Intracranial artery multi-class segmentation method based on anatomy and topology perception

Abstract

The invention relates to the field of image segmentation, and provides an intracranial artery multi-class segmentation method based on anatomy and topological perception, which comprises the following steps of: extracting a region-of-interest block containing a Willis ring from an acquired whole-brain image to obtain an input image and a real annotation image; and inputting the input image into a pre-constructed intracranial artery multi-class segmentation model to obtain a multi-class prediction probability graph. And calculating radius correction-based Dice loss, vascular rupture perception loss, adjacency relation perception loss and multi-class cross entropy loss based on the prediction probability graph, the real annotation image and the weight graph. And determining total segmentation loss according to the loss, and training the model to obtain a trained intracranial artery multi-class segmentation model. And extracting a target region-of-interest block containing the Willis ring from the obtained target whole-brain image, and inputting the target region-of-interest block into the trained model to obtain a segmentation result of the Willis ring multi-vessel segment on the target region-of-interest block, thereby realizing accurate segmentation of the Willis ring multi-class vessel segments.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a multi-category segmentation method of intracranial arteries based on anatomical and topological perception. Background Art

[0002] The brain is supplied by two internal carotid arteries and two vertebral arteries, which are connected by a central anastomosis at the base of the brain, the Circle of Willis (CoW). These blood vessels are rich in oxygen and nutrients and play an important role in the healthy and stable operation of the brain. Abnormalities of the cerebral arteries, such as hypoplasia, vascular stenosis, duplication and absence, may lead to the occurrence and development of cerebrovascular diseases, such as stroke, aneurysm, arteriovenous malformation and other vascular abnormalities. Recent studies have shown that variations in the geometric characteristics of certain cerebral arteries, such as increased carotid artery diameter and incomplete Circle of Willis, are also causes of neurovascular diseases. The manual annotation of the Circle of Willis structure consumes extremely high manpower, material and time costs. The Circle of Willis data obtained by manual annotation has the disadvantages of poor reproducibility, low efficiency and high complexity. In order to effectively assist doctors in the observation and diagnosis of neurovascular diseases, the automatic segmentation of different vascular segments in the vascular tree of the Circle of Willis is of great significance.

[0003] The existing segmentation models still have some defects for the task of vascular segmentation. First, the structure of the Willis loop contains multiple types of vascular segments such as the carotid artery, the anterior, middle and posterior cerebral arteries, and the anterior and posterior communicating arteries. The diameters and lengths of these vascular segments are quite different. As segmentation targets, the vascular segments with large diameters and long lengths account for a much larger proportion of voxels in the image than the vascular segments with small diameters and short lengths. This causes the problem of class imbalance in multi-class segmentation tasks, making the model's segmentation performance for large targets significantly better than that for small targets. Secondly, the loss function of the existing segmentation model mainly focuses on the correctness of the segmentation of a single voxel itself, ignoring the learning of the spatial neighborhood information of slender structures such as blood vessels, which will cause the segmented blood vessels to be broken. In addition, for the special connection structure of the Willis loop, the existing segmentation methods often ignore the prior knowledge that different types of vascular segments have specific connection relationships, resulting in the model making incorrect vascular pair connections when predicting, and the problem of vascular segment confusion, which makes the segmentation accuracy of multi-class vascular segments of the Willis loop low. Summary of the invention

[0004] The present invention provides a multi-category intracranial artery segmentation method based on anatomy and topology perception, which is used to solve the defect of low segmentation accuracy of multi-category vascular segments of the Willis circle in the prior art and realize accurate segmentation of multi-category vascular segments of the Willis circle. The technical solution proposed by the present invention is as follows: In a first aspect, the present invention provides a multi-category segmentation method of intracranial arteries based on anatomy and topology perception, comprising: Extracting a region of interest block including the circle of Willis from the acquired whole-brain image to obtain an input image and a true annotated image, and determining a weight of each vascular voxel based on the true annotated image to obtain a weight map; Inputting the input image into a pre-built intracranial artery multi-category segmentation model to obtain a multi-category prediction probability map; Calculating multiple segmentation losses based on the predicted probability map, the true annotated image, and the weight map, the multiple segmentation losses including radius-corrected Dice loss, vascular fracture perception loss, adjacency perception loss, and multi-category cross entropy loss; Determining an overall segmentation loss according to the multiple segmentation losses, and training the intracranial artery multi-category segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-category segmentation model; A target region of interest block including the circle of Willis is extracted from the acquired target whole-brain image, and the target region of interest block is input into the trained intracranial artery multi-category segmentation model to predict and obtain a segmentation result of multiple blood vessel segments of the circle of Willis on the target region of interest block.

[0005] Optionally, the method further comprises: The coordinates of the target region of interest block on the target whole-brain image are extracted, and the segmentation result of the multi-vascular segment of the circle of Willis is restored from the target region of interest block to the whole-brain space according to the coordinates to obtain the segmentation result of the multi-vascular segment of the circle of Willis on the whole brain.

[0006] Optionally, the intracranial artery multi-category segmentation model is a three-dimensional U-shaped network, and the three-dimensional U-shaped network includes a pair of encoders and decoders: The encoder comprises a plurality of down-sampling modules, wherein the plurality of down-sampling modules are sequentially connected and gradually reduce the resolution; The decoder comprises a plurality of up-sampling modules, wherein the plurality of up-sampling modules are sequentially connected and gradually increase the resolution; The number of the up-sampling modules is the same as the number of the down-sampling modules, and the up-sampling modules are connected to the down-sampling modules at corresponding positions.

[0007] Optionally, the radius-corrected Dice loss is determined in the following manner: Multiplying the weight map with the true annotated image and the predicted probability map respectively to obtain a weighted true annotated image and a weighted predicted probability map; Based on the weighted true annotated image and the weighted predicted probability map, the radius-corrected Dice loss is calculated using a Dice coefficient formula.

[0008] Optionally, the blood vessel rupture perception loss is determined by: For each voxel, determining the category of the voxel with the maximum probability in the predicted probability map to obtain the target category; One-hot encoding is respectively created for the predicted probability map and the real labeled image according to the target category to obtain a first encoded image and a second encoded image; Using a 3D grouped convolution operation to locally sum the first encoded image and the second encoded image of each category, respectively, to obtain a first convolution map and a second convolution map of each category; The weighted mean square error between the first convolution map and the second convolution map of each category is determined according to the weight map to obtain the blood vessel rupture perception loss.

[0009] Optionally, the adjacency perceived loss is determined in the following manner: For each voxel, determining the category of the voxel with the maximum probability in the predicted probability map to obtain a predicted category image; For each category, extract a segmentation map according to the category and the predicted category image, binarize and dilate the segmentation map, and multiply the dilated result with the predicted category image to obtain an expanded neighborhood category overlay map; Extracting the blood vessel segment categories existing in the expanded neighborhood category overlay map to obtain an adjacent category list, comparing the adjacent category list with a pre-established prior knowledge list, determining false positive categories and false negative categories, and obtaining a false positive category list and a false negative category list; wherein the false positive category is a neighbor category that appears in the adjacent category list but not in the prior knowledge list, and the false negative category is a neighbor category that appears in the prior knowledge list but not in the adjacent category list; Determine a false positive adjacent error voxel set according to the false positive category list, the predicted category image and the true annotated image, and determine a false negative adjacent error voxel set according to the false negative category list, the predicted category image and the true annotated image; The false positive adjacent error voxel set, the false negative adjacent error voxel set and the predetermined second encoded image are all multiplied by the predicted probability map and the weight map to obtain a probability weighted map of the false positive adjacent error voxel set, a probability weighted map of the false negative adjacent error voxel set and a probability weighted map of the true positive voxel set, respectively; wherein the second encoded image is a one-hot encoded image of the true annotated image; The adjacency relationship perception loss is calculated using a Dice coefficient formula according to the probability weighted graph of the false positive adjacent error voxel set, the probability weighted graph of the false negative adjacent error voxel set, and the probability weighted graph of the true positive voxel set.

[0010] In a second aspect, the present invention further provides a multi-category intracranial artery segmentation device based on anatomical and topological perception, comprising the following modules: A weight determination module is used to extract a region of interest block containing the circle of Willis from the acquired whole-brain image to obtain an input image and a true annotated image, and determine the weight of each vascular voxel based on the true annotated image to obtain a weight map; A first prediction module, used for inputting the input image into a pre-built intracranial artery multi-category segmentation model to obtain a multi-category prediction probability map; A loss determination module, configured to calculate a plurality of segmentation losses based on the predicted probability map, the true annotated image and the weight map, wherein the plurality of segmentation losses include a radius-corrected Dice loss, a vascular fracture-aware loss, an adjacency-aware loss and a multi-category cross entropy loss; A model training module, used for determining an overall segmentation loss according to the multiple segmentation losses, and training the intracranial artery multi-category segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-category segmentation model; The second prediction module is used to extract a target region of interest block including the circle of Willis from the acquired target whole-brain image, and input the target region of interest block into the trained intracranial artery multi-category segmentation model to predict and obtain a segmentation result of multiple blood vessel segments of the circle of Willis on the target region of interest block.

[0011] In a third aspect, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the multi-category segmentation method of intracranial arteries based on anatomical and topological perception as described in the first aspect above is implemented.

[0012] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-category segmentation method of intracranial arteries based on anatomical and topological perception as described in the first aspect above.

[0013] In a fifth aspect, the present invention further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the multi-category segmentation method of intracranial arteries based on anatomical and topological perception as described in the first aspect above.

[0014] Based on the above technical solution, the beneficial effects of the present invention compared with the prior art are as follows: The multi-category segmentation method of intracranial arteries based on anatomical and topological perception provided by the present invention solves the problem of category imbalance by assigning weights to each vascular voxel based on real annotated images and generating a weight map. Vessels with smaller radii or key vascular segments will be given higher weights to pay more attention to them during the segmentation process, so that the contribution of large and small vessels to the loss function can be balanced, the segmentation performance of the model for small vascular segments is improved, and the problem of inaccurate segmentation caused by category imbalance is reduced. The vascular fracture perception loss introduces considerations of vascular continuity, so that the model can learn the spatial neighborhood information of the vessel, thereby reducing the situation of segmented vascular fractures. This significantly reduces the incidence of vascular fractures and improves the integrity and continuity of vascular segmentation. The adjacency perception loss utilizes the anatomical structure information of the vessel, especially the adjacency relationship between the vessels, to ensure that the segmentation results conform to the expected anatomical structure. By introducing the prior knowledge of vascular anatomical connections, the model can learn that there is a specific adjacency relationship between vascular segments of different categories, thereby reducing the problem of confusion and erroneous connection of vascular segments. This improves the accuracy of vascular segment classification, makes the segmentation results more consistent with anatomical features, and realizes accurate segmentation of multi-category vascular segments of the Willis circle.

[0015] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0016] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a flowchart of the multi-category segmentation method of intracranial arteries based on anatomical and topological perception provided by the present invention.

[0019] Figure 2 It is a structural schematic diagram of a three-dimensional U-shaped network provided by the present invention.

[0020] Figure 3 It is a schematic diagram of the model training phase provided by the present invention.

[0021] Figure 4It is a schematic diagram of the structure of the intracranial artery multi-category segmentation device based on anatomical and topological perception provided by the present invention.

[0022] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] Combine the following Figure 1-Figure 3 The present invention describes the multi-class segmentation method of intracranial arteries based on anatomy and topology awareness.

[0025] This multi-category segmentation method of intracranial arteries based on anatomy and topology perception enables the model to pay attention to the special connection relationship characteristics between vascular segments of the Circle of Willis and the spatial neighborhood characteristics of the blood vessels, and directly obtain multi-category segmentation results. Figure 1 As shown, the method includes the following: Step S110: extracting a region of interest block including the circle of Willis from the acquired whole-brain image to obtain an input image and a true annotated image, and determining a weight of each vascular voxel based on the true annotated image to obtain a weight map.

[0026] First, whole-brain imaging data is obtained from medical imaging devices (such as MRI or CTA). Using image processing technology, the region of interest block containing the circle of Willis is extracted from the whole-brain image. The circle of Willis is the main vascular structure at the base of the brain, consisting of the anterior and posterior communicating arteries and the terminal parts of the anterior cerebral artery, middle cerebral artery, posterior cerebral artery and basilar artery on both sides. Generate true labeled images for the extracted ROI by manual annotation or using existing annotation datasets. These labeled images indicate the boundaries of different vascular categories in the ROI. Based on the true labeled images, a weight is assigned to each vascular voxel (i.e., vascular pixel or voxel) to generate a weight map. The weight is determined based on the thickness of the blood vessel.

[0027] Specifically, in order to train the multi-class segmentation model of intracranial arteries, refer to Figure 3 As shown in the figure, we first need to prepare a set of normal human time-of-flight magnetic resonance angiography (TOF-MRA) data containing Willis ring labels. These data have been manually annotated to clarify the boundaries of the Willis ring and its vascular segments.

[0028] Given that the whole-brain imaging data is large and complex, and the Willis circle only occupies a small part of the key area of ​​the whole-brain imaging, it is necessary to extract the region of interest (ROI) containing the Willis circle from the whole-brain imaging to improve the training efficiency and accuracy of the model. The ROI extraction process is as follows: First, the template is defined: on the standard TOF-MRA image in the standard Montreal Neurological Institute (MNI) space, a cube of appropriate size is used to manually frame the ROI area containing the circle of Willis. This cube template will serve as a reference for subsequent ROI extraction.

[0029] Next, registration is performed: the TOF-MRA image defined in the standard space is registered to the actual input image space using a registration algorithm. The purpose of registration is to ensure that the ROI template in the standard space can be accurately mapped to each input image, thereby extracting the corresponding ROI area.

[0030] ROI cropping: Based on the registration results, the ROI region is cropped from the input image. At the same time, the corresponding real annotated image is cropped using the corresponding annotated data. These cropped ROI images and annotated images will be used as input data for model training.

[0031] Through the above steps, the ROI containing the Willis circle can be efficiently extracted from the whole brain image, and the input image and real labeled image for network training can be prepared. This not only reduces the redundant information that the model needs to process, but also improves the segmentation accuracy of the Willis circle and its vascular segments.

[0032] Step S120: input the input image into a pre-built intracranial artery multi-category segmentation model to obtain a multi-category prediction probability map.

[0033] Pre-build a multi-class segmentation model for intracranial arteries. This model can be a deep learning model, such as a three-dimensional U-shaped network (refer to Figure 2 The model has an architecture (as shown in Figure 1) that can process three-dimensional imaging data and output multi-category prediction probability maps. The input image (i.e., the ROI containing the ring of Willis) is input into the model. The model outputs a multi-category prediction probability map, in which each voxel corresponds to a probability distribution representing a different vascular category. The multi-category prediction probability map contains a total of C channels, each of which is an image of the same size as the true annotated image, where the value of the voxel on the image of the i-th channel represents the predicted probability that the location belongs to the i-th category.

[0034] Step S130, calculating multiple segmentation losses based on the predicted probability map, the real annotated image and the weight map, wherein the multiple segmentation losses include radius-corrected Dice loss, vascular rupture perception loss, adjacency perception loss and multi-category cross entropy loss.

[0035] Considering that in the same input image, small blood vessels are much smaller in diameter and volume than large blood vessels, it is very easy to have class imbalance problems in the multi-class segmentation learning process. Therefore, this application constructs a radius-corrected Dice loss function, which introduces radius information in the calculation of the loss function to increase the influence of small-radius blood vessels in the loss function calculation, and guides the model to focus on the feature learning of small-radius blood vessels. The radius-corrected Dice loss takes into account the changes in the blood vessel radius to improve the segmentation accuracy of small blood vessels and fine structures.

[0036] Existing segmentation loss functions mainly focus on the accuracy of voxel classification itself, but as a target with strong spatial topological characteristics, the spatial neighborhood information of blood vessels should also be focused on in learning. The way to maintain the spatial connectivity of blood vessels in manual labeling is to label them frame by frame by moving the neighborhood. Inspired by the manual labeling method, the present invention constructs a vascular rupture perception loss function, which uses neighborhood information to quickly identify voxel areas where vascular ruptures occur. This loss function is designed to detect and penalize vascular ruptures, that is, discontinuities or interruptions of blood vessels in the model prediction results.

[0037] For the special connection structure of the Willis circle, existing segmentation methods often ignore the prior knowledge that different categories of vascular segments have specific anatomical connections, which leads to incorrect vascular pair connections when the model is predicted. Therefore, this application proposes a loss function based on adjacency perception, which incorporates category adjacency into network supervised learning to improve automatic segmentation performance. This loss function utilizes the anatomical structure information of the blood vessels, especially the adjacency relationship between blood vessels, to ensure that the segmentation results conform to the expected anatomical structure.

[0038] The above multi-class cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the true annotation.

[0039] Reference Figure 3 As shown, the dotted lines drawn from the loss function point to the two input images used for comparison in the loss function, and the dotted lines in the image pointing to the loss function represent that the image participates in the calculation of the loss function. This application calculates the multi-category cross entropy loss based on the obtained multi-category prediction probability map and the true annotation result (i.e. the above-mentioned true annotation image). And Dice loss based on radius correction . Calculate the vessel radius weight map based on the true annotation results. This weight map is used to calculate the radius-corrected Dice loss. Perform convolution operations on the multi-category prediction probability map and the true annotation results to obtain the prediction neighborhood fusion results and the true annotation neighborhood results. Calculate the vessel rupture perception loss based on the obtained prediction neighborhood fusion results and the true annotation neighborhood results. . The adjacency-aware Dice loss is calculated based on the multi-category prediction results and the true annotation results in the multi-category prediction probability graph, that is, the adjacency-aware loss mentioned above .

[0040] Step S140: determining an overall segmentation loss according to the multiple segmentation losses, and training the intracranial artery multi-category segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-category segmentation model.

[0041] The weighted sum of the above multiple loss functions is used to obtain the overall segmentation loss Train a multi-class segmentation model for intracranial arteries using a back-propagation algorithm and an optimizer such as Adam or SGD to minimize the overall segmentation loss. Evaluate the performance of the model on the validation set and perform parameter tuning or model optimization as needed.

[0042] The trained multi-category segmentation model of intracranial arteries can achieve end-to-end segmentation of multi-category vascular segments of the Circle of Willis, and retain the connectivity of the predicted blood vessels to the greatest extent, and the predicted blood vessels conform to the particularity of the connection between vascular segments that is unique to the Circle of Willis. The intracranial artery is a complex network structure, and there are specific connection relationships between different vascular segments. The multi-category segmentation model of the intracranial artery of the present invention can capture these relationships to ensure the correctness and completeness of the segmentation results. By introducing topological perception, the model can learn the connection relationship between vascular segments, thereby reducing erroneous connections in the segmentation results, which is crucial to maintaining the integrity and accuracy of the vascular network. The present invention makes full use of the anatomical knowledge and topological structure information of the intracranial artery, realizes anatomical perception and topological perception, and improves the accuracy and reliability of the segmentation results.

[0043] Step S150, extracting a target region of interest block including the circle of Willis from the acquired target whole-brain image, inputting the target region of interest block into the trained intracranial artery multi-category segmentation model to predict and obtain a segmentation result of multiple blood vessel segments of the circle of Willis on the target region of interest block.

[0044] The target region of interest (ROI) containing the circle of Willis is extracted from the new whole-brain image, i.e., the above-mentioned target whole-brain image. The target ROI is input into the trained multi-class segmentation model of intracranial arteries to obtain the segmentation results of multiple blood vessel segments. The prediction results can then be visualized, quantitatively evaluated, and analyzed to verify the accuracy and reliability of the model.

[0045] The multi-category segmentation method of intracranial arteries based on anatomical and topological perception provided by the present invention solves the problem of class imbalance by assigning weights to each vascular voxel based on real annotated images and generating a weight map. Blood vessels with smaller radii or key vascular segments are given higher weights to give more attention during the segmentation process, so that the contribution of large and small blood vessels to the loss function is balanced, improving the segmentation performance of the model for small blood vessel segments and reducing the problem of inaccurate segmentation caused by class imbalance.

[0046] The present invention designs a vessel breakage perception loss. By introducing the consideration of vessel continuity, the model can learn the spatial neighborhood information of the vessel, thereby reducing the occurrence of segmented vessel breakage, significantly reducing the incidence of vessel breakage, and improving the integrity and continuity of vessel segmentation. The adjacency perception loss utilizes the anatomical structure information of the vessel, especially the adjacency relationship between the vessels, to ensure that the segmentation result conforms to the expected anatomical structure. By introducing the prior knowledge of anatomy, the model can learn that there is a specific connection relationship between different categories of vessel segments, thereby reducing the problem of confusion and wrong connection of vessel segments. This improves the accuracy of vessel segment classification and makes the segmentation result more consistent with anatomical features. Moreover, different from the traditional idea of ​​segmentation first and then labeling, the present invention directly designs a multi-category segmentation model that can simultaneously complete segmentation and classification tasks. Through the integrated design, the model can simultaneously consider the category information of different vessel segments during the segmentation process, thereby reducing the accumulation of segmentation errors and improving the accuracy of overall segmentation and classification, especially for small vessels that are prone to segmentation and breakage. The effect is more significant.

[0047] In an optional embodiment, the intracranial artery multi-category segmentation model is a three-dimensional U-shaped network, and the three-dimensional U-shaped network includes a pair of encoders and decoders: The encoder includes a plurality of down-sampling modules, which are connected in sequence and gradually reduce the resolution; the decoder includes a plurality of up-sampling modules, which are connected in sequence and gradually increase the resolution; the number of the up-sampling modules is the same as the number of the down-sampling modules, and the up-sampling modules are connected to the down-sampling modules at corresponding positions.

[0048] Reference Figure 2 As shown, the encoder consists of M downsampling modules Composition, modules are arranged according to arrive The downsampling modules are connected in sequence and the resolution is gradually reduced. Each downsampling module can contain a convolution layer, an activation function and a pooling layer (or a downsampling layer) to extract image features and reduce spatial dimensions. The decoder consists of M upsampling modules. Composition, modules are arranged according to arrive The upsampling modules are connected in sequence and the resolution is gradually improved. Each upsampling module can include an upsampling layer (or a transposed convolution layer), a convolution layer, and an activation function to map the extracted features back to the original image space. There are m paths between the encoder and the decoder, which connect the downsampling modules at the corresponding positions. And upsampling module The features in the encoder are directly transferred to the corresponding positions in the decoder through skip connections, which helps to preserve spatial details and alleviate the gradient vanishing problem. This design enables the network to better learn the global and local features of the image. Figure 2 The number of modules in the encoder / decoder For example.

[0049] The present invention uses the downsampling module in the encoder, and the network can gradually extract low-level, medium-level and high-level features of the image. These features are crucial for subsequent segmentation tasks. The upsampling module in the decoder uses the features extracted by the encoder to gradually restore the resolution of the image and generate segmentation results. Due to the presence of jump connections, the decoder can make full use of the feature information in the encoder, thereby improving the accuracy of segmentation. The three-dimensional U-shaped network architecture shows strong robustness when processing three-dimensional medical images. It can adapt to intracranial arteries of different sizes and shapes and accurately segment complex structures.

[0050] In an optional embodiment, considering that in the same input image, small blood vessels are much smaller in diameter and volume than large blood vessels, it is very easy to have a class imbalance problem in the multi-class segmentation learning process. In view of the problem of class imbalance in the existing segmentation network for blood vessel segments with smaller diameters, the present invention proposes a loss function based on blood vessel radius correction, which can enhance the model's learning of the characteristics of blood vessels with smaller diameters, ensure the segmentation performance of small blood vessels, and alleviate the class imbalance problem. By introducing radius information in the calculation of the loss function, the influence of small radius blood vessels in the calculation of the loss function is increased, and the model is guided to focus on the feature learning of small radius blood vessels. The radius-corrected Dice loss is determined in the following way: S210: Multiply the weight map with the true annotated image and the predicted probability map respectively to obtain a weighted true annotated image and a weighted predicted probability map.

[0051] First, use the real annotated image to complete the calculation of the blood vessel radius on the CPU and save it as a radius image According to the calculated radius, the weight of each vascular voxel is calculated to obtain the weight map : For vascular voxels, the larger the radius, the closer the weight value is to 1, and the smaller the radius, the closer it is to 2, while the background voxels (i.e., the voxels belonging to the background area) obtain a uniform weight value of 1. The weight value calculation formula is:

[0052] in, represents the maximum radius value among all voxels belonging to the Willis ring, Represents the minimum radius value among all voxels belonging to the Willis ring. and They represent the maximum value function and the minimum value function respectively. represents voxel, represents a Willis ring, Representing voxels Belongs to the Willis ring, Representing voxels The radius value. Representing voxels The weight value of represents the background area, Representing voxels Belongs to the background area.

[0053] The calculated weight map is added to the calculation of Dice loss. The calculation method is as follows:

[0054] in, express Coefficient function, The real labeled image is a binary image or label image, in which the value of each voxel indicates whether the location belongs to the Willis circle area and the category of the blood vessel segment in the Willis circle area. represents the predicted probability map, Represents a logical AND operation.

[0055] It is the intersection of the true labeled image and the predicted probability map. Represents the sum of the voxel values ​​in the intersection of the true labeled image and the predicted probability map. represents the sum of the voxel values ​​belonging to the Willis ring in the true labeled image, Represents the sum of voxel values ​​of the predicted probability map.

[0056] The calculated weight map Compared with the real labeled images and predicted probability plot Multiplying, the calculation method is corrected to:

[0057] in, represents the weighted true labeled image, Represents the weighted predicted probability map.

[0058] The weight map is multiplied with the true annotated image and the predicted probability map respectively to obtain the weighted true annotated image and the weighted predicted probability map. The voxel values ​​in the true annotated image and the predicted probability map are adjusted according to the weight map, so that the voxel values ​​in the key areas can be enhanced.

[0059] S220: Based on the weighted real labeled image And the weighted predicted probability map , use the Dice coefficient formula to calculate the radius-corrected Dice loss .

[0060]

[0061] in, represents the Dice loss based on radius correction, express Coefficient function, Represents a logical AND operation. represents the sum of the voxel values ​​of the intersection of the weighted true annotation image and the weighted predicted probability map, represents the weighted sum of the voxel values ​​of the true labeled image, Represents the sum of the voxel values ​​of the weighted predicted probability map.

[0062] The present invention introduces a weight map and radius information based on the radius-corrected Dice loss to help the model pay more attention to the radius changes of the artery, so that the model can more accurately identify and segment the key areas of the intracranial artery, thereby improving the accuracy and reliability of the segmentation results. The introduction of the weight map and radius information enables the model to better adapt to intracranial arteries of different sizes and shapes, as well as complex anatomical structures and topological relationships. This enhances the robustness of the model, enabling it to maintain stable segmentation performance under various circumstances. The radius-corrected Dice loss is used as a loss function to guide the training process of the model. By minimizing the loss function, the model can gradually learn a more accurate segmentation strategy, thereby improving training efficiency and convergence speed.

[0063] In an optional embodiment, the existing segmentation loss function mainly focuses on the accuracy of the classification of the voxels themselves, but as a target with strong spatial topological characteristics, the spatial neighborhood information of the blood vessel should also be focused on in learning. The way to maintain the spatial connectivity of blood vessels in manual labeling is to label them frame by frame by moving the neighborhood. Drawing on the idea of ​​manual labeling, in order to address the problem that the existing segmentation network does not extract enough vascular neighborhood information, resulting in segmented blood vessel breakage, the present invention proposes a vascular breakage perception loss function combined with neighborhood information, enhances the learning of the features of blood vessels that are prone to breakage, and improves the problem of vascular breakage. The present invention uses neighborhood information to quickly identify voxel areas where vascular breakage occurs. The vascular breakage perception loss is determined in the following way: S310, for each voxel, determining the category of the voxel with the maximum probability in the predicted probability map to obtain a target category; For a voxel with coordinates (x, y, z), find the category with the maximum probability in the predicted probability map. This category is the target category of the voxel. :

[0064] in, Indicates the target category, that is, the category index with the highest predicted probability at the coordinate (x, y, z). When performing one-hot encoding, this index is used to determine which category channel has a non-zero voxel value. Represents the category, that is, the category index that the current voxel may belong to. Indicates that the voxel at coordinate (x, y, z) belongs to the category The predicted probability of . It means finding the category with the maximum probability in the predicted probability map.

[0065] S320: Create one-hot encoding for the predicted probability map and the real labeled image respectively according to the target category to obtain a first encoded image and a second encoded image.

[0066] According to the target category, the predicted probability map and the true labeled image are one-hot encoded respectively. One-hot encoding is a method of representing category labels, which converts a category label into a vector in which only one position is 1 (indicating the target category) and the rest are 0. In this way, the predicted probability map and the true labeled image are converted into the first encoded image and the second encoded image, which respectively represent the predicted and true category distributions.

[0067] Specifically, for the predicted probability map Create a one-hot encoding with C channel axes from the real labeled image T:

[0068] in, and Represents the predicted probability map and the coordinates in the real labeled image respectively. Category is One-hot encoding of voxels, and Represent the first encoded image and the second encoded image respectively. Indicates that the voxel at coordinate (x, y, z) belongs to the category The predicted probability of . Indicates that at the coordinate (x, y, z), the category index of the predicted voxel is . Indicates other. Indicated in coordinates The actual labeled category index. Indicates that in the real labeled image Central coordinates The value is the category The voxels in In the case of The voxels at the corresponding position has non-zero values ​​(i.e., predicted probabilities). In the example, only the real annotations are The voxels at the corresponding position The top position is 1, and the other positions are 0.

[0069] S330, using a 3D grouped convolution operation to locally sum the first encoded image and the second encoded image of each category respectively, to obtain a first convolution map and a second convolution map of each category; The first and second encoded images of each category are locally summed using 3D grouped convolution operations. 3D grouped convolution can extract features in three dimensions (depth, height, and width). Through local summation, the first and second local feature maps of each category can be obtained, which respectively represent the distribution characteristics of the predicted and true categories in the local area.

[0070] Specifically, 3D grouped convolution operations are used to and Apply a 3×3×3 convolution, and each weight in the convolution kernel is 1. This means that for each category of and Calculate the sum of the 3×3×3 regions respectively. This operation is equivalent to performing local summation on the one-hot encoded image of each category to obtain the first convolution map And the second convolutional graph :

[0071]

[0072] in, Indicated in The offset in the axis direction, from -1 to 1, a total of three positions (including the original position ). Indicated in The offset in the axis direction also ranges from -1 to 1, and there are also three positions. Indicated in The offset in the axis direction is also from -1 to 1, and there are also three positions. Indicates that the coordinates in the first convolution graph are and the eigenvalues ​​of voxels of category c, Indicates that the coordinates in the second convolution graph are and the eigenvalues ​​of voxels of class c.

[0073] S340. Determine the weighted mean square error between the first convolution map and the second convolution map of each category according to the weight map to obtain the blood vessel rupture perception loss.

[0074] calculate and The weighted mean square error (MSE) between , the weight comes from the radius-based weight map calculated in step S210 above , X, Y, Z are the number of voxels along the x, y, and z axes respectively, and C is the number of channel axes:

[0075] in, Indicated in coordinates The weight value on .

[0076] The present invention can accurately represent and extract the category distribution and local features in the vascular rupture task through one-hot encoding and 3D group convolution operations. This enables the model to better distinguish voxels of different categories and capture the key features of vascular rupture, thereby improving the accuracy of perception. The introduction of the weight map enables the model to better adapt to vascular rupture tasks in different situations. By adjusting the weight value, the influence of key areas can be highlighted, thereby enhancing the robustness and generalization ability of the model. The vascular rupture perception loss is used as a loss function to guide the training process of the model. By minimizing the loss function, the model can gradually learn a more accurate perception strategy, thereby improving the training efficiency and convergence speed.

[0077] In an optional embodiment, for the special connection structure of the Willis circle, the existing segmentation methods often ignore the prior knowledge that different categories of vascular segments have specific anatomical connections, which causes the model to make incorrect vascular pair connections when predicting. In view of the problem that the existing multi-category segmentation network does not learn the connection relationship between the vascular segments of the Willis circle enough, resulting in vascular segment confusion, the present invention proposes an adjacency-aware loss function, which uses the prior inter-class connection relationship to guide the network to learn the distinction between categories, improve the vascular segment confusion problem, and improve the automatic segmentation performance. The adjacency-aware loss is determined in the following way: S410 . For each voxel, determine the category of the voxel with the maximum probability in the predicted probability map, and obtain a predicted category image.

[0078] For each voxel with coordinates (x,y,z), find its The category with the largest probability in is the target category , thereby generating a predicted category image This image reflects the model's prediction for each voxel class.

[0079]

[0080] in, Indicates the target category. Indicates that the voxel at coordinate (x, y, z) belongs to the category The predicted probability of . This formula represents the generated predicted category image The median coordinate is The voxel value is the category corresponding to the maximum predicted probability.

[0081] S420 . For each category, extract a segmentation map according to the category and the predicted category image, binarize and dilate the segmentation map, and multiply the dilated result with the predicted category image to obtain a dilated neighborhood category overlay map.

[0082] For each category , extract the corresponding segmentation map from the predicted category image ,right The binarization process is performed, and the voxels belonging to the category are marked as 1, and the rest of the voxels are marked as 0. The binary segmentation map is expanded to expand the range of the category area, and the result map after the expansion operation is The original segmentation category image is the above predicted category image Multiply them together to get the expanded neighborhood category overlay map , this image reflects the category information within each voxel and its dilated neighborhood.

[0083]

[0084]

[0085]

[0086] in, Indicates category Segmentation map of the predicted category image In the case of hour, At this coordinate is true if true, otherwise it is zero.

[0087] Express Binarization is performed, that is, The process of converting to a system containing only two values ​​(such as 0 and 1) is Non-zero values ​​are converted to 1, and zero values ​​remain unchanged. Represents the binarized Perform expansion operation. is the result graph after the dilation operation, which contains the expanded categories The bounds of the object. To expand the neighborhood category overlay map, the result map after the expansion operation is And the predicted category image Multiplication is used to mark those with the category voxels adjacent to the target, and these voxels still belong to a certain category.

[0088] S430, extracting the blood vessel segment categories existing in the expanded neighborhood category overlay map to obtain an adjacent category list, comparing the adjacent category list with a pre-established prior knowledge list, determining false positive categories and false negative categories, and obtaining a false positive category list and a false negative category list; wherein the false positive category is a neighbor category that appears in the adjacent category list but not in the prior knowledge list, and the false negative category is a neighbor category that appears in the prior knowledge list but not in the adjacent category list.

[0089] Extract the expanded neighborhood category overlay map The vascular segment categories existing in the are saved as a list, which is the above adjacent category list , and the prior knowledge list of adjacent vascular segments of category c In contrast, Appears in but not in The neighbor category that appears is a false positive category, and a list of false positive categories is obtained. . And in Appears in but not in The neighbor category that appears is the false negative category, and the false negative category list is obtained. :

[0090]

[0091]

[0092] in, Is used from Function to extract a list of unique values ​​in . Represents the logical AND operation, Indicates not belonging.

[0093] S440, determining a set of false positive adjacent error voxels according to the false positive category list, the predicted category image and the true annotated image, where these voxels are incorrectly predicted to be adjacent to certain categories. And determining a set of false negative adjacent error voxels according to the false negative category list, the predicted category image and the true annotated image, where these voxels are not correctly predicted to be adjacent to certain categories.

[0094] The set of voxels with false positive adjacency errors and the set of voxels with false negative adjacency errors are calculated according to the following conditions:

[0095] in, represents the exclusive-or operation, Represents a logical AND operation. Category and Merge along the channel axis to get the false positive adjacent error voxel set and false negative neighbor error voxel set . Representing and predicting categories There are categories with incorrect adjacency relationships, that is, categories that are incorrectly associated with Adjacency occurs, or is incorrectly not associated with the category Adjacency relationship occurs.

[0096] Indicated in Category, in Coordinates Whether the voxel belongs to the false positive adjacent error voxel set. If the condition is met, the value is 1, otherwise it is 0.

[0097] Indicated in Category, in Coordinates Whether the voxel belongs to the false negative adjacent error voxel set. If the condition is met, the value is 1, otherwise it is 0.

[0098] Indicated in coordinates The voxels are predicted to be of class .

[0099] Indicated in coordinates The voxels are predicted to be of class .

[0100] Indicated in coordinates The true category label of the voxel is .

[0101] S450, multiplying the false positive adjacent error voxel set, the false negative adjacent error voxel set and the predetermined second coded image by the predicted probability map and the weight map to obtain a probability weighted map of the false positive adjacent error voxel set, a probability weighted map of the false negative adjacent error voxel set and a probability weighted map of the true positive voxel set, respectively; wherein the second coded image is a one-hot coded image of the true annotated image, and the method for determining the second coded image may refer to the description of S310-S320 above; Will , , and predicted probability plot and the weight map calculated in the above step S210 Multiplication:

[0102]

[0103]

[0104] in, A probability weighted graph representing the set of false positive neighboring error voxels, A probability weighted map representing the set of false negative neighboring error voxels, A probability weighted map representing the set of true positive voxels.

[0105] S460: Calculate the adjacency relationship perception loss using the Dice coefficient formula according to the probability weighted graph of the false positive adjacent error voxel set, the probability weighted graph of the false negative adjacent error voxel set, and the probability weighted graph of the true positive voxel set, to evaluate the consistency of the predicted category image and the true annotated image in terms of anatomical adjacency.

[0106] Use the Dice coefficient formula to calculate the final loss function, which is the above-mentioned adjacency perception loss :

[0107] in, , , They represent the sum of the voxel values ​​in the probability weighted map of the true positive voxel set, the probability weighted map of the false positive adjacent error voxel set, and the probability weighted map of the false negative adjacent error voxel set, respectively.

[0108] Overall segmentation loss The calculation formula is as follows:

[0109] in, and It is a hyperparameter that can be adjusted according to the effect to adapt to different data sets.

[0110] The present invention can accurately identify false positive categories and false negative categories by introducing a priori knowledge list and comparing it with the adjacent category list, thereby correcting the erroneous adjacency relationship in the predicted category image. This helps to improve the accuracy of anatomical relationships and provide a reliable basis for subsequent medical analysis and diagnosis. By considering the constraints of anatomical adjacency relationships, the method can guide the model to learn feature representations that are more consistent with anatomical structures, which helps to enhance the generalization ability of the model so that it can still maintain good performance when facing different patients and different imaging conditions. The adjacency relationship perception loss, as part of the loss function, can guide the training process of the model. By minimizing the loss function, the model can gradually learn more accurate anatomical adjacency relationships, thereby improving training efficiency and convergence speed.

[0111] In an optional embodiment, after the segmentation result of the multi-vessel segment of the Willis circle on the target region of interest is predicted, referring to Figure 1 As shown, the method also includes: S160, extracting the coordinates of the target region of interest block on the target whole-brain image, and restoring the segmentation result of the multi-vascular segment of the Circle of Willis from the target region of interest block to the whole-brain space according to the coordinates, to obtain the segmentation result of the multi-vascular segment of the Circle of Willis on the whole brain.

[0112] In the model inference stage, for the task of multi-vessel segmentation of the Circle of Willis in the whole-brain time-of-flight-magnetic resonance angiography (TOF-MRA) images, first, the region of interest (ROI) of the Circle of Willis is located in the whole-brain TOF-MRA image using registration. This step aims to quickly identify and extract a small part of the image area containing the key vascular structure of the Circle of Willis to reduce the amount of calculation and complexity of subsequent processing. Then, based on the positioning results, the whole-brain image is cropped to obtain an image block containing only the Circle of Willis ROI, that is, the target region of interest block mentioned above. This cropping process requires accurate recording of the coordinate information of the ROI on the original whole-brain image so that the segmentation result can be restored to the whole-brain space later.

[0113] The cropped target region of interest block is input into the trained intracranial artery multi-category segmentation model (using the above-mentioned three-dimensional U-shaped network architecture). This model has been trained with a large amount of labeled data and can accurately identify and segment the multi-vascular segments of the Willis circle. The model outputs the segmentation results of the multi-vascular segments of the Willis circle on the ROI, which are presented in the form of a multi-channel probability map or a multi-value segmentation map. After obtaining the segmentation results on the ROI, the previously saved ROI coordinate information on the original whole-brain image is used to accurately restore the segmentation results to the whole-brain space, ensuring that the segmented vascular segments can be accurately mapped back to their original positions in the whole-brain image. This result not only retains the fine segmentation information on the ROI, but also accurately reflects the position and morphology of the vascular segments in the whole brain.

[0114] The present invention first locates and crops the ROI of the whole brain image, and then performs segmentation prediction on the cropped ROI. This can significantly reduce the amount of calculation and processing time, which is particularly important for the processing of large-scale image data. Focusing on the segmentation task of the ROI can make the model more focused on the feature learning of the vascular structure of the Willis circle, thereby improving the accuracy and robustness of the segmentation.

[0115] The multi-category segmentation device for intracranial arteries based on anatomical and topological perception provided by the present invention is described below. The multi-category segmentation device for intracranial arteries based on anatomical and topological perception described below and the multi-category segmentation method for intracranial arteries based on anatomical and topological perception described above can be referenced to each other.

[0116] The multi-category intracranial artery segmentation device based on anatomical and topological perception provided by the present invention refers to Figure 4 As shown, including: The weight determination module 510 is used to extract the region of interest block including the circle of Willis from the acquired whole brain image, obtain the input image and the real annotated image, and determine the weight of each vascular voxel based on the real annotated image to obtain a weight map; A first prediction module 520 is used to input the input image into a pre-built intracranial artery multi-category segmentation model to obtain a multi-category prediction probability map; A loss determination module 530, configured to calculate a plurality of segmentation losses based on the predicted probability map, the true annotated image and the weight map, wherein the plurality of segmentation losses include a radius-corrected Dice loss, a vascular fracture-aware loss, an adjacency-aware loss and a multi-class cross entropy loss; A model training module 540 is used to determine an overall segmentation loss according to the multiple segmentation losses, and train the intracranial artery multi-category segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-category segmentation model; The second prediction module 550 is used to extract a target region of interest block including the circle of Willis from the acquired target whole-brain image, input the target region of interest block into the trained intracranial artery multi-category segmentation model for prediction, and obtain a segmentation result of multiple blood vessel segments of the circle of Willis on the target region of interest block.

[0117] Figure 5 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 5As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the multi-category segmentation method of intracranial arteries based on anatomical and topological perception.

[0118] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0119] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-category segmentation method of intracranial arteries based on anatomical and topological perception provided by the above-mentioned methods.

[0120] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the multi-category intracranial artery segmentation method based on anatomical and topological perception provided by the above-mentioned methods.

[0121] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0122] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-class segmentation method for intracranial arteries based on anatomy and topology perception, characterized in that: include: Extracting a region of interest block including the circle of Willis from the acquired whole-brain image to obtain an input image and a true annotated image, and determining a weight of each vascular voxel based on the true annotated image to obtain a weight map; Inputting the input image into a pre-built intracranial artery multi-category segmentation model to obtain a multi-category prediction probability map; Calculating multiple segmentation losses based on the predicted probability map, the true annotated image, and the weight map, the multiple segmentation losses including radius-corrected Dice loss, vascular fracture perception loss, adjacency perception loss, and multi-category cross entropy loss; Determining an overall segmentation loss according to the multiple segmentation losses, and training the intracranial artery multi-category segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-category segmentation model; A target region of interest block including the circle of Willis is extracted from the acquired target whole-brain image, and the target region of interest block is input into the trained intracranial artery multi-category segmentation model to predict and obtain a segmentation result of multiple blood vessel segments of the circle of Willis on the target region of interest block.

2. The method for multi-class segmentation of intracranial arteries based on anatomy and topology perception according to claim 1, characterized in that: The method further comprises: The coordinates of the target region of interest block on the target whole-brain image are extracted, and the segmentation result of the multi-vascular segment of the circle of Willis is restored from the target region of interest block to the whole-brain space according to the coordinates to obtain the segmentation result of the multi-vascular segment of the circle of Willis on the whole brain.

3. The method for multi-class segmentation of intracranial arteries based on anatomy and topology perception according to claim 1, characterized in that: The intracranial artery multi-category segmentation model is a three-dimensional U-shaped network, which includes a pair of encoders and decoders: The encoder comprises a plurality of down-sampling modules, wherein the plurality of down-sampling modules are sequentially connected and gradually reduce the resolution; The decoder comprises a plurality of up-sampling modules, wherein the plurality of up-sampling modules are connected in sequence and gradually increase the resolution; The number of the up-sampling modules is the same as the number of the down-sampling modules, and the up-sampling modules are connected to the down-sampling modules at corresponding positions.

4. The method for multi-class segmentation of intracranial arteries based on anatomy and topology perception according to claim 1, characterized in that: The radius-corrected Dice loss is determined in the following way: Multiplying the weight map with the true annotated image and the predicted probability map respectively to obtain a weighted true annotated image and a weighted predicted probability map; Based on the weighted true annotated image and the weighted predicted probability map, the radius-corrected Dice loss is calculated using a Dice coefficient formula.

5. The method for multi-class segmentation of intracranial arteries based on anatomy and topology perception according to claim 1, characterized in that: The vessel rupture perception loss is determined in the following manner: For each voxel, determining the category of the voxel with the maximum probability in the predicted probability map to obtain the target category; One-hot encoding is respectively created for the predicted probability map and the real labeled image according to the target category to obtain a first encoded image and a second encoded image; Using a 3D grouped convolution operation to locally sum the first encoded image and the second encoded image of each category, respectively, to obtain a first convolution map and a second convolution map of each category; The weighted mean square error between the first convolution map and the second convolution map of each category is determined according to the weight map to obtain the blood vessel rupture perception loss.

6. The method for multi-class segmentation of intracranial arteries based on anatomy and topology perception according to claim 1, characterized in that: The adjacency-aware loss is determined as follows: For each voxel, determining the category of the voxel with the maximum probability in the predicted probability map to obtain a predicted category image; For each category, extract a segmentation map according to the category and the predicted category image, binarize and dilate the segmentation map, and multiply the dilated result with the predicted category image to obtain an expanded neighborhood category overlay map; Extracting the blood vessel segment categories existing in the expanded neighborhood category overlay map to obtain an adjacent category list, comparing the adjacent category list with a pre-established prior knowledge list, determining false positive categories and false negative categories, and obtaining a false positive category list and a false negative category list; wherein the false positive category is a neighbor category that appears in the adjacent category list but not in the prior knowledge list, and the false negative category is a neighbor category that appears in the prior knowledge list but not in the adjacent category list; Determine a false positive adjacent error voxel set according to the false positive category list, the predicted category image and the true annotated image, and determine a false negative adjacent error voxel set according to the false negative category list, the predicted category image and the true annotated image; The false positive adjacent error voxel set, the false negative adjacent error voxel set and the predetermined second encoded image are all multiplied by the predicted probability map and the weight map to obtain a probability weighted map of the false positive adjacent error voxel set, a probability weighted map of the false negative adjacent error voxel set and a probability weighted map of the true positive voxel set, respectively; wherein the second encoded image is a one-hot encoded image of the true annotated image; The adjacency relationship perception loss is calculated using a Dice coefficient formula according to the probability weighted graph of the false positive adjacent error voxel set, the probability weighted graph of the false negative adjacent error voxel set, and the probability weighted graph of the true positive voxel set.

7. A multi-category intracranial artery segmentation device based on anatomical and topological perception, characterized in that: include: A weight determination module is used to extract a region of interest block containing the circle of Willis from the acquired whole-brain image to obtain an input image and a true annotated image, and determine the weight of each vascular voxel based on the true annotated image to obtain a weight map; A first prediction module, used for inputting the input image into a pre-built intracranial artery multi-category segmentation model to obtain a multi-category prediction probability map; A loss determination module, configured to calculate a plurality of segmentation losses based on the predicted probability map, the true annotated image and the weight map, wherein the plurality of segmentation losses include a radius-corrected Dice loss, a vascular fracture-aware loss, an adjacency-aware loss and a multi-category cross entropy loss; A model training module, used for determining an overall segmentation loss according to the multiple segmentation losses, and training the intracranial artery multi-category segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-category segmentation model; The second prediction module is used to extract a target region of interest block including the circle of Willis from the acquired target whole-brain image, and input the target region of interest block into the trained intracranial artery multi-category segmentation model to predict and obtain a segmentation result of multiple blood vessel segments of the circle of Willis on the target region of interest block.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the multi-category segmentation method of intracranial arteries based on anatomical and topological perception is implemented as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the multi-category segmentation method of intracranial arteries based on anatomical and topological perception is implemented as claimed in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the multi-category segmentation method of intracranial arteries based on anatomical and topological perception is implemented as claimed in any one of claims 1 to 6.

Images