Automatic blood vessel segmentation method and system for CT (Computed Tomography) image

Abstract

The invention relates to an automatic blood vessel segmentation method and system for a CT image. The method comprises the following steps: acquiring an enhanced CT image and a plain-scan CT image at the same position; inputting the enhanced CT image to a pre-trained first blood vessel automatic segmentation model to obtain a first blood vessel segmentation result; and registering the enhanced CT image and the plain-scan CT image to mark and map the first blood vessel segmentation result to the plain-scan CT image so as to obtain a second blood vessel segmentation result. The method has the advantages that a deep learning model (such as a registration network based on U-Net or Transform) is utilized to calculate a deformation field, and in combination with multi-scale feature enhancement and regularization strategies, the precision and stability of enhanced CT and plain-scan CT image registration are improved, and artifacts and local distortion in the deformation field are reduced; in order to solve the problem that the contrast difference of a plain scanning CT image and an enhanced CT image is significant, a compensation strategy of multi-modal texture and intensity distribution is introduced, such as adversarial loss and structural similarity index (SSIM) optimization, so that registration is more robust among different modals.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method, system, computer equipment and computer-readable storage medium for automatic blood vessel segmentation in CT images. Background Art

[0002] With the rapid development of medical imaging technology, more and more doctors are using it to diagnose patients' conditions and propose treatment plans. Common medical imaging technologies include computed tomography (CT) and magnetic resonance imaging (MRI). Medical image processing is a key technology that helps doctors understand patients' conditions. Imaging technology can help doctors more intuitively understand the patient's internal pathology and make accurate and effective diagnoses.

[0003] Medical image segmentation, which separates diseased areas, organs at risk, and radiotherapy targets from raw medical images, is a crucial component of medical image processing and plays a vital role in enabling doctors to accurately understand patients' conditions, make diagnoses, and prescribe treatment plans. Currently, medical image segmentation is primarily performed manually by experienced physicians. This is not only time-consuming and laborious, but its accuracy is also heavily dependent on the physician's experience and skill. Therefore, research on efficient and accurate automatic medical image segmentation methods can significantly improve diagnostic efficiency and accuracy.

[0004] Existing medical image segmentation technologies mainly focus on manual segmentation and semi-automatic segmentation. In the task of vascular segmentation, professional physicians are usually required to manually outline the vascular contours through enhanced CT images to obtain accurate annotations. Although this method is accurate, it is time-consuming and labor-intensive, especially for complex vascular structures. In addition, in recent years, deep learning has made breakthrough progress in the field of medical image target recognition and segmentation, and relevant scholars have also proposed a variety of image segmentation models based on deep learning. However, the target areas in medical images often have the characteristics of low contrast, blurred boundaries, and insignificant features of small targets. Existing network models are usually difficult to handle such complexity and diversity, and the segmentation accuracy is often low. In particular, it is difficult to accurately and effectively identify and segment targets with blurred boundaries and small sizes.

[0005] Currently, existing network models in related technologies are usually difficult to deal with such complexity and diversity, and the segmentation accuracy is often low. In particular, for segmentation targets with blurred boundaries and small sizes, it is difficult to accurately and effectively identify and segment them. No effective solutions have yet been proposed. Summary of the Invention

[0006] The purpose of this application is to address the deficiencies in the prior art and to provide a method and system for automatic blood vessel segmentation in CT images, so as to at least solve the problems in the related art, such as the existing network models are usually difficult to be so complex and diverse, the segmentation accuracy is often low, and it is difficult to accurately and effectively identify and segment targets with blurred boundaries and small sizes.

[0007] To achieve the above objectives, the technical solutions adopted in this application are:

[0008] In a first aspect, the present invention provides a method for automatic segmentation of blood vessels in CT images, comprising:

[0009] Acquire enhanced CT images and plain CT images of the same location;

[0010] Inputting the enhanced CT image into the pre-trained first vessel automatic segmentation model to obtain the first vessel segmentation result;

[0011] The enhanced CT image and the plain CT image are registered to map the first blood vessel segmentation result annotation to the plain CT image to obtain a second blood vessel segmentation result.

[0012] In some embodiments, registering the enhanced CT image and the plain CT image to map the first blood vessel segmentation result annotation to the plain CT image to obtain the second blood vessel segmentation result includes:

[0013] Standardizing the enhanced CT images and the plain CT images to obtain standard enhanced CT images and standard plain CT images;

[0014] The deformation field is generated by nonlinearly registering standard enhanced CT images with standard plain CT images using a deep learning registration network (VoxelMorph).

[0015] The first blood vessel segmentation result is mapped to the standard plain scan CT image based on the deformation field to obtain the second blood vessel segmentation result.

[0016] In some embodiments, standardizing the enhanced CT image and the plain CT image to obtain a standard enhanced CT image and a standard plain CT image includes:

[0017] Normalizing the enhanced CT image and the plain CT image to obtain a first sub-enhanced CT image and a first sub-plain CT image;

[0018] resampling the first sub-enhanced CT image and the first sub-plain CT image to obtain a second sub-enhanced CT image and a second sub-plain CT image;

[0019] The second sub-enhanced CT image and the second sub-plain CT image are cropped to obtain a third sub-enhanced CT image and a third sub-plain CT image;

[0020] The third sub-enhanced CT image and the third sub-plain CT image are subjected to histogram equalization processing to obtain a standard enhanced CT image and a standard plain CT image.

[0021] In some of these embodiments,

[0022] Mapping the first blood vessel segmentation result to a standard plain scan CT image based on the deformation field to obtain a second blood vessel segmentation result includes:

[0023] Based on the spatial transformation of the deformation field, the first vessel segmentation result is mapped to the standard plain scan CT image through bilinear interpolation to obtain the mapping result;

[0024] Performing morphological processing on the mapping result to obtain a second blood vessel segmentation result;

[0025] The second vessel segmentation results are self-supervised and optimized based on the mapping results and standard enhanced CT images.

[0026] In some of these embodiments,

[0027] The self-supervised optimization of the second vessel segmentation results based on the mapping results and the standard enhanced CT images includes:

[0028] Mapping the first blood vessel segmentation result to the spatial position of the standard plain scan CT image to obtain the initial pseudo-labeling;

[0029] Inputting the initial pseudo-annotation into the first blood vessel automatic segmentation model to train and obtain the second blood vessel automatic segmentation model;

[0030] Input the standard plain scan CT image into the second blood vessel automatic segmentation model and obtain the second blood vessel segmentation result;

[0031] According to the second blood vessel segmentation result, a high-confidence area is selected as the optimized pseudo-annotation to obtain a reference pseudo-annotation;

[0032] Based on the reference pseudo-annotations as training data, the second vessel automatic segmentation model is gradually improved to achieve self-supervised optimization.

[0033] In some embodiments, the training method of the first blood vessel automatic segmentation model includes:

[0034] Acquire a first training image and first labeled data;

[0035] Preprocessing the first training image and the first labeled data to obtain a second training image and second labeled data;

[0036] The second training image and the second labeled data are input into the nnU-Net automatic segmentation network to obtain a first blood vessel automatic segmentation model.

[0037] In some embodiments, the training method of the first blood vessel automatic segmentation model further includes:

[0038] Obtain the file format and target resolution of the second training image;

[0039] Based on the file format and target resolution, select the applicable nnU-Net automatic segmentation network.

[0040] In some embodiments, the training method of the first blood vessel automatic segmentation model further includes:

[0041] Preprocessing the first training image to obtain the second training image includes:

[0042] performing normalization processing on the first training image to obtain a first intermediate training image;

[0043] performing cropping processing on the first intermediate training image to obtain a second intermediate training image;

[0044] Perform histogram equalization processing on the second intermediate training image to obtain a second training image.

[0045] In a second aspect, the present invention provides a system for automatic segmentation of blood vessels for CT images, comprising:

[0046] a first acquisition module, configured to acquire an enhanced CT image and a plain scan CT image at the same position;

[0047] an automatic segmentation module, the automatic segmentation module being used to input the enhanced CT image into a pre-trained first blood vessel automatic segmentation model to obtain a first blood vessel segmentation result;

[0048] A registration and mapping module is used to register the enhanced CT image and the plain CT image to map the first blood vessel segmentation result annotation to the plain CT image to obtain a second blood vessel segmentation result.

[0049] In some embodiments, the registration mapping module includes:

[0050] An image processing submodule, wherein the image processing submodule is used to perform standardization processing on the enhanced CT image and the plain scan CT image to obtain a standard enhanced CT image and a standard plain scan CT image;

[0051] A registration submodule, wherein the registration submodule is used to perform nonlinear registration of the standard enhanced CT image and the standard plain scan CT image through a deep learning registration network (VoxelMorph) to generate a deformation field;

[0052] A mapping submodule is used to map the first blood vessel segmentation result to a standard plain scan CT image based on the deformation field to obtain a second blood vessel segmentation result.

[0053] In some embodiments, the automatic blood vessel segmentation system further comprises:

[0054] a second acquisition module, the second acquisition module being used to acquire the first training image and the first annotation data;

[0055] a preprocessing module, configured to preprocess the first training image and the first annotated data to obtain a second training image and second annotated data;

[0056] A training module is used to input the second training image and the second labeled data into the nn-Unet automatic segmentation network to obtain a first blood vessel automatic segmentation model.

[0057] In some embodiments, further comprising:

[0058] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-described automatic blood vessel segmentation method when executing the computer program.

[0059] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned automatic blood vessel segmentation method.

[0060] Compared with related technologies, the embodiment of the present application provides a method and system for automatic segmentation of blood vessels in CT images. It uses a deep learning model (such as a registration network based on U-Net or Transformer) to calculate the deformation field, combines multi-scale feature enhancement and regularization strategies to improve the accuracy and stability of the registration of enhanced CT and plain CT images, and reduces artifacts and local distortions in the deformation field. To address the problem of significant contrast differences between plain and enhanced CT images, it introduces compensation strategies for multimodal texture and intensity distribution, such as adversarial loss and structural similarity index (SSIM) optimization, to make the registration more robust between different modalities. It uses the mapped annotations as the initial pseudo-annotations, and dynamically improves the annotation quality through pseudo-annotation generation and iterative optimization mechanisms. The training process integrates self-supervised learning tasks (such as image reconstruction and contrastive learning) to further enhance the generalization ability of the segmentation model. It introduces a segmentation network architecture that combines graph neural networks (GNNs) and transformers. GNNs are used to learn the global topological relationship of blood vessels and solve the modeling problems of branch refinement and long path structures. Transformers are used to capture long-range dependency features and improve the anti-interference ability of segmentation against complex backgrounds and noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0062] Figure 1 is a flow chart (1) of a method for automatic blood vessel segmentation according to an embodiment of the present application;

[0063] Figure 2 is a flowchart (II) of the method for automatic blood vessel segmentation according to an embodiment of the present application;

[0064] Figure 3 Flowchart (3) of the automatic blood vessel segmentation method according to an embodiment of the present application;

[0065] Figure 4 Flowchart (4) of the automatic blood vessel segmentation method according to an embodiment of the present application;

[0066] Figure 5 Flowchart (5) of the automatic blood vessel segmentation method according to an embodiment of the present application;

[0067] Figure 6 Flowchart (6) of the automatic blood vessel segmentation method according to an embodiment of the present application;

[0068] Figure 7 Flowchart (VII) of the method for automatic segmentation of blood vessels according to an embodiment of the present application;

[0069] Figure 8 Flowchart (eight) of the automatic blood vessel segmentation method according to an embodiment of the present application;

[0070] Figure 9 Flowchart (IX) of the automatic blood vessel segmentation method according to an embodiment of the present application;

[0071] Figure 10 10 is a flowchart of a method for automatic blood vessel segmentation according to an embodiment of the present application;

[0072] Figure 11 1 is a framework diagram of an automatic blood vessel segmentation system according to an embodiment of the present application (I);

[0073] Figure 12 2 is a framework diagram of the automatic blood vessel segmentation system according to an embodiment of the present application;

[0074] Figure 13 3 is a framework diagram of the automatic blood vessel segmentation system according to an embodiment of the present application;

[0075] Figure 14 4 is a framework diagram of the automatic blood vessel segmentation system according to an embodiment of the present application.

[0076] The figures are marked as follows: 1110, first acquisition module; 1120, automatic segmentation module; 1130, registration and mapping module; 1131, image processing submodule; 1132, registration submodule; 1133, mapping submodule; 1140, second acquisition module; 1150, preprocessing module; 1151, first preprocessing submodule; 1152, second preprocessing submodule; 1153, third preprocessing submodule; 1154, acquisition submodule; 1155, selection submodule; 1160, training module. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely used to explain this application and are not intended to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without making any creative efforts are within the scope of protection of this application.

[0078] Obviously, the drawings described below are merely examples or embodiments of the present application. Those skilled in the art can, without inventive effort, apply the present application to other similar scenarios based on these drawings. Furthermore, it is also understood that, although the effort involved in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, changes in design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as an insufficiency of the content disclosed in this application.

[0079] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0080] Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by a person of ordinary skill in the technical field to which this application belongs. The words "one", "a", "the" and the like used in this application do not indicate a limit on quantity and may indicate the singular or plural. The terms "include", "comprise", "have" and any variations thereof used in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or units (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units that are inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The word "multiple" used in this application means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0081] like Figure 1 As shown, the present invention provides a method for automatic segmentation of blood vessels for CT images, comprising:

[0082] Step S1000: Acquire an enhanced CT image and a plain CT image of the same position;

[0083] Step S2000: inputting the enhanced CT image into a pre-trained first blood vessel automatic segmentation model to obtain a first blood vessel segmentation result;

[0084] Step S3000 : registering the enhanced CT image and the plain CT image to map the first blood vessel segmentation result annotation to the plain CT image to obtain a second blood vessel segmentation result.

[0085] It should be noted that enhanced CT is performed after intravenous injection of an iodinated contrast agent, in addition to a fixed CT scan. The contrast agent enhances the density difference between blood vessels and lesions, making lesions more visible. Enhanced CT images can dynamically observe blood perfusion in lesions and are suitable for assessing dynamic changes in tumors, inflammation, vascular diseases, and other conditions.

[0086] It's important to note that plain scan CT is a basic CT scan method that uses X-rays to scan the body without the use of contrast agents, producing two-dimensional images. It's primarily used for preliminary screening and diagnosis, revealing differences in tissue density and suitable for examinations of the head, chest, abdomen, and other areas.

[0087] like Figure 2 As shown, in step S2000, the training method of the first blood vessel automatic segmentation model includes:

[0088] Step S2100: Acquire a first training image and first labeled data;

[0089] Step S2200: pre-process the first training image and the first annotated data to obtain a second training image and second annotated data;

[0090] Step S2300: input the second training image and the second labeled data into the nnU-Net automatic segmentation network to obtain a first blood vessel automatic segmentation model.

[0091] It should be noted that the first training image in step S2100 is an enhanced CT image; the first annotation data includes manual delineation of blood vessels on the first training image by professional physicians to confirm the annotation of accurate segmentation.

[0092] It should be noted that, in step S2200, the preprocessing includes but is not limited to normalization processing, cropping processing, histogram equalization processing, etc.

[0093] It should be noted that, in step S2200, the pre-processing also includes format conversion processing.

[0094] like Figure 3 As shown, step S2200 includes:

[0095] Step S2210: performing normalization processing on the first training image to obtain a first intermediate training image;

[0096] Step S2220: cropping the first intermediate training image to obtain a second intermediate training image;

[0097] Step S2230: Perform histogram equalization processing on the second intermediate training image to obtain a second training image.

[0098] It should be noted that, before step S2210, step S2200 further includes:

[0099] Perform format conversion processing on the first training image and the first labeled data.

[0100] Specifically, before normalizing the first training image, the first training image and the first labeled data need to be converted into a standard format that can be processed by nnU-Net to meet the input requirements of the model, that is, the first training image and the first labeled data are converted into the NIfTI format commonly used by nnU-Net.

[0101] It should be noted that, in step S2210, the grayscale value of the first training image is remapped to the target range. The specific formula is as follows:

[0102]

[0103] Among them, I new is the grayscale value of the pixel in the new range after normalization and linear mapping; I old is the grayscale value of a pixel in the first training image; I min is the minimum grayscale value of all pixels in the first training image; I max is the maximum grayscale value of all pixels in the first training image; TargetMax and TargetMin are the minimum and maximum values ​​of the target grayscale value range.

[0104] It should be noted that in step S2220, by selecting appropriate window width and window level, the image grayscale value is cropped, and the part exceeding the range is truncated. The specific processing standards are as follows:

[0105]

[0106] Where I(x,y) represents the pixel value of the first training image at the coordinate (x,y); WL (Window Level) is the parameter of the window level; WW (Window Width) is the parameter of the window width; I clipped(x, y) is the pixel value at the coordinate (x, y) after the first training image is cropped.

[0107] Among them, when When , that is, the pixel value is less than the window position minus half the window width, then, That is, the pixel value is set to the window position minus half the window width.

[0108] Among them, when When , that is, the pixel value is less than the window position plus half the window width, then, That is, the pixel value is set to the window position plus half the window width.

[0109] Among them, when When the pixel value is within the window width, I clipped (x, y) = I(x, y), that is, the pixel value remains unchanged.

[0110] It should be noted that in step S2230, the image grayscale value distribution is adjusted to make it more uniform and enhance the local contrast. The specific formula is as follows:

[0111]

[0112] Where CDF(I) is the cumulative distribution function.

[0113] like Figure 4 As shown, step S2200 also includes:

[0114] Step S2240: Obtain the file format and target resolution of the second training image;

[0115] Step S2250: Select an applicable nnU-Net automatic segmentation network based on the file format and target resolution.

[0116] It should be noted that the selection of the nn-Unet automatic segmentation network must meet the following criteria:

[0117] 1. Encoding and decoding network selection: Based on the type and target resolution of the first training image and the first annotated data, select the appropriate nnU-Net network architecture, such as the 2D U-Net model, the 3d_fullres model (a 3D U-Net operating at high image resolution), or the 3d_lowres→3d_cascade_fullres model (a 3D U-Net cascade, where a 3D U-Net first operates on a low-resolution image, and then a second high-resolution 3D U-Net refines the predictions of the former). The 2D U-Net model is applicable to both 2D and 3D datasets; the 3d_fullres model is only applicable to 3D datasets; and the 3d_lowres→3d_cascade_fullres model is only applicable to 3D datasets with large image sizes.

[0118] 2. Loss function combination: Use the simple average of Cross Entropy Loss and Dice Loss to optimize the model.

[0119] Cross-entropy loss includes binary cross-entropy loss (Binary Cross-Entropy Loss) and multi-classification cross-entropy loss (Categorical Cross-Entropy Loss);

[0120] Among them, for the two-class problem, it is equivalent to the binary cross entropy loss, the formula is:

[0121]

[0122] y is the true label, which takes the value of 0 or 1; is the probability value predicted by the model, ranging from 0 to 1.

[0123] Among them, for multi-category problems (i.e., a sample can belong to one of multiple classes), the formula for cross entropy loss is:

[0124]

[0125] N is the number of samples. C is the number of classes. i，c is the true label of the i-th sample in category c (in one-hot encoding form). i，c is the probability that the model predicts that the i-th sample belongs to category c.

[0126] about The index (called Dice similarity coefficient (DSC) when applied to Boolean data) is the most commonly used metric for evaluating segmentation accuracy and is calculated as:

[0127]

[0128] The Dice loss can therefore be defined as:

[0129] Loss Dice =1-DSC or

[0130] Among them, A and B represent the predicted results and the true label area respectively.

[0131] 3. Optimizer Configuration: Use the SGD optimizer with an initial learning rate of 1e-2, a momentum of 0.99, a weight decay of 3e-5, and enable Nesterov momentum. Combined with the PolyLRScheduler, a polynomial decay strategy is used to adjust the learning rate over the entire 1000 training epochs.

[0132] 4. Training configuration: Each epoch consists of 250 training iterations and 50 validation iterations, and the oversampling ratio of foreground samples is set to 0.33. These configurations are implemented by calling the configure_optimizers method to ensure accurate replication of the nnUNet training and learning rate adjustment strategy.

[0133] like Figure 5 As shown, step S3000 includes:

[0134] Step S3100: Standardize the enhanced CT image and the plain CT image to obtain a standard enhanced CT image and a standard plain CT image;

[0135] Step S3200: Perform nonlinear registration of the standard enhanced CT image and the standard plain scan CT image using a deep learning registration network (VoxelMorph) to generate a deformation field;

[0136] Step S3300: Map the first blood vessel segmentation result to a standard plain scan CT image based on the deformation field to obtain a second blood vessel segmentation result.

[0137] like Figure 6 As shown, step S3100 includes:

[0138] Step S3110: normalize the enhanced CT image and the plain CT image to obtain a first sub-enhanced CT image and a first sub-plain CT image;

[0139] Step S3120: resampling the first sub-enhanced CT image and the first sub-plain CT image to obtain a second sub-enhanced CT image and a second sub-plain CT image;

[0140] Step S3130: cropping the second sub-enhanced CT image and the second sub-plain CT image to obtain a third sub-enhanced CT image and a third sub-plain CT image;

[0141] Step S3140: Perform histogram equalization processing on the third sub-enhanced CT image and the third sub-plain CT image to obtain a standard enhanced CT image and a standard plain CT image.

[0142] It should be noted that the normalization processing method for the enhanced CT image and the plain scan CT image in step S3110 is the same as the processing method in step S2210 and will not be repeated here.

[0143] It should be noted that, in step S3120, the first sub-enhanced CT image and the first sub-plain CT image are resampled to the same voxel size and spatial resolution to ensure that each voxel represents the same physical size in the first sub-enhanced CT image and the first sub-plain CT image.

[0144] It should be noted that, since the first training image is a reference image (ie, can be used as a standard), no resampling is required.

[0145] It should be noted that the cropping method for the second sub-enhanced CT image and the second sub-plain CT image in step S3130 is the same as the cropping method in step S2220 and will not be repeated here.

[0146] It should be noted that the method of performing histogram equalization processing on the third sub-enhanced CT image and the third sub-plain CT image in step S3140 is the same as that in step S2230 and will not be repeated here.

[0147] It should be noted that VoxelMorph is an image registration model based on convolutional neural networks (CNN). VoxelMorph can learn complex (nonlinear) deformation fields to achieve accurate image alignment.

[0148] It should be noted that, in the present invention, the TensorFlow framework is used to implement the VoxelMorph network; in addition, the network structure and code can be adjusted according to specific task requirements.

[0149] It should be noted that VoxelMorph uses unsupervised learning and does not require the actual deformation field as a label. By minimizing the loss function, the network learns to align the standard plain CT image to the standard enhanced CT image.

[0150] like Figure 7As shown, it should be noted that, in step S3200, a compensation strategy for multimodal texture and intensity distribution is introduced, such as adversarial loss and structural similarity index (SSIM) optimization, to make the registration more robust between different modalities.

[0151] Specifically, a generative adversarial network (GAN) framework was used to convert standard plain CT images into "pseudo-standard enhanced CT images." Modality alignment was performed before registration to bring the intensity distribution of the standard plain CT images closer to that of the standard enhanced CT images, thereby reducing the impact of modality differences on registration.

[0152] More specifically, step S3200 includes:

[0153] Step S3210: input the standard plain scan CT image into the Generative Adversarial Network (GAN) framework to obtain a pseudo-standard enhanced CT image;

[0154] Step S3220: Constraining the geometric information of the pseudo-standard enhanced CT image and the standard plain scan CT image using a loss function;

[0155] Step S3230: Register the pseudo-standard enhanced CT image with the standard enhanced CT image to generate a deformation field.

[0156] It should be noted that the loss function is as follows:

[0157] Similarity loss (L sim ): It is encouraged that the registered standard enhanced CT images and the standard plain CT images be similar in appearance.

[0158]

[0159] Among them, F represents standard enhanced CT images; M represents standard plain CT images; represents the deformation field, which is used to describe how the standard plain CT image is transformed to align with the standard enhanced CT image;

[0160] Smoothness loss (L smooth ): Encourage the smoothness of the deformation field and avoid unreasonable deformation:

[0161]

[0162] Among them, Ω represents the spatial domain where the image is located; Representing the deformation field The gradient at point p;

[0163] Total loss function:

[0164]

[0165] or

[0166] L=L sim +λ*L smooth ;

[0167] Here, λ is the regularization parameter.

[0168] Fighting Loss:

[0169]

[0170] in, Represents the adversarial loss value; is the expectation operator; I enhanced is the enhanced real image; D is the discriminator, G is the generator; I plain For the original simple image.

[0171] like Figure 8 As shown, step S3200 also includes:

[0172] Step S3240: using SSIM to measure the local similarity between the registered pseudo-standard enhanced CT image and the standard enhanced CT image, so as to optimize the similarity metric in the registration algorithm;

[0173] Step S3250: Reduce the difference in grayscale value distribution between the pseudo-standard enhanced CT image and the standard enhanced CT image through a standardization strategy.

[0174] It should be noted that the SSIM in step S3240 is defined as follows:

[0175]

[0176] Where x and y are pseudo-standard enhanced CT images and standard enhanced CT images, respectively; μ x , μ y are the pixel means of pseudo-standard enhanced CT images and standard enhanced CT images, respectively; are the pixel variances of pseudo-standard enhanced CT images and standard enhanced CT images respectively; σ xy is the covariance between pseudo-standard enhanced CT images and standard enhanced CT images; C1 and C2 are constants.

[0177] It should be noted that the Z-score normalization formula in step S3250 is as follows:

[0178]

[0179] Where μ is the image mean and σ is the standard deviation.

[0180] It is important to note that optimization algorithms such as stochastic gradient descent (SGD) are used to adjust network parameters to minimize the loss function. Through training, the network learns to predict the deformation field that can align a standard plain CT image to a standard enhanced CT image.

[0181] In addition, after the registration is completed, the Dice coefficient can be used to evaluate the degree of overlap of anatomical structures in the registered images. The calculation formula is as follows:

[0182] ;

[0183] A Dice coefficient of 1 indicates complete overlap and ideal registration.

[0184] A Dice coefficient of 0 indicates no overlap and registration failure.

[0185] like Figure 9 As shown, step S3300 includes:

[0186] Step S3310: Based on the spatial transformation of the deformation field, the first blood vessel segmentation result is mapped to the standard plain scan CT image by bilinear interpolation to obtain a mapping result;

[0187] Step S3320: performing morphological processing on the mapping result to obtain a second blood vessel segmentation result;

[0188] Step S3330: Perform self-supervised optimization on the second blood vessel segmentation result based on the mapping result and the standard enhanced CT image.

[0189] It should be noted that in step S3310, the deformation field T(x, y, z) is a vector field, and the vector corresponding to each voxel position (x, y, z) points to the matching position in the standard enhanced CT image, which is specifically defined as follows:

[0190]

[0191] Where u, v, and w represent the displacements in the x, y, and z directions, respectively.

[0192] It should be noted that the deformation field is calculated using a deep learning model (such as a registration network based on U-Net or Transformer), combined with multi-scale feature enhancement and regularization strategies, to improve the accuracy and stability of the registration of standard enhanced CT images to standard plain scan CT images, and reduce artifacts and local distortions in the deformation field.

[0193] like Figure 10 As shown, step S3330 includes:

[0194] Step S3331: Map the first blood vessel segmentation result to the spatial position of the standard plain scan CT image to obtain an initial pseudo-label;

[0195] Step S3332: input the initial pseudo-labeling into the first blood vessel automatic segmentation model to train and obtain a second blood vessel automatic segmentation model;

[0196] Step S3333: input the standard plain scan CT image into the second blood vessel automatic segmentation model to obtain the second blood vessel segmentation result;

[0197] Step S3334: selecting high-confidence regions according to the second blood vessel segmentation result as optimized pseudo-labels to obtain reference pseudo-labels;

[0198] Step S3335: Using the reference pseudo-labels as training data, gradually improve the second blood vessel automatic segmentation model to achieve self-supervised optimization.

[0199] It should be noted that in step S3331, non-rigid registration (such as B-spline, Demons, or deep learning-based VoxelMorph) is used to align the standard enhanced CT image and the standard plain scan CT image.

[0200] It should be noted that, in step S3331, the first segmentation result is mapped to the spatial position of the standard plain scan CT image using the deformation field.

[0201] It should be noted that pseudo-labeling errors that may occur during the mapping process can be corrected by using morphological operations (such as closing operations and hole filling) to correct discontinuities or artifacts in the segmentation results and by applying region growing algorithms to eliminate small pseudo-labeling areas.

[0202] It should be noted that in order to improve the quality of pseudo-labeling, model training and iterative optimization mechanisms are used to gradually improve the accuracy of labeling.

[0203] It should be noted that in step S3334, uncertainty estimation (such as Dropout based on Monte Carlo sampling) is used to evaluate the reliability of the model prediction to screen high-confidence areas as optimized pseudo-labels.

[0204] It should be noted that while using reference pseudo-annotations for self-supervised optimization, manual correction can also be used, and professional physicians can review and correct the initial pseudo-annotations as high-quality data for further model training.

[0205] It should be noted that the steps shown in the above process or the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0206] This embodiment provides an automatic blood vessel segmentation system for CT images. This system is used to implement the above-mentioned embodiments and preferred embodiments. Details already described are not repeated here. As used below, the terms "module," "unit," "subunit," etc. may refer to a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation using hardware, or a combination of software and hardware, is also possible and contemplated.

[0207] like Figure 11 As shown, a blood vessel automatic segmentation system for CT images includes a first acquisition module 1110, an automatic segmentation module 1120, and a registration and mapping module 1130. The first acquisition module 1110 is used to acquire an enhanced CT image and a plain CT image of the same location; the automatic segmentation module 1120 is used to input the enhanced CT image into a pre-trained first blood vessel automatic segmentation model to obtain a first blood vessel segmentation result; and the registration and mapping module 1130 is used to register the enhanced CT image and the plain CT image to map the first blood vessel segmentation result annotations to the plain CT image to obtain a second blood vessel segmentation result.

[0208] Specifically, if Figure 12 As shown, the registration and mapping module 1130 includes an image processing submodule 1131, a registration submodule 1132, and a mapping submodule 1133. The image processing submodule 1131 is used to perform standardization processing on the enhanced CT image and the plain CT image to obtain a standard enhanced CT image and a standard plain CT image; the registration submodule 1132 is used to perform nonlinear registration of the standard enhanced CT image and the standard plain CT image using a deep learning registration network (VoxelMorph) to generate a deformation field; and the mapping submodule 1133 is used to map the first segmentation result to the registered standard plain CT image based on the deformation field, thereby achieving automatic blood vessel segmentation in the standard plain CT image.

[0209] More specifically, the image processing submodule 1131 includes a first image processing unit, a second image processing unit, a third image processing unit, and a fourth image processing unit. The first image processing unit is configured to perform normalization processing on the enhanced CT image and the plain CT image to obtain a first sub-enhanced CT image and a first sub-plain CT image; the second image processing unit is configured to perform resampling processing on the first sub-enhanced CT image and the first sub-plain CT image to obtain a second sub-enhanced CT image and a second sub-plain CT image; the third image processing unit is configured to perform cropping processing on the second sub-enhanced CT image and the second sub-plain CT image to obtain a third sub-enhanced CT image and a third sub-plain CT image; and the fourth image processing unit is configured to perform histogram equalization processing on the third sub-enhanced CT image and the third sub-plain CT image to obtain a standard enhanced CT image and a standard plain CT image.

[0210] More specifically, the registration submodule 1132 includes an input unit, a constraint unit, and a generation unit. The input unit is used to input the standard plain CT image into the Generative Adversarial Network (GAN) framework to obtain a pseudo-standard enhanced CT image; the constraint unit is used to constrain the geometric information of the pseudo-standard enhanced CT image and the standard plain CT image using a loss function; and the generation unit is used to register the pseudo-standard enhanced CT image with the standard enhanced CT image to generate a deformation field.

[0211] Furthermore, the registration submodule 1132 also includes a measurement unit and a normalization unit. The measurement unit is used to measure the local similarity between the registered pseudo-standard enhanced CT image and the standard enhanced CT image using SSIM to optimize the similarity metric in the registration algorithm; the normalization unit is used to reduce the difference in grayscale value distribution between the pseudo-standard enhanced CT image and the standard enhanced CT image through a normalization strategy.

[0212] More specifically, the mapping submodule 1133 includes a mapping unit, a mapping processing unit, and an optimization unit. The mapping unit is configured to map the first vessel segmentation result to a standard plain scan CT image using bilinear interpolation based on the spatial transformation of the deformation field to obtain a mapping result. The mapping processing unit is configured to perform morphological processing on the mapping result to obtain a second vessel segmentation result. The optimization unit is configured to perform self-supervised optimization of the second vessel segmentation result based on the mapping result and the standard enhanced CT image.

[0213] Furthermore, the optimization unit further includes a mapping subunit, a first output subunit, a second output subunit, a screening subunit, and a training subunit. The mapping subunit is configured to map the first vessel segmentation result to the spatial position of a standard plain scan CT image to obtain an initial pseudo-annotation; the first output subunit is configured to input the initial pseudo-annotation into the first vessel automatic segmentation model to train and obtain a second vessel automatic segmentation model; the second output subunit is configured to input the standard plain scan CT image into the second vessel automatic segmentation model to obtain a second vessel segmentation result; the screening subunit is configured to screen high-confidence regions based on the second vessel segmentation result as optimized pseudo-annotations to obtain reference pseudo-annotations; and the training subunit is configured to gradually improve the second vessel automatic segmentation model based on the reference pseudo-annotations as training data to achieve self-supervised optimization.

[0214] like Figure 13As shown, the automatic blood vessel segmentation system further includes a second acquisition module 1140, a preprocessing module 1150, and a training module 1160. The second acquisition module 1140 is configured to acquire a first training image and first annotated data; the preprocessing module 1150 is configured to preprocess the first training image and first annotated data to obtain a second training image and second annotated data; and the training module 1160 is configured to input the second training image and second annotated data into the nn-Unet automatic segmentation network to obtain a first automatic blood vessel segmentation model.

[0215] Specifically, if Figure 14 As shown, the pre-processing module 1150 includes a first pre-processing sub-module 1151, a second pre-processing sub-module 1152, and a third pre-processing sub-module 1153. The first pre-processing sub-module 1151 is used to perform normalization processing on the first training image to obtain a first intermediate training image; the second pre-processing sub-module 1152 is used to perform cropping processing on the first intermediate training image to obtain a second intermediate training image; and the third pre-processing sub-module 1153 is used to perform histogram equalization processing on the second intermediate training image to obtain a second training image.

[0216] Furthermore, the pre-processing module 1150 further includes an acquisition submodule 1154 and a selection submodule 1155. The acquisition submodule 1154 is used to acquire the file format and target resolution of the second training image; and the selection submodule 1155 is used to select an applicable nnU-Net automatic segmentation network based on the file format and target resolution.

[0217] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0218] In addition, the blood vessel automatic segmentation method of the embodiment of the present application can be implemented by a computer device. The components of the computer device may include but are not limited to a processor and a memory storing computer program instructions.

[0219] In some embodiments, the processor may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0220] In some embodiments, the memory may include a large capacity memory for data or instructions. By way of example and not limitation, the memory may include a hard disk drive (HDD), a floppy disk drive, a solid state drive (SSD), a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory may include removable or non-removable (or fixed) media. Where appropriate, the memory may be inside or outside the data processing device. In a specific embodiment, the memory is a non-volatile memory. In a specific embodiment, the memory includes a read-only memory (ROM) and a random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM) or a flash memory (FLASH), or a combination of two or more of these. Under appropriate circumstances, the RAM can be a static random access memory (SRAM) or a dynamic random access memory (DRAM), where the DRAM can be a fast page mode dynamic random access memory (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0221] The memory may be used to store or cache various data files required for processing and / or communication, as well as possible computer program instructions executed by the processor.

[0222] The processor reads and executes computer program instructions stored in the memory to implement any one of the methods for automatic segmentation of blood vessels in CT images in the above embodiments.

[0223] In some embodiments, the computer device may further include a communication interface and a bus, wherein the processor, the memory, and the communication interface are connected via the bus and communicate with each other.

[0224] The communication interface is used to enable communication between the various units, devices, units, and / or devices in the embodiments of the present application. The communication interface can also enable data communication with other components such as external devices, image / data acquisition equipment, databases, external storage, and image / data processing workstations.

[0225] A bus, which includes hardware, software, or both, couples components of a computer device to each other. Buses include, but are not limited to, at least one of the following: a data bus, an address bus, a control bus, an expansion bus, and a local bus. By way of example and not limitation, a bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of the above. Where appropriate, a bus may include one or more buses. Although embodiments herein describe and illustrate a particular bus, this application contemplates any suitable bus or interconnect.

[0226] The computer device can execute the blood vessel automatic segmentation method in the embodiment of the present application.

[0227] In addition, in conjunction with the automatic blood vessel segmentation method in the above embodiments, embodiments of the present application may provide a computer-readable storage medium for implementation. The computer-readable storage medium stores computer program instructions; when executed by a processor, the computer program instructions implement any of the automatic blood vessel segmentation methods for CT images in the above embodiments.

[0228] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0229] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for automatic segmentation of blood vessels in CT images, characterized in that: include: Acquire enhanced CT images and plain CT images of the same location; Inputting the enhanced CT image into the pre-trained first vessel automatic segmentation model to obtain the first vessel segmentation result; The enhanced CT image and the plain CT image are registered to map the first blood vessel segmentation result annotation to the plain CT image to obtain a second blood vessel segmentation result.

2. The blood vessel automatic segmentation method according to claim 1, characterized in that: Registering the enhanced CT image and the plain CT image to map the first vessel segmentation result annotation to the plain CT image to obtain the second vessel segmentation result includes: Standardizing the enhanced CT images and the plain CT images to obtain standard enhanced CT images and standard plain CT images; The standard enhanced CT images and the standard plain CT images are nonlinearly registered using a deep learning registration network to generate a deformation field. The first blood vessel segmentation result is mapped to the standard plain scan CT image based on the deformation field to obtain the second blood vessel segmentation result.

3. The blood vessel automatic segmentation method according to claim 2, characterized in that: Standardization processing of enhanced CT images and plain CT images to obtain standard enhanced CT images and standard plain CT images includes: Normalizing the enhanced CT image and the plain CT image to obtain a first sub-enhanced CT image and a first sub-plain CT image; resampling the first sub-enhanced CT image and the first sub-plain CT image to obtain a second sub-enhanced CT image and a second sub-plain CT image; The second sub-enhanced CT image and the second sub-plain CT image are cropped to obtain a third sub-enhanced CT image and a third sub-plain CT image; performing histogram equalization processing on the third sub-enhanced CT image and the third sub-plain CT image to obtain a standard enhanced CT image and a standard plain CT image; and / or Mapping the first blood vessel segmentation result to a standard plain scan CT image based on the deformation field to obtain a second blood vessel segmentation result includes: Based on the spatial transformation of the deformation field, the first vessel segmentation result is mapped to the standard plain scan CT image through bilinear interpolation to obtain the mapping result; Performing morphological processing on the mapping result to obtain a second blood vessel segmentation result; The second vessel segmentation results are self-supervised and optimized based on the mapping results and standard enhanced CT images.

4. The blood vessel automatic segmentation method according to claim 3, characterized in that: The self-supervised optimization of the second vessel segmentation results based on the mapping results and the standard enhanced CT images includes: Mapping the first blood vessel segmentation result to the spatial position of the standard plain scan CT image to obtain the initial pseudo-labeling; Inputting the initial pseudo-annotation into the first blood vessel automatic segmentation model to train and obtain the second blood vessel automatic segmentation model; Input the standard plain scan CT image into the second blood vessel automatic segmentation model and obtain the second blood vessel segmentation result; According to the second blood vessel segmentation result, a high-confidence area is selected as the optimized pseudo-annotation to obtain a reference pseudo-annotation; Based on the reference pseudo-annotations as training data, the second vessel automatic segmentation model is gradually improved to achieve self-supervised optimization.

5. The automatic blood vessel segmentation method according to any one of claims 1 to 4, characterized in that: The training method of the first blood vessel automatic segmentation model includes: Acquire a first training image and first labeled data; Preprocessing the first training image and the first labeled data to obtain a second training image and second labeled data; The second training image and the second labeled data are input into the nn-Unet automatic segmentation network to obtain a first blood vessel automatic segmentation model.

6. The blood vessel automatic segmentation method according to claim 5, characterized in that: The training method of the first blood vessel automatic segmentation model also includes: Obtain the file format and target resolution of the second training image; Select the appropriate nnU-Net automatic segmentation network based on the file format and target resolution; and / or Preprocessing the first training image to obtain the second training image includes: performing normalization processing on the first training image to obtain a first intermediate training image; performing cropping processing on the first intermediate training image to obtain a second intermediate training image; Perform histogram equalization processing on the second intermediate training image to obtain a second training image.

7. A blood vessel automatic segmentation system for CT images, characterized in that: include: a first acquisition module, configured to acquire an enhanced CT image and a plain scan CT image at the same position; an automatic segmentation module, the automatic segmentation module being used to input the enhanced CT image into a pre-trained first blood vessel automatic segmentation model to obtain a first blood vessel segmentation result; A registration and mapping module is used to register the enhanced CT image and the plain CT image to map the first blood vessel segmentation result annotation to the plain CT image to obtain a second blood vessel segmentation result.

8. The automatic blood vessel segmentation system according to claim 7, characterized in that: The registration mapping module includes: An image processing submodule, wherein the image processing submodule is used to perform standardization processing on the enhanced CT image and the plain scan CT image to obtain a standard enhanced CT image and a standard plain scan CT image; A registration submodule, wherein the registration submodule is used to perform nonlinear registration of the standard enhanced CT image and the standard plain scan CT image through a deep learning registration network to generate a deformation field; a mapping submodule, configured to map the first blood vessel segmentation result to a standard plain scan CT image based on the deformation field to obtain a second blood vessel segmentation result; and / or The automatic blood vessel segmentation system further includes: a second acquisition module, the second acquisition module being used to acquire the first training image and the first annotation data; a preprocessing module, configured to preprocess the first training image and the first annotated data to obtain a second training image and second annotated data; A training module is used to input the second training image and the second labeled data into the nn-Unet automatic segmentation network to obtain a first blood vessel automatic segmentation model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the blood vessel automatic segmentation method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the blood vessel automatic segmentation method according to any one of claims 1 to 6 is implemented.

Images