CT image vascular tree, coronary artery tree segmentation result optimization method and system

Abstract

The present invention provides a method and system for optimizing the segmentation results of CT image vascular trees and coronary artery trees. The method comprises the following steps: obtaining a segmentation result image of the vascular tree based on the original CT image; multiplying the segmented image with the original CT image by a region of interest acquisition unit to obtain a CT image containing only the vascular tree; and dividing the image into N non-overlapping image blocks by an image division unit; a random mask unit having an adaptive mask ratio; performing a normalized linear transformation on the CT values of all non-masked image blocks by a normalization unit to obtain a preprocessed segmentation result image; using a loss function combining MSE and clDice to train an optimized network model in a network model training module, and obtaining weight information of the trained optimized network model; and inputting the preprocessed segmentation result image as an input image into the trained optimized network model in a network model inference module, combining the weight information to obtain an optimized segmentation result image.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging equipment, and in particular to a method and system for optimizing segmentation results of CT image vascular trees and coronary artery trees. Background Art

[0002] With the improvement of imaging speed and scanning accuracy of CT (computed tomography) equipment, coronary three-dimensional reconstruction technology based on CTA (CT angiography) has been widely used in cardiac examination and disease diagnosis. Accurate visualization and analysis of coronary artery conditions can provide doctors with coronary artery contours, facilitating the analysis of conditions such as vascular stenosis, calcification, and plaques. It is an important clinical method for early screening of coronary atherosclerotic heart disease, and its key prerequisite is the segmentation and extraction of coronary arteries.

[0003] Currently, the segmentation of coronary arteries can often identify most of the coronary artery tree. However, due to objective factors such as the complexity and strong randomness of CT angiography images, the segmentation and extraction of the coronary arteries of the heart is difficult. The main reasons are: the coronary artery structure is complex, with many small branches and blood vessels; the coronary artery grayscale is uneven, and the distal parts of the blood vessels are small and have blurred boundaries; the coronary artery contains a variety of lesions; and the motion artifacts of the heart are unavoidable, all of which affect the imaging of the coronary arteries, often resulting in incomplete segmented coronary artery trees and over-segmentation such as adhesion to veins.

[0004] Existing coronary artery segmentation methods can be broadly categorized as fully automatic, semi-automatic, and interactive. Both semi-automatic and interactive segmentation methods require manual intervention, are tedious, and are highly subjective. Regardless of the segmentation method, over-segmentation and under-segmentation are common, resulting in low segmentation accuracy and continuity for small branches. Summary of the Invention

[0005] In order to overcome the above technical defects, the purpose of the present invention is to provide a deep learning-based CT image vascular tree segmentation result optimization method, system and CT image coronary artery tree segmentation result optimization method to improve the continuity and accuracy of the segmentation results.

[0006] The present invention discloses a method for optimizing the segmentation results of a CT image vascular tree based on deep learning, comprising the following steps: obtaining a segmentation result image of the vascular tree based on an original CT image, multiplying the segmented image with the original CT image to obtain a CT image containing only the vascular tree, and dividing the image into N non-overlapping image blocks; presetting a mask ratio a, randomly sampling the N image blocks using a uniform distribution strategy, retaining N*(1-a) image blocks, which are recorded as non-masked image blocks, and recording the remaining N*a image blocks as masked image blocks; calculating the average value of the non-masked image blocks, which is recorded as m; when m=0, randomly selecting n image blocks with non-zero average values from the masked image blocks and exchanging them with n image blocks in the non-masked image blocks to obtain a preprocessed segmentation result image; establishing an optimization network model, training the optimization network model using a loss function until the optimization network model converges; and obtaining weight information of the trained optimization network model; inputting the preprocessed segmentation result image as an input image into the trained optimization network model, and combining the weight information to obtain an optimized segmentation result image.

[0007] Preferably, after obtaining the pre-processed segmentation result image, the method further includes: performing a normalized linear transformation on the CT values of all non-masked image blocks, and the CT values Among them, v mean is the mean of the CT values v of the non-masked image block, v std is the standard deviation of the CT value v of the non-masked image block.

[0008] Preferably, when m=0, randomly selecting n image blocks with non-zero average values from the mask image blocks and exchanging them with n image blocks in the non-mask image blocks further comprises: a is 50%-75%.

[0009] Preferably, when m=0, randomly selecting n image blocks with non-zero average values from the mask image blocks and exchanging them with n image blocks in the non-mask image blocks further includes: n is less than or equal to 25% of the total number of all image blocks with non-zero average values in the mask image blocks.

[0010] Preferably, the loss function includes an MSE loss function, which is used to evaluate the mean of the sum of squares of errors between the preprocessed segmentation result image output by the optimization network model and the segmentation result image annotated by the physician; Among them, x is the segmentation result image after the physician's annotation, and y is the preprocessed segmentation result image output by the optimized network model.

[0011] Preferably, the MSE loss function is only used for mask image blocks.

[0012] Preferably, the loss function also includes the clDice loss function,

[0013] Among them, V L represents the segmentation result image after the physician’s annotation, V P represents the pre-processed segmentation result image output by the optimized network model; S P and S L Represents the V P and V L The skeleton structure extracted from T prec (S P , V L ) is the topological accuracy, indicating S P In V L The proportion of T sens (S L , V P ) is the topological sensitivity, indicating S L In V P The proportion in .

[0014] Preferably, the clDice loss function is used for masked image blocks and non-masked image blocks.

[0015] The present invention also discloses a method for optimizing the segmentation results of the coronary artery tree in a CT image based on deep learning, which adopts the above-mentioned method for optimizing the segmentation results of the vascular tree in a CT image.

[0016] The present invention also discloses a CT image vascular tree segmentation result optimization system based on deep learning, including an image processing module, a network model training module and a network model inference module; the image processing module includes a region of interest acquisition unit, an image division unit, a random mask unit and a normalization unit; the segmentation result image of the vascular tree is obtained based on the original CT image, the segmented image is multiplied by the region of interest acquisition unit with the original CT image to obtain a CT image containing only the vascular tree, and the image is divided into N non-overlapping image blocks by the image division unit; the mask ratio a is preset by the random mask unit, and the N image blocks are randomly sampled using a strategy that obeys uniform distribution, and N*(1-a) image blocks are retained and recorded as non-masked image blocks, and the remaining N*a image blocks are Denoted as mask image block; the random mask unit calculates the average value of the non-masked image block, which is denoted as m; when m=0, randomly select n image blocks with non-zero average values from the mask image block and exchange them with n image blocks in the non-masked image block, and perform normalized linear transformation on the CT values of all non-masked image blocks through the normalization unit to obtain the preprocessed segmentation result image; establish an optimization network model, and use the loss function to train the optimization network model in the network model training module until the optimization network model converges; and obtain the weight information of the trained optimization network model; in the network model inference module, input the preprocessed segmentation result image as an input image into the trained optimization network model, and combine the weight information to obtain the optimized segmentation result image.

[0017] Compared with the existing technology, the above technical solution has the following beneficial effects:

[0018] 1. The present invention optimizes the coronary artery segmentation results through a masked autoencoder (the random mask unit), automatically compensating for the segmentation discontinuity caused by over-segmentation and under-segmentation, and ensuring the integrity and accuracy of the extraction of the main and small branches of the coronary artery tree;

[0019] 2. The mask probability of the masked autoencoder is selected adaptively to ensure that the coronary tree exists in both masked and unmasked data;

[0020] 3. The optimized network model based on deep learning adopts a self-supervised learning method and a loss function that combines MSE and clDice, where MSE only acts on the mask image block and clDice acts on all image blocks. clDice further constrains the connectivity of the network output results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of the method for optimizing the segmentation results of the CT image blood vessel tree provided by the present invention;

[0022] Figure 2 This is a module diagram of the system for optimizing the segmentation results of the CT image blood vessel tree provided by the present invention;

[0023] Figure 3 A schematic structural diagram of the image processing module provided by the present invention;

[0024] Figure 4 A schematic diagram of the workflow of the random mask unit provided by the present invention;

[0025] Figure 5 A schematic diagram of the structure of the network model training module provided by the present invention;

[0026] Figure 6 A schematic diagram of the structure of the network model inference module provided by the present invention;

[0027] Figure 7 A schematic diagram of optimization results of a preferred embodiment provided by the present invention;

[0028] Figure 8 This is a schematic diagram of the optimization results of the preferred embodiment provided by the present invention.

[0029] Among them: 100-image processing module, 101-region of interest acquisition unit, 102-image division unit, 103-random mask unit, 104-normalization unit, 200-network model training module, 201-encoder, 202-decoder, 300-network model inference module, DETAILED DESCRIPTION

[0030] The advantages of the present invention are further described below with reference to the accompanying drawings and specific embodiments.

[0031] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure, as detailed in the appended claims.

[0032] The terms used in this disclosure are for the purpose of describing specific embodiments only and are not intended to limit the disclosure. As used in this disclosure and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0033] It should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining."

[0034] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0035] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal communication between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0036] In the following description, the suffixes such as "module", "component" or "unit" used to represent elements are only used to facilitate the description of the present invention and have no specific meaning. Therefore, "module" and "component" can be used interchangeably.

[0037] See attached Figure 1 The present invention discloses a method for optimizing segmentation results of a CT image blood vessel tree based on deep learning, comprising the following steps:

[0038] S1, obtain a three-dimensional medical image of a vascular tree obtained by CT scanning; the data involved are all raw data and stored in the DICOM data format;

[0039] S2, obtaining a segmented and annotated image of the vascular tree; the annotated image involved is obtained by combining the comprehensive annotation results of multiple professional physicians;

[0040] S3, obtaining a segmentation result image of the vascular tree; the segmentation image involved may be an automatic segmentation, semi-automatic segmentation, or interactive segmentation result, and the segmentation method is not limited;

[0041] S4, multiplying the segmentation result image of the vascular tree with the original CT image to obtain a CT image containing only the vascular tree, and dividing the image into N non-overlapping image blocks;

[0042] S5, preset the mask ratio a, adopt the strategy of obeying uniform distribution, randomly sample N image blocks, retain N*(1-a) image blocks, record them as non-masked image blocks, and record the remaining N*a image blocks as masked image blocks;

[0043] S6, calculating the average value of the non-masked image block, denoted as m; when m=0, randomly selecting n image blocks with non-zero average values from the masked image block and exchanging them with n image blocks in the non-masked image block to obtain the preprocessed segmentation result image;

[0044] S7, performing normalized linear transformation on the CT values of all non-masked image blocks;

[0045] S8, establishing an optimized network model for CT image vascular tree segmentation results;

[0046] S9, using the selected loss function until the training optimization network model converges, obtaining the weight of the entire network and saving it;

[0047] S10, loading the trained weight file into the trained optimization network model, and inputting the pre-processed segmentation result image to complete the optimization of the segmentation result of the vascular tree.

[0048] The loss functions for training the optimization network model include MSE loss function and clDice loss function.

[0049] The MSE loss function is used to evaluate the mean of the sum of squares of the errors between the preprocessed segmentation result image output by the optimization network model and the segmentation result image annotated by the physician; Where x is the segmentation result image after the physician's annotation, and y is the preprocessed segmentation result image output by the optimized network model. The MSE value range is [0, +∞]. When the predicted value (i.e., the preprocessed segmentation result image output by the optimized network model) and the true value (i.e., the segmentation result image after the physician's annotation) are completely consistent, the value is 0; when the error between the predicted value (i.e., the preprocessed segmentation result image output by the optimized network model) and the true value (i.e., the segmentation result image after the physician's annotation) is greater, the MSE value is larger.

[0050] In the present invention, the MSE loss function is applied only to the masked image block. Masked tokens are taken from all tokens generated by decoder 202 and input into the fully connected layer. The output channels are mapped to the pixel size of one image block (P×P×C), that is, the size at this time is (N', P×P×C). The predicted value and the masked image block are used as the true value, and the MSE loss function is calculated.

[0051] The clDice loss function is a metric for evaluating the connectivity of two samples based on extracting the skeletons of intersecting tubular or existing structures. Among them, V L represents the segmentation result image after the physician’s annotation, V P Represents the preprocessed segmentation result image output by the optimized network model; S P and S L Represents the V P and V L The skeleton structure extracted from T prec (S P , V L ) is the topological accuracy, indicating S P In V L The proportion of T sens (S L , V P ) is the topological sensitivity, indicating S L In V P The proportion in .

[0052] In the present invention, the clDice loss function is applied to all image blocks (i.e., masked and unmasked image blocks). All tokens generated by the decoder 202 are input into the fully connected layer, and the output channels are mapped to the number of pixels (P×P×C) of one image block, that is, the size at this time is (N, P×P×C) as the predicted value, and all image blocks are used as the true value.

[0053] See attached Figure 2 The present invention also provides a deep learning-based CT image vascular tree segmentation result optimization system for implementing the above method, including an image processing module 100, a network model training module 200 and a network model reasoning module 300.

[0054] See attached Figure 3 The image processing module 100 includes a region of interest acquisition unit 101, an image segmentation unit 102, a random mask unit 103, and a normalization unit 104. The selection of the mask probability in the random mask unit 103 is adaptive to ensure that the vessel tree exists in both masked and unmasked data.

[0055] The region of interest acquisition unit 101 mainly multiplies the segmentation result of the vascular tree (for training data, the segmentation result refers to the segmentation result obtained by combining the comprehensive annotation results of multiple professional doctors) with the original CT image to obtain a CT image x∈R containing only the vascular tree H×W×C , the image spatial resolution is H×W;

[0056] The image division unit 102 is mainly used to divide the image x into flat image blocks. Where P×P is the size of each image block, N is the number of image blocks,

[0057] The random mask unit 103 generally refers to randomly sampling N image blocks according to a preset mask ratio a and adopting a uniform distribution strategy, retaining N*(1-a) image blocks as non-masked image blocks, and the remaining N*a image blocks as masked image blocks.

[0058] See attached Figure 4 Since the input of random masking unit 103 in the present invention has been processed by region of interest acquisition unit 101, to ensure that the unmasked image blocks contain valid information, it is necessary to calculate the average value of the unmasked image blocks, denoted as m. When m = 0, n image blocks with non-zero averages are randomly selected from the masked image blocks and exchanged with n image blocks in the unmasked image blocks. Through multiple experiments, it is generally found that a is selected between 50% and 75%, and n is no greater than 25% of the total number of image blocks with non-zero averages in the masked image blocks.

[0059] The normalization unit 104 is mainly used to perform linear transformation on the CT value v of the non-masked image block to facilitate the convergence of the subsequent network training process. Among them, v mean is the mean of the CT values v of the non-masked image block, v std is the standard deviation of the CT value v of the non-masked image block.

[0060] See attached Figure 5 , the network model training module 200 includes an encoder 201 and a decoder 202. The input image refers to the image block processed by the random mask unit 103, and the encoder 201 only processes the non-masked image block. The encoder 201 can be ViT or ResNet or other backbone, etc. Here, ViT is taken as an example, and a trainable linear projection is used to map the vectorized xp to a potential D-dimensional embedding space. The mapped result is called a token, and its size is (N, D). In order to encode the spatial information of the image block, a specific position is embedded in the token (position embedding) to retain the position information, thereby adding position information to each image block. See formula in, is the mapping projection, E pos ∈R N×D is the embedding position. Position embedding is shared by all images and can be learned. Encoder 201 mainly consists of L layers of multi-head self-attention mechanism (MSA) and multi-layer perceptron (MLP) modules. Therefore, the output of layer l is as follows: l =MSA(LN(z l-1 ))+z l-1 ;z l =MLP(LN(z′ l ))+z′ l . Among them, LN(·) represents the normalization operation, z l Represents an encoded image.

[0061] The decoder 202 not only needs to process the non-masked tokens that have been encoded by the encoder 201, but also needs to process the masked tokens. Masked tokens are learnable vectors shared by all masked image blocks. As in the encoder 201, the masked tokens also add position information, and each masked image block corresponds to a position embedding, whose size is ('N', D), where 'N' is the number of masked image blocks. The masked tokens are copied 'N' times so that each masked image block corresponds to a masked token. Therefore, the non-masked tokens that have been encoded by the encoder 201 and the masked tokens with added position information are spliced together in the order that they originally corresponded in the image block form as the input of the decoder 202. When the dimension of the non-masked tokens that have been encoded by the encoder 201 is inconsistent with the input dimension of the decoder 202, it is necessary to use a linear mapping to map the dimension to the required dimension.

[0062] See attached Figure 6 The input of the network model inference module 300 consists of the segmentation result of the vascular tree (i.e., the preprocessed segmentation result image) and weight information. The network model inference module 300 is mainly used to load the trained weight information into the deep learning network model and optimize the input vascular tree segmentation result. The vascular tree segmentation result can be input into the optimized network model with the weight file loaded for inference through the image processing module 100. It can effectively compensate for the discontinuous positions in the initial segmentation result and suppress false positives caused by reasons such as venous adhesion, thereby completing the optimization of the initial segmentation result. The entire process is simple and easy to operate, takes a short time, is a fully automatic process, does not require additional manual operation, and can effectively improve the integrity and accuracy of the vascular tree segmentation result.

[0063] See attached Figure 7-8 This is an optimization example of a preferred embodiment of the deep learning-based CT image vessel tree segmentation optimization method of the present invention. The vessel tree segmentation results used in the optimized network model are derived by combining the comprehensive annotation results of multiple professional physicians. The predicted segmentation results output by the trained optimized network model are also compared with the comprehensive annotation results of multiple professional physicians.

[0064] The present invention also discloses a method for optimizing the segmentation results of the coronary artery tree in a CT image based on deep learning, which uses the above-mentioned method for optimizing the segmentation results of the CT image blood vessel tree to optimize the coronary artery tree.

[0065] Based on deep learning model technology, the present invention trains a model to optimize the segmentation results of CT image vascular trees (such as coronary artery trees) through a loss function that combines the MSE loss function and the clDice loss function. It automatically compensates for the segmentation discontinuity problem caused by over-segmentation and under-segmentation, and ensures the integrity and accuracy of the extraction of the main branches and small branches of the coronary artery tree.

[0066] It should be noted that the embodiments of the present invention have better practicability and do not impose any form of limitation on the present invention. Any technician familiar with the field may use the technical content disclosed above to change or modify it into an equivalent effective embodiment. However, any modification or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for optimizing segmentation results of CT image vascular trees based on deep learning, characterized in that: The steps include: Obtaining a segmentation result image of the vascular tree based on the original CT image, multiplying the segmentation result image with the original CT image to obtain a CT image containing only the vascular tree, and dividing the image into N non-overlapping image blocks; The mask ratio a is preset, and a uniform distribution strategy is adopted to randomly sample N image blocks. N*(1-a) image blocks are retained and recorded as non-masked image blocks, and the remaining N*a image blocks are recorded as masked image blocks. Calculate the average value of the non-masked image blocks, denoted as m; When m=0, n image blocks with non-zero average values are randomly selected from the mask image blocks and exchanged with n image blocks in the non-mask image blocks to obtain the preprocessed segmentation result image; Establishing an optimized network model, training the optimized network model using a loss function until the optimized network model converges; and obtaining weight information of the trained optimized network model; Inputting the preprocessed segmentation result image as an input image into the trained optimization network model, and combining the weight information to obtain an optimized segmentation result image; in The loss function includes an MSE loss function, which is used to evaluate the mean of the sum of squares of errors between the preprocessed segmentation result image output by the optimization network model and the segmentation result image annotated by the physician; Wherein, x is the segmentation result image after the physician's annotation, y is the pre-processed segmentation result image output by the optimization network model; the MSE loss function is only used for the mask image block; The loss function also includes the clDice loss function, Among them, V L represents the segmentation result image after the physician’s annotation, V P represents the pre-processed segmentation result image output by the optimized network model; S P and S L Represents the V P and V L The skeleton structure extracted from T prec (S P ,V L ) is the topological accuracy, indicating S P In V L The proportion of T sens (S L ,V P ) is the topological sensitivity, indicating S L In V P The clDice loss function is used for masked image blocks and non-masked image blocks.

2. The method for optimizing segmentation results of a CT image blood vessel tree according to claim 1, characterized in that: After obtaining the pre-processed segmentation result image, the following steps are further included: Perform normalized linear transformation on the CT values of all non-masked image blocks, and the CT value Among them, v mean is the mean of the CT values v of the non-masked image block, v std is the standard deviation of the CT value v of the non-masked image block.

3. The method for optimizing segmentation results of a CT image blood vessel tree according to claim 1, characterized in that: When m=0, randomly selecting n image blocks with non-zero average values from the masked image blocks and exchanging them with n image blocks in the non-masked image blocks further includes: a is 50%-75%.

4. The method for optimizing segmentation results of a CT image blood vessel tree according to claim 3, characterized in that: When m=0, randomly selecting n image blocks with non-zero average values from the masked image blocks and exchanging them with n image blocks in the non-masked image blocks further includes: n is less than or equal to 25% of the total number of all image blocks whose average values are not zero in the mask image blocks.

5. A method for optimizing segmentation results of coronary artery trees in CT images based on deep learning, characterized in that: The method for optimizing segmentation results of a CT image blood vessel tree according to any one of claims 1 to 4 is adopted.

6. A deep learning-based CT image vascular tree segmentation result optimization system, characterized in that: Includes image processing module, network model training module and network model inference module; The image processing module includes a region of interest acquisition unit, an image division unit, a random mask unit and a normalization unit; Acquire a segmentation result image of the vascular tree based on the original CT image, multiply the segmentation result image with the original CT image by the region of interest acquisition unit to obtain a CT image containing only the vascular tree, and divide the image into N non-overlapping image blocks by the image division unit; The random mask unit presets a mask ratio a, adopts a uniform distribution strategy, randomly samples N image blocks, retains N*(1-a) image blocks, and records the remaining N*a image blocks as masked image blocks; The random mask unit calculates the average value of the non-masked image blocks, which is denoted as m; When m=0, n image blocks with non-zero average values are randomly selected from the mask image blocks and exchanged with n image blocks in the non-mask image blocks; the CT values of all the non-mask image blocks are normalized linearly transformed by the normalization unit to obtain a preprocessed segmentation result image; Establishing an optimized network model, training the optimized network model using a loss function in the network model training module until the optimized network model converges; and obtaining weight information of the trained optimized network model; In the network model inference module, the preprocessed segmentation result image is input as an input image into the trained optimization network model, and the optimized segmentation result image is obtained by combining the weight information; in The loss function includes an MSE loss function, which is used to evaluate the mean of the sum of squares of errors between the preprocessed segmentation result image output by the optimization network model and the segmentation result image annotated by the physician; Wherein, x is the segmentation result image after the physician's annotation, y is the pre-processed segmentation result image output by the optimization network model; the MSE loss function is only used for the mask image block; The loss function also includes the clDice loss function, Among them, V L represents the segmentation result image after the physician’s annotation, V P represents the pre-processed segmentation result image output by the optimized network model; S P and S L Represents the V P and V L The skeleton structure extracted from T prec (S P ,V L ) is the topological accuracy, indicating S P In V L The proportion of T sens (S L ,V P ) is the topological sensitivity, indicating S L In V P The clDice loss function is used for masked image blocks and non-masked image blocks.

Images