A blood vessel segmentation method and device based on a deep learning model

Abstract

The application provides a blood vessel segmentation method and device based on a deep learning model, which comprises the following steps: performing image transformation on a medical image to obtain a transformed image; taking the transformed image as a sample and the medical image as a label to form a pre-training sample; pre-training a deep learning model according to the pre-training sample; performing secondary training on the deep learning model by using a formal training sample to obtain a blood vessel segmentation model; and inputting a medical image to be segmented into the blood vessel segmentation model to obtain a segmented blood vessel image. The pre-training sample is generated by image transformation, the deep learning model is pre-trained, the pre-training process does not need manual labeling of the medical image sample, and the model training cost is reduced; the deep learning model is further trained on the basis of the pre-training by using the formal training sample labeled with blood vessel / non-blood vessel labels, the model can further identify the blood vessel structure, and the blood vessel segmentation precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a blood vessel segmentation method and device based on a deep learning model. BACKGROUND

[0002] Blood vessel segmentation refers to marking blood vessel pixels in medical images to provide important information support for subsequent diagnosis / treatment of patients. Blood vessel segmentation based on a deep learning model can greatly save doctors' time and improve their work efficiency.

[0003] In the prior art, a blood vessel segmentation method based on deep learning marks medical images manually and trains a deep learning model using the marked samples. However, manually marking three-dimensional medical images is extremely time-consuming and requires professional personnel to mark, that is, the marked samples of three-dimensional medical images are scarce, and the small amount of marked sample data also leads to low recognition accuracy of the blood vessel segmentation model after training. SUMMARY

[0004] The present application provides a blood vessel segmentation method and device based on a deep learning model to solve or at least partially solve the problem of scarce medical image marking samples and low recognition accuracy of the blood vessel segmentation model in the prior art.

[0005] The present application provides a blood vessel segmentation method based on a deep learning model, comprising:

[0006] performing image transformation on the medical image to obtain a transformed image;

[0007] using the transformed image as a sample and the medical image as a label to form a pre-training sample;

[0008] pre-training a deep learning model according to the pre-training sample;

[0009] performing secondary training on the deep learning model using a formal training sample to obtain a blood vessel segmentation model;

[0010] inputting a medical image to be segmented into the blood vessel segmentation model to obtain a segmented blood vessel image.

[0011] According to the blood vessel segmentation method based on a deep learning model provided by the present application, the image transformation at least includes one of the following: linear intensity transformation, local pixel reorganization, intensity inversion, outward filling, and inward filling.

[0012] According to the blood vessel segmentation method based on a deep learning model provided by the present application, the linear intensity transformation is implemented according to the following formula:

[0013] f(x)=kx+b;

[0014] Wherein, x represents the intensity of a certain position pixel point, f(x) represents the intensity value of the pixel point after linear intensity transformation, k is a positive integer, and b is a fixed constant.

[0015] According to the blood vessel segmentation method based on the deep learning model provided by the application, the local pixel recombination is realized according to the following formula:

[0016]

[0017] Wherein, x represents the position of any one pixel point in omega, L(x) represents the target position of the pixel recombination of the point, Ran represents random transformation in omega, omega represents the overall region of a medical image, and omega represents a sub-region (omega omega) of omega.

[0018] According to the blood vessel segmentation method based on the deep learning model provided by the application, the intensity inversion is realized according to the following formula:

[0019] p(x)=Max(I)-x+Min(I);

[0020] Wherein, x represents the intensity of a certain pixel point in the image I, p(x) represents the intensity value of the intensity inversion of the point, Max(I) represents the highest intensity value in the image I, and Min(I) represents the lowest intensity value in the image I.

[0021] According to the blood vessel segmentation method based on the deep learning model provided by the application, the deep learning model is specifically a U-net network model containing an attention mechanism.

[0022] According to the blood vessel segmentation method based on the deep learning model provided by the application, the U-net network model containing the attention mechanism comprises a first number of down-sampling layers, a second number of convolution layers and a first number of up-sampling layers connected in sequence, each down-sampling layer of each level is connected to the up-sampling layer of the same level through a jump structure, and an attention module is further arranged on each jump structure.

[0023] According to the blood vessel segmentation method based on the deep learning model provided by the application, before the medical image to be segmented is input into the blood vessel segmentation model, the method further comprises:

[0024] The medical image to be segmented is at least data enhanced in one of the following ways: gray scale scaling, rotation, inversion, and translation.

[0025] The application also provides a blood vessel segmentation device based on a deep learning model, comprising:

[0026] a transformation module, configured to perform image transformation on the medical image to obtain a transformed image;

[0027] a sample module, configured to take the transformed image as a sample and the medical image as a label to form a pre-training sample;

[0028] a pre-training module, configured to pre-train a deep learning model according to the pre-training sample;

[0029] a training module, configured to perform secondary training on the deep learning model by using a formal training sample to obtain a blood vessel segmentation model;

[0030] a segmentation module, configured to input an image to be segmented into the blood vessel segmentation model to obtain a segmented blood vessel image.

[0031] The application further provides a medical image segmentation model training method, comprising:

[0032] performing image transformation on the medical image to obtain a transformed image;

[0033] taking the transformed image as a sample and the medical image as a label to form a pre-training sample;

[0034] pre-training a deep learning model according to the pre-training sample;

[0035] performing secondary training on the deep learning model by using a formal training sample to obtain a medical image segmentation model.

[0036] The blood vessel segmentation method and device based on the deep learning model provided by the application generate a pre-training sample by image transformation, pre-train a deep learning model, make the model learn the biological tissue structure features and pixel features of the target part, have a good migration learning basis, and do not need to manually label the medical image sample in the pre-training process, thereby reducing the model training cost; the deep learning model is further trained on the basis of the pre-training by using the formal training sample with the blood vessel / non-blood vessel label, the model can further identify the blood vessel structure, and the blood vessel segmentation precision is improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0038] Figure 1 is a flowchart of the blood vessel segmentation method based on the deep learning model provided by the application;

[0039] Figure 2 is a structural schematic diagram of a U-net network model provided by the application, which comprises an attention mechanism;

[0040] Figure 3 is a structural schematic diagram of a blood vessel segmentation device based on a deep learning model provided by the application;

[0041] Figure 4 is a structural schematic diagram of an electronic device provided by the application. DETAILED DESCRIPTION

[0042] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] The blood vessel segmentation method and device based on a deep learning model of the present application will be described below in combination with Figures 1-4

[0044] Figure 1 The present application provides a flowchart of the blood vessel segmentation method based on a deep learning model, as shown in Figure 1 The method comprises the following steps:

[0045] S11, image transformation is performed on a medical image to obtain a transformed image;

[0046] S12, the transformed image is taken as a sample, and the medical image is taken as a label to form a pre-training sample;

[0047] S13, a deep learning model is pre-trained according to the pre-training sample;

[0048] S14, the deep learning model is twice trained by using a formal training sample to obtain a blood vessel segmentation model;

[0049] S15, a medical image to be segmented is input into the blood vessel segmentation model to obtain a segmented blood vessel image.

[0050] ​Specifically, the medical image can be a medical image collected and acquired by a computerized tomography (CT), an X-ray film, a magnetic resonance imaging (MRT), a positron emission tomography (PET) and the like for a target part, such as a head, a lung, an eye and the like. It should be noted that for the blood vessel segmentation of a two-dimensional medical image, ImageNet is usually used, which is a 2D dataset (i.e., a labeled sample). Training with such a dataset will limit the structure of the network and also make the network unable to obtain the spatial information contained in the 3D medical image. The three-dimensional medical image provides more abundant spatial information, which helps to provide more comprehensive and detailed information reference for doctors. However, it is difficult to label the blood vessels of the three-dimensional medical image, and it takes a long time, and usually the amount of labeled sample data is small. The medical image in the blood vessel segmentation method provided by the present application refers to a three-dimensional medical image, that is, the present application is for blood vessel segmentation of a three-dimensional medical image.

[0051] First, a pre-training sample is made by image transformation, and a deep learning model is trained. The medical image here is an unlabeled medical image, which is transformed and adjusted by image transformation to obtain a transformed image, so that the images before and after transformation have differences in biological tissue structure or pixel features. Then, the transformed image is used as a sample, and the medical image before transformation is used as a corresponding label to form a pre-training sample, which is used as training data to pre-train the deep learning model. It can be understood that the labels in the pre-training sample do not contain label information of "blood vessels" and "non-blood vessels", and do not need to be manually labeled, but only need to be obtained by convenient image transformation.

[0052] Pre-training enables the trained deep learning model to learn the biological tissue structure, pixel features and the like of the target part. The pre-trained model has a good basis for transfer learning and can be used for further training to complete different functional purposes and segment specific tissue structures (such as brain regions, blood vessels, tumors and the like). On the basis of pre-training, the deep learning model is trained again using formal training samples, so that the model can further identify the blood vessel structure in combination with the biological tissue structure and pixel features of the target part learned by pre-training. It can be understood that the medical image in the formal training sample here is also three-dimensional, and this medical image is labeled with blood vessels, i.e., contains "blood vessels" and "non-blood vessels" label information for each pixel point. The deep learning model is trained again using the formal training sample, so that the model can identify the blood vessel structure in the target part. The deep learning model after the second training is a blood vessel segmentation model, which can be used to process the medical image to be segmented and output the segmented blood vessel image.

[0053] In this embodiment, the pre-training samples are generated by image transformation, the deep learning model is pre-trained, so that the model learns the biological tissue structure features and pixel features of the target part, has a good basis for transfer learning, and the pre-training process does not require manual annotation of medical image samples, reducing the model training cost; the deep learning model is further trained based on the pre-training through the formal training samples labeled with blood vessel / non-blood vessel labels, so that the model can further identify the blood vessel structure and improve the blood vessel segmentation accuracy.

[0054] Based on any of the above embodiments, in one embodiment, the image transformation at least includes one of the following: linear intensity transformation, local pixel reorganization, intensity inversion, outward filling, inward filling.

[0055] Specifically, the medical image is adjusted by image transformation to obtain a transformed image, so that the transformed image and the image before transformation have differences in biological tissue structure or pixel features, and the deep learning model can learn the biological tissue structure features and pixel features of the target part in the process of "restoring" the transformed image to the original medical image.

[0056] The image transformation can be implemented in various ways, such as linear intensity transformation, local pixel reorganization, intensity inversion, outward filling, and inward filling. Among them, the linear intensity transformation refers to linear adjustment of the intensity of the pixels of the medical image, which can enable the pre-trained model to learn the pixel intensity information of the target part; the local pixel reorganization refers to rearranging the positions of the "image blocks" in the overall medical image, which can enable the pre-trained model to learn the tissue structure information of the target part; the intensity inversion refers to adjusting the intensity of each pixel in the medical image, so that the pixels with relatively high intensity become lower, and the pixels with relatively low intensity become higher, which can enable the pre-trained model to learn the contrast information between the tissue and the surrounding background; the outward filling refers to obscuring the edges of the medical image and then letting the deep learning model infer and fill, and the inward filling refers to obscuring some areas in the middle of the medical image and then letting the deep learning model infer and fill, which can enable the pre-trained model to learn the tissue structure information of the target part.

[0057] In this embodiment, various image transformation methods such as linear intensity transformation, local pixel reorganization, intensity inversion, outward filling, and inward filling are used to enable the pre-trained deep learning model to learn the biological tissue structure or pixel features of the target part, and have a good basis for transfer learning, improving the segmentation accuracy of the blood vessel structure.

[0058] Based on any of the above embodiments, in one embodiment, the linear intensity transformation is implemented according to the following formula:

[0059] f(x) = kx + b;

[0060] Wherein, x represents the intensity of a certain position pixel point, f(x) represents the intensity value of the pixel point after linear intensity transformation, k is a positive integer, and b is a fixed constant.

[0061] In the embodiment, the medical image is simply and conveniently linearly intensity transformed by the above formula, and the pre-trained model can learn the pixel intensity information of the target part.

[0062] Based on any one of the above embodiments, in an embodiment, the local pixel recombination is implemented according to the following formula:

[0063]

[0064] Wherein, x represents the position of any one pixel point in ω, L(x) represents the target position of the point after pixel recombination, Ran represents random transformation in ω, and Ω represents the overall area of a medical image, and ω represents a sub-area of Ω (ω∈Ω).

[0065] In the embodiment, the medical image is simply and conveniently locally pixel recombined by the above formula, and the pre-trained model can learn the tissue structure information of the target part.

[0066] Based on any one of the above embodiments, in an embodiment, the intensity inversion is implemented according to the following formula:

[0067] p(x) = Max(I) - x + Min(I);

[0068] Wherein, x represents the intensity of a certain pixel point in the image I, p(x) represents the intensity value of the point after intensity inversion, Max(I) represents the highest intensity value in the image I, and Min(I) represents the lowest intensity value in the image I.

[0069] In the embodiment, the medical image is simply and conveniently intensity inverted by the above formula, and the pre-trained model can learn the contrast information of the tissue and the surrounding background.

[0070] Based on any one of the above embodiments, in an embodiment, the deep learning model is specifically a U-net network model containing an attention mechanism.

[0071] Specifically, the medical image has the characteristics of small data volume, unclear distinction between foreground and background, and unclear distinction between different physiological tissues. Therefore, the deep learning model can adopt a network model more suitable for medical image processing, such as a convolutional neural network model, a fully convolutional neural network, and a U-net network. Further, the U-net network model containing an attention mechanism (Attention Gate, AG) is preferably adopted, which can use the attention mechanism to make the model pay more attention to the area with blood vessels and reduce the influence of the surrounding background area.

[0072] The U-net network model containing an attention mechanism adopted in this embodiment is more suitable for medical image segmentation and can reduce the influence of the background area and improve the blood vessel segmentation accuracy.

[0073] Based on any of the above embodiments, in one embodiment, the U-net network model containing an attention mechanism comprises a first number of down-sampling layers, a second number of convolutional layers, and a first number of up-sampling layers connected in sequence; each level of down-sampling layer is connected to the same level of up-sampling layer through a skip connection structure, and an attention module is further arranged on each skip connection structure.

[0074] Specifically, referring to Figure 2 In a preferred embodiment, the medical image is processed by three down-sampling layers, two convolutional layers, and three up-sampling layers after being input into the model, and finally outputs a label field (i.e., a blood vessel segmentation result). The three down-sampling layers are composed of two 3x3x3 convolutional layers and one 2x2x2 pooling layer; the intermediate convolutional layer is two 2x2x2 convolutional operations; the up-sampling part is composed of one 2x2x2 deconvolutional layer and two 3x3x3 convolutional layers; finally, one 1x1x1 convolution is performed to obtain the output label field. Before the down-sampled feature map enters the up-sampling module through the skip connection structure, it will first pass through the attention module (Attention Gate), which will scale it and then judge whether it is a pixel point of the target tissue structure (i.e., the blood vessel area to be focused on) pixel by pixel, so as to improve the sensitivity of the U-net network to the target tissue structure and suppress the sensitivity of the background area.

[0075] In this embodiment, the influence of the background area is reduced by introducing the attention mechanism, and the blood vessel segmentation accuracy is improved.

[0076] Based on any of the above embodiments, in one embodiment, before the medical image to be segmented is input into the blood vessel segmentation model, the following steps are further included:

[0077] The medical image to be segmented is data enhanced by at least one of the following methods: gray scale scaling, rotation, inversion, and translation.

[0078] Specifically, the purpose of data enhancement is to preprocess the medical image to be processed, adjust it to a form convenient for the blood vessel segmentation model to process, and avoid unnecessary interference factors. Gray scale scaling refers to scaling the gray scale (i.e., intensity) of the image to a certain gray scale interval. Gray scale scaling can avoid interference caused by different medical image acquisition devices and different image acquisition environments. Rotation and inversion refer to adjusting the direction of the medical image. Translation refers to adjusting the position of the target tissue structure in the medical image (for example, if the shooting position is low, the target tissue structure is adjusted to the middle region of the medical image).

[0079] In this embodiment, by performing data enhancement on the medical image to be segmented, the medical image is adjusted to a form convenient for the blood vessel segmentation model to process, unnecessary interference factors are avoided, and the blood vessel segmentation accuracy is improved.

[0080] Next, a medical image segmentation model training method provided by the present application is described.

[0081] The medical image segmentation model training method described below can be mutually referred to the blood vessel segmentation method based on the deep learning model described above. The medical image segmentation model provided by the present application

[0082] The segmentation model training method comprises:

[0083] Performing image transformation on the medical image to obtain a transformed image;

[0084] Taking the transformed image as a sample and the medical image as a label to form a pre-training sample;

[0085] 0According to the pre-training sample, the deep learning model is pre-trained;

[0086] The deep learning model is trained again using the formal training sample to obtain a medical image segmentation model.

[0087] Specifically, the medical image can be a medical image acquired by computed tomography (CT), X-ray film, magnetic resonance imaging (MRT), positron emission tomography (PET), etc. for the target part, such as the head, the lung, the eye, etc. Only

[0088] Segmenting a two-dimensional medical image will limit the structure of the network and also make the network unable to obtain the spatial information contained in the 3D medical image. Three-dimensional medical images provide more abundant spatial information, which helps to provide more comprehensive and detailed information for doctors. However, it is difficult to annotate a three-dimensional medical image, which takes a long time and usually has a small amount of annotated sample data. The medical image in the present method refers to a three-dimensional medical image, i.e., the present method is for three-dimensional medical

[0089] images. The medical image in the present method refers to a three-dimensional medical image, i.e., the present method is for three-dimensional medical

[0090] The image is segmented.

[0091] First, pre-training samples are created through image transformation to train the deep learning model. The medical images here are unannotated medical images; image transformation is used to transform the medical images...

[0092] The images are transformed and adjusted to obtain transformed images, resulting in differences in biological tissue structures or pixel features between the images before and after the transformation. Then, using the transformed images as samples and the original medical images as corresponding labels, pre-training samples are formed and used as training data to pre-train a deep learning model. It is understandable that the labels in the pre-training samples do not contain specific label information, such as "brain region" and "non-brain region," "blood vessel" and "non-blood vessel," or "tumor" and "non-tumor." The process of obtaining pre-training samples does not require manual annotation; it can be achieved simply through convenient image transformation.

[0093] Pre-training enables the trained deep learning model to learn the biological tissue structure and pixel features of the target site. The pre-trained model possesses a good foundation for transfer learning and can be used for further training to perform different functional applications, such as segmenting specific tissue structures (e.g., brain regions, blood vessels, tumors). Based on pre-training, the deep learning model is then trained a second time using formal training samples. This allows the model to combine the biological tissue structure and pixel features of the target site learned during pre-training to further identify specific tissue structures. It is understood that the medical images in the formal training samples are also three-dimensional and annotated, containing specific label information for each pixel, such as "brain region" and "non-brain region," "blood vessel" and "non-blood vessel," and "tumor" and "non-tumor," etc.

[0094] The deep learning model is then retrained using formal training samples, enabling it to identify specific tissue structures within the target area. The retrained deep learning model then becomes a medical image segmentation model (such as a brain region segmentation model, blood vessel segmentation model, tumor segmentation model, etc.) that meets the segmentation requirements of the corresponding tissue structures. It can be used to process medical images to be segmented and output segmented medical images. The specific segmentation target can be flexibly adjusted according to actual needs; no restrictions are imposed here.

[0095] The pre-training sample is generated through image transformation in the embodiment, the deep learning model is pre-trained, the model learns the biological tissue structure features and pixel features of the target part, has a good migration learning foundation, and the pre-training process does not need manual labeling of the medical image sample, thereby reducing the model training cost; the deep learning model is secondarily trained on the basis of the pre-training through the formal training sample labeled with specific labels, the model can further identify specific tissue structures, and the medical image segmentation precision is improved.

[0096] The deep learning model-based blood vessel segmentation device provided in the application is described below, and the deep learning model-based blood vessel segmentation device described below can be mutually corresponding with the deep learning model-based blood vessel segmentation device described above.

[0097] Figure 3 is a structural schematic diagram of the deep learning model-based blood vessel segmentation device provided in the application, as Figure 3 indicated, the device comprises:

[0098] The transformation module 31 is configured to perform image transformation on the medical image to obtain a transformed image.

[0099] The sample module 32 is configured to form a pre-training sample by taking the transformed image as a sample and taking the medical image as a label.

[0100] The pre-training module 33 is configured to pre-train a deep learning model according to the pre-training sample.

[0101] The training module 34 is configured to secondarily train the deep learning model by using a formal training sample to obtain a blood vessel segmentation model.

[0102] The segmentation module 35 is configured to input an image to be segmented into the blood vessel segmentation model to obtain a segmented blood vessel image.

[0103] The pre-training sample is generated through image transformation in the embodiment, the deep learning model is pre-trained, the model learns the biological tissue structure features and pixel features of the target part, has a good migration learning foundation, and the pre-training process does not need manual labeling of the medical image sample, thereby reducing the model training cost; the deep learning model is secondarily trained on the basis of the pre-training through the formal training sample labeled with specific labels, the model can further identify specific tissue structures, and the medical image segmentation precision is improved.

[0104] Figure 4 An example of an entity structure schematic diagram of an electronic device is shown in Figure 4As shown, the electronic device can include a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 complete communications with each other through the communications bus 440. The processor 410 can invoke the logic instructions in the memory 430 to execute all or part of the steps of the above-mentioned blood vessel segmentation method based on a deep learning model.

[0105] In addition, the logic instructions in the memory 430 described above can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0106] On the other hand, the present application also provides a computer program product, which includes a computer program stored on a non-transitory computer readable storage medium, and the computer program includes program instructions, when the program instructions are executed by a computer, the computer can execute all or part of the steps of the above-mentioned blood vessel segmentation method based on a deep learning model.

[0107] In yet another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement all or part of the steps of the above-mentioned blood vessel segmentation method based on a deep learning model.

[0108] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0109] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A blood vessel segmentation method based on a deep learning model, characterized in that, include: The medical image is transformed to obtain the transformed image; the image transformation includes at least one of the following: linear intensity transformation, local pixel recombination, intensity inversion, outward filling, and inward filling; Using the transformed images as samples and the medical images as labels, pre-training samples are formed; The deep learning model is pre-trained based on the pre-trained samples; The deep learning model is trained a second time using formal training samples to obtain a blood vessel segmentation model. The medical image to be segmented is input into the blood vessel segmentation model to obtain the segmented blood vessel image.

2. The deep learning-based blood vessel segmentation method according to claim 1, characterized in that, The linear intensity transformation is implemented according to the following formula: f(x) = kx + b; Where x represents the intensity of a pixel at a certain location, f(x) represents the intensity value of the pixel after linear intensity transformation, k is a positive integer, and b is a fixed constant.

3. The deep learning-based blood vessel segmentation method according to claim 1, characterized in that, The local pixel recombination is achieved according to the following formula: Where x represents the position of any pixel in ω, L(x) represents the target position of that pixel after pixel recombination, Ran represents random transformation within the range of ω, Ω represents the overall region of a medical image, and ω represents a sub-region of Ω (ω∈Ω).

4. The deep learning-based blood vessel segmentation method according to claim 1, characterized in that, The intensity reversal is achieved according to the following formula: p(x) = Max(I) - x + Min(I); Where x represents the intensity of a pixel in image I, p(x) represents the intensity value after the intensity of that pixel is inverted, Max(I) represents the highest intensity value in image I, and Min(I) represents the lowest intensity value in image I.

5. The deep learning-based blood vessel segmentation method according to claim 1, characterized in that, The deep learning model is specifically a U-net network model that includes an attention mechanism.

6. The deep learning-based blood vessel segmentation method according to claim 5, characterized in that, The U-net network model with an attention mechanism includes a first number of downsampling layers, a second number of convolutional layers, and a first number of upsampling layers connected in sequence. Each downsampling layer is connected to the upsampling layer at the same level through a jumper structure, and each jumper structure is also equipped with an attention module.

7. The deep learning-based blood vessel segmentation method according to claim 6, characterized in that, Before inputting the medical image to be segmented into the blood vessel segmentation model, the process also includes: At least one of the following methods is used to augment the medical image to be segmented: grayscale scaling, rotation, inversion, and translation.

8. A blood vessel segmentation device based on a deep learning model, characterized in that, include: The transformation module is used to perform image transformation on medical images to obtain transformed images; the image transformation includes at least one of the following: linear intensity transformation, local pixel recombination, intensity inversion, outward filling, and inward filling; The sample module is used to form pre-training samples using the transformed image as a sample and the medical image as a label; The pre-training module is used to pre-train the deep learning model based on the pre-training samples; The training module is used to perform secondary training on the deep learning model using formal training samples to obtain a blood vessel segmentation model. The segmentation module is used to input the image to be segmented into the blood vessel segmentation model to obtain the segmented blood vessel image.

9. A method for training a medical image segmentation model, characterized in that, include: The medical image is transformed to obtain the transformed image; the image transformation includes at least one of the following: linear intensity transformation, local pixel recombination, intensity inversion, outward filling, and inward filling; Using the transformed images as samples and the medical images as labels, pre-training samples are formed; The deep learning model is pre-trained based on the pre-trained samples; The deep learning model is trained a second time using formal training samples to obtain a medical image segmentation model.

Images