types, enhance all aortic images to be analyzed, input the aortic images to be analyzed of each image type after the enhancement process into the first extraction model to obtain the corresponding lesion area, then extract a key area of ​​a predetermined size from each lesion area, use the trained first recognition model to identify each key area in the aortic image to be analyzed, obtain a lesion score for each lesion type, calculate a first lesion score for each lesion type of the aortic image to be analyzed, and determine the corresponding lesion type based on the first lesion score.

[0029] A third aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned complex aortic lesion intelligent analysis method.

[0030] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0031] In the technical solution provided in the present application, a number of aortic images of multiple image types containing different lesion types are collected, lesion information is collected as label information, and the lesion aortic images are annotated based on the label information; lesion areas and key areas are extracted, first training data is generated, and a first recognition model is trained; aortic images to be analyzed of multiple image types of the patient are obtained and enhanced, each key area in the aortic image to be analyzed is identified using the trained first recognition model, a lesion score for each lesion type is obtained, a first lesion score for each lesion type of the aortic image to be analyzed is calculated, and the lesion type is determined based on the first lesion score.

[0032] By combining multiple image types, characteristic information of aortic lesions can be obtained from different angles, thereby more comprehensively reflecting the complexity of the lesions; through the fusion of multimodal data, the system can utilize the advantages of different modalities, make up for the shortcomings of a single modality, and significantly improve the accuracy of lesion diagnosis; by extracting the lesion area, it also combines multi-dimensional information such as texture features and grayscale features to make lesion identification more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0034] Figure 1 A schematic diagram of an embodiment of a complex aortic lesion intelligent analysis method in an embodiment of the present application;

[0035] Figure 2 is a schematic diagram of an embodiment of a polygon drawn based on the normalized eigenvalues ​​of the first region in an embodiment of the present application;

[0036] Figure 3 is a schematic diagram of an embodiment of a polygon drawn based on the normalized eigenvalues ​​of the second region in an embodiment of the present application;

[0037] Figure 4 This is a schematic diagram of an embodiment of an intelligent analysis system for complex aortic lesions in an embodiment of the present application. DETAILED DESCRIPTION

[0038] Embodiments of the present application provide a method, system and storage medium for intelligent analysis of complex aortic lesions. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0039] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of a complex aortic lesion intelligent analysis method includes:

[0040] Step S1, collecting a number of aortic images of multiple image types containing different lesion types, each image type of aortic image includes a normal aortic image and a diseased aortic image, collecting corresponding lesion information as label information, the label information includes lesion type and lesion range, and annotating the diseased aortic image based on the label information.

[0041] Specifically, traditional diagnosis of aortic lesions mainly relies on medical imaging technology, such as CTA, MRA and echocardiography. However, traditional methods have single modality limitations and are difficult to fully reflect the complex characteristics of aortic lesions. In order to accurately analyze complex aortic lesions, several aortic images of multiple image types containing different lesion types are collected. The image types include CTA, ARI, MRA, echocardiography and PET-CT image data. The lesion types include Stranford B type aortic dissection, complex aortic aneurysm and aortic stenosis. The aortic images of each image type collected include normal aortic images and diseased aortic images with lesions. For the diseased aortic images, the corresponding lesion information is also collected as label information. The label information includes lesion type and lesion range. The lesion range refers to the location and size of the lesion in the diseased aortic image. The diseased aortic image is annotated based on the label information to facilitate the training of subsequent models.

[0042] Step S2, using the first extraction model to extract the lesion area from the aortic image, extracting the key area of ​​a preset first size from the lesion area, combining the key area and the corresponding lesion information based on the annotation information to generate the first training data, and training the first recognition model based on the first training data. The first recognition model can output the corresponding lesion type and lesion score according to the key area of ​​the aortic image, and the lesion score represents the lesion probability of the corresponding lesion type.

[0043] Specifically, in order to facilitate the subsequent first recognition model to identify the type of lesion, the first extraction model is used to extract the lesion area from the collected aortic image. The specific extraction method will be explained in detail later. After the lesion area is extracted, the extracted lesion area may include a fuzzy area at the boundary, that is, it includes an area where a lesion must have occurred, and it may also include a boundary area between the lesion area and the non-lesion area. In order to improve the recognition accuracy of the subsequent first recognition model, a key area of ​​a preset first size is extracted from the lesion area. The preset first size can be, for example, a key area of ​​256 pixels belonging to the middle position of the lesion area. The key area is generally in the middle position of the lesion area. After the key areas are identified, the key areas are combined with the marking information of the diseased aortic image to generate the first training data, and the first recognition model is trained based on the first training data. The first recognition model can output the corresponding lesion type and lesion score according to the key areas of the input aortic image. For example, when the aortic image with the key areas marked is input, the first recognition model can output the corresponding lesion type and lesion score. For example, image A includes three key areas, and there are three lesion types a, b and c. For each key area, the first recognition model can output the lesion score of the key area being lesion type a, for example 0.75, the lesion score of the lesion type b, for example 0.6, and the lesion score of the lesion type c, for example 0.5.

[0044] Step S3, obtaining the patient's aortic images to be analyzed of multiple image types, enhancing all the aortic images to be analyzed, inputting the enhanced aortic images of each image type to be analyzed into the first extraction model to obtain the corresponding lesion area, then extracting the key area of ​​the predetermined size from each lesion area, using the trained first recognition model to identify each key area in the aortic image to be analyzed, obtaining the lesion score of each lesion type, calculating the first lesion score of each lesion type of the aortic image to be analyzed, and determining the corresponding lesion type based on the first lesion score.

[0045] Specifically, in order to improve the accuracy of identifying lesion types, multiple image types of aortic images to be analyzed of the patient are obtained. Some aortic images may not be clear enough, so the aortic images to be analyzed are first enhanced to improve the quality of the aortic images to be analyzed, which helps to obtain the lesion area more clearly in the subsequent process and identify the lesion type, thereby improving the accuracy of identification. The aortic images to be analyzed of each image type after enhancement are input into the first extraction model to obtain the corresponding lesion area, and then the key areas of a predetermined size are extracted from the lesion area. The first recognition model trained in the previous step is used to identify each key area in the aortic image to be analyzed to obtain the lesion score of each lesion type, and then the first lesion score of each lesion type of the aortic image to be analyzed is calculated based on the recognition result of the first recognition model. The specific process of calculating the first lesion score will be explained in detail later. The corresponding first lesion score is obtained for each image type of aortic image to be analyzed, and the final lesion type is determined by combining the first lesion scores of all image types.

[0046] Through the above method, multiple image types and artificial intelligence technologies can be effectively utilized to provide more accurate and efficient support for the analysis and diagnosis of aortic lesions.

[0047] In a specific embodiment, the lesion area is extracted from the aorta image using the first extraction model, and the following steps are further performed:

[0048] The brightness value of each pixel in the aortic image is obtained, the frequency distribution of each brightness value is counted, the maximum frequency and the minimum frequency are obtained, and a first value is preset. The frequency of each brightness value is subtracted from the minimum frequency to obtain a first result value, the maximum frequency is subtracted from the minimum frequency to obtain a second result value, the first result value is divided by the second result value to obtain a third result value, the third result value is multiplied by the first value to obtain a fourth result value, the fourth result value is used as the normalized frequency of the corresponding brightness value, a histogram of the brightness value of the aortic image is generated based on the normalized frequency, a plurality of different mathematical models are used to fit the normalized frequency distribution, the parameters of each mathematical model are adjusted, and the plurality of mathematical models are combined to fit the distribution of the normalized frequency, and the pixels in the aortic image are divided into different groups based on the mathematical model after the parameters are adjusted, the pixels divided into the same group are regarded as belonging to the same area, and the plurality of different areas are analyzed to identify the diseased area.

[0049] Specifically, in order to accurately extract the diseased area in the aortic image, the brightness value of each pixel in the aortic image is first obtained, and the frequency distribution of each brightness value is counted, for example, the distribution of brightness values ​​from 0 to 255 is counted, the maximum frequency and the minimum frequency are obtained, and a first numerical value is preset. The frequency of each brightness value is subtracted from the minimum frequency to obtain a first result value, the maximum frequency is subtracted from the minimum frequency to obtain a second result value, the first result value is divided by the second result value to obtain a third result value, the third result value is multiplied by the first numerical value to obtain a fourth result value, and the fourth result value is used as the normalized frequency of the corresponding brightness value. The frequency distribution is normalized by the above method to enhance the contrast of the image, and then the image is generated based on the normalized frequency. A histogram of the brightness values ​​of the aortic image is formed, and then a plurality of different mathematical models are used to fit the normalized frequency distribution, and the parameters of each mathematical model are adjusted. The plurality of mathematical models are combined to fit the normalized frequency distribution. For example, a mixed Gaussian distribution method can be used to divide the normalized frequency distribution into a plurality of overall distributions. First, the initial Gaussian distribution parameters are selected, and the parameters of the Gaussian distribution are adjusted through an iterative optimization algorithm. The normalized frequency distribution is fitted using a linear combination of a plurality of Gaussian distributions, and finally a plurality of overall distributions are determined. Each normal distribution corresponds to an area in the aortic image, and pixels belonging to the same area are divided into the same group. The divided multiple different areas are further analyzed to identify the lesion area.

[0050] In a specific embodiment, analyzing multiple different regions to identify the lesion region specifically includes the following steps:

[0051] Extract the brightness value of the pixel from each area, calculate the average brightness value and the first standard deviation of each area, and also calculate the second value of each area, use the brightness average, the first standard deviation and the second value as eigenvalues, normalize the eigenvalues, and draw the characteristic polygon of each area based on the normalized eigenvalues. First, set an origin, draw multiple line segments along a first number of directions with the origin, each line segment represents an eigenvalue, mark the normalized eigenvalue on the corresponding line segment for each area, connect all the eigenvalues ​​corresponding to each area in sequence to form a polygon, calculate the area of ​​each polygon, obtain the first polygon whose area is greater than the preset first threshold, calculate the second standard deviation of the eigenvalue corresponding to each first polygon, obtain the first polygon whose second standard deviation is greater than the second threshold, which is called the second polygon, and take the area corresponding to the second polygon as the lesion area.

[0052] Specifically, in order to identify the diseased area, a series of eigenvalues ​​are extracted from each area, and the eigenvalues ​​include the brightness average, the first standard deviation, the irregularity of the area shape and the second value. The second value represents the texture feature (such as smoothness or roughness) of each area. The specific method for calculating the second value will be explained in detail later. The multiple eigenvalues ​​are normalized, and the characteristic variable image of each area is drawn based on the normalized eigenvalues. The first number is the total number of eigenvalues ​​4. Table 1 is a table of the eigenvalues ​​of the two areas and the corresponding maximum and minimum values.

[0053] Eigenvalue First Area Second area Maximum Minimum Average brightness 120 150 200 100 First standard deviation 10 20 30 5 Second value 0.5 0.8 0.9 0.4 Irregularity of area shape 0.3 0.6 0.7 0.2

[0054] Table 1

[0055] In fact, there are multiple regions. For the sake of clarity, this embodiment lists only two regions. After normalizing the eigenvalues ​​of the first region, the normalized eigenvalues ​​are 0.2, 0.2, 0.2, and 0.2, respectively. After normalizing the eigenvalues ​​of the second region, the corresponding normalized eigenvalues ​​are 0.5, 0.6, 0.8, and 0.8, respectively. The normalized eigenvalues ​​are marked on the corresponding line segments, such as Figure 2 As shown in the figure, a polygon is drawn based on the normalized eigenvalues ​​of the first region. First, an origin O is set, and line segments are drawn along the four directions. The normalized eigenvalues ​​0.2, 0.2, 0.2, 0.2 are marked on the corresponding line segments, and then all the eigenvalues ​​after normalization are connected in sequence to form a polygon, as shown in the figure. Figure 3 As shown in FIG. 1 , polygons are drawn based on the normalized eigenvalues ​​of the second region. After drawing the polygons corresponding to each region, by comparing the characteristic polygon maps of different regions (such as lesion regions and normal tissue regions), their characteristic differences can be intuitively seen. For example, the characteristic polygon map of the lesion region may be larger than that of the normal region in some features (such as texture roughness or irregularity), and the characteristic polygon map of the normal region may be closer to the center in some features (such as average brightness or regional area). The area of ​​each polygon is calculated. The area size can reflect the comprehensive strength of the regional features. The larger the area of ​​the polygon map, the higher the characteristic value of the region. For example, the characteristic polygon map of the lesion region is larger. The area of ​​the characteristic polygon map may be larger than that of the normal area, which may mean that the features of the lesion area are more significant. Therefore, the first polygon with an area greater than the first threshold is obtained. The shape of the polygon map can reflect the relationship between different features. For example, if the characteristic polygon map of the lesion area is more prominent in texture roughness, it may indicate that the texture of the lesion area is more complex. For each first polygon, the second standard deviation of its corresponding eigenvalue is calculated. The second standard deviation is used to measure the discreteness of the eigenvalue. A second standard deviation threshold (second threshold) is set, and the first polygon with a second standard deviation greater than the threshold (called the second polygon) is screened out, and the area corresponding to the second polygon is marked as the lesion area.

[0056] In a specific embodiment, calculating the second value of each region includes the following steps:

[0057] A small area containing a preset number of pixels is obtained from each area, and the central pixel of each small area is obtained. For other pixels in the small area, the grayscale value is compared with that of the central pixel. If the grayscale value of other pixels is greater than or equal to the grayscale value of the central pixel, it is marked as 1, otherwise it is marked as 0. The marked values ​​corresponding to all other pixels contained in each small area are combined into a binary number, the binary numbers of all small areas are counted, the total number of all non-repeating binary numbers is obtained, and the total number is used as the second value.

[0058] Specifically, in order to calculate the second value of each area, that is, the texture feature value of each area, a small area containing a preset number of pixels (for example, 3*3) is obtained from each area, the central pixel of each small area is obtained, and the pixels other than the central pixel in the small area are called other pixels. The grayscale values ​​of other pixels and the central pixel are compared. If the grayscale value of other pixels is greater than or equal to the grayscale value of the central pixel, it is marked as 1, otherwise it is marked as 0. The corresponding mark values ​​of all other pixels contained in each small area are sorted from left to right and from top to bottom according to the position of the pixels in the small area and combined into a binary number, such as 11101100. The binary numbers of all small areas are counted to obtain the total number of binary numbers in all small areas. The binary values ​​corresponding to different small areas may be the same. The same binary value represents that the small areas have the same texture characteristics. Therefore, the total number of all non-repeating binary numbers is counted, and the total number is used as the second value.

[0059] In a specific embodiment, calculating the first lesion score of each lesion type of the aorta image to be analyzed specifically includes the following steps:

[0060] For each lesion type, the lesion score of each key area is obtained, the average score of the lesion score is calculated, and the standard deviation of the average score is calculated as the third standard deviation. If the third standard deviation is greater than or equal to the preset third threshold, the area ratio of the lesion area corresponding to each key area to the aortic image to be analyzed is also calculated, all area ratios are normalized, and the normalized area ratio is used as the first weight of the corresponding lesion score. Based on the first weight and the lesion score, the weighted average score of the lesion score is calculated, and the weighted average score is used as the first lesion score of the lesion type. If the third standard deviation is less than the third threshold, the average score is used as the first lesion score of the lesion type.

[0061] Specifically, in order to more accurately calculate the first lesion score of each lesion type of the aortic image to be analyzed, the lesion score of each key area is obtained for each lesion type. Assuming that there are 4 lesion types and 3 key areas, as shown in Table 2, the first recognition model outputs the corresponding lesion score of each of the 3 key areas of the aortic image to be analyzed of a certain image type.

[0062] Lesion type a Lesion type b Lesion type c Lesion type Key Area 1 0.8 0.5 0.4 0.45 Key Area 2 0.3 0.85 0.35 0.6 Key Area 3 0.4 0.5 0.45 0.7

[0063] Table 2

[0064] For each lesion type (such as aortic dissection, aortic aneurysm, etc.), the lesion score of each key area is obtained. In order to evaluate the central tendency and dispersion of the lesion score and determine whether it is necessary to introduce the area ratio as a weight, the average score (0.5, 0.62, 0.4, 0.58) and the third standard deviation (0.0283) are calculated for each lesion type. The dispersion of the lesion score is determined according to the standard deviation of the score, and it is determined whether it is necessary to introduce the area ratio as a weight to more accurately determine the lesion type. A preset third threshold (assumed to be 0.02) is set to determine the third threshold. Are the three standard deviations large enough? If the third standard deviation is greater than the third threshold, it is considered that the lesion score has a large degree of discreteness and it is necessary to introduce an area ratio as a weight. For each key area, the area ratio of the lesion area to the aortic image to be analyzed is calculated. For each key area, the area ratio of the lesion area to the aortic image to be analyzed is calculated. All area ratios are normalized and the weighted average of the lesion scores is calculated based on the normalized area ratio as the weight as the first lesion score. Otherwise, it is considered that the lesion score has a small degree of discreteness and the average score is directly used as the first lesion score.

[0065] The above method can more accurately calculate the first lesion score and improve the accuracy of subsequent lesion type judgment by comprehensively considering the two factors of lesion score and area ratio.

[0066] In a specific embodiment, determining the corresponding lesion type based on the first lesion score specifically includes the following steps:

[0067] Calculate the first lesion score of the lesion type of the aortic image to be analyzed for each image type, set a corresponding second weight for each image type according to historical data, calculate the comprehensive lesion score of each lesion type of the multiple image types based on the second weight and the first lesion score, obtain the maximum comprehensive lesion score among all comprehensive lesion scores, and determine whether the maximum comprehensive lesion score is greater than or equal to a preset fourth threshold. If so, use the lesion type corresponding to the maximum comprehensive lesion score as the lesion type of the corresponding patient. If not, have professionals further review the results of each step.

[0068] Specifically, for each image type, a first lesion score for each lesion type is calculated. Based on historical diagnostic data, the accuracy and reliability of each image type in the diagnosis of different lesion types are analyzed. For example, CTA images may be more sensitive to certain lesion types (such as aortic aneurysms), while MRI images may be clearer for calcified lesions. A second weight is set for each image type. In order to comprehensively consider the first lesion scores of multiple image types, a comprehensive lesion score for each lesion type is calculated. Based on the second weight and the first lesion score, the comprehensive lesion score for each lesion type of the multiple image types is calculated, and the maximum comprehensive lesion score among all comprehensive lesion scores is obtained. It is determined whether the maximum comprehensive lesion score is greater than or equal to a preset fourth threshold. The fourth threshold represents the minimum confidence level of the diagnosis result. If so, the diagnosis result is accepted, and the lesion type corresponding to the maximum comprehensive lesion score is used as the lesion type of the corresponding patient. If not, the diagnosis result is unreliable and further examination may be required.

[0069] The above method takes into account the first lesion scores of multiple image types by calculating the comprehensive lesion score, and by setting weights for each image type, it can more comprehensively reflect the severity and characteristics of the lesion. This method avoids the risk of misdiagnosis that may be caused by a single image type.

[0070] In a specific embodiment, all aorta images to be analyzed are enhanced, specifically comprising the following steps:

[0071] Prepare multiple first-category images and second-category images in advance, associate the first-category images and the second-category images with each other to generate an image combination, generate a first model based on image combination training, the first model can generate the second-category images based on the first-category images, after the first model generates the second-category images, use the second-category images generated by the first model as third-category images, compare the third-category images with the corresponding second-category images, calculate the difference between the third-category images and the second-category images, if the difference is greater than or equal to a preset fifth threshold, adjust the model parameters of the first model based on the difference, and repeat this step until the difference is less than the fifth threshold.

[0072] Specifically, in order to enhance the aortic image to be analyzed, improve the quality of the active image to be analyzed, and help to more accurately determine the type of lesion later, multiple first-category images and second-category images are prepared in advance, the first-category images refer to low-resolution images, and the second-category images refer to high-resolution images. The low-resolution images (first-category images) and the corresponding high-resolution images (second-category images) are paired to form image combinations. These image combinations will be used to train the first model, and the first-category images are input into the first model. By learning the feature mapping relationship of the images, an attempt is made to generate a result as close to the second-category image as possible. The difference between the generated third-category image and the real second-category image is also calculated, and a preset fifth threshold is set to determine whether the generated third-category image is close enough to the real second-category image. If the difference is greater than or equal to the preset fifth threshold, it means that the gap between the generated third-category image and the real second-category image is still large, and the first model needs to be further optimized. The model parameters are adjusted according to the difference to reduce the difference between the generated image and the real image. The above method adjusts the parameters of the first model by the difference to optimize the performance of the first model until the generated third-category image is close enough to the real second-category image (high-resolution image).

[0073] In a specific embodiment, generating a first model based on image combination training specifically includes the following steps:

[0074] In the process of training and generating the first model, the feature maps of the first type of image are divided into several groups, each group is convolved using multiple convolution kernels of different sizes to obtain an intermediate feature map, all the intermediate feature maps are spliced ​​together to generate the third type of image, the weight of each group is generated based on the difference, the intermediate feature map of each group is multiplied by the corresponding weight to obtain a new intermediate feature map, and the new intermediate feature maps are combined to generate the third type of image.

[0075] Specifically, in order to generate high-resolution images that can provide clearer details so as to help accurately diagnose the type of lesions, in the process of training and generating the first model, the feature maps of the first type of images are divided into several groups, and each group uses multiple convolution kernels of different sizes to perform convolution operations to obtain intermediate feature maps. Through group convolution and convolution kernels of different sizes, more diverse features are extracted, and the expression ability of the model is improved. All intermediate feature maps are spliced ​​together to generate the third type of image, and the weight of each group is generated based on the difference. The intermediate feature map of each group is multiplied by the corresponding weight to obtain a new intermediate feature map, and the new intermediate feature map is combined to generate the third type of image. The weight is dynamically adjusted according to the difference between the generated image and the real image, which can better highlight important features and suppress unimportant features. Through the combination and reconstruction of weighted feature maps, the generated high-resolution image is closer to the real image and has richer details. Combined with dynamic weight adjustment and difference feedback, the stability and convergence of the training process are ensured.

[0076] The above method extracts more diverse features through group convolution and dynamic weight adjustment, and dynamically optimizes the performance of the generator according to the difference between the generated image and the real image. This method not only improves the quality of the generated image, but also has an efficient training process and wide applicability.

[0077] The above describes a complex aortic lesion intelligent analysis method in the embodiment of the present application. The following describes a complex aortic lesion intelligent analysis system in the embodiment of the present application. Figure 4 In the embodiment of the present application, an intelligent analysis system for complex aortic lesions includes:

[0078] An image collection module is used to collect a number of aortic images of multiple image types including different lesion types, wherein the aortic images of each image type include normal aortic images and lesion aortic images, collect corresponding lesion information as label information, wherein the label information includes lesion type and lesion range, and annotate the lesion aortic images based on the label information;

[0079] A model training module, used to extract a lesion area from the aorta image using a first extraction model, extract a key area of ​​a preset first size from the lesion area, combine the key area and the corresponding lesion information based on the annotation information to generate first training data, and train a first recognition model based on the first training data, wherein the first recognition model can output a corresponding lesion type and a lesion score according to the key area of ​​the aorta image, and the lesion score represents a lesion probability of the corresponding lesion type;

[0080] The lesion analysis module is used to obtain aortic images to be analyzed of multiple image types of the patient, enhance all the aortic images to be analyzed, input the aortic images to be analyzed of each image type after the enhancement process into the first extraction model to obtain the corresponding lesion area, then extract the key area of ​​the predetermined size from each lesion area, use the trained first recognition model to identify each key area in the aortic image to be analyzed, obtain the lesion score of each lesion type, calculate the first lesion score of each lesion type of the aortic image to be analyzed, and determine the corresponding lesion type based on the first lesion score.

[0081] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of a complex aortic lesion intelligent analysis method.

[0082] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0083] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0084] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent analysis method for complex aortic lesions, characterized in that: The method comprises: Step S1, collecting a number of aortic images of multiple image types including different lesion types, wherein each image type of aortic image includes a normal aortic image and a lesion aortic image, collecting corresponding lesion information as label information, wherein the label information includes a lesion type and a lesion range, and annotating the lesion aortic image based on the label information; Step S2, extracting a lesion area from the aorta image using a first extraction model, extracting a key area of ​​a preset first size from the lesion area, combining the key area and corresponding lesion information based on the annotation information to generate first training data, training a first recognition model based on the first training data, wherein the first recognition model can output a corresponding lesion type and a lesion score according to the key area of ​​the aorta image, wherein the lesion score represents a lesion probability of the corresponding lesion type; Step S3, obtaining aortic images to be analyzed of multiple image types of the patient, enhancing all the aortic images to be analyzed, inputting the aortic images to be analyzed of each image type after the enhancement into the first extraction model to obtain the corresponding lesion area, then extracting a key area of ​​a predetermined size from each lesion area, using the trained first recognition model to identify each key area in the aortic image to be analyzed, obtaining a lesion score for each lesion type, calculating a first lesion score for each lesion type of the aortic image to be analyzed, and determining the corresponding lesion type based on the first lesion score.

2. The method according to claim 1, characterized in that The first extraction model is used to extract the lesion area from the aorta image, including: The brightness value of each pixel in the aortic image is obtained, the frequency distribution of each brightness value is counted, the maximum frequency and the minimum frequency are obtained, and a first value is preset. The frequency of each brightness value is subtracted from the minimum frequency to obtain a first result value, the maximum frequency is subtracted from the minimum frequency to obtain a second result value, the first result value is divided by the second result value to obtain a third result value, the third result value is multiplied by the first value to obtain a fourth result value, the fourth result value is used as the normalized frequency of the corresponding brightness value, and a histogram of the brightness value of the aortic image is generated based on the normalized frequency. A plurality of different mathematical models are used to fit the normalized frequency distribution, and the parameters of each of the mathematical models are adjusted. The plurality of mathematical models are combined to fit the distribution of the normalized frequency. Based on the mathematical model after adjusting the parameters, the pixels in the aortic image are divided into different groups, and the pixels divided into the same group are regarded as belonging to the same area. The multiple different areas are analyzed to identify the diseased area.

3. The method according to claim 2, characterized in that Several different regions were analyzed to identify the diseased area, including: Extract the brightness value of the pixel from each area, calculate the average brightness value and the first standard deviation of each area, and also calculate the second value of each area, use the brightness average, the first standard deviation and the second value as eigenvalues, normalize the eigenvalues, and draw the characteristic polygon of each area based on the normalized eigenvalues. First, set an origin, draw multiple line segments along a first number of directions with the origin, each line segment represents an eigenvalue, mark the normalized eigenvalue on the corresponding line segment for each area, connect all the eigenvalues ​​corresponding to each area in sequence to form a polygon, calculate the area of ​​each polygon, obtain the first polygon whose area is greater than the preset first threshold, calculate the second standard deviation of the eigenvalues ​​corresponding to each of the first polygons, obtain the first polygon whose second standard deviation is greater than the second threshold, and call it the second polygon, and use the area corresponding to the second polygon as the lesion area.

4. The method according to claim 3, characterized in that Calculate the second value for each region, including: A small area containing a preset number of pixels is obtained from each area, and the central pixel of each of the small areas is obtained. For other pixels in the small areas, the grayscale value is compared with that of the central pixel. If the grayscale value of other pixels is greater than or equal to the grayscale value of the central pixel, it is marked as 1, otherwise it is marked as 0. The marked values ​​corresponding to all other pixels contained in each of the small areas are combined into a binary number, and the binary numbers of all the small areas are counted to obtain the total number of all non-repeating binary numbers, and the total number is used as the second value.

5. The method according to claim 1, characterized in that Calculate the first lesion score for each lesion type of the aortic image to be analyzed, including: For each type of lesion, the lesion score of each key area is obtained, the average score of the lesion scores is calculated, and the standard deviation of the average score is calculated as the third standard deviation. If the third standard deviation is greater than or equal to a preset third threshold, the area ratio of the lesion area corresponding to each key area to the aorta image to be analyzed is also calculated, all the area ratios are normalized, and the normalized area ratio is used as the first weight corresponding to the lesion score. Based on the first weight and the lesion score, a weighted average score of the lesion score is calculated, and the weighted average score is used as the first lesion score of the lesion type. If the third standard deviation is less than the third threshold, the average score is used as the first lesion score of the lesion type.

6. The method according to claim 1, characterized in that Determining a corresponding lesion type based on the first lesion score includes: Calculate the first lesion score of the lesion type of the aortic image to be analyzed for each image type, set a corresponding second weight for each image type according to historical data, calculate the comprehensive lesion score of each lesion type of multiple image types based on the second weight and the first lesion score, obtain the maximum comprehensive lesion score among all the comprehensive lesion scores, and determine whether the maximum comprehensive lesion score is greater than or equal to a preset fourth threshold. If so, use the lesion type corresponding to the maximum comprehensive lesion score as the lesion type of the corresponding patient. If not, have professionals further review the results of each step.

7. The method according to claim 1, characterized in that All aortic images to be analyzed were enhanced, including: Prepare multiple first-category images and second-category images in advance, associate the first-category images and the second-category images with each other to generate an image combination, and generate a first model based on the image combination training. The first model can generate the second-category images based on the first-category images. After the first model generates the second-category images, use the second-category images generated by the first model as third-category images, compare the third-category images with the corresponding second-category images, and calculate the difference between the third-category images and the second-category images. If the difference is greater than or equal to a preset fifth threshold, the first model adjusts the model parameters based on the difference, and repeats this step until the difference is less than the fifth threshold.

8. The method according to claim 7, characterized in that Generating a first model based on the image combination training includes: In the process of training and generating the first model, the feature maps of the first category of images are divided into several groups, each group uses multiple convolution kernels of different sizes to perform convolution operations to obtain intermediate feature maps, all the intermediate feature maps are spliced ​​together to generate the third category of images, and the weight of each group is generated based on the difference. The intermediate feature map of each group is multiplied by the corresponding weight to obtain a new intermediate feature map, and the new intermediate feature maps are combined to generate the third category of images.

9. A complex aortic lesion intelligent analysis system, used to implement a complex aortic lesion intelligent analysis method according to any one of claims 1 to 8, characterized in that: The system comprises: An image collection module is used to collect a number of aortic images of multiple image types including different lesion types, wherein the aortic images of each image type include normal aortic images and lesion aortic images, collect corresponding lesion information as label information, wherein the label information includes lesion type and lesion range, and annotate the lesion aortic images based on the label information; A model training module, configured to extract a lesion area from an aortic image using a first extraction model, extract a key area of ​​a preset first size from the lesion area, combine the key area and corresponding lesion information based on the annotation information to generate first training data, and train a first recognition model based on the first training data, wherein the first recognition model can output a corresponding lesion type and a lesion score according to the key area of ​​the aortic image, wherein the lesion score represents a lesion probability of the corresponding lesion type; A lesion analysis module is used to obtain aortic images to be analyzed of a patient of multiple image types, enhance all aortic images to be analyzed, input the aortic images to be analyzed of each image type after the enhancement process into the first extraction model to obtain the corresponding lesion area, then extract a key area of ​​a predetermined size from each lesion area, use the trained first recognition model to identify each key area in the aortic image to be analyzed, obtain a lesion score for each lesion type, calculate a first lesion score for each lesion type of the aortic image to be analyzed, and determine the corresponding lesion type based on the first lesion score.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, a complex aortic lesion intelligent analysis method as described in any one of claims 1-8 is implemented.

Images