Backbone identification method and related system for target network in medical images

Abstract

The present disclosure provides a trunk recognition method and a related system for a target network in a medical image, wherein the trunk recognition method comprises obtaining a first image to be recognized, the first image being a medical image and comprising a target network having a trunk and branches, recognizing a branch region in the target network, and recognizing a segment in the target network based on the branch region to obtain a segment set; recognizing the trunk of the target network in a second image by a first model based on deep learning to obtain a prediction result corresponding to the second image, and determining the confidence that the segment in the segment set belongs to the trunk based on the prediction result, wherein the difference between the trunk and the branch in the training image of the first model is greater than a first preset difference; selecting a segment from the segment set as an initial trunk based on the confidence; and determining a segment identifier of the segment in the segment set based on the initial trunk and the geometric parameters of the segment in the segment set, wherein the segment identifier is used to distinguish whether it is a trunk. According to the present disclosure, the accuracy of trunk recognition can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and in particular to a backbone recognition method for target networks in medical images and a related system. Background Art

[0002] Corneal nerve segmentation technology is of great significance for assisting the diagnosis of corneal diseases (e.g., diabetic neuropathy) and the study of corneal physiological functions. For example, the parameters of various corneal nerves can be calculated based on the corneal neural network obtained by segmentation. Corneal nerves are a vital component of corneal tissue. They are not only responsible for sensing external stimuli, but also participate in the nutrition supply and metabolic regulation of the cornea. Therefore, accurately segmenting corneal nerve images, especially distinguishing the main trunk, is a very challenging task.

[0003] In recent years, with the rapid development of deep learning technology, neural network-based corneal nerve segmentation methods have gradually become mainstream. These methods learn the characteristic representation of corneal nerves by training a large amount of image data, thereby achieving automatic segmentation of corneal nerves.

[0004] However, although the existing corneal nerve segmentation technology has achieved certain results, the trunk and branches of the corneal nerves often have great similarities in grayscale, morphology and direction, which makes them difficult to distinguish in the image. Summary of the invention

[0005] The present disclosure is proposed in view of the above situation, and its purpose is to provide a backbone recognition method and related system for target networks in medical images that can improve the accuracy of backbone recognition.

[0006] To this end, a first aspect of the present disclosure provides a method for identifying a trunk of a target network in a medical image, comprising: acquiring a first image to be identified, the first image being a medical image and comprising a target network having a trunk and branches, identifying branch regions in the target network, and identifying segments in the target network based on the branch regions to obtain a segment set; identifying the trunk of the target network in a second image by using a first model based on deep learning to obtain a prediction result corresponding to the second image, and determining a confidence level that a segment in the segment set belongs to the trunk based on the prediction result, wherein a difference between the trunk and the branches in a training image of the first model is greater than a first preset difference; selecting a segment from the segment set as an initial trunk based on the confidence level; and determining a segment identifier of the segment in the segment set based on the initial trunk and geometric parameters of the segments in the segment set, wherein the segment identifier is used to distinguish whether it is a trunk.

[0007] In the first aspect of the present disclosure, a first model based on deep learning obtains the trunk confidence of a segment, an initial trunk is determined by the trunk confidence of the segment, and a segment identifier of the segment is determined by combining the geometric parameters of the initial trunk and the segment. The first model can be used to automatically identify the trunk without relying on the first model to accurately identify all trunks, that is, the accuracy of trunk identification can be improved while reducing the accuracy requirements of the first model by combining the geometric parameters of the segment. In addition, the geometric parameters usually remain relatively stable under different acquisition conditions, thereby improving the robustness of trunk identification. In addition, the difference between the trunk and branches in the training image of the first model is greater than the first preset difference, which can reduce the interference of the training image in which the trunk and branches are not obvious on the first model to distinguish the trunk and branches. Compared with identifying branches or unclear trunks, the accuracy of identifying obvious trunks is higher, which makes it easier to identify the most obvious trunk, which is beneficial to improve the accuracy of subsequent initial trunk identification.

[0008] In addition, in the method for identifying the backbone of a target network in a medical image involved in the first aspect of the present disclosure, optionally, a binary image of the first image is obtained through a second model, the binary image is used to distinguish between the background and the target network, the segment set is the segment of the target network in the binary image, and the second image is the binary image. In this case, the second model focuses on identifying the target network in the first image, and the first model focuses on identifying the backbone of the target network, which facilitates targeted optimization of the identification target, thereby improving the accuracy of target identification. In addition, the binary image can eliminate the interference of some content that is not related to the identification backbone, and can improve the accuracy of the first model in identifying the backbone.

[0009] In addition, in the method for identifying the backbone of a target network in a medical image according to the first aspect of the present disclosure, optionally, a target skeleton is determined based on the binary image, a skeleton segment is obtained based on the target skeleton and the branch area, and a corresponding pixel area in the binary image is identified based on the skeleton segment to determine the segment of the target network in the binary image. In this case, obtaining the segment based on the target skeleton makes the key features that are conducive to segment identification more prominent, thereby improving the accuracy of segment identification.

[0010] In addition, in the backbone identification method for the target network in the medical image involved in the first aspect of the present disclosure, optionally, before the branch area is used to identify the skeleton segment, the branch area in the target skeleton is expanded, and the expanded branch area is used to determine a new branch area. In this case, the branch area that is misprocessed as a segment can be re-identified as a branch area, thereby minimizing the difference between the target skeleton and the actual skeleton, improving the accuracy of the target skeleton, and thereby improving the accuracy of segment identification.

[0011] In addition, in the backbone identification method for a target network in a medical image involved in the first aspect of the present disclosure, optionally, a segment is selected from the segment set as the initial backbone based on the confidence and the length of the segment. Thus, combining the length of the segment can reduce the possibility of the segment being misidentified as a backbone and then misselected as an initial backbone.

[0012] In addition, in the backbone identification method for the target network in the medical image involved in the first aspect of the present disclosure, optionally, if the confidence is greater than the first preset confidence and the length of the segment is greater than the first preset length, or the confidence is greater than the second preset confidence and the length of the segment is greater than the second preset length, the segment is used as the initial backbone, wherein the first preset confidence is greater than the second preset confidence, and the first preset length is less than the second preset length. In this case, the influence of the backbone confidence and length on the initial backbone identification can be balanced, reducing the risk of deviation in the initial backbone identification due to the restriction of a single parameter under specific conditions, thereby improving the accuracy of the initial backbone identification.

[0013] In addition, in the method for identifying the backbone of a target network in a medical image according to the first aspect of the present disclosure, optionally, based on the initial backbone and the geometric parameters of the segments in the segment set, the segment identification of the segments in the segment set is determined in an iterative manner, and in each iteration: a branch region set corresponding to the backbone segment is obtained; a branch segment set corresponding to each branch region in the branch region set is obtained; and the segment identification of the segments in the branch segment set is determined based on the backbone segment and the geometric parameters of the segments in the branch segment set, wherein the backbone segment in the first iteration is the initial backbone, and in the iterations after the first iteration, the backbone segment is the segment that has been determined as the backbone. Thus, other backbones can be identified gradually from the initial backbone based on the structure of the target network.

[0014] In addition, in the trunk identification method for the target network in the medical image involved in the first aspect of the present disclosure, optionally, the number of segments in the branch segment set is not less than 3, and the geometric parameters include parallelism; if there are two segments in the branch segment set as the trunk, then the segment identification of the other segments in the branch segment set except the two segments is updated as non-trunk; and if there are no two segments in the branch segment set as the trunk, and the difference in parallelism between the two segments except the trunk segment and the trunk segment is not less than the second preset difference, then the segment identification of the segment with the largest parallelism with the trunk segment in the branch segment set is updated as the trunk, and the segment identification of the other segments other than the trunk segment is updated as non-trunk. In this case, considering that most of the connected trunks will not bend too much, the segment identification of most segments can be determined by determining the geometric relationship between the trunk segments and the segments through parallelism. In addition, when the parallelism of a segment relative to the trunk segment is too close, determining the trunk based solely on the parallelism with the trunk segment may be biased. Combining the difference in parallelism can improve the accuracy of considering the segment with the largest parallelism as the trunk.

[0015] In addition, in the trunk identification method for a target network in a medical image involved in the first aspect of the present disclosure, optionally, the geometric parameter also includes width, and if the difference in parallelism between two segments other than the trunk segment and the trunk segment is less than the second preset difference, the segment identification of the segment with the largest width in the branch segment set and not the trunk segment is updated to the trunk, and the segment identification of another segment not the trunk segment is updated to the non-trunk. Thus, the segment identification of the segment that is too close in parallelism to the trunk segment can be determined in combination with the width.

[0016] A second aspect of the present disclosure provides a backbone recognition system for a target network in a medical image, comprising at least one processing circuit, wherein the at least one processing circuit is configured to: execute the backbone recognition method for a target network in a medical image involved in the first aspect of the present disclosure.

[0017] According to the present disclosure, a backbone identification method for a target network in a medical image and a related system are provided, which can improve the accuracy of backbone identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present disclosure will now be explained in further detail, by way of example only, with reference to the accompanying drawings.

[0019] Figure 1 is an exemplary schematic diagram showing an exemplary environment for recognizing corneal nerve images according to an example of the present disclosure.

[0020] Figure 2is an exemplary flow chart showing an identification method involved in the examples of the present disclosure.

[0021] Figure 3A is a schematic diagram showing a first image involved in the example of the present disclosure.

[0022] Figure 3B is a schematic diagram showing a binary image involved in the example of the present disclosure.

[0023] Figure 3C is a schematic diagram showing an image corresponding to a target skeleton involved in the example of the present disclosure.

[0024] Figure 3D is a schematic diagram showing branch areas involved in the examples of the present disclosure.

[0025] Figure 4 is an exemplary flow chart showing the process of obtaining segments in a binary image according to an example of the present disclosure.

[0026] Figure 5 is an exemplary flow chart showing the process of acquiring segment sets and trunk confidences based on a binary image according to an example of the present disclosure.

[0027] Figure 6 is an exemplary flow chart showing each iteration involved in the examples of the present disclosure.

[0028] Figure 7 FIG. 1 is an exemplary flow chart showing the process of identifying a trunk in a branch segment concentration according to an example of the present disclosure. DETAILED DESCRIPTION

[0029] Hereinafter, with reference to the accompanying drawings, preferred embodiments of the present disclosure are described in detail. In the following description, the same symbols are given to the same components, and repeated descriptions are omitted. In addition, the accompanying drawings are only schematic diagrams, and the ratio of the dimensions of the components to each other or the shapes of the components, etc. may be different from the actual ones. It should be noted that the terms "including" and "having" in the present disclosure and any variations thereof, such as a process, method, system, product or device including or having a series of steps or units, are not necessarily limited to those steps or units clearly listed, but may include or have other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] First, relevant terms involved in the present disclosure are introduced.

[0031] "Medical image" may refer to an image that can reflect the structure or function of a tissue region or organ region of a living body. For example, medical images may include but are not limited to fundus images, corneal nerve images, lung images, stomach images, chest images, and brain images. Therefore, the scheme involved in the present disclosure can be used to identify small targets in medical images. For example, for corneal nerve images, the corneal nerve, the trunk and / or branches of the corneal nerve can be identified.

[0032] "Target network" may refer to the network structure of a target object in a medical image. The target object may be a target to be identified in a medical image. For a medical image, the target object may be a corresponding tissue or organ. For example, if the target object is a blood vessel, the medical image may be a fundus image, the target network may be a vascular network, and the vascular network may be a network structure of the blood vessel. For example, if the target object is a corneal nerve, the medical image may be a corneal nerve image, the target network may be a corneal neural network, and the corneal neural network may be a network structure of the corneal nerve. In the present disclosure, the target network may have a trunk and branches. The trunk in the target network may be identified using the scheme involved in the examples of the present disclosure.

[0033] "Neural network" can refer to a computing model that mimics the way neurons in the human brain are connected. By learning the characteristics of training data, the model can be used to recognize images, predict outcomes, or perform other complex tasks. Models based on deep learning are also a type of neural network.

[0034] "Annotation result" may refer to the label of the image determined after the image is annotated. It may also be called the gold standard. In the field of artificial intelligence, the annotation result can be used to indicate what the neural network needs to predict or learn. In addition, "annotation" may refer to the process of manually labeling or annotating a data set (such as training data). Annotation may include operations such as classification, bounding box labeling, or semantic segmentation of images, text, or other types of data. For example, for corneal nerve images, the labels provided by the annotation may be labels for target areas on the corneal nerve image (such as corneal nerve areas, trunk areas, or branch areas).

[0035] The cornea referred to in the present disclosure may refer to a transparent dome-shaped structure at the front of the eye. The cornea plays a vital role in focusing light onto the retina.

[0036] The examples disclosed herein relate to a method for backbone recognition of a target network in a medical image, which determines the segment identification of a segment by combining the backbone confidence obtained by a first model based on deep learning and the geometric parameters of the segment, and can improve the accuracy of backbone recognition while reducing the accuracy requirement of the first model. In addition, the method for backbone recognition of a target network in a medical image involved in the examples disclosed herein may also be referred to as a backbone recognition method, a recognition method, or a detection method, etc. In some examples, the medical image may be a corneal nerve image, and accordingly, the recognition method may also be referred to as a backbone recognition method for a corneal neural network in a corneal nerve image.

[0037] For the convenience of description below, some examples are described using the target object as the corneal nerve, and the medical image is further described using the corneal nerve image as an example. It should be noted that this does not limit the present disclosure, and unless there is a contradiction, the relevant description is also applicable to other types of images including the target network. In addition, although the same terms as the recognition process are mentioned in the training process, unless there is a contradiction, it may refer to the training image or the corresponding content in the training image. For example, for the target network, in the recognition stage it may refer to the target network corresponding to the first image to be recognized, and in the training stage it may refer to the target network corresponding to the training image, that is, the target network in the first image used for training.

[0038] Examples of the present disclosure will be described in detail below with reference to the accompanying drawings. Figure 1 is an exemplary schematic diagram showing an exemplary environment for recognizing corneal nerve images according to an example of the present disclosure.

[0039] refer to Figure 1 In some examples, the recognition environment may include a collection device 6 and a computing device 7. The collection device 6 may be used to collect image data of a cornea 8 of a subject. For example, the collection device 6 may be a confocal microscope. The collection device 6 may transmit the collected image data to the computing device 7.

[0040] In some examples, an observer 9 (eg, a doctor) may observe the cornea 8 of the subject through the acquisition device 6 to adjust the imaging field of view and / or identify the state of the cornea 8 .

[0041] In addition, the computing device 7 can receive the image data and identify the backbone in the image data by executing the recognition method involved in the example of the present disclosure. In addition, the computing device 7 can be any device with computing capabilities. For example, the computing device 7 can be a cloud server, a personal computer, a mainframe, and a distributed computing system.

[0042] In some examples, the acquisition device 6 may also operate as an independent device and have an integrated computing device 7. That is, the acquisition device 6 may also identify the backbone in the image data by executing the identification method involved in the example of the present disclosure.

[0043] Figure 2 is an exemplary flow chart showing an identification method involved in the examples of the present disclosure. Figure 3A is a schematic diagram showing a first image involved in the example of the present disclosure. Figure 3B is a schematic diagram showing a binary image involved in the example of the present disclosure. Figure 3C is a schematic diagram showing an image corresponding to a target skeleton involved in the example of the present disclosure. Figure 3D is a schematic diagram showing the branch area involved in the example of the present disclosure. The identification method involved in the example of the present disclosure can be implemented by at least a part of the components in the above-mentioned computing device 7. For example, the identification method can be implemented by the processor of the computing device 7.

[0044] refer to Figure 2 , the recognition method may include: obtaining a segment set corresponding to a first image to be recognized (step S110), predicting a second image related to the first image through a first model based on deep learning to determine the confidence that the segment in the segment set belongs to the trunk (step S120), selecting a segment from the segment set as an initial trunk based on the confidence that the segment belongs to the trunk (step S130), and determining the segment identification of the segment in the segment set based on the initial trunk and the geometric parameters of the segment in the segment set (step S140). Generally speaking, the trunk and branches of small targets (such as corneal nerves) often have great similarities in grayscale, morphology and direction, and it is generally difficult for the model to accurately recognize all trunks. In this case, the first model based on deep learning obtains the trunk confidence of the segment, determines the initial trunk by the trunk confidence of the segment, and determines the segment identification of the segment by combining the initial trunk and the geometric parameters of the segment. The first model can be used to automatically identify the trunk but does not rely on the first model to accurately identify all trunks. That is, the combination of the geometric parameters of the segment can improve the accuracy of trunk identification while reducing the accuracy requirements of the first model. In addition, the geometric parameters usually remain relatively stable under different acquisition conditions, thereby improving the robustness of trunk identification.

[0045] Continue to refer Figure 2 In some examples, in step S110, the first image may include a target network. For example, the first image may be a corneal nerve image, and the corneal nerve image may include a corneal neural network. In addition, as described above, the target network may have a trunk and branches.

[0046] In some examples, the first image may be a medical image. Thus, the target network of small targets, the trunk and / or branches of the target network, etc., can be identified using the scheme involved in the examples of the present disclosure. In some examples, the first image may be acquired by the acquisition device 6. For example, for the relevant tissue of the fundus, the acquisition device 6 may include a fundus camera, a confocal microscope, an optic nerve analyzer, or a fundus angiography machine. As an example, Figure 3A The first image is a schematic diagram of an image of corneal nerves acquired by a confocal microscope.

[0047] In some examples, the first image may be a color image or a grayscale image. Preferably, the first image may be a grayscale image. In this case, the grayscale image can reduce the amount of data to be processed, thereby speeding up image processing. In addition, the grayscale image reflects the information in the image through brightness changes, which can highlight the structure of the target network.

[0048] In addition, the segment set may include segments of the target network. A segment may be a target object segment (eg, a nerve segment or a blood vessel segment) between branch regions or between a branch region and an end region in the target network.

[0049] In some examples, in step S110 , a binary image of the first image may be acquired, and subsequent correlation processing (described later) may be performed based on the binary image.

[0050] In addition, binary images can be used to distinguish between background and target. That is, binary images can be used to distinguish between background and target network. Binary images generally have only two pixel values. For example, the two pixel values ​​are 0 and 1, respectively, 1 can represent the target in the first image, and 0 can represent the background in the first image. In some examples, for corneal nerve images, the two pixel values ​​can represent the background and corneal nerves, respectively.

[0051] In some examples, a binary image of the first image may be obtained through the second model. Here, for the purpose of describing the binary image, the second model is mentioned first, and the first model is described later.

[0052] In addition, the second model can be any model that can convert the first image into a binary image. Preferably, the second model can be a deep learning model. Thus, the characteristics of the target network can be automatically identified. For example, the first image with a size of 384×384 can be input into the second model to obtain a binary image.

[0053] In some examples, the second model may be based on UNET. Preferably, the second model may be U2NETP. U2NETP is a lightweight improved version of U2Net, with a smaller model size and fewer parameters, and can process large images and batch data faster while improving segmentation effects. As an example, Figure 3Bshows the U2NETP Figure 3A The binary image is obtained by segmenting the corneal nerve image in .

[0054] In some examples, during training, the input of the second model may be a training image including the target network, and the output of the second model may be a binary image of the training image. In some examples, the annotation result of the training image of the second model may be a label for the corresponding area of ​​the target object in the training image. In some examples, the annotation label of the training image of the second model may include background and target object. For example, for a corneal nerve image, the annotation label may include background and corneal nerve.

[0055] In some examples, data enhancement may be performed on the training images. That is, the training images used to train the second model may include the training images and the data-enhanced training images. Thus, the sample size can be expanded and the generalization ability of the first model can be improved. In some examples, the data enhancement method may include at least one of size normalization, cropping, flipping, rotating, increasing color disturbance, increasing brightness disturbance, and increasing contrast disturbance.

[0056] In addition, size normalization may be to adjust the size of the training images to the same size, thereby facilitating input into the first model for training.

[0057] In addition, the cropped image may be randomly cropped after increasing the training image to a specified size. Thus, it is possible to avoid the first model from being overly dependent on the edge information of the target network and to increase the classification capability of the model for different targets.

[0058] In addition, flipping and rotating can increase the diversity of training data, improve the generalization ability of the first model, and reduce the overfitting risk of the first model.

[0059] In addition, adding color, brightness and contrast disturbances can increase the diversity of training data, and can simulate images collected under different conditions, thereby improving the robustness and generalization ability of the first model. In addition, corneal nerves generally vary in brightness, and increasing brightness and contrast disturbances can especially improve the generalization ability of the first model to identify corneal nerves. In some examples, the above-mentioned disturbances can be random. In this way, the diversity of training data can be further increased.

[0060] In addition, for the second model, the present disclosure also provides a training example in which the training image is a corneal nerve image, which does not represent a limitation of the present disclosure.

[0061] Specifically, more than 200 corneal nerve images from 37 different subjects can be collected as training images, and the contours of the corneal nerves can be annotated using annotation tools under the guidance of professional physicians and stored in json format. The training images are divided into training set, validation set and test set in a ratio of 8:1:1 and used as training data.

[0062] In addition, the second model can be U2NETP, the optimizer can be an SGD (Stochastic Gradient Descent) optimizer, the learning rate optimization strategy can be a Polynomial Decay strategy, and the loss function can be a cross entropy function, wherein the parameter settings of the SGD optimizer can be: momentum (momentum) = 0.9, weight_decay (weight decay) = 4E-5, and the parameter settings of the Polynomial Decay strategy can be: learning_rate (initial learning rate) = 0.01, end_lr (final learning rate) = 0, power (decay exponent) = 0.9. The final mIoU (mean Intersection over Union) is greater than 0.8 (for example, 0.839), and the single inference time is less than 0.1 seconds, which is suitable for deployment in most computing devices 7.

[0063] In some examples, the trained second model can be converted into an inference model and exported as an ONNX (OpenNeural Network Exchange) general inference model. In this case, enabling the model to be exchanged and deployed between different frameworks and platforms can improve the flexibility of model deployment and be applicable to various application scenarios.

[0064] In addition, the segment set may be derived from the first image or an image further processed from the first image (e.g., a binary image). That is, the segment set may be derived from any image related to the first image and capable of acquiring segments of the target network. In some examples, the segment set may be segments of the target network in the binary image. Thus, segments with clear boundaries may be obtained.

[0065] In some examples, segments with a width less than a preset width can be screened out from the segment set. In this way, noise points can be filtered. Specifically, the area corresponding to the noise point may be identified as a segment, but the width of the noise point is generally not large, and part of the noise points can be filtered out by judging the width.

[0066] In some examples, a branch region in the target network in the first image may be identified, and a segment in the target network may be identified based on the branch region to obtain a segment set.

[0067] In addition, a branch region may be a region with three or more segments around it. That is, a branch region may be a pixel point connecting at least three segments. In addition, a region with only one segment around it may be considered as an end region. That is, an end region may be a pixel point connected to only one segment. In addition, a region with only two segments around it may be considered as a segment region (i.e., a region on a segment).

[0068] In addition, a branch region may include at least one pixel. In some examples, a branch region including only one pixel may also be referred to as a branch point. In addition, an endpoint region including only one pixel may also be referred to as an endpoint. In addition, a segment region including only one pixel may also be referred to as a segment point.

[0069] In some examples, a target skeleton (i.e., a skeleton of a target object) corresponding to the first image may be obtained, and a branch region may be identified based on the target skeleton. In some examples, in the target skeleton, a point with 3 or more pixels around it may be considered a branch region. In addition, a point with only 1 pixel around it may be considered an endpoint region. A point with only 2 pixels around it may be considered a segment region.

[0070] In addition, there are many methods for identifying branch areas. For example, the corresponding pixel points can be found from the target skeleton as branch areas according to the characteristics of the branch area. Preferably, the Sobel operator can be used to find the branch area in the target skeleton. In this way, the time complexity can be reduced, and the efficiency of the search can be improved. For example, the size of the convolution kernel of the Sobel operator can be 3×3, the center of the convolution kernel can be 10, and the edge can be 1. The image corresponding to the target skeleton is subjected to 2D convolution using the convolution kernel, and points with 3 or more pixels around are regarded as branch areas (that is, value (result of the convolution operation)>=13). In some examples, points with only 1 pixel around can also be regarded as endpoint areas (that is, value=11), and points with only 2 pixels around can be regarded as segment areas (that is, value=12). In addition, the convolution operation can be to multiply the pixels at the corresponding positions of the convolution kernel and the target skeleton and then add them.

[0071] In some examples, a target skeleton may be determined based on a binary image of the first image. In some examples, the binary image may be thinned to obtain the target skeleton. Specifically, the target network in the binary image may be thinned to a width of one pixel to form the target skeleton, and the basic topological structure of the target object shape may be kept unchanged. For example, for a corneal nerve image, the binary image may be thinned to obtain a corneal nerve skeleton. As an example of a target skeleton, Figure 3C Shows the Figure 3B Schematic representation of the corneal neural skeleton obtained after thinning of the binary image.

[0072] In some examples, a thinning algorithm may be used to thin the binary image to extract the target skeleton. In some examples, the thinning algorithm may include but is not limited to the Zhang-Suen algorithm, the Guo-Hall algorithm, the Thinning algorithm, the Medial Axis Transform (MAT) algorithm, the Voronoi Skeletonization algorithm, or the Distance Transform Based Methods algorithm.

[0073] The inventors have found that due to the limitations of the thinning algorithm itself, the actual single branch area may be thinned into a shorter line segment, and some small branch segments that do not actually exist may be generated. As an example, Figure 3C The area D11 in FIG. 1 shows a small branch segment that does not actually exist. This area should be a branch area, and should be a branch point on the target skeleton. Therefore, it is necessary to identify the small branch segment to determine whether it is a branch area.

[0074] In some examples, the branch regions may be processed before being used to identify segments or skeleton segments, thereby improving the accuracy of subsequent processing. In addition, the processed branch regions may be used to determine new branch regions for subsequent processing (e.g., segment identification or skeleton segment identification).

[0075] In some examples, a dilation operation can be used to re-identify a small branch segment as a branch region. Specifically, processing the branch region may include dilating the branch region in the target skeleton, and determining a new branch region using the dilated branch region. In this case, the branch region that was misprocessed as a segment can be re-identified as a branch region (that is, the region that does not belong to the segment can be identified), thereby minimizing the difference between the target skeleton and the actual skeleton, improving the accuracy of the target skeleton, and thereby improving the accuracy of segment recognition.

[0076] In some examples, the size of the expanded structural element may be 3×3. This facilitates the expansion of a target skeleton with a width of 1 pixel. In some examples, the number of dilation iterations may be 3. In this case, it is convenient to convert small branch segments back into branch regions (i.e., connect two branch regions), but some originally shorter segments will not be converted into branch regions due to excessive expansion. For example, if two branch regions are very close, they will generally be connected together after three expansions. If they are not connected together after three expansions, it may indicate that the distance between the two branch regions is relatively far.

[0077] As described above, the processed branch regions may be used to determine new branch regions. In some examples, the branch regions after eliminating small branch segments may be used as new branch regions.

[0078] In some examples, the intersection of the processed branch region and the target skeleton can be taken as the new branch region. That is, the portion of the processed branch region that does not exceed the target skeleton is taken as the new branch region. Taking the expansion operation as an example, the intersection of the expanded branch region and the target skeleton can be taken as the new branch region. In addition, as described above, the branch region can be identified based on the target skeleton. As an example, Figure 3D A schematic diagram of the branching areas in the corneal neural skeleton after dilation operation is shown, where the green parts are the branching areas.

[0079] Figure 4 is an exemplary flow chart showing the process of obtaining segments in a binary image according to an example of the present disclosure.

[0080] In some examples, a segment set can be obtained based on a binary image of the first image. Specifically, a branch region of a target network in the binary image can be identified, and segments in the target network can be identified based on the branch region to obtain a segment set. To this end, the disclosed example also provides an exemplary process for obtaining segments in a binary image, where segments are obtained based on a target skeleton, so that key features that are conducive to segment identification are more prominent, thereby improving the accuracy of segment identification. Specifically, refer to Figure 4 , the exemplary process includes:

[0081] Step S210: determining the target skeleton based on the binary image.

[0082] Step S220: Obtain skeleton segments based on the target skeleton and the branch regions.

[0083] As described above, in some examples, branch regions can be identified based on the target skeleton (see Figure 3D ). In some examples, the target skeleton can be segmented according to branch areas and connectivity to obtain skeleton segments.

[0084] In some examples, the target skeleton may be disconnected at the branch region to obtain the skeleton segment. In some examples, the pixel value of the branch region may be set to the pixel value corresponding to the background (eg, set to 0) to disconnect the target skeleton at the branch region.

[0085] In some examples, after disconnection at the branch region, the connectivity of the target skeleton can be detected to obtain the skeleton segment. That is, the connected pixels can be regarded as the pixels of a skeleton segment. For example, the eight connected regions of the pixels can be detected to determine the connectivity.

[0086] Step S230: Identify the corresponding pixel area in the binary image based on the skeleton segment to determine the segment of the target network in the binary image. That is, the pixel area corresponding to the skeleton segment in the binary image can be identified.

[0087] In some examples, a pixel point corresponding to a skeleton segment may be searched in a binary image, and then adjacent pixel points belonging to the target object may be determined based on the pixel point, thereby determining a corresponding pixel region in the binary image.

[0088] In some examples, a distance transform may be used to identify pixel regions in a binary image that correspond to skeleton segments. Specifically, a distance transform algorithm may be used to find pixel points in a binary image that correspond to skeleton segments.

[0089] Return to reference Figure 2 In some examples, in step S110, the geometric parameters of the segment may also be obtained. In some examples, the geometric parameters may include at least one of length, area, width, angle, and parallelism. The specific parameters may be selected as needed. For example, in the process of trunk identification, the corresponding geometric parameters may be selected to identify the trunk according to the morphological characteristics of the target network.

[0090] In addition, parallelism can represent the degree of parallelism between two segments. In some examples, parallelism can be represented by the angle between the two segments, and the smaller the angle, the more parallel the two segments are, and the greater the parallelism. For example, the parallelism between a segment and a trunk segment (described later) can be represented by the angle between the segment and the trunk segment.

[0091] In some examples, in step S110, the segments, branch regions, and geometric parameters of the segments may also be stored in a storage medium, thereby facilitating subsequent processing and reading.

[0092] Continue to refer Figure 2 In some examples, in step S120, the first model predicts the second image by identifying the first model with respect to the trunk of the target network in the second image to obtain a prediction result corresponding to the second image. In some examples, the confidence that the segment in the segment set belongs to the trunk can be determined based on the prediction result.

[0093] In addition, the first model can be used to identify a trunk in the second image. In some examples, the first model can also be used to identify branches in the second image.

[0094] In addition, the prediction result corresponding to the second image can be the probability that the segment belongs to the corresponding category. The category can be a trunk or a branch. In some examples, the first model can output the probability that the segment belongs to the trunk. In some examples, the first model can also output the probability that the segment belongs to the branch.

[0095] In addition, the confidence of a segment may be the probability that the segment belongs to the corresponding category. The confidence that a segment belongs to a trunk (hereinafter referred to as trunk confidence) may be the probability that the segment belongs to a trunk. The confidence that a segment belongs to a branch (hereinafter referred to as branch confidence) may be the probability that the segment belongs to a branch.

[0096] In some examples, the prediction result may include the category to which each pixel in the second image belongs. In some examples, the probability that the segment belongs to the corresponding category may be determined based on the proportion of pixels in the segment in the corresponding category. In some examples, the probability that the segment belongs to the trunk and / or branch may be determined based on the proportion of the trunk and / or branch of pixels in the segment.

[0097] In addition, the first model may be any model capable of identifying a trunk and / or a branch. Preferably, the first model may be based on deep learning, thereby being able to automatically identify features of a trunk and / or a branch.

[0098] In some examples, the difference between the trunk and branches in the training image of the first model can be greater than the first preset difference. That is, a first image in which the appearance difference between the trunk and branches is obvious can be selected as a training image. In this case, the interference of the training image in which the trunk and branches are not obvious on the first model in distinguishing the trunk and branches can be reduced, and the accuracy of identifying the obvious trunk is higher than that of identifying branches or unclear trunks (that is, the more obvious the trunk, the higher the accuracy of being identified as the trunk), which makes it easier to identify the most obvious trunk, which is beneficial to improve the accuracy of subsequent initial trunk identification. That is, the first model is not required to accurately identify all trunks, but the accuracy of obvious trunk identification must be improved.

[0099] In addition, the present disclosure does not specifically limit the method for measuring the degree of difference between the trunk and the branches. Manual judgment can be used (for example, the annotator's eyes can directly distinguish the trunk and the branches, which can be regarded as a difference greater than the first preset difference). Computer programs can also be used to determine some thresholds for measuring the differences based on statistical analysis methods as the first preset difference. An initial first preset difference can also be set, and the first preset difference can be adjusted through the performance of the first model to improve the accuracy of the first model in identifying obvious trunks.

[0100] In some examples, when annotating the training image of the first model, the trunk may be annotated first, and then the branches. Thus, it is convenient to annotate the branches with reference to the annotation of the trunk. In some examples, the annotation labels of the training image of the first model may include background, trunk, and branches.

[0101] In addition, the second image can be any image related to the first image and including the target network. In some examples, the second image can be the above-mentioned binary image. In this case, the second model focuses on identifying the target network in the first image, and the first model focuses on identifying the backbone of the target network, which facilitates targeted optimization of the identified target, thereby improving the accuracy of target recognition. In addition, when training the first model and the second model, the annotator can also focus on one aspect of annotation, which can reduce the difficulty of annotation. In addition, the binary image can eliminate the interference of some content that is not related to the identification of the backbone, and can improve the accuracy of the first model in identifying the backbone.

[0102] In some examples, when the second image is the above-mentioned binary image, the first model and the second model can be different deep learning models. In this case, it is convenient to select corresponding models for images of different forms, which is conducive to improving overall performance. In some examples, the network structure of the first model can be simpler than the network structure of the second model. In this case, combining the first model and the second model can improve the efficiency of the recognition backbone. For example, the second model can be U2NETP, and the first model can be UNET.

[0103] In addition, for the first model, the present disclosure also provides a training example in which the training image is a binary image of a corneal nerve image, which does not represent a limitation of the present disclosure.

[0104] Specifically, more than 100 binary images can be labeled with trunks and branches, and only images with obvious differences between trunks and branches are selected to train the first model. The training images are divided into a training set, a validation set, and a test set in a ratio of 8:1:1 and used as training data.

[0105] In addition, the first model can be UNET, the optimizer can be an SGD optimizer, the learning rate optimization strategy can be a PolynomialDecay strategy, and the loss function can be a DiceLoss loss function, wherein the parameter settings of the SGD optimizer can be: momentum=0.9, weight_decay=4E-5, and the parameter settings of the PolynomialDecay strategy can be: learning_rate=0.01, end_lr=0, power=0.9.

[0106] Figure 5 is an exemplary flow chart showing the process of acquiring segment sets and trunk confidences based on a binary image according to an example of the present disclosure.

[0107] As described above, in some examples, a binary image of the first image can be obtained through the second model, the segment set can be a segment in the binary image, and the second image can be a binary image. To this end, the example of the present disclosure also provides an exemplary process around the binary image and combined with step S110 and step S120. Figure 5 , the exemplary process may include:

[0108] Step S310: obtaining a binary image of the first image through the second model.

[0109] Step S320: Acquire segments in the binary image as a segment set.

[0110] Step S330: predicting the binary image through the first model to determine the backbone confidence of the segments in the segment set. Specifically, the backbone of the target network in the binary image is identified through the first model to obtain a prediction result corresponding to the binary image, and the backbone confidence of the segments in the segment set is determined based on the prediction result.

[0111] The inventors found that there is no clear rule for distinguishing trunks and branches, and it is impossible to complete the discrimination work by relying solely on model reasoning. Therefore, the inventors combined the trunk confidence of the segment and the geometric parameters of the segment to identify the trunk in the segment set.

[0112] Return to reference Figure 2 In some examples, in step S130, a preset number of segments that are most likely to be the trunk can be selected from the segment set as the initial trunk based on the trunk confidence. That is, the first several segments that are most likely to be the trunk can be selected from the segment set as the initial trunk based on the trunk confidence.

[0113] In some examples, the preset number may be 1. In this case, for an initial backbone, it is convenient to iterate the backbone identification process starting from the segment most likely to be the backbone, thereby improving the accuracy of backbone identification.

[0114] In some examples, the preset number may be greater than 1. In this case, for multiple initial trunks, it is convenient to start the iterative trunk identification process from multiple locations, which can improve the efficiency of trunk identification and improve the accuracy of trunk identification when combining multiple identification results. For example, if the identification results corresponding to two initial trunks are the same, it can further indicate the accuracy of the identification result.

[0115] In some examples, the segments can be selected in combination with the backbone confidence and the geometric parameters of the segments. Here, the geometric parameters can be any parameters that can enhance the possibility of the segments being the backbone and are related to the shape of the segments. In some examples, the segments can be selected in combination with the length of the segments. Specifically, the segments can be selected from the segment set as the initial backbone based on the backbone confidence and the length. Thus, the possibility of the segments being misidentified as the backbone and then misselected as the initial backbone can be reduced in combination with the length of the segments.

[0116] In some examples, the trunk confidence and length can be considered comprehensively. After research, the inventor found that if the trunk confidence of the segment is high, the length of the segment is not required to be too long. If the trunk confidence of the segment is low, the length of the segment is required to be as long as possible. When selecting a segment in combination with the trunk confidence and the length of the segment, the trunk confidence and length can be negatively correlated. Specifically, if the trunk confidence is greater than the first preset confidence and the length of the segment is greater than the first preset length, or the trunk confidence is greater than the second preset confidence and the length of the segment is greater than the second preset length, the segment can be used as the initial trunk, wherein the first preset confidence can be greater than the second preset confidence, and the first preset length can be less than the second preset length. In this case, the influence of the trunk confidence and length on the initial trunk recognition can be balanced, reducing the risk that a single parameter may be restricted by specific conditions and cause deviations in the initial trunk recognition, thereby improving the accuracy of the initial trunk recognition.

[0117] In addition, the first preset reliability, the second preset reliability, the first preset length, and the second preset length can be preset values ​​adjusted based on experience, experiments, or specific needs. For example, it can be an empirical value adjusted according to the morphological characteristics of the target object. For example, for the corneal nerve image, the first preset reliability can be 0.8, the second preset reliability can be 0.4, the first preset length can be 10 pixels, and the second preset length can be 200 pixels. It should be noted that this does not represent a limitation to the present disclosure, and can be adjusted according to the morphological characteristics of the target object.

[0118] Return to reference Figure 2 In some examples, in step S140, segment identifiers of segments in the segment set (i.e., segment identifiers of segments other than the initial backbone) can be determined in an iterative manner based on the initial backbone and the geometric parameters of the segments in the segment set. Thus, other backbones can be identified gradually from the initial backbone based on the structure of the target network.

[0119] In some examples, it is possible to start from the initial trunk and repeat the following steps: determine the segment identification of the segment in the adjacent segment set by combining the geometric parameters of the trunk segment and the adjacent segment set adjacent to the trunk segment, until the segment identification of the segment in the segment set is finally determined. That is, starting from the initial trunk, a new trunk can be determined in the adjacent segment set, and iterative judgment can be performed based on this, until the segment identification of the segment in the segment set is finally determined. In addition, the adjacent segment set can be determined by the branch area corresponding to the trunk segment (i.e., the branch segment set described later).

[0120] In addition, the segment identification can be used to distinguish whether it is a trunk. In some examples, the segment identification can include a trunk and a non-trunk. In some examples, the segment identification can include a trunk and a branch. In some examples, the segment identification of a segment can be initialized as a non-trunk, and in the process of identifying the trunk, the segment identification can be updated to identify the segment identified as the trunk.

[0121] In some examples, a segment identification of a segment as a non-trunk or branch may indicate that the segment is a branch. Thus, the branch can be identified.

[0122] In addition, the trunk segment may be a segment of the trunk. The trunk segment of the first iteration may be an initial trunk, and in iterations after the first iteration, the trunk segment may be a segment that has been determined as the trunk (that is, in the iteration process, the segment is identified as a segment of the trunk).

[0123] Figure 6 is an exemplary flow chart showing each iteration involved in the examples of the present disclosure.

[0124] Figure 7 FIG. 1 is an exemplary flow chart showing the process of identifying a trunk in a branch segment concentration according to an example of the present disclosure.

[0125] As described above, in some examples, the segment identities of the segments in the segment set may be determined iteratively based on the initial trunk and the geometric parameters of the segments in the segment set. Figure 6 , in some examples, in each iteration:

[0126] Step S410: Acquire a branch region set corresponding to the trunk segment.

[0127] In addition, the branch region set corresponding to the trunk segment may include branch regions connected to the trunk segment. For each trunk segment, the pixels near its two ends may belong to the end region or the branch region, and the branch region set corresponding to the trunk segment may be the branch region therein.

[0128] Step S420: Obtain a branch segment set corresponding to each branch region in the branch region set.

[0129] In addition, the branch segment set corresponding to each branch region may include segments connected to the branch region. In some examples, the branch regions in the branch region set may be traversed, and the branch segment set corresponding to each branch region may be obtained.

[0130] In some examples, the branch segment set with a segment number exceeding a preset number can be filtered. In this case, the influence of the branch segment set with an unexpected number of segments on the iteration process can be reduced. In addition, the preset number can be determined according to the morphological characteristics of the target network. For example, for the corneal nerve image, the number of segments in the branch segment set can be no less than 3.

[0131] Step S430: determining the segment identifiers of the segments in the branch segment set (ie, the segment identifiers of the segments other than the trunk segment) based on the geometric parameters of the trunk segment and the segments in the branch segment set.

[0132] As described above, the trunk segment may be a segment of the trunk. In some examples, when it is determined that one of the segments in the branch segment set is already the trunk (e.g., the initial trunk), the geometrical parameters of the segment may be used to determine the geometrical relationship between the trunk segments and the segments and / or the shape of the segment itself, and then the segment identification of the segment may be determined using the geometrical relationship between the trunk segments and the segments and / or the shape of the segment itself.

[0133] In some examples, the segment identification of a segment can be determined by its parallelism with the trunk segment. To this end, the example disclosed herein also takes the case where the number of segments in the branch segment set is not less than 3 and the geometric parameters include parallelism as an example, and provides an exemplary process for identifying the trunk in the branch segment set. Considering that most of the connected trunks will not bend too much, the geometric relationship between the trunk segments and the segments is determined by parallelism to determine the segment identification of the segment, and the segment identification of most of the segments can be determined. In addition, when the parallelism of the segment relative to the trunk segment is too close, there may be deviations in determining the trunk based only on the size of the parallelism with the trunk segment. Combining the difference in parallelism can improve the accuracy of considering the segment with the largest parallelism as the trunk.

[0134] refer to Figure 7 , the exemplary process may include:

[0135] Step S510: Determine whether there are two segments in the branch segment set that are the main trunk.

[0136] In some examples, if there are two segments as the trunk, step S520 is performed, otherwise step S530 is performed.

[0137] Step S520: updating the segment identifiers of the segments other than the two segments in the branch segment set as non-trunk.

[0138] Step S530: Determine whether the difference between the parallelism of the two segments other than the trunk segment and the trunk segment is not less than a second preset difference. That is, the two segments and the trunk segment may correspond to two parallelisms, and determine whether the difference between the two parallelisms is not less than the second preset difference.

[0139] In some examples, if the difference is not less than the second preset difference, step S540 may be performed. In some examples, if the difference is less than the second preset difference, the judgment may continue to be made based on the width of the segment. Specifically, the geometric parameter may also include the width, and the segment identification of the segment with the largest width in the branch segment set and the non-trunk segment is updated to the trunk, and the segment identification of the segment of another non-trunk segment is updated to the non-trunk. Thus, the segment identification of the segment that is too close in parallelism to the trunk segment can be determined in combination with the width.

[0140] In addition, the second preset difference can be any value that can distinguish that the parallelism of two segments relative to the trunk segment is too close. In some examples, for the parallelism represented by the angle (that is, the angle between the trunk segment and the segment), the second preset difference can be 20° to 30°. For example, the second preset difference can be 20°, 21°, 22°, 23°, 24°, 25°, 26°, 27°, 28°, 29° or 30°, etc.

[0141] Step S540: Update the segment identifier of a segment in the branch segment set that has the largest parallelism with the trunk segment (i.e., the most parallel) to the trunk segment as the trunk, and update the segment identifiers of the segments of other non-trunk segments to non-trunk segments. For example, if the number of segments in the branch segment set is 3, the segment identifier of a segment in the branch segment set that has the largest parallelism with the trunk segment can be updated to the trunk, and the segment identifiers of the segments of other non-trunk segments can be updated to non-trunk.

[0142] In addition, the present disclosure also provides an example of using a computer program to identify the trunk in the branch segment set, which should be noted that it does not limit the present disclosure. Specifically, the branch area in the branch area set is pushed into a stack, and if the stack is not empty, a branch area is obtained by popping the stack. For this branch area, if there are already two segments in the branch segment set as the trunk, and the remaining segments are branches, if the segment identification of the remaining segments is not determined, a judgment is made based on parallelism. If the parallelism is not close, a segment that is most similar to the trunk segment is selected. If the parallelism is close, the thickest segment is selected as the trunk according to the width.

[0143] In addition, the example of the present disclosure also provides a parameter calculation method, which can use the above-mentioned recognition method to obtain the target network, trunk and / or branch for the first image to be recognized, and calculate the target parameters based on the target network, trunk and / or branch. This facilitates visualization of target object recognition and can automatically calculate various parameters, thereby achieving the purpose of auxiliary analysis (such as diagnosis).

[0144] In some examples, for corneal nerve images, the target parameters may include at least one of corneal nerve fiber density (Corneal Nerve Fiber Density, CNFD), corneal nerve branch density (Corneal Nerve Branch Density, CNBD), corneal nerve fiber length (Corneal Nerve Fiber Length, CNFL), corneal nerve total branch density (Corneal Nerve Total Branch Density, CTBD), corneal nerve fiber area (Corneal Nerve Fiber Area, CNFA) and corneal nerve fiber width (Corneal Nerve Fiber Width, CNFW).

[0145] The examples of the present disclosure also disclose a backbone recognition system for a target network in a medical image, and the backbone recognition system may include at least one processing circuit. The at least one processing circuit is configured to perform one or more steps in the above-mentioned recognition method or parameter calculation method. It should be noted that, unless there is a contradiction, the above-mentioned description of the recognition method or parameter calculation method is also applicable to the backbone recognition system.

[0146] The examples of the present disclosure also disclose a computer-readable storage medium, which can store at least one instruction, and when the at least one instruction is executed by a processor, one or more steps in the above-mentioned identification method or parameter calculation method are implemented. Wherein, the computer-readable storage medium can include but is not limited to any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a micro drive and a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nano system (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.

[0147] Although the present disclosure is specifically described above in conjunction with the accompanying drawings and examples, it is to be understood that the above description does not limit the present disclosure in any form. Those skilled in the art may modify and change the present disclosure as needed without departing from the essential spirit and scope of the present disclosure, and these modifications and changes all fall within the scope of the present disclosure.

Claims

1. A method for identifying a backbone of a target network in a medical image, characterized in that: include: Acquire a first image to be identified, the first image is a medical image and includes a target network having a trunk and branches, the target network refers to a network structure of a target object in the medical image, identify a branch area in the target network, the branch area is an area with 3 or more segments around it, identify the segments in the target network based on the branch area to obtain a segment set; identify the trunk of the target network in the second image by using a first model based on deep learning to obtain a prediction result corresponding to the second image, and determine the confidence that the segment in the segment set belongs to the trunk based on the prediction result, wherein the difference between the trunk and the branch in the training image of the first model is greater than a first preset difference, the second image is an image related to the first image and includes the target network, and the prediction result is the probability that the segment belongs to the trunk; A segment is selected from the segment set as an initial backbone based on the confidence level; and a segment identifier of the segment in the segment set is determined based on the initial backbone and geometric parameters of the segment in the segment set, wherein the segment identifier is used to distinguish whether it is a backbone.

2. The backbone identification method according to claim 1, characterized in that: A binary image of the first image is obtained through a second model, the binary image is used to distinguish between the background and the target network, the branch area of ​​the target network in the binary image is identified, and the segments of the target network in the binary image are identified based on the branch area in the binary image to obtain the segment set, wherein the segment set is the segments of the target network in the binary image; the second image is the binary image.

3. The backbone identification method according to claim 2, characterized in that: A target skeleton is determined based on the binary image, skeleton segments are acquired based on the target skeleton and the branch regions, and corresponding pixel regions in the binary image are identified based on the skeleton segments to determine segments of the target network in the binary image.

4. The backbone identification method according to claim 3, characterized in that: Before using the branch region to identify the skeleton segment, the branch region in the target skeleton is expanded, and a new branch region is determined using the expanded branch region.

5. The backbone identification method according to any one of claims 1 to 4, characterized in that: A segment is selected from the segment set as the initial backbone based on the confidence and the length of the segment.

6. The backbone identification method according to claim 5, characterized in that: If the confidence is greater than the first preset confidence and the length of the segment is greater than the first preset length, or the confidence is greater than the second preset confidence and the length of the segment is greater than the second preset length, the segment is used as the initial trunk, wherein the first preset confidence is greater than the second preset confidence, and the first preset length is less than the second preset length.

7. The backbone identification method according to claim 1, characterized in that: Based on the initial trunk and the geometric parameters of the segments in the segment set, the segment identifiers of the segments in the segment set are determined in an iterative manner, in each iteration: Acquire a branch region set corresponding to a trunk segment, wherein the branch region set includes branch regions connected to the trunk segment; Acquire a branch segment set corresponding to each branch region in the branch region set, wherein the branch segment set includes segments connected to each branch region; and The segment identifier of the segment in the branch segment set is determined based on the geometric parameters of the trunk segment and the segments in the branch segment set, wherein the trunk segment in the first iteration is the initial trunk, and in iterations after the first iteration, the trunk segment is a segment that has been determined to be the trunk.

8. The backbone identification method according to claim 7, characterized in that: The number of segments in the branch segment set is not less than 3, and the geometric parameters include parallelism; if there are two segments in the branch segment set that are the trunk, the segment identifiers of the other segments in the branch segment set except the two segments are updated to non-trunk; and if there are not two segments in the branch segment set that are the trunk, and the difference in parallelism between the two segments except the trunk segment and the trunk segment is not less than a second preset difference, the segment identifier of the segment in the branch segment set that has the largest parallelism with the trunk segment is updated to the trunk, and the segment identifiers of the other segments that are not the trunk segments are updated to non-trunk.

9. The backbone identification method according to claim 8, characterized in that: The geometric parameters also include width. If the difference in parallelism between two segments other than the trunk segment and the trunk segment is less than the second preset difference, the segment identifier of the segment with the largest width in the branch segment set and which is not the trunk segment is updated to the trunk, and the segment identifier of another segment that is not the trunk segment is updated to non-trunk.

10. A backbone recognition system for target networks in medical images, characterized in that: The method comprises at least one processing circuit, wherein the at least one processing circuit is configured to: execute the backbone identification method according to any one of claims 1 to 9.

Images