Method, system, medium, and electronic device for automatic segmentation of blood vessels surrounding organs

Abstract

The present invention provides a method, system, medium, and electronic device for automatically segmenting blood vessels around organs. The method comprises the following steps: obtaining a user's enhanced abdominal CT image and extracting an image around a target organ; constructing multiple angiography-driven perturbation images based on the image around the target organ; constructing perturbation features that simulate false correlations based on the image around the target organ; and training an automatic blood vessel segmentation model based on the image around the target organ, the perturbation images, and the perturbation features, thereby obtaining a blood vessel segmentation result around the target organ based on the trained automatic blood vessel segmentation model. The method, system, medium, and electronic device for automatically segmenting blood vessels around organs constructed by the present invention construct an automatic blood vessel segmentation model based on image-level and feature-level causal intervention schemes, effectively enhancing the generalization performance of the automatic blood vessel segmentation model and improving the reliability of the blood vessel segmentation results.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and in particular to a method, system, medium, and electronic device for automatically segmenting blood vessels surrounding an organ. Background Art

[0002] Pancreatic cancer is a highly lethal malignancy with a poor prognosis, with a five-year survival rate of less than 9%. Currently, surgical resection is the only cure for pancreatic cancer. One of the main reasons for poor postoperative prognosis is cancer recurrence caused by positive tumor resection margins. Therefore, accurate preoperative assessment of surgical resectability is crucial for achieving good survival outcomes. This assessment primarily depends on the extent of tumor involvement with the peripancreatic vasculature. These vessels include the superior mesenteric artery (SMA), the celiac artery (TC), and part of the portal venous system (PPS), which comprises the superior mesenteric vein, portal vein, and splenic vein. However, current preoperative assessments by physicians exhibit considerable interobserver variability, even among experienced physicians. There is a significant clinical need for objective and reliable techniques for analyzing surgical resectability. Accurate automated segmentation of the peripancreatic vessels is a crucial prerequisite.

[0003] Vascular segmentation can help understand the physiological morphology, anatomical structure, and blood flow status of blood vessels, which is of great significance for disease diagnosis and preoperative planning. Deep learning-based methods can adaptively learn multi-dimensional abstract features relevant to the task, and with high accuracy and good performance, have become the mainstream method for automatic vascular segmentation.

[0004] Deep learning-based vascular segmentation methods can be broadly categorized as those based on multi-scale feature learning and those leveraging prior information about vascular geometry. Deep learning methods based on multi-scale feature fusion, such as Yan et al., designed an attention-guided cascade module and a multi-scale feature fusion module to guide the model to capture useful information from features at different scales, achieving more refined vascular segmentation performance. He et al. employed a dense biased connection approach to fuse features from multiple receptive fields and multiple resolutions to address the multi-scale vascular segmentation problem. In recent years, several methods have focused on leveraging prior information about vascular geometry to improve vascular segmentation performance. For example, Qin et al. proposed a spatially aware feature recalibration module and an attention distillation module to enhance the model's representation of tubular structures; while Mou et al. applied self-attention mechanisms to channel and spatial dimensions to develop a segmentation model for curved structures. Other work has also utilized graph neural networks to model vascular structural information to assist in segmentation. These studies attempt to improve vascular segmentation performance through sophisticated network design. Despite demonstrating promising performance, these methods have not explored the generalization capabilities of these models, hindering their clinical application.

[0005] In addition, Dima et al. used 3D U-Net to perform binary segmentation of the peripancreatic arteries. Mahmoudi et al. proposed a hybrid model including 2D Attention U-Net and Texture Attention U-Net for the segmentation of the superior mesenteric vein and superior mesenteric artery. Yao et al. proposed using a self-learning method to segment pancreatic cancer and peripancreatic blood vessels without manual annotation of their private datasets. These methods all achieved good segmentation results. However, the above work is based on a single medical center, and the generalization performance is unknown, that is, the segmentation performance in multiple other centers (hospitals) is unknown. Therefore, the clinical use value of these methods needs further exploration.

[0006] Numerous studies have focused on improving the generalization performance of models. This involves learning a generalizable model from limited data that can be directly deployed on unseen real-world data. Generalized segmentation methods can be broadly categorized into those based on data augmentation and those based on general knowledge learning. Data augmentation-based generalization methods consider variations in image appearance distribution as the primary culprit for declining model generalization performance and aim to simulate the appearance distribution of images likely to be encountered in unseen real-world data through data augmentation. For example, Zhang et al. proposed a deep stacked transformation method that applies a series of transformations to each image during network training to simulate the appearance distribution of images likely to be encountered on unseen real-world data, thereby improving model generalization performance. Lv et al. further utilized the Sinkhorn distance measurement on the unit sphere to increase the diversity of appearance distributions in data augmentation. Zhou et al. proposed that the network preserves image appearance information at a shallow level and increases the diversity of appearance distributions during training by randomly mixing appearance distributions of different data at the feature level. Methods based on general knowledge learning generally train models using data from multiple datasets, improving generalization performance by guiding the model to learn general knowledge independent of the dataset. For example, Dou et al. and Liu et al. divided data from different datasets into meta-training and meta-test sets, using a meta-learning framework to learn common knowledge and improve the model's robustness to data from different sources. Wang et al. constructed a domain knowledge base to store common knowledge across different datasets and then leveraged this knowledge to enhance the general representation of currently unknown data. Since the primary difference between datasets lies in the image appearance distribution, both approaches essentially seek causal-invariant information under different image appearance distributions. While these methods have achieved promising results, those based on common knowledge learning typically require the collection of multiple datasets, significantly increasing the entry level. While methods based on data augmentation do not require a dataset, their augmentation methods are highly random, resulting in images that may not meet real-world requirements and may corrupt information, interfering with model decision-making. Based on causal theory, this is because deep learning models essentially model correlations between input and output, including both causal and spurious correlations. Spurious correlations are interfering factors that have high correlation coefficients but are not causally related to the segmentation labels. It has been observed that the main factor hindering the generalization performance of vessel segmentation is the appearance distribution of the image data. Summary of the Invention

[0007] In view of the shortcomings of the prior art described above, the purpose of the present invention is to provide a method, system, medium, and electronic device for automatic segmentation of blood vessels around organs, which constructs an automatic blood vessel segmentation model based on image-level and feature-level causal intervention schemes, effectively enhancing the generalization performance of the automatic blood vessel segmentation model and improving the reliability of the organ peripheral blood vessel segmentation results.

[0008] In a first aspect, the present invention provides a method for automatic segmentation of blood vessels around organs, the method comprising the following steps: acquiring an enhanced CT image of the abdomen of a user and extracting an image around a target organ; constructing a plurality of angiography-driven perturbation images based on the image around the target organ; constructing perturbation features simulating false correlations based on the image around the target organ; training an automatic blood vessel segmentation model based on the image around the target organ, the perturbation image, and the perturbation features, so as to obtain a blood vessel segmentation result around the target organ based on the trained automatic blood vessel segmentation model.

[0009] In one implementation of the first aspect, extracting an image surrounding the target organ includes the following steps:

[0010] Truncating the grayscale value of the abdominal enhanced CT image to a first preset interval;

[0011] The abdominal enhanced CT image after grayscale truncation is mapped to a second preset interval using a maximum-minimum normalization algorithm;

[0012] Acquiring a target organ region in the mapped abdominal enhanced CT image;

[0013] The image within the circumscribed cube of the target organ region is used as the target organ peripheral image.

[0014] In an implementation of the first aspect, constructing a plurality of contrast-driven disturbance images based on the target organ periphery image includes the following steps:

[0015] Training an image appearance distribution perturbation image generation model, the image appearance distribution perturbation image generation model comprising a first image generator, a second image generator, a first image discriminator, and a second image discriminator; the first image generator is used to generate a migration image based on a source domain image and a migration variable, the second image generator is used to generate the source domain image based on the migration image and the migration variable, the first image discriminator is used to discriminate between the source domain image and the migration image; the second image discriminator is used to discriminate between the migration image and the target domain image;

[0016] The target organ peripheral image in the venous phase is used as the source domain image, and migration images corresponding to different migration variables in the arterial phase and the delayed phase are generated based on the first image generator, and the migration images are used as the perturbation images.

[0017] In an implementation of the first aspect, the loss function of the image appearance distribution perturbation image generation model is a weighted sum of an adversarial loss function and a cycle-consistent loss function; the adversarial loss function is a weighted sum of an adversarial loss function of the first image discriminator and an adversarial loss function of the second image discriminator; and the cycle-consistent loss function is a loss function of the first generator and the second generator.

[0018] In an implementation of the first aspect, constructing a disturbance feature simulating false correlation based on the target organ surrounding image includes the following steps:

[0019] Extracting features of the target organ's surrounding image, and constructing a prototype and a probability map of the target organ's surrounding image based on the features;

[0020] Extracting prototypes of different blood vessels to construct a prototype pool, and extracting random prototypes based on the prototype pool;

[0021] Calculating the difference between the random prototype and the prototype as a characteristic disturbance direction;

[0022] Calculating a characteristic disturbance amount based on the characteristic disturbance direction and the probability map;

[0023] The disturbance feature is constructed based on the feature disturbance amount and the feature.

[0024] In an implementation of the first aspect, training a blood vessel automatic segmentation model based on the target organ peripheral image, the perturbation image library, and the perturbation features includes the following steps:

[0025] Based on the target organ peripheral image, randomly selecting a paired perturbation image from the plurality of perturbation images;

[0026] The target organ surrounding image, the paired perturbation image, and the perturbation feature are input as three tributaries into the automatic blood vessel segmentation model for training; wherein the three tributaries share network weight parameters and use the same label for the blood vessel segmentation results; the target organ surrounding image and the tributaries corresponding to the paired perturbation image are constrained for consistency using a prototype contrast loss function.

[0027] In one implementation of the first aspect, the performance evaluation of the trained automatic blood vessel segmentation model is further included, and the performance evaluation parameters include one or more combinations of Dice similarity coefficient, average surface distance, Hausdorff distance, average skeleton distance, skeleton recall rate, and skeleton precision rate.

[0028] In a second aspect, the present invention provides an automatic segmentation system for blood vessels surrounding organs, the system comprising an acquisition module, a first construction module, a second construction module, and a training module;

[0029] The acquisition module is used to acquire the user's abdominal enhanced CT image and extract the target organ peripheral image;

[0030] The first construction module is used to construct a plurality of contrast-driven disturbance images based on the target organ peripheral image;

[0031] The second construction module is used to construct a disturbance feature simulating false correlation based on the target organ surrounding image;

[0032] The training module is used to train a blood vessel automatic segmentation model based on the target organ surrounding image, the disturbance image and the disturbance feature, so as to obtain a blood vessel segmentation result around the target organ based on the trained blood vessel automatic segmentation model.

[0033] In a third aspect, the present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for automatic segmentation of blood vessels surrounding organs.

[0034] In a fourth aspect, the present invention provides an electronic device, comprising: a processor and a memory;

[0035] The memory is used to store computer programs;

[0036] The processor is used to execute the computer program stored in the memory, so that the electronic device performs the above-mentioned method for automatically segmenting blood vessels around organs.

[0037] As described above, the method, system, medium, and electronic device for automatically segmenting blood vessels surrounding organs of the present invention have the following beneficial effects:

[0038] (1) The image appearance distribution is modeled as a false correlation factor that affects generalization. The causal mechanism of the vascular segmentation task is analyzed, and targeted causal intervention schemes at both the image and feature levels are proposed to suppress the false correlation within the vascular automatic segmentation model and strengthen the causal characteristics of the vascular automatic segmentation model, thereby enhancing the generalization performance of the vascular automatic segmentation model.

[0039] (2) It provides a new paradigm for suppressing non-causal interference in the segmentation framework and achieving stability and generalization of segmentation;

[0040] (3) The dataset required for the automatic vascular segmentation model comes from data collected in clinical practice. The dataset is large and the evaluation performance is excellent, which ensures the effectiveness and reliability of the automatic vascular segmentation model.

[0041] (4) It is particularly suitable for the application scenario of pancreatic peripheral blood vessel segmentation, which helps in the diagnosis and treatment of pancreatic related diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Shown is a flow chart of an embodiment of the method for automatic segmentation of blood vessels surrounding an organ according to the present invention;

[0043] Figure 2 Shown is a schematic structural diagram of an automatic blood vessel segmentation model according to an embodiment of the present invention;

[0044] Figure 3 Shown is a schematic structural diagram of an image appearance distribution perturbation model according to an embodiment of the present invention;

[0045] Figure 4 A schematic diagram showing a false correlation feature simulation disturbance in one embodiment of the present invention;

[0046] Figure 5 Shown is a visualization diagram of the segmentation results of the automatic blood vessel segmentation model of the present invention and other advanced automatic blood vessel segmentation models on an experimental dataset in one embodiment;

[0047] Figure 6 Shown is a visualization diagram of the segmentation results of the automatic blood vessel segmentation model of the present invention and other advanced automatic blood vessel segmentation models on an independent validation set in one embodiment;

[0048] Figure 7 Shown is a schematic structural diagram of an automatic segmentation system for blood vessels surrounding organs according to one embodiment of the present invention;

[0049] Figure 8 FIG. 1 is a schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0051] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0052] The present invention's method, system, medium, and electronic device for automatic segmentation of peripheral blood vessels in organs employs a contrast-driven image appearance distribution perturbation model for abdominal enhanced CT images, simulating, at the image level, the appearance distribution variations caused by contrast factors such as varying contrast agent metabolic rates and different physician acquisition times. Consistency constraints are then used to suppress network encoding of related appearance distribution content, promoting stable feature learning. Furthermore, using vascular prototypes corresponding to different patients as a guide, feature migration is performed on the underlying features of the original image, simulating feature value deviations caused by spurious correlation factors. Feature-level interventions are then constructed, and the use of identical label constraints improves the network's robustness to deviating features, thereby enhancing the generalization performance of the automatic blood vessel segmentation model.

[0053] The technical solutions in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0054] like Figure 1 and Figure 2 As shown, in one embodiment, the method for automatically segmenting blood vessels surrounding an organ of the present invention includes the following steps:

[0055] Step S1: Obtain a user's abdominal enhanced CT image and extract the target organ's peripheral image.

[0056] Specifically, first, an abdominal enhanced CT image of the user is acquired, where the layer thickness of the abdominal enhanced CT image is in the range of 0.8-1.5 mm; then, an image surrounding the target organ in the abdominal enhanced CT image is extracted to automatically segment the blood vessels surrounding the target organ based on the image surrounding the target organ.

[0057] In one embodiment, extracting the target organ peripheral image includes the following steps:

[0058] 11) Truncating the grayscale value of the abdominal enhanced CT image to a first preset interval, such as the interval [-100, 240].

[0059] 12) Using a maximum-minimum normalization algorithm, the abdominal enhanced CT image after grayscale truncation is mapped to a second preset interval, such as the interval [0, 1].

[0060] 13) Acquire the target organ region in the mapped abdominal enhanced CT image using an organ segmentation model or other organ localization methods.

[0061] 14) The image within the circumscribed cube of the target organ region is used as a region of interest, where the region of interest is the image surrounding the target organ.

[0062] Step S2: constructing a plurality of contrast-driven disturbance images based on the target organ peripheral image.

[0063] Specifically, the present invention simulates various image appearance distributions caused by imaging factors such as different contrast agent metabolism rates and different doctor acquisition times at the image level based on the image appearance distribution perturbation image generation model, and constructs imaging-related perturbation images as image-level intervention.

[0064] In one embodiment, constructing a plurality of contrast-driven disturbance images based on the target organ periphery image comprises the following steps:

[0065] 21) Training an image appearance distribution perturbation image generation model, the image appearance distribution perturbation image generation model comprising a first image generator, a second image generator, a first image discriminator and a second image discriminator; the first image generator is used to generate a migration image based on a source domain image and a migration variable, the second image generator is used to generate the source domain image based on the migration image and the migration variable, the first image discriminator is used to discriminate between the source domain image and the migration image; the second image discriminator is used to discriminate between the migration image and the target domain image.

[0066] Among them, the process of enhanced CT imaging going through the arterial phase, venous phase and delayed phase can be modeled as style transfer, that is, while retaining the content of the image (the human anatomical structure remains unchanged), the image appearance distribution is changed according to the laws related to contrast agent metabolism. By taking the abdominal enhanced CT images of the same user at different phases as the starting point (source domain) and end point (target domain) of style transfer to train the generator, the generator can learn the laws of image appearance style changes related to angiography. Based on the laws learned by the generator, synthetic images with continuous intermediate styles between arterial phase images and venous phase images, and between venous phase images and delayed phase images can be generated to simulate the image appearance distribution caused by angiography factors such as different contrast agent metabolism rates or different acquisition times. For example Figure 3 As shown, the image appearance distribution perturbation image generation model includes a first image generator G ST , the second image generator G TS , the first image discriminator D S and the second image discriminator D T . Set X S ,X T Denote the source domain and target domain respectively. The migration variable z∈[0,1] is introduced to characterize the correlation between the appearance distribution of the generated image and the appearance distribution of the source domain image, that is, the correlation between the migration image and the source domain image. When z=0, the migration image presents the appearance distribution of the source domain image; when z=1, the migration image presents the appearance distribution of the target domain image. In the image appearance distribution perturbation image generation model input source domain image x s And the migration variable z, you can get a migration image with an intermediate style

[0067] The first image discriminator D S and the second image discriminator D T Used to distinguish and x S , and the target domain image x T , the corresponding adversarial losses are:

[0068]

[0069]

[0070] in, represents the mathematical expectation; x S with x T denote images from the source domain and target domain respectively; P(·) denotes the probability distribution, x S ~P(X s ) represents the source domain image x S Obey the probability distribution P(X s ), x T ~P(X T ) represents the source domain image x T Obey the probability distribution P(x T ). Therefore, the adversarial loss function of the image appearance distribution perturbation model is the weighted sum of the adversarial loss function of the first image discriminator and the adversarial loss function of the second image discriminator, that is, At the same time, the image appearance distribution perturbation image generation model applies a cycle-consistent loss function L cyc To ensure that the semantic content of the image is well preserved in the migrated image, that is, G TS G ST Generated Restore to x S ,Right now Wherein, ||·||1 represents the L1 norm. The cycle-consistent loss function is a loss function for the first generator and the second generator.

[0071] In summary, the loss function of the image appearance distribution perturbation model is the weighted sum L of the adversarial loss function and the cycle-consistent loss function. gen =L adv +λL cyc , where λ is a hyperparameter that balances the two losses.

[0072] 22) Using the target organ peripheral image in the venous phase as the source domain image, based on the first image generator, generate migration images corresponding to different migration variables that migrate to the arterial phase and delayed phase, and use the migration images as the perturbation images. Among them, using the trained image appearance distribution perturbation model, taking the venous phase abdominal enhanced CT image as the migration starting point, the original data is style-transferred in the direction of the arterial phase and delayed phase to generate migration images with different contrast enhancement effects and store them in the perturbation image library, where the migration variable z takes values ​​at a fixed interval γ in the range of [0,θ]. θ is the upper bound of the migration degree, which limits the interference image to the enhancement effect of the venous phase to prevent interference with model training. For simple representation, the perturbation image library is composed of X z ={x z |z∈[-θ,θ]} represents. When z∈[-θ,0], x z Represents the image generated by migration to the arterial phase; when z∈[0,θ], x z Represents the image generated by migrating to the delay period.

[0073] Step S3: constructing disturbance features simulating false correlation based on the target organ surrounding image.

[0074] Specifically, in addition to contrast factors, there are other potential factors that affect the image appearance distribution. At the same time, there is no guarantee that the influence of contrast factors on features can be completely removed through causal intervention at the image level. Figure 4 As shown, the model is further guided by simulating perturbations of false correlated features to suppress feature deviations caused by image appearance distribution information, and robust decision making is achieved through feature-level intervention.

[0075] In one embodiment, constructing a disturbance feature simulating false correlation based on the target organ surrounding image comprises the following steps:

[0076] 31) Extracting features of the target organ peripheral image, and constructing a prototype and a probability map of the target organ peripheral image based on the features.

[0077] The target organ image is input into a weight-sharing encoder, and then the prototype and probability map of the target organ surrounding image are constructed based on the features output by the encoder. For the underlying features of the target organ surrounding image, the prototype of the target organ surrounding image is calculated using a masked average pooling operation. The subscript c is the category, is the category set, f up The underlying features are upsampled to restore the input size, y is the segmentation label, N is the total prime number; I[·] is an indicator function, if the parameter is true, the output value is 1, otherwise the output is 0. Probability map P coarse =softmax(W θf), where W θ is the weight matrix of point-by-point convolution, and f is the underlying feature.

[0078] 32) Extracting prototypes of different blood vessels to construct a prototype pool, and extracting random prototypes based on the prototype pool.

[0079] A prototype pool is pre-built, and the prototype pool contains prototypes of different blood vessels. Preferably, the prototypes obtained in step 31) are stored in the prototype pool constructed based on the queue for subsequent use in false correlation feature simulation perturbation. In addition, prototypes of other data are randomly extracted from the prototype pool.

[0080] 33) Calculate the difference between the random prototype and the prototype as the characteristic disturbance direction.

[0081] Among them, the characteristic perturbation direction

[0082] 34) Calculate the characteristic disturbance amount based on the characteristic disturbance direction and the probability map.

[0083] Among them, the characteristic disturbance Where C, D, H, and W represent the channel, depth, height, and width of the underlying features, respectively.

[0084] 35) Constructing the disturbance feature based on the characteristic disturbance amount and the feature.

[0085] Among them, a disturbance of random degree λ is applied to the feature to simulate the feature deviation caused by different false correlation factors, thereby obtaining the disturbance feature, that is, Here, f′ is the perturbation feature, and λ is the perturbation level, which is randomly sampled from [0, α], where α is the upper limit of the perturbation level. During subsequent model training, to stabilize training, the perturbation level is gradually increased, λ = 0.05 + current training round / total training rounds, λ∈[0,1].

[0086] Step S4: training a blood vessel automatic segmentation model based on the target organ surrounding image, the perturbation image library, and the perturbation features, so as to obtain a blood vessel segmentation result around the target organ based on the trained blood vessel automatic segmentation model.

[0087] Specifically, training a blood vessel automatic segmentation model based on the target organ surrounding image, the perturbation image library, and the perturbation features includes the following steps:

[0088] 41) Based on the target organ peripheral image, randomly select a paired perturbation image from the multiple perturbation images.

[0089] 42) The target organ peripheral image, the paired perturbation image and the perturbation feature are input as three tributaries into the automatic blood vessel segmentation model for training; wherein the three tributaries share network weight parameters and use the same label for the blood vessel segmentation results; the target organ peripheral image and the tributaries corresponding to the paired perturbation image are constrained for consistency using a prototype contrast loss function.

[0090] In one embodiment, the automatic blood vessel segmentation model includes three tributaries, each of which includes an encoder and a decoder. The target organ surrounding image and the paired perturbation image are input into the encoder input of the first tributary and the second tributary respectively; the features output by the encoders of the first tributary and the second tributary are subjected to a prototype contrast loss function to achieve consistency constraints; guided by the information of the target organ surrounding image of other data, the features output by the encoder of the third tributary are perturbed, and the perturbed features are input into the corresponding decoder, and implicit consistency constraints are imposed by sharing decoder weights and using the same segmentation labels. It should be noted that, in fact, since the three encoders and decoders share weights, the encoders and decoders of the three tributaries are actually one encoder and decoder. After training, the automatic blood vessel segmentation model only needs to leave the encoder and decoder to complete the automatic segmentation of the blood vessels around the target organ.

[0091] Among them, according to the input target organ surrounding image x, from the perturbation database X z Randomly sample paired perturbation images x with migration degree z z . Replace x and x z Input the two streams of the network with shared weights, extract prototypes from the underlying features of the two streams respectively, and use the prototype comparison loss function to impose consistency constraints on the feature level. The prototype comparison consistency loss function is in, is the prototype of the same category (positive sample), is the set of corresponding negative samples, i.e., the prototypes of other categories. '·' represents the dot product, and τ>0 is the temperature factor.

[0092] In addition, the segmentation results of the target organ surrounding image and the perturbation image are constrained by the same label. Segmentation loss function By Cross-Entropy loss (CE) cross entropy loss and DSC loss The specific formula is as follows:

[0093]

[0094]

[0095] LSEG =L CE +L DSC

[0096] Among them, y and Represent the label and blood vessel segmentation results respectively, N is the total prime number, y i and They represent the i-th voxel respectively.

[0097] In addition, the probability map P coarse After upsampling to restore the input size, DSC loss is used As an auxiliary loss function To constrain.

[0098] Simultaneously, a prototype of other data is randomly extracted from the prototype pool. The difference between the random prototype and the prototype is used as the feature perturbation direction to randomly perturb the feature to obtain the perturbed feature f′. This perturbed feature f′ is input into the decoder for subsequent prediction. By sharing decoder weights and imposing implicit consistency constraints using the same segmentation labels, the automatic vessel segmentation model is more robust to feature deviations introduced by various spurious correlation factors, ultimately improving the stability of the model's decision making.

[0099] Therefore, by perturbing both the image and the features, we construct causal interventions at two levels (image and feature), systematically suppressing spurious correlations within the model and strengthening the model's causal characteristics, thereby enhancing the model's generalization performance. Specifically, image-level causal intervention strengthens the model's ability to capture key information when the image appearance distribution changes, thereby improving the stability of the model's features. Conversely, feature-level causal intervention enhances the model's robustness to feature deviations, ultimately improving the stability of the model's decisions.

[0100] During the training of the automatic vessel segmentation model, a five-fold cross-validation strategy was used, dividing the data into five fixed independent folds. Four of the five folds were used for training, and the remaining fold was used for testing. A 3DU-Net variant was used as the segmentation backbone network, in which two-layer convolutional residual blocks were used to replace the original convolutional layers, and group normalization was used instead of batch normalization. The training batch size was 1, the ADAM optimizer was used, and the initial learning rate was 1×10 -4 , and as the training progresses, the i+1 =lr i ×((1-current training round / total training rounds) γ) with an exponential decrease, where γ is set to 0.9. During training, image patches of size (64, 120, 120) are randomly sampled from the region of interest as training images. Furthermore, data augmentation techniques are introduced throughout the training process, including random rotation (between -15° and 15°), random mirroring, random stretching, and the addition of random Gaussian noise. During network training, the error between the segmentation results and the labels is calculated based on the loss function, and the network parameters are adjusted using the directional propagation algorithm.

[0101] After the blood vessel automatic segmentation model is trained, the abdominal enhanced CT image to be processed is input into the blood vessel automatic segmentation model, and the blood vessel segmentation result can be obtained through the encoder and decoder.

[0102] In one embodiment, the method for automatic segmentation of blood vessels surrounding an organ of the present invention further includes performing a performance evaluation on the trained automatic blood vessel segmentation model, wherein the performance evaluation parameters include one or more combinations of the following: Dice similarity coefficient (DSC), average surface distance (ASD), Hausdorff distance (HD), mean skeleton distance (MSD), skeleton recall (SR), and skeleton precision (SP). Specifically defined as follows:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] where y and Represent the label and blood vessel segmentation results respectively; a∈y represents the pixel a in the label y, represents pixel b in the vessel segmentation result y; δ(·) represents the skeletonization operation. Specifically, we first use the skeletonize_3d function in the skimage.morphology library to obtain a preliminary skeleton, which may be disconnected in the thinner parts of the vessel. Then, we use the (3,3,3) core to dilate the obtained skeleton to maintain the connectivity of the vessel skeleton.

[0110] Compared with existing studies on automatic blood vessel segmentation, the superiority of the present invention's method for automatic segmentation of blood vessels around organs can be demonstrated through experimental results on experimental datasets. The specific analysis of the experimental results is as follows:

[0111] (1) Segmentation result analysis

[0112] In order to fully evaluate the performance of the automatic vascular segmentation model of the present invention, ten repeated five-fold cross-validation experiments were conducted on the experimental dataset. The experimental results are shown in Table 1, including DSC (Dice similarity), HD (Hausdorff distance), ASD (average surface distance), MSD (average skeleton distance), SR (skeleton recall rate), and SP (skeleton precision rate). The automatic vascular segmentation model showed accurate segmentation performance in the ten repeated experiments. The average DSC, HD, ASD, MSD, SR, and SP of the three types of blood vessels, namely the superior mesenteric artery (SMA), the celiac trunk artery (TC), and the partial portal vein system (PPS) composed of the superior mesenteric vein, portal vein, and splenic vein, were 81.77%, 15.57 mm, 1.11 mm, 1.62 mm, 90.84%, and 91.38%, respectively. The numerical fluctuations of each indicator were small, indicating that the automatic vascular segmentation model has stable segmentation performance.

[0113] Table 1. Automatic segmentation performance of the pancreatic blood vessels of the automatic blood vessel segmentation model of the present invention on the experimental dataset

[0114]

[0115] (2) Ablation experiment analysis

[0116] To verify the necessity and superiority of all strategies, we conducted ablation experiments on the automatic blood vessel segmentation model of our invention. The ablation test results are shown in Table 2, where BL represents the baseline model, IPM represents the perturbed image operation, FPM represents the perturbed feature operation, and ours represents the automatic blood vessel segmentation model of our invention, namely BL+IPM+FPM. Compared to the baseline model, adding perturbed images and corresponding consistency constraints during training improves the average DSC, SR, and SP of the three blood vessel types by 1.57%, 2.35%, 1.48%, 8.81% (1.77mm), 13.29% (0.19mm), and 13.21% (0.28mm), respectively. Furthermore, the perturbed features and corresponding consistency constraints significantly improve the average DSC, SR, and SP of the three blood vessel types by 1.88%, 2.39%, 2.01%, 14.29% (2.87mm), 17.48% (0.25mm), and 16.98% (0.36mm), respectively. Furthermore, while the SR of the final PPS model is 0.53% lower than that of the BL+PPM model, its SP is 0.89% higher. The SR of the TC model is 0.34% lower than that of the BL+IPM model, but its SP is 3.06% higher. Therefore, in summary, the average values ​​of the three types of vascular indicators of the vascular automatic segmentation model equipped with two levels of perturbation-related operations during training reach the optimal level, indicating that the two levels of perturbation and consistency constraints can be well compatible and contribute consistent performance improvements.

[0117] Table 2. Model ablation results of the automatic blood vessel segmentation model of the present invention on the experimental dataset

[0118]

[0119] (3) Generalization performance analysis

[0120] To verify the generalization performance of the proposed automatic vascular segmentation model and the effectiveness of all strategies for generalization tasks, tests were conducted on three independent test sets from medical disciplines. The results are shown in Tables 3 and 4. Equipped with two layers of perturbation-related operations, the automatic vascular segmentation model achieved improved segmentation performance for all three types of vessels on the three independent test sets. Furthermore, the combined model achieved average DSC improvements of 2.35%, 3.47%, and 4.9%, and average MSD improvements of 27.75% (0.53 mm), 30.09% (0.65 mm), and 33.47% (0.84 mm), respectively, on the three independent test sets. These experimental results demonstrate that the proposed two modules can effectively improve the generalization performance of the automatic vascular segmentation model on unseen datasets, demonstrating the stability and generalization of the proposed causal-invariant driven intervention scheme for vascular segmentation.

[0121] Table 3. Results of the multi-center generalized ablation experiment of the automatic blood vessel segmentation model of the present invention (DSC, %)

[0122]

[0123] Table 4. Results of the multi-center generalized ablation experiment of the automatic blood vessel segmentation model of the present invention (MSD, %)

[0124]

[0125] (4) Comparison with advanced vessel segmentation algorithms

[0126] The present invention is compared with existing advanced blood vessel segmentation algorithms, including 3D Unet, 3DResUNet, Vnet, TSNet, CS 2 Net, LVSNet. The results of the experimental dataset are shown in Table 5. The 3D segmentation results are visualized as follows Figure 5 As shown. Then, the generalization performance of the vessel segmentation model is compared. Based on the independent validation set, the proposed method is further compared with two state-of-the-art generalization methods, BigAug and MixStyle. The results are shown in Tables 6 and 7, and the visualization results are shown in Figure 6 As shown in the figure, it should be noted that since these two networks were not specifically developed for vascular segmentation, for a fair comparison, the method of the aforementioned article was replicated based on the baseline network of the present invention. In the experimental dataset and independent validation set, the vascular segmentation method of the present invention achieved optimal performance in all indicators for all blood vessels. The visualization results intuitively demonstrate that the vascular segmentation method of the present invention can more accurately segment vascular regions. Both quantitative and qualitative results show that the vascular segmentation method of the present invention achieves optimal performance in peripancreatic vascular segmentation.

[0127] Table 5. Comparison of the automatic blood vessel segmentation model of the present invention with advanced blood vessel segmentation algorithms on the experimental dataset

[0128]

[0129]

[0130] Table 6. Comparison of the automatic blood vessel segmentation model of the present invention with advanced blood vessel segmentation algorithms on an independent validation set (DSC, %)

[0131]

[0132] Table 7. Comparison of the automatic blood vessel segmentation model of the present invention with advanced blood vessel segmentation algorithms on an independent validation set (MSD, mm)

[0133]

[0134]

[0135] The protection scope of the method for automatic segmentation of blood vessels around organs described in the embodiment of the present invention is not limited to the execution order of the steps listed in this embodiment. All solutions implemented by adding, subtracting, or replacing steps in the prior art based on the principles of the present invention are included in the protection scope of the present invention.

[0136] An embodiment of the present invention further provides an automatic segmentation system for blood vessels surrounding organs. The automatic segmentation system for blood vessels surrounding organs can implement the automatic segmentation method for blood vessels surrounding organs described in the present invention. However, the implementation device of the automatic segmentation system for blood vessels surrounding organs described in the present invention includes but is not limited to the structure of the automatic segmentation system for blood vessels surrounding organs listed in this embodiment. Any structural deformation and replacement of the existing technology made according to the principles of the present invention are included in the protection scope of the present invention.

[0137] like Figure 7 As shown, in one embodiment, the automatic segmentation system for blood vessels surrounding organs of the present invention includes an acquisition module 71 , a first construction module 72 , a second construction module 73 and a training module 74 .

[0138] The acquisition module 71 is used to acquire the user's abdominal enhanced CT image and extract the target organ peripheral image.

[0139] The first construction module 72 is connected to the acquisition module 71 and is used to construct a plurality of angiography-driven disturbance images based on the target organ peripheral image.

[0140] The second construction module 73 is connected to the acquisition module 71 and is used to construct disturbance features simulating false correlation based on the image around the target organ.

[0141] The training module 74 is connected to the first construction module 72 and the second construction module 73, and is used to train the automatic blood vessel segmentation model based on the target organ surrounding image, the disturbance image and the disturbance feature, so as to obtain the blood vessel segmentation result around the target organ based on the trained automatic blood vessel segmentation model.

[0142] The structures and principles of the acquisition module 71 , the first construction module 72 , the second construction module 73 and the training module 74 correspond one-to-one to the steps in the above-mentioned method for automatic segmentation of blood vessels around organs, and therefore will not be described in detail here.

[0143] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices or methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of modules / units is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or units can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules or units, which can be electrical, mechanical or other forms.

[0144] Modules / units described as separate components may or may not be physically separate, and components displayed as modules / units may or may not be physical modules, that is, they may be located in one place or distributed across multiple network elements. Some or all of the modules / units may be selected based on actual needs to achieve the objectives of the embodiments of the present invention. For example, the functional modules / units in various embodiments of the present invention may be integrated into a single processing module, each module / unit may exist physically separately, or two or more modules / units may be integrated into a single module / unit.

[0145] Those skilled in the art should further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0146] The present invention also provides a computer-readable storage medium. A person skilled in the art will appreciate that all or part of the steps in the method for automatically segmenting blood vessels surrounding an organ in the above-described embodiment can be performed by instructing a processor through a program. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as a random access memory, a read-only memory, a flash memory, a hard disk, a solid-state drive, a magnetic tape, a floppy disk, an optical disc, and any combination thereof. The storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a digital video disc (DVD)), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0147] An embodiment of the present invention further provides an electronic device comprising a processor and a memory.

[0148] The memory is used to store computer programs.

[0149] The memory includes various media that can store program codes, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0150] The processor is connected to the memory and is used to execute the computer program stored in the memory, so that the electronic device executes the above-mentioned method for automatically segmenting blood vessels around organs.

[0151] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0152] like Figure 8As shown, the electronic device of the present invention is in the form of a general-purpose computing device. Components of the electronic device may include, but are not limited to: one or more processors or processing units 81, a memory 82, and a bus 83 connecting different system components (including the memory 82 and the processing unit 81).

[0153] Bus 83 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures. Examples of these architectures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MAC) bus, an Enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

[0154] Electronic devices typically include a variety of computer system readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, removable and non-removable media.

[0155] The memory 82 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 821 and / or cache memory 822. The electronic device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system 823 may be used to read and write non-removable, non-volatile magnetic media ( Figure 8 Not shown, often called a "hard drive"). Although Figure 8 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk"), and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 83 via one or more data medium interfaces. Memory 82 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of various embodiments of the present invention.

[0156] A program / utility 824 having a set (at least one) of program modules 8241 may be stored, for example, in memory 82. Such program modules 8241 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data, each of which, or some combination thereof, may include an implementation of a network environment. Program modules 8241 generally implement the functions and / or methods of the embodiments described herein.

[0157] The electronic device may also communicate with one or more external devices (e.g., keyboards, pointing devices, displays, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (e.g., network cards, modems, etc.). Such communication may be performed via input / output (I / O) interface 84. Furthermore, the electronic device may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 85. Figure 8 As shown, the network adapter 85 communicates with other modules of the electronic device via the bus 83. It should be understood that, although not shown in the figures, other hardware and / or software modules may be used in conjunction with the electronic device, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0158] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. A method for automatically segmenting blood vessels around an organ, characterized by: The method comprises the following steps: Obtain the user's abdominal enhanced CT image and extract the surrounding image of the target organ; constructing a plurality of contrast-driven disturbance images based on the target organ periphery image; Constructing a disturbance feature simulating false correlation based on the peripheral image of the target organ; Training a blood vessel automatic segmentation model based on the target organ surrounding image, the disturbance image, and the disturbance feature, so as to obtain a blood vessel segmentation result around the target organ based on the trained blood vessel automatic segmentation model; Constructing a plurality of contrast-driven disturbance images based on the target organ peripheral image comprises the following steps: Training an image appearance distribution perturbation image generation model, the image appearance distribution perturbation image generation model comprising a first image generator, a second image generator, a first image discriminator, and a second image discriminator; the first image generator is used to generate a migration image based on a source domain image and a migration variable, the second image generator is used to generate the source domain image based on the migration image and the migration variable, the first image discriminator is used to discriminate between the source domain image and the migration image; the second image discriminator is used to discriminate between the migration image and the target domain image; Using the target organ periphery image in the venous phase as the source domain image, generating migration images corresponding to different migration variables in the arterial phase and the delayed phase based on the first image generator, and using the migration images as the perturbation images; Constructing a disturbance feature simulating false correlation based on the target organ surrounding image comprises the following steps: Extracting features of the target organ's surrounding image, and constructing a prototype and a probability map of the target organ's surrounding image based on the features; Extracting prototypes of different blood vessels to construct a prototype pool, and extracting random prototypes based on the prototype pool; Calculating the difference between the random prototype and the prototype as a characteristic disturbance direction; Calculating a characteristic disturbance amount based on the characteristic disturbance direction and the probability map; Constructing the disturbance feature based on the characteristic disturbance amount and the feature; Training a blood vessel automatic segmentation model based on the target organ peripheral image, the perturbation image and the perturbation feature comprises the following steps: Based on the target organ peripheral image, randomly selecting a paired perturbation image from the plurality of perturbation images; The target organ surrounding image, the paired perturbation image, and the perturbation feature are input as three tributaries into the automatic blood vessel segmentation model for training; wherein the three tributaries share network weight parameters and use the same label for the blood vessel segmentation results; the target organ surrounding image and the tributaries corresponding to the paired perturbation image are constrained for consistency using a prototype contrast loss function.

2. The method for automatic segmentation of blood vessels surrounding an organ according to claim 1, wherein: Extracting the surrounding image of the target organ includes the following steps: Truncating the grayscale value of the abdominal enhanced CT image to a first preset interval; The abdominal enhanced CT image after grayscale truncation is mapped to a second preset interval using a maximum-minimum normalization algorithm; Acquiring a target organ region in the mapped abdominal enhanced CT image; The image within the circumscribed cube of the target organ region is used as the target organ peripheral image.

3. The method for automatic segmentation of blood vessels surrounding an organ according to claim 1, wherein: The loss function of the image appearance distribution perturbation image generation model is the weighted sum of the adversarial loss function and the cycle-consistent loss function; the adversarial loss function is the weighted sum of the adversarial loss function of the first image discriminator and the adversarial loss function of the second image discriminator; the cycle-consistent loss function is the loss function of the first image generator and the second image generator.

4. The method for automatic segmentation of blood vessels surrounding an organ according to claim 1, wherein: It also includes performance evaluation of the trained blood vessel automatic segmentation model, and the performance evaluation parameters include one or more combinations of Dice similarity coefficient, average surface distance, Hausdorff distance, average skeleton distance, skeleton recall rate, and skeleton precision rate.

5. An automatic segmentation system for blood vessels surrounding organs, characterized by: The system includes an acquisition module, a first construction module, a second construction module and a training module; The acquisition module is used to acquire the user's abdominal enhanced CT image and extract the target organ peripheral image; The first construction module is used to construct a plurality of contrast-driven disturbance images based on the target organ peripheral image; The second construction module is used to construct a disturbance feature simulating false correlation based on the target organ surrounding image; The training module is used to train a blood vessel automatic segmentation model based on the target organ surrounding image, the disturbance image and the disturbance feature, so as to obtain a blood vessel segmentation result around the target organ based on the trained blood vessel automatic segmentation model; Constructing a plurality of contrast-driven disturbance images based on the target organ peripheral image comprises the following steps: Training an image appearance distribution perturbation image generation model, the image appearance distribution perturbation image generation model comprising a first image generator, a second image generator, a first image discriminator, and a second image discriminator; the first image generator is used to generate a migration image based on a source domain image and a migration variable, the second image generator is used to generate the source domain image based on the migration image and the migration variable, the first image discriminator is used to discriminate between the source domain image and the migration image; the second image discriminator is used to discriminate between the migration image and the target domain image; Using the target organ periphery image in the venous phase as the source domain image, generating migration images corresponding to different migration variables in the arterial phase and the delayed phase based on the first image generator, and using the migration images as the perturbation images; Constructing a disturbance feature simulating false correlation based on the target organ surrounding image comprises the following steps: Extracting features of the target organ's surrounding image, and constructing a prototype and a probability map of the target organ's surrounding image based on the features; Extracting prototypes of different blood vessels to construct a prototype pool, and extracting random prototypes based on the prototype pool; Calculating the difference between the random prototype and the prototype as a characteristic disturbance direction; Calculating a characteristic disturbance amount based on the characteristic disturbance direction and the probability map; Constructing the disturbance feature based on the characteristic disturbance amount and the feature; Training a blood vessel automatic segmentation model based on the target organ peripheral image, the perturbation image and the perturbation feature comprises the following steps: Based on the target organ peripheral image, randomly selecting a paired perturbation image from the plurality of perturbation images; The target organ surrounding image, the paired perturbation image, and the perturbation feature are input as three tributaries into the automatic blood vessel segmentation model for training; wherein the three tributaries share network weight parameters and use the same label for the blood vessel segmentation results; the target organ surrounding image and the tributaries corresponding to the paired perturbation image are constrained for consistency using a prototype contrast loss function.

6. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for automatically segmenting blood vessels surrounding an organ according to any one of claims 1 to 4 is implemented.

7. An electronic device, characterized in that: include: processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, so as to enable the electronic device to perform the method for automatic segmentation of blood vessels surrounding an organ according to any one of claims 1 to 4.

Images