A method and system for identifying liver cancer blood supply arteries in DSA angiography images

Through the cyclic consistency generation adversarial network, the blurred image is converted into clear images in DSA contrast images, solving the radiation problem caused by multiple contrasts, and achieving the effect of efficiently identifying the blood-supplying artery of liver cancer in a single contrast, reducing radiation damage.

CN118799351BActive Publication Date: 2024-11-29XIEHE HOSPITAL ATTACHED TO TONGJI MEDICAL COLLEGE HUAZHONG SCI & TECH UNIV
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411284351.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-11-29
Estimated Expiration
2044-09-13

AI Technical Summary

Technical Problem

In the prior art, in order to identify multiple blood-supply arteries of liver cancer, multiple DSA angiography operations are required, resulting in a large amount of radiation during the identification of blood-supply arteries, resulting in radiation damage.

Method used

The cyclic consistency generation adversarial network (cycleGan) is used to convert the blurred DSA contrast images into clear images, and multiple DSA contrast images are obtained through a single contrast, and the tumor tissue blood supply artery in the liver tissue model is marked by the edge detection algorithm. Blood vessel contours to achieve image clarity improvement.

Benefits of technology

High-definition recognition of all blood supply arteries in a single radiography is achieved, reducing the dose of inclination radiation, and ensuring the recognition effect of blood supply arteries and avoiding radiation damage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118799351B_ABST
    Figure CN118799351B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of medical image recognition technology, and in particular to a method and system for identifying liver cancer blood supply arteries in DSA angiography images, comprising the following steps: dividing the clarity of the vascular contours of the tumor tissue blood supply arteries according to the edge detection algorithm of each target image, and obtaining a clear target image and a fuzzy target image; using a cycle consistency generative adversarial network cycleGan to convert the fuzzy target image into a clear target image according to the clear target image; and marking the vascular contours of the tumor tissue blood supply arteries in the liver tissue model in all clear target images by the edge detection algorithm. The present invention enables all images obtained by a single angiography to present high definition, and thus can identify all blood supply arteries, achieve low radiation dose of angiography rays, and at the same time ensure the recognition effect of the blood supply arteries.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for identifying liver cancer blood supply arteries in DSA angiography images. Background Art

[0002] Blocking the blood supply arteries of liver cancer can prevent liver cancer cells from obtaining nutrients from the blood, leading to ischemia, hypoxia and lack of oxygen and nutrient supply and death of liver cancer cells. Therefore, the effect of liver cancer cell destruction depends on the effect of identifying the blood supply arteries, that is, when all blood supply arteries are identified, all liver cancer blood supply arteries can be embolized to prevent tumor survival, recurrence and metastasis. When all blood supply arteries cannot be identified, all liver cancer blood supply arteries cannot be embolized, which will lead to tumor survival, recurrence and metastasis.

[0003] In order to identify multiple blood supply arteries of liver cancer and improve the identification effect of blood supply arteries in the existing technology, multiple angiography operations are required. However, multiple angiography operations will cause the patient to be exposed to a large amount of radiation during the identification of blood supply arteries, resulting in radiation damage. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for identifying blood supply arteries for liver cancer in DSA angiography images, so as to solve the technical problem that in order to identify multiple blood supply arteries for liver cancer and improve the identification effect of blood supply arteries in the prior art, multiple angiography operations are required, but multiple angiography operations will cause the patient to be exposed to a large amount of radiation during the identification of blood supply arteries, resulting in radiation damage.

[0005] In order to solve the above technical problems, the present invention specifically provides the following technical solutions:

[0006] A method for identifying a liver cancer blood supply artery in a DSA angiography image comprises the following steps:

[0007] A single DSA angiography is performed on the liver tissue model using a DSA angiography device to obtain multiple DSA angiography images;

[0008] In each DSA angiography image, the contour of the tumor tissue in the liver tissue model and the vascular contours of the hepatic artery and aorta in the liver tissue model are marked by an edge detection algorithm;

[0009] Extracting a local image between the contour of the tumor tissue and the vascular contours of the hepatic artery and the aorta in each DSA angiography image, and obtaining a target image for identifying the blood supply artery of the tumor tissue in the liver tissue model;

[0010] The clarity of the blood vessel contours of the blood supply arteries of the tumor tissues is divided according to the edge detection algorithm in each of the target images, so as to obtain a clear target image and a blurred target image;

[0011] The cycle consistency generative adversarial network cycleGan is used to convert the blurred target image into a clear target image according to the clear target image;

[0012] The vascular contours of the tumor tissue supplying arteries in the liver tissue model are marked in all clear target images using an edge detection algorithm.

[0013] As a preferred solution of the present invention, the method for dividing the clear target image and the blurred target image includes:

[0014] Perform edge detection on the blood vessel contour of the blood supply artery of the tumor tissue in each target image through an edge detection algorithm to obtain the blood vessel contour of the blood supply artery of the tumor tissue in each target image;

[0015] In each target image, the gradient amplitude of the blood vessel contour of the blood supply artery of the tumor tissue along the horizontal direction and the gradient amplitude along the vertical direction are calculated;

[0016] The square sum of the gradient amplitude of the blood vessel contour of the blood supply artery of the tumor tissue in each target image along the horizontal direction and the gradient amplitude along the vertical direction is calculated as the clarity value of the target image;

[0017] Marking a target image whose clarity value is higher than a preset threshold as a clear target image;

[0018] The target image whose clarity value is lower than or equal to a preset threshold is marked as a blurred target image.

[0019] As a preferred solution of the present invention, a method for converting a blurred target image into a clear target image according to a clear target image by using a cycle-consistent generative adversarial network cycleGan includes:

[0020] The blurred target image and the clear target image are used as two input items of the cycle-consistency generative adversarial network cycleGan, and the image conversion model is obtained by training the generator and the discriminator;

[0021] The image conversion model is: {Lh=G H (L), LHl=G L (Lh), D L (L), D L (Hl)}, {Hl = G L (H), Hlh=G H (Hl), D H (H), D H (Lh)};

[0022] In the formula, L is the blurred target image, H is the clear target image, and Lh is the HThe clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator.

[0023] As a preferred solution of the present invention, the loss function of the image conversion model formed by training the cycle consistency generative adversarial network cycleGan is: Ls=L GAN (G H ,D H ,L,H)+L GAN (G L ,D L ,L,H)+λ[L cyc (G H ,G L )+L lpips (L,Lhl,Hlh,k)]; where Ls is the loss function, L GAN (G H , D H , L,H) is the generator G H and the discriminator D H The loss function, L GAN (G L , D L , L,H) is the generator G L and the discriminator D L The loss function, L cyc (G H , G L ) is the objective visual consistency loss function, L lpips (L, Lhl, Hlh, k) is the subjective visual consistency loss function, L is the blurred target image, H is the clear target image, and Lh is the loss function generated by the generator G H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G LThe blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator of cyc (G H ,G L )+L lpips The weight value of (L, Lhl, Hlh, k);

[0024] in, ; ; ; Where, L i For the i A blurred target image, H i For the i A clear target image, Lh i For the generator G H According to the blurred target image L i The generated clear target image, LHl i For the generator G L According to the clear target image Lh i The generated blurred target image, Hl i For the generator G L According to the clear target image H i The generated blurred target image, Hlh i For the generator G H According to the blurred target image Hl i The generated clear target image, F is the sigmoid model, d(L i , LHl i ) is L i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(L i , Hlh i ) is L i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, d(H i , LHl i ) is H i Features extracted by VGG16 network and LHl iThe feature distance between features extracted by the VGG16 network, d(H i , Hlh i ) is H i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, n is the total number of blurred target images, and m is the total number of clear target images. and is the L2 norm, min is the minimization operator, and max is the maximization operator.

[0025] As a preferred solution of the present invention, the weight value λ is related to the clarity value of the blurred target image L, where λ=-(1-P(L))logP(L); wherein P(L) is the clarity value of the blurred target image L.

[0026] As a preferred solution of the present invention, the generator G H and the generator G L and the discriminator D H and the discriminator D L Synchronous adversarial training is adopted.

[0027] As a preferred solution of the present invention, a size normalization process is performed on a plurality of DSA angiography images obtained in a single angiography.

[0028] As a preferred solution of the present invention, the evaluation indicators of the edge detection algorithm include accuracy, recall rate, F1 score and average angular error.

[0029] As a preferred solution of the present invention, the edge detection algorithm adopts the Canny algorithm.

[0030] As a preferred embodiment of the present invention, the present invention provides a system for identifying a blood supply artery for liver cancer in a DSA angiography image, which is applied to the method for identifying a blood supply artery for liver cancer in a DSA angiography image, and the system comprises:

[0031] An image acquisition unit, used for performing a single DSA angiography on the liver tissue model through a DSA angiography device to obtain multiple DSA angiography images;

[0032] An image preprocessing unit, used for marking the contour of the tumor tissue in the liver tissue model and the vascular contours of the hepatic artery and the aorta in the liver tissue model by an edge detection algorithm in each DSA angiography image;

[0033] and for extracting a local image between the contour of the tumor tissue and the vascular contours of the hepatic artery and the aorta in each DSA angiography image, so as to obtain a target image for identifying the blood supply artery of the tumor tissue in the liver tissue model;

[0034] The clarity of the blood vessel contours of the blood supply arteries of the tumor tissues is divided according to the edge detection algorithm in each of the target images, so as to obtain a clear target image and a blurred target image;

[0035] An image conversion unit, used for converting a blurred target image into a clear target image according to a clear target image by using a cycle consistency generative adversarial network cycleGan;

[0036] The target recognition unit is used to mark the vascular contours of the blood supply arteries of the tumor tissue in the liver tissue model by edge detection algorithm in all clear target images.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention adopts a cycle consistency generative adversarial network (CycleGan) to convert a fuzzy target image into a clear target image according to a clear target image, thereby improving the clarity of a fuzzy image obtained by a single angiography, making all images obtained by a single angiography present high definition, and then all blood supply arteries can be identified, achieving low angiography radiation dose while ensuring the identification effect of the blood supply arteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0040] Figure 1 A flow chart of a method for automatically identifying liver cancer supplying arteries in DSA angiography images provided by an embodiment of the present invention;

[0041] Figure 2 A block diagram of a system for automatically identifying liver cancer supplying arteries in DSA angiography images provided by an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of a liver tissue model provided by an embodiment of the present invention;

[0043] Figure 4 A DSA angiography image including a blurred target image provided by an embodiment of the present invention;

[0044] Figure 5 A DSA angiography image converted from a blurred target image to a clear target image provided by an embodiment of the present invention;

[0045] Figure 6 A DSA angiography image with the vascular contours of the blood supply arteries of the tumor tissue in the liver tissue model marked is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0047] like Figure 1 As shown, the present invention provides a method for identifying liver cancer blood supply arteries in DSA angiography images, comprising the following steps:

[0048] The liver tissue model (such as Figure 3 A single angiography is performed to obtain multiple DSA angiography images;

[0049] In each DSA angiography image, the contour of the tumor tissue in the liver tissue model and the vascular contours of the hepatic artery and aorta in the liver tissue model are marked by an edge detection algorithm. Generally, the tumor tissue, hepatic artery and aorta are relatively large, and their contours can be accurately obtained directly through edge recognition.

[0050] Extracting a local image between the contour of the tumor tissue and the vascular contours of the hepatic artery and the aorta in each DSA angiography image, and obtaining a target image for identifying the blood supply artery of the tumor tissue in the liver tissue model;

[0051] The clarity of the vascular contours of the blood supply arteries of the tumor tissues is divided according to the edge detection algorithm of each target image, and a clear target image and a fuzzy target image are obtained;

[0052] CycleGan is used to generate adversarial networks that blur the target image (such as Figure 4 As shown) the image is converted into a clear target image according to the clear target image (as shown Figure 5 shown);

[0053] In all clear target images, the vascular contours of the tumor tissue supplying arteries in the liver tissue model are marked by edge detection algorithms (such as Figure 6 as shown).

[0054] The present invention divides all images obtained from a single angiography and determines images with insufficient clarity (i.e., blurred target images) among all images in the single angiography, wherein it is difficult to accurately identify the vascular contours of the blood supply arteries of the tumor tissue in the images with insufficient clarity, whereas the vascular contours of the blood supply arteries of the tumor tissue can be accurately identified in images with sufficient clarity (i.e., clear target images).

[0055] In order to improve the clarity of an image with insufficient clarity (i.e., a blurred target image) in a single angiography, the present invention uses a cycle consistency generative adversarial network, cycleGan, to perform image style transfer, and uses an image with sufficient clarity (i.e., a clear target image) as the target style for conversion, and performs image conversion / migration on the image with insufficient clarity (i.e., a blurred target image), thereby converting the image with insufficient clarity (i.e., a blurred target image) into an image with sufficient clarity (i.e., a clear target image), achieving clarity improvement, so that the converted image with insufficient clarity (i.e., a blurred target image) can accurately identify the vascular contours of the blood supply arteries of the tumor tissue.

[0056] In the present invention, an image with sufficient clarity in a single angiography (i.e., a clear target image) is used as the target style for image conversion / migration, so that the converted image with insufficient clarity (i.e., a blurred target image) continues the visual style of the clear image of the single angiography, and can ensure that the converted image with insufficient clarity (i.e., a blurred target image) is mixed with the original image with sufficient clarity (i.e., a clear target image) to give the embolization operator a consistent visual sense, ensure visual unity and coherence, help the operator's visual perception, ensure the coherence of the operation, and avoid operation pauses caused by visual adaptation.

[0057] In the present invention, only fuzzy images need to be processed in a single angiography, and clear images do not need to be processed, thereby reducing the amount of data processing and improving the efficiency of blood supply artery image recognition.

[0058] The methods for dividing clear target images and blurred target images include:

[0059] Perform edge detection on the blood vessel contour of the blood supply artery of the tumor tissue in each target image through an edge detection algorithm to obtain the blood vessel contour of the blood supply artery of the tumor tissue in each target image;

[0060] In each target image, the gradient amplitude of the blood vessel contour of the blood supply artery of the tumor tissue along the horizontal direction and the gradient amplitude along the vertical direction are calculated;

[0061] The square sum of the gradient amplitude of the blood vessel contour of the blood supply artery of the tumor tissue in each target image along the horizontal direction and the gradient amplitude along the vertical direction is calculated as the clarity value of the target image;

[0062] Marking a target image whose clarity value is higher than a preset threshold as a clear target image;

[0063] The target image whose clarity value is lower than or equal to a preset threshold is marked as a blurred target image.

[0064] The method of converting a blurred target image into a clear target image according to a clear target image by using a cycle consistency generative adversarial network CycleGan includes:

[0065] The blurred target image and the clear target image are used as two input items of the cycle consistency generative adversarial network cycleGan, and the image conversion model is obtained by training the generator and the discriminator; the image conversion model is: {Lh=G H (L), LHl=G L (Lh), D L (L), D L (Hl)}, {Hl = G L (H), Hlh = G H (Hl), D H (H), D H (Lh)}; where L is the blurred target image, H is the clear target image, and Lh is the H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator.

[0066] The CycleGAN network inherits the idea of ​​adversarial training of GAN, and uses a dual training and learning method to achieve mapping between the source domain and the target domain without a paired relationship. This allows CycleGAN to achieve the feature of migration without the need for paired data sets. The adversarial training of the GAN model is transformed into a cyclic adversarial mode. The model is transformed from the original one generator and one discriminator to the current two generator networks and two discriminator networks. A cyclic consistency loss function is added to constrain the structural similarity between the generated images and the real images.

[0067] The loss function of the image conversion model constructed by training the cycle-consistent generative adversarial network cycleGan is: Ls=L GAN (G H ,D H ,L,H)+L GAN (G L ,D L ,L,H)+λ[L cyc (G H ,G L )+L lpips (L,Lhl,Hlh,k)]; where Ls is the loss function, L GAN (G H , D H , L,H) is the generator G H and the discriminator D H The loss function, L GAN (G L , D L , L,H) is the generator G L and the discriminator D L The loss function, L cyc (G H , G L ) is the objective visual consistency loss function, L lpips (L, Lhl, Hlh, k) is the subjective visual consistency loss function, L is the blurred target image, H is the clear target image, and Lh is the loss function generated by the generator G H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator of cyc (G H ,G L )+L lpips The weight value of (L, Lhl, Hlh, k);

[0068] in, ; ; ; Where, L i For the i A blurred target image, H i For the i A clear target image, Lh i For the generator G H According to the blurred target image L i The generated clear target image, LHl i For the generator G L According to the clear target image Lh i The generated blurred target image, Hl i For the generator G L According to the clear target image H i The generated blurred target image, Hlh i For the generator G H According to the blurred target image Hl i The generated clear target image, F is the sigmoid model, d(L i , LHl i ) is L i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(L i , Hlh i ) is L i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, d(H i , LHl i ) is H i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(H i , Hlh i ) is H i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, n is the total number of blurred target images, and m is the total number of clear target images. and is the L2 norm, min is the minimization operator, and max is the maximization operator. lpips The feature distance in is measured by the learning-perceptual image block similarity LPIPS. Therefore, feature stacks are extracted in different convolutional layers (here, the D layer is taken as an example), and the feature stacks in different channels are normalized. The feature normalization result of x in the D layer is recorded as U D, and the feature normalization result of y in layer D is recorded as V D , and then scale the activation channel by vector, use the distance calculated by L2 norm, and average it in space to get the feature distance d(x,y), Where H D is the feature height in layer D, W D is the feature width in layer D, wl is the vector of scaling activation channels, U hw D is the eigenvalue of x at the position of height h and width w in layer D, V hw D is the eigenvalue of y at the position of height h and width w in layer D, is the L2 norm.

[0069] In the present invention, d(L i , LHl i ), d(L i , Hlh i ), d(H i , LHl i ), d(H i , Hlh i ), and d(x, y) is replaced accordingly to complete the calculation. The present invention comprises the adversarial loss function of the generative adversarial network cycleGan and the cycle consistency loss function, wherein the adversarial loss function L GAN (G H , D H , L,H) and L GAN (G L , D L , L,H) constrains the two generators and the two discriminators so that they can achieve the best results during the training process, ensuring that the generator and the discriminator evolve with each other, thereby ensuring that the generator can produce more realistic pictures.

[0070] The cycle consistency loss function of the present invention adopts L cyc (G H , G L ) is the objective visual consistency loss function, L lpips (L, Lhl, Hlh, k) is the subjective visual consistency loss function, where L cyc (G H , G L) is the objective visual consistency loss, which means that the output clear image of the generator and the input blurred image are only different in style, but the substantive content is the same, that is, after the blurred target image is converted, only the clarity of the blood vessel contour is improved, and the content information of the blood vessel contour is lost the least during the conversion process, or is retained to the greatest extent. At this time, the objective image information content is retained during the conversion process, and human subjective visual perception is not considered. lpips (L, Lhl, Hlh, k) is the subjective visual consistency loss, which means that the output clear image of the generator conforms to human visual perception, that is, after the conversion of the blurred target image, only the clarity of the blood vessel contour is improved, while the content information of the blood vessel contour is lost minimally during the conversion process, or is retained to the greatest extent while still conforming to human visual perception, making the converted image more acceptable to people and helping operators to grasp the image information more quickly. At this time, the subjective image visual content is restored during the conversion process.

[0071] In the present invention, L lpips (L,Lhl,Hlh,k) is based on the learned perceptual image patch similarity (LPIPS), also known as "perceptual loss", which is used to measure the difference between two images for quantification. This metric learns the reverse mapping of the generated image to the ground truth, forcing the generator to learn the reverse mapping of reconstructing the real image from the fake image and prioritizes the perceptual similarity between them. LPIPS is more consistent with human perception than traditional methods (such as L2 / PSNR, SSIM, FSIM).

[0072] Therefore, the cycle consistency loss of the present invention can improve the clarity of the converted blurred target image while retaining the image information to the greatest extent and conforming to human visual perception.

[0073] The lower the definition of the image, the easier it is to lose information in the image conversion process in order to ensure the clarity improvement, thus causing invalid image conversion. Therefore, the present invention adds weights to the cycle consistency loss. When the definition value of the blurred target image L is lower, the cycle consistency loss [L cyc (G H ,G L )+L lpips (L,Lhl,Hlh,k)] is larger, the cycle consistency loss [L cyc (G H ,G L )+L lpips (L,Lhl,Hlh,k)] is associated with the sharpness value of the blurred target image L, i.e., the cycle consistency loss [Lcyc (G H ,G L )+L lpips (L, Lhl, Hlh, k)] is adaptively adjusted as the definition value of the blurred target image L is adjusted. The lower the definition of the blurred target image, the more emphasis is placed on the cycle consistency loss [L cyc (G H ,G L )+L lpips (L, Lhl, Hlh, k)] is trained to ensure that the original image information is retained to the greatest extent possible during the clarity improvement process and that it conforms to human visual perception to the greatest extent possible, thereby avoiding ignoring the retention of image information during the clarity improvement process.

[0074] Specifically, the weight value λ is related to the clarity value of the blurred target image L, where λ=-(1-P(L))logP(L);

[0075] Where P(L) is the clarity value of the blurred target image L. Generator G H and the generator G L and the discriminator D H and the discriminator D L Synchronous adversarial training is adopted.

[0076] The size of multiple DSA angiography images obtained in a single angiography was normalized.

[0077] Evaluation metrics for edge detection algorithms include precision, recall, F1 score, and mean angular error.

[0078] The edge detection algorithm adopts the Canny algorithm, which is a multi-stage edge detection method including filtering, enhancement, detection and non-maximum suppression. The Canny algorithm can detect real weak edges and is insensitive to noise, and is suitable for the scenes requiring high precision and high reliability of the present invention.

[0079] like Figure 2 As shown, the present invention provides a system for identifying blood supply arteries for liver cancer in DSA angiography images, which is applied to a method for identifying blood supply arteries for liver cancer in DSA angiography images. The system includes:

[0080] An image acquisition unit, used for performing a single DSA angiography on the liver tissue model through a DSA angiography device to obtain multiple DSA angiography images;

[0081] An image preprocessing unit, used for marking the contour of the tumor tissue in the liver tissue model and the vascular contours of the hepatic artery and the aorta in the liver tissue model by an edge detection algorithm in each DSA angiography image;

[0082] and for extracting a local image between the contour of the tumor tissue and the vascular contours of the hepatic artery and the aorta in each DSA angiography image, so as to obtain a target image for identifying the blood supply artery of the tumor tissue in the liver tissue model;

[0083] The clarity of the vascular contours of the blood supply arteries of the tumor tissues is divided according to the edge detection algorithm of each target image, and a clear target image and a fuzzy target image are obtained;

[0084] An image conversion unit, used for converting a blurred target image into a clear target image according to a clear target image by using a cycle consistency generative adversarial network cycleGan;

[0085] The target recognition unit is used to mark the vascular contours of the blood supply arteries of the tumor tissue in the liver tissue model by edge detection algorithm in all clear target images.

[0086] The present invention adopts a cycle consistency generative adversarial network (CycleGan) to convert a fuzzy target image into a clear target image according to a clear target image, thereby improving the clarity of a fuzzy image obtained by a single angiography, making all images obtained by a single angiography present high definition, and then all blood supply arteries can be identified, achieving low angiography radiation dose while ensuring the identification effect of the blood supply arteries.

[0087] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present application.

Claims

1. A method for identifying blood supply arteries of liver cancer in DSA angiography images, characterized in that: The following steps are involved: A single DSA angiography is performed on the liver tissue model using a DSA angiography device to obtain multiple DSA angiography images; In each DSA angiography image, the contour of the tumor tissue in the liver tissue model and the vascular contours of the hepatic artery and aorta in the liver tissue model are marked by an edge detection algorithm; Extracting a local image between the contour of the tumor tissue and the vascular contours of the hepatic artery and the aorta in each DSA angiography image, and obtaining a target image for identifying the blood supply artery of the tumor tissue in the liver tissue model; The clarity of the blood vessel contours of the blood supply arteries of the tumor tissues is divided according to the edge detection algorithm in each of the target images, so as to obtain a clear target image and a blurred target image; The cycle consistency generative adversarial network cycleGan is used to convert the blurred target image into a clear target image according to the clear target image; The vascular contours of the tumor tissue supplying arteries in the liver tissue model are marked by edge detection algorithms in all clear target images; The method of converting a blurred target image into a clear target image according to a clear target image by using a cycle consistency generative adversarial network CycleGan includes: The blurred target image and the clear target image are used as two input items of the cycle-consistency generative adversarial network cycleGan, and the image conversion model is obtained by training the generator and the discriminator; The image conversion model is: {Lh=G H (L), LHl=G L (Lh), D L (L), D L (Hl)}, {Hl = G L (H), Hlh = G H (Hl), D H (H), D H (Lh)}; In the formula, L is the blurred target image, H is the clear target image, and Lh is the H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator of The loss function of training the cycle-consistent generative adversarial network cycleGan to form an image conversion model is: Ls=L GAN (G H ,D H ,L,H)+L GAN (G L ,D L ,L,H)+λ[L cyc (G H ,G L )+L lpips (L,Lhl,Hlh,k)]; In the formula, Ls is the loss function, L GAN (G H , D H , L,H) is the generator G H and the discriminator D H The loss function, L GAN (G L ,D L , L,H) is the generator G L and the discriminator D L The loss function, L cyc (G H , G L ) is the objective visual consistency loss function, L lpips (L, Lhl, Hlh, k) is the subjective visual consistency loss function, L is the blurred target image, H is the clear target image, and Lh is the loss function generated by the generator G H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator of cyc (G H ,G L )+L lpips The weight value of (L, Lhl, Hlh, k); in, ; ; ; Where, L i For the i A blurred target image, H i For the i A clear target image, Lh i For the generator G H According to the blurred target image L i The generated clear target image, LHl i For the generator G L According to the clear target image Lh i The generated blurred target image, Hl i For the generator G L According to the clear target image H i The generated blurred target image, Hlh i For the generator G H According to the blurred target image Hl i The generated clear target image, F is the sigmoid model, d(L i , LHl i ) is L i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(L i , Hlh i ) is L i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, d(H i , LHl i ) is H i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(H i , Hlh i ) is H i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, n is the total number of blurred target images, and m is the total number of clear target images. and is the L2 norm, min is the minimization operator, and max is the maximization operator.

2. The method for identifying the blood supply artery of liver cancer in DSA angiography images according to claim 1, characterized in that: The method for dividing the clear target image and the blurred target image includes: Perform edge detection on the blood vessel contour of the blood supply artery of the tumor tissue in each target image through an edge detection algorithm to obtain the blood vessel contour of the blood supply artery of the tumor tissue in each target image; In each target image, the gradient amplitude of the blood vessel contour of the blood supply artery of the tumor tissue along the horizontal direction and the gradient amplitude along the vertical direction are calculated; The square sum of the gradient amplitude of the blood vessel contour of the blood supply artery of the tumor tissue in each target image along the horizontal direction and the gradient amplitude along the vertical direction is calculated as the clarity value of the target image; Marking a target image whose clarity value is higher than a preset threshold as a clear target image; The target image whose clarity value is lower than or equal to a preset threshold is marked as a blurred target image.

3. The method for identifying the blood supply artery of liver cancer in DSA angiography images according to claim 1, characterized in that: The weight value λ is related to the clarity value of the blurred target image L, where λ = -(1-P(L))logP(L); Where P(L) is the clarity value of the blurred target image L.

4. The method for identifying the blood supply artery of liver cancer in DSA angiography images according to claim 1, characterized in that: Generator G H and the generator G L and the discriminator D H and the discriminator D L Synchronous adversarial training is adopted.

5. The method for identifying liver cancer blood supply arteries in DSA angiography images according to claim 1, characterized in that: The size of multiple DSA angiography images obtained in a single angiography was normalized.

6. The method for identifying the blood supply artery of liver cancer in DSA angiography images according to claim 1, characterized in that: The evaluation metrics of the edge detection algorithm include precision, recall, F1 score, and mean angular error.

7. The method for identifying the blood supply artery of liver cancer in DSA angiography images according to claim 1, characterized in that: The edge detection algorithm adopts the Canny algorithm.

8. A system for identifying liver cancer blood supply arteries in DSA angiography images, characterized in that: A method for identifying a liver cancer blood supply artery in a DSA angiography image as described in any one of claims 1 to 7, the system comprising: An image acquisition unit, used for performing a single DSA angiography on the liver tissue model through a DSA angiography device to obtain multiple DSA angiography images; An image preprocessing unit, used for marking the contour of the tumor tissue in the liver tissue model and the vascular contours of the hepatic artery and the aorta in the liver tissue model by an edge detection algorithm in each DSA angiography image; and for extracting a local image between the contour of the tumor tissue and the vascular contours of the hepatic artery and the aorta in each DSA angiography image, so as to obtain a target image for identifying the blood supply artery of the tumor tissue in the liver tissue model; The clarity of the blood vessel contours of the blood supply arteries of the tumor tissues is divided according to the edge detection algorithm in each of the target images, so as to obtain a clear target image and a blurred target image; An image conversion unit, used for converting a blurred target image into a clear target image according to a clear target image by using a cycle consistency generative adversarial network cycleGan; A target recognition unit, used for marking the vascular contours of the blood supply arteries of the tumor tissue in the liver tissue model by using an edge detection algorithm in all clear target images; The method of converting a blurred target image into a clear target image according to a clear target image by using a cycle consistency generative adversarial network CycleGan includes: The blurred target image and the clear target image are used as two input items of the cycle-consistency generative adversarial network cycleGan, and the image conversion model is obtained by training the generator and the discriminator; The image conversion model is: {Lh=G H (L), LHl=G L (Lh), D L (L), D L (Hl)}, {Hl = G L (H), Hlh = G H (Hl), D H (H), D H (Lh)}; In the formula, L is the blurred target image, H is the clear target image, and Lh is the H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator of The loss function of training the cycle-consistent generative adversarial network cycleGan to form an image conversion model is: Ls=L GAN (G H ,D H ,L,H)+L GAN (G L ,D L ,L,H)+λ[L cyc (G H ,G L )+L lpips (L,Lhl,Hlh,k)]; where Ls is the loss function, L GAN (G H , D H , L,H) is the generator G H and the discriminator D H The loss function, L GAN (G L , D L , L,H) is the generator G L and the discriminator D L The loss function, L cyc (G H , G L ) is the objective visual consistency loss function, L lpips (L, Lhl, Hlh, k) is the subjective visual consistency loss function, L is the blurred target image, H is the clear target image, and Lh is the loss function generated by the generator G H The clear target image G is generated based on the blurred target image L. H is the generator of the clear target image generated from the blurred target image, G L is a generator of a blurred target image generated from a clear target image, LH1 is a generator generated by the generator G L The blurred target image D is generated based on the clear target image Lh. L For the generator G L The discriminator of Hl is generated by the generator G L The blurred target image H is generated based on the clear target image H, and Hlh is generated by the generator G H The clear target image D is generated based on the blurred target image H1. H For the generator G H The discriminator of cyc (G H ,G L )+L lpips The weight value of (L, Lhl, Hlh, k); in, ; ; ; Where, L i For the i A blurred target image, H i For the i A clear target image, Lh i For the generator G H According to the blurred target image L i The generated clear target image, LHl i For the generator G L According to the clear target image Lh i The generated blurred target image, Hl i For the generator G L According to the clear target image H i The generated blurred target image, Hlh i For the generator G H According to the blurred target image Hl i The generated clear target image, F is the sigmoid model, d(L i , LHl i ) is L i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(L i , Hlh i ) is L i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, d(H i , LHl i ) is H i Features extracted by VGG16 network and LHl i The feature distance between features extracted by the VGG16 network, d(H i , Hlh i ) is H i Features extracted by VGG16 network and Hlh i The feature distance between features extracted by the VGG16 network, n is the total number of blurred target images, and m is the total number of clear target images. and is the L2 norm, min is the minimization operator, and max is the maximization operator.

Citation Information

Patent Citations

  • CycleGAN-based image training network structure ArcGAN and method

    CN110222837A

  • System and method for converting image style

    CN111161132A

  • Image motion blur removing method based on improved cyclic generative adversarial network

    CN112634163A

  • Cyclic adversarial neural network unmanned rolling recognition defogging method

    CN113870139A

  • Liver three-dimensional ultrasonic and CT image registration initialization method and device and electronic equipment

    CN114494364A