Anatomical- and Topology-Aware Multi-Class Segmentation Method for Intracranial Arteries

Abstract

The present invention relates to the field of image segmentation, and provides a multi-class segmentation method for intracranial arteries based on anatomy and topology perception, including: extracting a region of interest block containing the circle of Willis from the acquired whole-brain image to obtain an input image and a ground truth annotation image. Inputting the input image into a pre-constructed multi-class segmentation model for intracranial arteries to obtain a multi-class prediction probability map. Calculating a radius-corrected Dice loss, a vascular breakage perception loss, an adjacency relationship perception loss, and a multi-class cross-entropy loss based on the prediction probability map, the ground truth annotation image, and a weight map. Determining an overall segmentation loss according to these losses and training the model to obtain a trained multi-class segmentation model for intracranial arteries. Extracting a target region of interest block containing the circle of Willis from the acquired target whole-brain image, and inputting it into the trained model to obtain the segmentation result of multiple vascular segments of the circle of Willis on the target region of interest block, realizing the accurate segmentation of multiple-class vascular segments of the circle of Willis.

Description

Technical Field

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

[0002] The brain is supplied with blood by two internal carotid arteries and two vertebral arteries, which are connected through the central anastomosis at the bottom 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, can all 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 an increase in the diameter of the carotid artery and an incomplete Circle of Willis, are also risk factors for neurovascular diseases. The manual annotation of the Circle of Willis structure consumes extremely high human, material, and time costs. The Circle of Willis data obtained by relying on manual annotation has disadvantages such as poor reproducibility, low efficiency, and high complexity. In order to effectively assist doctors in observing and diagnosing neurovascular diseases, the automatic segmentation of different vascular segments within the Circle of Willis vascular tree is of great significance.

[0003] Existing segmentation models still have some deficiencies for the vascular segmentation task. First, the Circle of Willis structure contains multiple types of vascular segments, such as the carotid artery, anterior, middle, and posterior cerebral arteries, and anterior and posterior communicating arteries. There are significant differences in the diameter and length between these vascular segments. The vascular segments with a large diameter and long length, as segmentation targets, occupy a much larger proportion of voxels in the image than those with a small diameter and short length. This results in the problem of class imbalance in the multi-class segmentation task, making the segmentation performance of the model for large targets significantly better than that for small targets. Second, the loss functions of existing segmentation models mainly focus on the segmentation correctness of single voxels themselves, ignoring the learning of the spatial neighborhood information of slender structures such as blood vessels, which will lead to the segmentation of blood vessels being discontinuous. In addition, for the special connection structure of the Circle of Willis, existing segmentation methods often ignore the prior knowledge that there are specific connection relationships between different types of vascular segments, resulting in incorrect vascular pair connections and the problem of vascular segment confusion during model prediction, which makes the segmentation accuracy of multi-class vascular segments of the Circle of Willis low. Summary of the Invention

[0004] The present invention provides a multi-class segmentation method for intracranial arteries based on anatomical and topological perception to solve the defect of low segmentation accuracy of multi-class vascular segments of the Circle of Willis in the prior art and achieve accurate segmentation of multi-class vascular segments of the Circle of Willis. The technical solutions proposed by the present invention are as follows:

[0005] In a first aspect, the present invention provides a multi-class segmentation method for intracranial arteries based on anatomical and topological perception, including:

[0006] Extract a region of interest block containing the circle of Willis from the acquired whole-brain image to obtain an input image and a ground truth image, and determine the weight of each vascular voxel based on the ground truth image to obtain a weight map;

[0007] Input the input image into a pre-constructed multi-class segmentation model for intracranial arteries to obtain a multi-class prediction probability map;

[0008] Calculate multiple segmentation losses based on the prediction probability map, the ground truth image, and the weight map, where the multiple segmentation losses include a Dice loss based on radius correction, a vascular break perception loss, an adjacency relationship perception loss, and a multi-class cross-entropy loss;

[0009] Determine an overall segmentation loss according to the multiple segmentation losses, and train the multi-class segmentation model for intracranial arteries based on the overall segmentation loss to obtain a trained multi-class segmentation model for intracranial arteries;

[0010] Extract a target region of interest block containing the circle of Willis from the acquired target whole-brain image, and input the target region of interest block into the trained multi-class segmentation model for intracranial arteries to predict the segmentation result of multiple vascular segments of the circle of Willis on the target region of interest block.

[0011] Optionally, the method further includes:

[0012] Extract the coordinates of the target region of interest block on the target whole-brain image, and restore the segmentation result of multiple vascular segments 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 multiple vascular segments of the circle of Willis on the whole brain.

[0013] Optionally, the multi-class segmentation model for intracranial arteries is a three-dimensional U-shaped network, and the three-dimensional U-shaped network includes a pair of encoders and decoders:

[0014] The encoder includes a plurality of downsampling modules, and the plurality of downsampling modules are connected in sequence and gradually reduce the resolution;

[0015] The decoder includes a plurality of upsampling modules, and the plurality of upsampling modules are connected in sequence and gradually increase the resolution;

[0016] The number of upsampling modules is the same as the number of downsampling modules, and the upsampling modules are connected to the corresponding downsampling modules at corresponding positions.

[0017] Optionally, the Dice loss based on radius correction is determined by the following method:

[0018] Multiply the weight map with the ground truth image and the predicted probability map respectively to obtain the weighted ground truth image and the weighted predicted probability map;

[0019] Based on the weighted ground truth image and the weighted predicted probability map, use the Dice coefficient formula to calculate the Dice loss based on radius correction.

[0020] Optionally, the blood vessel break perception loss is determined in the following manner:

[0021] For each voxel, determine the class with the maximum probability in the predicted probability map to obtain the target class;

[0022] Create one-hot encodings for the predicted probability map and the ground truth image respectively according to the target class to obtain the first encoded image and the second encoded image;

[0023] Use 3D grouped convolution operations to perform local summation on the first encoded image and the second encoded image for each class to obtain the first convolution map and the second convolution map for each class;

[0024] Determine the weighted mean square error between the first convolution map and the second convolution map for each class according to the weight map to obtain the blood vessel break perception loss.

[0025] Optionally, the adjacency relation perception loss is determined in the following manner:

[0026] For each voxel, determine the class with the maximum probability in the predicted probability map to obtain the predicted class image;

[0027] For each class, extract the segmentation map according to the class and the predicted class image, binarize and dilate the segmentation map, and multiply the dilated result with the predicted class image to obtain the dilated neighborhood class superposition map;

[0028] Extract the blood vessel segment classes existing in the dilated neighborhood class superposition map to obtain the adjacency class list, compare the adjacency class list with the pre-established prior knowledge list to determine the false positive classes and the false negative classes, and obtain the false positive class list and the false negative class list; among them, the false positive class is the neighbor class that appears in the adjacency class list but does not appear in the prior knowledge list, and the false negative class is the neighbor class that appears in the prior knowledge list but does not appear in the adjacency class list;

[0029] Determine the false positive adjacency error voxel set according to the false positive class list, the predicted class image and the ground truth image, and determine the false negative adjacency error voxel set according to the false negative class list, the predicted class image and the ground truth image;

[0030] Multiply the false positive adjacent error voxel set, the false negative adjacent error voxel set, and a pre-determined second encoded image with the prediction probability map and the weight map respectively 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; wherein, the second encoded image is a one-hot encoded image of the true annotation image.

[0031] According to the probability weighted map of the false positive adjacent error voxel set, the probability weighted map of the false negative adjacent error voxel set, and the probability weighted map of the true positive voxel set, use the Dice coefficient formula to calculate the adjacent relationship perception loss.

[0032] In a second aspect, the present invention further provides an intracranial artery multi-class segmentation device based on anatomy and topology perception, including the following modules:

[0033] A weight determination module, configured to extract a region of interest block containing the Willis circle from the acquired whole brain image to obtain an input image and a true annotation image, and determine the weight of each vascular voxel based on the true annotation image to obtain a weight map.

[0034] A first prediction module, configured to input the input image into a pre-constructed intracranial artery multi-class segmentation model to obtain a multi-class prediction probability map.

[0035] A loss determination module, configured to calculate multiple segmentation losses based on the prediction probability map, the true annotation image, and the weight map, and the multiple segmentation losses include a Dice loss based on radius correction, a vascular break perception loss, an adjacent relationship perception loss, and a multi-class cross-entropy loss.

[0036] A model training module, configured to determine an overall segmentation loss according to the multiple segmentation losses, and train the intracranial artery multi-class segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-class segmentation model.

[0037] A second prediction module, configured to extract a target region of interest block containing the Willis circle from the acquired target whole brain image, and input the target region of interest block into the trained intracranial artery multi-class segmentation model to predict the segmentation result of multiple vascular segments of the Willis circle on the target region of interest block.

[0038] In a third aspect, the present invention further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, and when the processor executes the computer program, it implements the intracranial artery multi-class segmentation method based on anatomy and topology perception as described in the first aspect above.

[0039] Fourthly, the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method for multi-class segmentation of intracranial arteries based on anatomy and topology perception as described in the first aspect above.

[0040] Fifthly, the present invention further provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the method for multi-class segmentation of intracranial arteries based on anatomy and topology perception as described in the first aspect above.

[0041] Based on the above technical solutions, the beneficial effects of the present invention compared with the prior art are as follows:

[0042] The method for multi-class segmentation of intracranial arteries based on anatomy and topology perception provided by the present invention solves the problem of class imbalance by assigning weights to each vascular voxel based on real annotation images to generate a weight map. Vessels with a smaller radius or key vascular segments will be assigned higher weights to receive more attention during the segmentation process, balancing the contributions of large and small vessels to the loss function, improving the segmentation performance of the model for small vascular segments, and reducing the problem of inaccurate segmentation caused by class imbalance. The vessel breakage perception loss takes into account the continuity of vessels by introducing it, enabling the model to learn the spatial neighborhood information of vessels, thereby reducing the occurrence of broken vessels in the segmented vessels. This significantly reduces the incidence of vessel breakage and improves the integrity and continuity of vessel segmentation. The adjacency relationship perception loss utilizes the anatomical structure information of vessels, especially the adjacency relationship between vessels, to ensure that the segmentation result conforms to the expected anatomical structure. By introducing the prior knowledge of vascular anatomical connections, the model can learn that there are specific adjacency relationships between different classes of vascular segments, thereby reducing the problems of vascular segment confusion and incorrect connection. This improves the accuracy of vascular segment classification, makes the segmentation result more in line with anatomical features, and realizes the accurate segmentation of multi-class vascular segments of the Willis circle.

[0043] Other features and advantages of the present invention will be described in the subsequent description. Moreover, some of them will become obvious from the description, or can be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the description, claims, and drawings.

[0044] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and detailed descriptions are provided in conjunction with the accompanying drawings as follows. Description of the Drawings

[0045] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0046] Figure 1 It is a schematic flowchart of the multi-class segmentation method for intracranial arteries based on anatomy and topology perception provided by the present invention.

[0047] Figure 2 It is a schematic structural diagram of the three-dimensional U-shaped network provided by the present invention.

[0048] Figure 3 It is a schematic diagram of the model training stage provided by the present invention.

[0049] Figure 4 It is a schematic structural diagram of the multi-class segmentation device for intracranial arteries based on anatomy and topology perception provided by the present invention.

[0050] Figure 5 It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0052] The following combines Figures 1-3 to describe the multi-class segmentation method for intracranial arteries based on anatomy and topology perception of the present invention.

[0053] The multi-class segmentation method for intracranial arteries based on anatomy and topology perception can enable the model to pay attention to the special connection relationship characteristics between the vascular segments of the circle of Willis and the spatial neighborhood characteristics of the blood vessels, and directly obtain multi-class segmentation results. Referring to Figure 1 as shown, the method includes the following:

[0054] Step S110: Extract a region of interest block containing the circle of Willis from the obtained whole-brain image to obtain an input image and a ground truth image, and determine the weight of each vascular voxel based on the ground truth image to obtain a weight map.

[0055] First, obtain whole-brain image data from medical imaging devices such as MRI or CTA. Using image processing techniques, extract the region of interest (ROI) block containing the circle of Willis 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 ground truth annotation images for the extracted ROI through manual annotation or by using existing annotation datasets. These annotation images indicate the boundaries of different vascular categories in the ROI. Based on the ground truth annotation images, assign weights to each vascular voxel (i.e., vascular pixel or voxel) to generate a weight map. The weights are determined based on the thickness of the blood vessels.

[0056] Specifically, for training the intracranial artery multi-class segmentation model, refer to Figure 3 As shown, first, a set of normal time-of-flight magnetic resonance angiography (TOF-MRA) data containing the circle of Willis labels needs to be prepared. These data have been manually annotated to clarify the boundaries of the circle of Willis and its various vascular segments.

[0057] Given that the whole-brain image data is huge and complex, and the circle of Willis only occupies a small key area of the whole-brain image, it is necessary to extract the region of interest (ROI) block containing the circle of Willis from the whole-brain image to improve the training efficiency and accuracy of the model. The process of extracting the ROI is as follows:

[0058] First, perform template definition: On the standard TOF-MRA image in the standard Montreal Neurological Institute (MNI) space, manually frame the ROI area containing the circle of Willis using a cube of appropriate size. This cube template will be used as a reference for subsequent ROI extraction.

[0059] Next, perform registration: Use the registration algorithm to register the TOF-MRA image defined in the standard space to the actual input image space. The purpose of registration is to ensure that the ROI template in the standard space can be accurately mapped to each input image, so as to extract the corresponding ROI area.

[0060] ROI cropping: According to the registration result, crop the ROI area from the input image. At the same time, use the corresponding annotation data to crop the corresponding ground truth annotation image. These cropped ROI images and annotation images will be used as input data for model training.

[0061] Through the above steps, it is possible to efficiently extract the ROI containing the circle of Willis from the whole-brain images and prepare the input images and ground truth annotation images for network training. This not only reduces the redundant information that the model needs to process but also improves the segmentation accuracy of the circle of Willis and its vascular segments by the model.

[0062] Step S120: Input the input image into a pre-constructed multi-class segmentation model for intracranial arteries to obtain a multi-class prediction probability map.

[0063] Pre-construct a multi-class segmentation model for intracranial arteries. This model can be a deep learning model, such as the architecture of a three-dimensional U-shaped network (as shown in Figure 2 ), which is capable of processing three-dimensional image data and outputting a multi-class prediction probability map. Input the input image (i.e., the ROI containing the circle of Willis) into the model. The model outputs a multi-class prediction probability map, where each voxel corresponds to a probability distribution representing belonging to different vascular classes. The multi-class prediction probability map contains a total of C channels, and each channel contains an image of the same size as the ground truth annotation image. The value of the voxel on the image in the i-th channel represents the prediction probability that the position belongs to the i-th class.

[0064] Step S130: Calculate multiple segmentation losses based on the prediction probability map, the ground truth annotation image, and the weight map. The multiple segmentation losses include the Dice loss based on radius correction, the blood vessel breakage perception loss, the adjacency relationship perception loss, and the multi-class cross-entropy loss.

[0065] Considering that in the same input image, small blood vessels are much smaller in diameter and volume than large blood vessels, and it is very easy to have a class imbalance problem in the multi-class segmentation learning process. Therefore, the present application constructs a Dice loss function based on radius correction. 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, guiding the model to focus on the feature learning of small-radius blood vessels. The Dice loss based on radius correction takes into account the change of blood vessel radius to improve the segmentation accuracy of small blood vessels and fine structures.

[0066] Existing segmentation loss functions mainly focus on the class classification accuracy of voxels themselves. However, as an object with strong spatial topological features, the spatial neighborhood information of blood vessels should also be emphasized in learning. The way to maintain the spatial connectivity of blood vessels in manual annotation is to annotate frame by frame through the way of neighborhood movement. Inspired by the manual annotation method, the present invention constructs a blood vessel breakage perception loss function, which can quickly identify the voxel regions where blood vessel breakage occurs using neighborhood information. This loss function aims to detect and punish the situation of blood vessel breakage, that is, the discontinuity or interruption of blood vessels in the model prediction results.

[0067] For the special connection structure of the circle of Willis, existing segmentation methods often ignore the prior knowledge that different types of vascular segments have specific anatomical connections, resulting in incorrect vascular pair connections during model prediction. Therefore, this application proposes a loss function based on adjacency relation perception, incorporating category adjacency relations into network supervised learning to improve automatic segmentation performance. This loss function utilizes the anatomical structure information of blood vessels, especially the adjacency relations between blood vessels, to ensure that the segmentation results conform to the expected anatomical structure.

[0068] The above multi-class cross-entropy loss is used to measure the difference between the probability distribution predicted by the model and the ground truth annotation.

[0069] Refer to Figure 3 As shown, the dotted line pointed by the loss function indicates the two input graphs used for comparison in this loss function, and the dotted line pointing to the loss function in the image represents that the image participates in the calculation of this loss function. This application calculates the multi-class cross-entropy loss and the Dice loss based on radius correction respectively based on the obtained multi-class prediction probability map and the ground truth annotation result (i.e., the above ground truth annotation image). Calculate the blood vessel radius weight map based on the ground truth annotation result, and this weight map is used for the calculation of the Dice loss based on radius correction. Perform convolution operations on the multi-class prediction probability map and the ground truth annotation result respectively to obtain the predicted neighborhood fusion result and the ground truth annotation neighborhood result. Calculate the blood vessel breakage perception loss based on the obtained predicted neighborhood fusion result and the ground truth annotation neighborhood result. Calculate the adjacency relation perception Dice loss based on the multi-class prediction results and the ground truth annotation result in the multi-class prediction probability map, that is, the above adjacency relation perception loss .

[0070] Step S140: Determine the overall segmentation loss according to the multiple segmentation losses, and train the intracranial artery multi-class segmentation model based on the overall segmentation loss to obtain the trained intracranial artery multi-class segmentation model.

[0071] Sum the above multiple loss functions with weights to obtain the overall segmentation loss . Use the backpropagation algorithm and an optimizer (such as Adam or SGD) to minimize the overall segmentation loss, thereby training the intracranial artery multi-class segmentation model. Evaluate the performance of the model on the validation set and make parameter adjustments or model optimizations as needed.

[0072] The trained intracranial artery multi-class segmentation model can achieve end-to-end segmentation of the multi-class vascular segments of the circle of Willis, and to the greatest extent preserve the connectivity of the predicted blood vessels, and the predicted blood vessels conform to the particularity of the inter-segment connection of the circle of Willis specific blood vessels. The intracranial artery is a complex network structure, and there are specific connection relationships between different vascular segments. The intracranial artery multi-class segmentation model of the present invention can capture these relationships to ensure the correctness and integrity of the segmentation results. By introducing topological awareness, the model can learn the connection relationships between vascular segments, thereby reducing incorrect connections in the segmentation results, which is crucial for 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 to achieve anatomical awareness and topological awareness, and improve the accuracy and reliability of the segmentation results.

[0073] Step S150: Extract a target region of interest block containing the circle of Willis from the obtained target whole-brain image, and input the target region of interest block into the trained intracranial artery multi-class segmentation model to predict the segmentation result of the multi-vascular segments of the circle of Willis on the target region of interest block.

[0074] Extract a target region of interest (ROI) containing the circle of Willis from the new whole-brain image, i.e., the above-mentioned target whole-brain image. Input the target ROI into the trained intracranial artery multi-class segmentation model to obtain the segmentation result of the multi-vascular segments. Then, the prediction result can be visualized, quantitatively evaluated and analyzed to verify the accuracy and reliability of the model.

[0075] The intracranial artery multi-class segmentation method based on anatomical and topological awareness provided by the present invention solves the problem of class imbalance by assigning weights to each vascular voxel based on the real annotation image to generate a weight map. Blood vessels with a smaller radius or key vascular segments are given higher weights to pay more attention during the segmentation process, so that the contributions of large blood vessels and small blood vessels to the loss function are balanced, improving the segmentation performance of the model for small vascular segments and reducing the problem of inaccurate segmentation caused by class imbalance.

[0076] The present invention designs a vascular rupture perception loss. By introducing the consideration of vascular continuity, the model can learn the spatial neighborhood information of blood vessels, thereby reducing the occurrence of ruptures in the segmented blood vessels, significantly reducing the incidence of vascular ruptures, and improving the integrity and continuity of vascular segmentation. The adjacency relationship perception loss utilizes the anatomical structure information of blood vessels, especially the adjacency relationship between blood vessels, to ensure that the segmentation results conform to the expected anatomical structure. By introducing prior anatomical knowledge, the model can learn that there are specific connection relationships between different categories of vascular segments, thereby reducing the problems of vascular segment confusion and incorrect connection. This improves the accuracy of vascular segment classification and makes the segmentation results more in line with anatomical features. Moreover, different from the traditional idea of first segmenting and then annotating, the present invention directly designs a multi-class segmentation model that can simultaneously complete the segmentation and classification tasks. Through the integrated design, the model can simultaneously consider the category information of different vascular segments during the segmentation process, thereby reducing the accumulation of segmentation errors and improving the overall accuracy of segmentation and classification. Especially for small blood vessels that are prone to segmentation rupture, the effect is more significant.

[0077] In an optional embodiment, the intracranial artery multi-class segmentation model is a three-dimensional U-shaped network, and the three-dimensional U-shaped network includes a pair of encoders and decoders:

[0078] The encoder includes a plurality of downsampling modules, and the plurality of downsampling modules are connected in sequence and gradually reduce the resolution; the decoder includes a plurality of upsampling modules, and the plurality of upsampling modules are connected in sequence and gradually increase the resolution; the number of upsampling modules is the same as the number of downsampling modules, and the upsampling modules are connected to the corresponding downsampling modules at corresponding positions.

[0079] Referring to Figure 2 as shown, the encoder is composed of M downsampling modules which are connected in sequence from to and gradually reduce the resolution. Each downsampling module may include a convolutional layer, an activation function, and a pooling layer (or a downsampling layer) for extracting features of the image and reducing the spatial dimension. The decoder is composed of M upsampling modules which are connected in sequence from to and gradually increase the resolution. Each upsampling module may include an upsampling layer (or a transposed convolutional layer), a convolutional layer, and an activation function for mapping the extracted features back to the original image space. There are m paths between the encoder and the decoder, and these paths connect the corresponding downsampling modules and upsampling modules By using skip connections to directly transfer features from the encoder to corresponding positions in the decoder, it helps to preserve spatial details and alleviate the vanishing gradient problem. This design enables the network to better learn the global and local features of the image. Figure 2 in terms of the number of modules in the encoder / decoder as an example.

[0080] Through the downsampling module in the encoder of the present invention, the network can gradually extract low-level, mid-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 the segmentation result. Due to the existence of skip 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.

[0081] In an optional embodiment, considering that in the same input image, small blood vessels are much smaller than large blood vessels in terms of diameter and volume, and it is very easy to have a class imbalance problem during the multi-class segmentation learning process. Aiming at the problem of class imbalance in existing segmentation networks for small-diameter blood vessel segments, the present invention proposes a loss function based on blood vessel radius correction, which can enhance the model's learning of small-diameter blood vessel features, ensure the segmentation performance of small blood vessels, and alleviate the class imbalance problem. By introducing radius information into the calculation of the loss function, the influence of small-radius blood vessels in the loss function calculation is increased, guiding the model to focus on the feature learning of small-radius blood vessels. The radius-corrected Dice loss is determined as follows:

[0082] S210. Multiply the weight map with the ground truth image and the predicted probability map respectively to obtain the weighted ground truth image and the weighted predicted probability map.

[0083] First, calculate the blood vessel radius on the CPU using the ground truth image and save it as a radius image According to the calculated radius, calculate the weight of each blood vessel voxel to obtain a weight map : For blood vessel voxels, the larger the radius of the voxel, the closer the weight value is to 1, and the smaller the radius, the closer it is to 2, while background voxels (i.e., voxels belonging to the background region) obtain a unified weight value of 1. The weight value calculation formula is:

[0084]

[0085] where represents the maximum radius value among all voxels belonging to the circle of Willis, Represents the minimum radius value among all voxels belonging to the Circle of Willis. and represent the maximum function and the minimum function respectively. represents a voxel, represents the Circle of Willis, represents a voxel belonging to the Circle of Willis, represents a voxel 's radius value. represents a voxel 's weight value, represents the background region, represents a voxel belonging to the background region.

[0086] Add the calculated weight map to the calculation of the Dice loss. The original Dice loss is calculated as follows:

[0087]

[0088] where, represents the coefficient function, is the ground truth image, which is a binary image or label image, where the value of each voxel indicates whether the position belongs to the Circle of Willis region and the category of the blood vessel segment in the Circle of Willis region. represents the predicted probability map, represents the logical AND operation.

[0089] is the intersection of the ground truth image and the predicted probability map. represents the sum of the voxel values in the intersection of the ground truth image and the predicted probability map. represents the sum of the voxel values belonging to the Circle of Willis in the ground truth image, represents the sum of the voxel values of the predicted probability map.

[0090] Multiply the calculated weight map separately with the ground truth image and the predicted probability map , then the calculation method is modified to:

[0091]

[0092] where, represents the weighted ground truth image, represents the weighted predicted probability map.

[0093] Multiply the weight map with the ground truth annotation image and the predicted probability map respectively to obtain the weighted ground truth annotation image and the weighted predicted probability map. Adjusting the voxel values in the ground truth annotation image and the predicted probability map according to the weight map can enhance the voxel values in the key regions.

[0094] S220. Based on the weighted ground truth annotation image and the weighted predicted probability map , use the Dice coefficient formula to calculate the Dice loss based on radius correction .

[0095]

[0096] where represents the Dice loss based on radius correction, represents the coefficient function, represents the logical AND operation. represents the sum of the voxel values of the intersection of the weighted ground truth annotation image and the weighted predicted probability map, represents the sum of the voxel values of the weighted ground truth annotation image, represents the sum of the voxel values of the weighted predicted probability map.

[0097] The Dice loss based on radius correction in the present invention helps the model to better focus on the radius changes of arteries by introducing the weight map and radius information, so that the model can more accurately identify and segment the key regions of intracranial arteries, 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 in various situations. The Dice loss based on radius correction, as a loss function, can guide the training process of the model. By minimizing this loss function, the model can gradually learn a more accurate segmentation strategy, thereby improving the training efficiency and convergence speed.

[0098] In an alternative embodiment, existing segmentation loss functions mainly focus on the classification accuracy of voxels themselves. However, as blood vessels are objects with strong spatial topological features, their spatial neighborhood information should also be emphasized in learning. In manual annotation, the way to maintain the spatial connectivity of blood vessels is to annotate frame by frame through neighborhood movement. Drawing on the idea of manual annotation, in view of the problem that existing segmentation networks are insufficient in extracting blood vessel neighborhood information, resulting in broken segmented blood vessels, the present invention proposes a blood vessel breakage perception loss function that combines neighborhood information to enhance the learning of features at easily broken parts of blood vessels and improve the problem of blood vessel breakage. The present invention can quickly identify the voxel regions where blood vessel breakage occurs using neighborhood information. The blood vessel breakage perception loss is determined in the following manner:

[0099] S310. For each voxel, determine the target class by obtaining the class with the highest probability in the predicted probability map of the voxel;

[0100] For a voxel with coordinates (x, y, z), find the class with the highest probability in its predicted probability map, and this class is the target class of the voxel :

[0101]

[0102] wherein, represents the target class, that is, the class index with the highest predicted probability at coordinates (x, y, z). When performing one-hot encoding, according to this index, it is determined which class channel of the voxels is non-zero. represents the class, that is, the class index to which the current voxel may belong. represents that the voxel at coordinates (x, y, z) belongs to class of the predicted probability. represents finding the class with the highest probability in the predicted probability map.

[0103] S320. Create one-hot encodings for the predicted probability map and the ground truth annotation image respectively according to the target class, to obtain a first encoded image and a second encoded image.

[0104] According to the target class, perform one-hot encoding on the predicted probability map and the ground truth annotation image respectively. One-hot encoding is a method for representing class labels. It converts a class label into a vector in which only one position is 1 (representing the target class), and the remaining positions are 0. In this way, the predicted probability map and the ground truth annotation image are converted into a first encoded image and a second encoded image, which respectively represent the predicted and true class distributions.

[0105] Specifically, for the predicted probability map Create one-hot encoding with the channel axis number C for the predicted probability map P and the ground truth annotation image T:

[0106]

[0107] Wherein, and respectively represent the one-hot encoding of the voxel with coordinates and category in the predicted probability map and the ground truth annotation image. and respectively represent the first encoded image and the second encoded image. represents the predicted probability that the voxel at coordinates (x, y, z) belongs to category . represents that at coordinates (x, y, z), the category index of the predicted voxel is . represents others. represents the category index of the ground truth annotation at coordinates . represents the voxel with the value of category at coordinates in the ground truth annotation image . In , only the voxel predicted as category has a non-zero value (i.e., the predicted probability) at the corresponding position . In , only the voxel with the ground truth annotation as category is 1 at the corresponding position , and 0 at other positions.

[0108] S330. Use 3D grouped convolution operations to perform local summation on the first encoded image and the second encoded image of each category respectively, to obtain the first convolution map and the second convolution map of each category;

[0109] Use 3D grouped convolution operations to perform local summation on the first encoded image and the second encoded image of each category respectively. Through 3D grouped convolution, features can be extracted in three dimensions (depth, height, and width). Through local summation, the first local feature map and the second local feature map of each category can be obtained, which respectively represent the distribution characteristics of the predicted and true categories in the local region.

[0110] Specifically, use 3D grouped convolution operations to apply 3×3×3 convolution to and respectively. Each weight value 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 a local sum on the one-hot encoded images of each category to obtain the first convolutional map and the second convolutional map :

[0111]

[0112]

[0113] where represents the offset in the axis direction, ranging from -1 to 1, with a total of three positions (including the original position ). represents the offset in the axis direction, also ranging from -1 to 1, with three positions as well. represents the offset in the axis direction, also ranging from -1 to 1, with three positions too. represents the eigenvalue of the voxel with coordinates and category c in the first convolutional map, represents the eigenvalue of the voxel with coordinates and category c in the second convolutional map.

[0114] S340. Determine the weighted mean square error between the first convolutional map and the second convolutional map of each category according to the weight map to obtain the vascular rupture perception loss.

[0115] Calculate and to obtain the above-mentioned vascular rupture perception loss by calculating the weighted mean square error (Mean Square Error, MSE) , where the weights come from the radius-based weight map calculated in the above step S210 , where X, Y, and Z are the number of voxels along the x, y, and z axis directions respectively, and C is the number of channels:

[0116]

[0117] where represents the weight value at the coordinate .

[0118] Through one-hot encoding and 3D grouped convolution operations, the present invention can accurately represent and extract the class distribution and local features in the blood vessel rupture task. This enables the model to better distinguish different classes of voxels and capture the key features of blood vessel rupture, thereby improving the accuracy of perception. The introduction of the weight map enables the model to better adapt to the blood vessel rupture task under different circumstances. By adjusting the weight values, the influence of key regions can be highlighted, thereby enhancing the robustness and generalization ability of the model. The blood vessel rupture perception loss is used as the loss function to guide the training process of the model. By minimizing this loss function, the model can gradually learn a more accurate perception strategy, thereby improving the training efficiency and convergence speed.

[0119] In an optional embodiment, for the special connection structure of the circle of Willis, existing segmentation methods often ignore the prior knowledge that different classes of blood vessel segments have specific anatomical connections, resulting in incorrect blood vessel pair connections during model prediction. Aiming at the problem that existing multi-class segmentation networks lack learning of the connection relationship between blood vessel segments of the circle of Willis and there is confusion of blood vessel segments, the present invention proposes an adjacency relationship perception loss function, which uses the prior inter-class connection relationship to guide the network to learn the discrimination between classes, improves the problem of blood vessel segment confusion, and improves the automatic segmentation performance. The adjacency relationship perception loss is determined in the following manner:

[0120] S410. For each voxel, determine the class with the highest probability of the voxel in the predicted probability map to obtain a predicted class image.

[0121] For each voxel with coordinates (x, y, z), find its class with the highest probability in the predicted probability map, that is, the target class , thereby generating a predicted class image . This image reflects the model's prediction of the class of each voxel.

[0122]

[0123] Wherein, represents the target class. represents the predicted probability that the voxel at coordinates (x, y, z) belongs to class . This formula indicates that the voxel value at coordinates in the generated predicted class image is the class corresponding to the predicted maximum probability.

[0124] S420. For each class, extract a segmentation map according to the class and the predicted class image, binarize and dilate the segmentation map, and multiply the dilated result by the predicted class image to obtain a dilated neighborhood class superposition map.

[0125] For each category , extract the corresponding segmentation map from the predicted category image , and perform binarization on it, marking the voxels belonging to this category as 1 and the remaining voxels as 0. Perform a dilation operation on the binarized segmentation map to expand the range of the category region, and multiply the resulting map after the dilation operation by the original segmentation category image, i.e., the above-mentioned predicted category image to obtain the dilated neighborhood category overlay map , which reflects the category information of each voxel and its dilated neighborhood.

[0126]

[0127]

[0128]

[0129] Among them, represents the segmentation map of category . In the predicted category image , only when the predicted category of a voxel is equal to , the value of at this coordinate is true, otherwise it is zero.

[0130] represents performing binarization on , that is, the process of converting into a process that only contains two values (such as 0 and 1), converting the non-zero values in to 1 and keeping the zero values unchanged. represents performing a dilation operation on the binarized . is the resulting map after the dilation operation, which contains the boundaries of the expanded category objects. is the dilated neighborhood category overlay map, obtained by multiplying the resulting map after the dilation operation by the predicted category image to mark those voxels adjacent to the target of category and these voxels still belong to a certain category.

[0131] S430. Extract the vascular segment categories existing in the superimposed inflated neighborhood category map to obtain an adjacency category list. Compare the adjacency category list with a pre-established prior knowledge list to determine false positive categories and false negative categories, and obtain a false positive category list and a false negative category list. Among them, a false positive category is a neighbor category that appears in the adjacency category list but does not appear in the prior knowledge list, and a false negative category is a neighbor category that appears in the prior knowledge list but does not appear in the adjacency category list.

[0132] Extract the superimposed inflated neighborhood category map for the existing vascular segment categories, and save them as a list, i.e., the above adjacency category list , and compare it with the prior knowledge list of the adjacent vascular segment categories of category c . Among those that appear but do not appear in , the neighbor categories are false positive (False Positive) categories, and a false positive category list is obtained. And among those that appear but do not appear in , the neighbor categories are false negative (False Negative) categories, and a false negative category list is obtained:

[0133]

[0134]

[0135]

[0136] Among them, is a function used to extract a unique value list from . represents a logical AND operation, represents not belonging to.

[0137] S440. Determine a false positive adjacent error voxel set based on the false positive category list, the predicted category image, and the true annotation image. These voxels are wrongly predicted to be adjacent to certain categories. And determine a false negative adjacent error voxel set based on the false negative category list, the predicted category image, and the true annotation image. These voxels fail to be correctly predicted to be adjacent to certain categories.

[0138] Calculate the occurrence of the false positive adjacent error voxel set and the occurrence of the false negative adjacent error voxel set according to the following conditions:

[0139]

[0140] Among them, represents an exclusive OR operation. Represents a logical AND operation. By merging the categories and along the channel axis respectively, the false positive adjacent error voxel set and the false negative adjacent error voxel set can be obtained. Indicates the category with an adjacent relationship error with the predicted category , that is, having an adjacent relationship with the category by mistake, or failing to have an adjacent relationship with the category by mistake.

[0141] Indicates whether the voxel at coordinate under the category belongs to the false positive adjacent error voxel set. If the condition is met, the value is 1, otherwise it is 0.

[0142] Indicates whether the voxel at coordinate under the category belongs to the false negative adjacent error voxel set. If the condition is met, the value is 1, otherwise it is 0.

[0143] Indicates that the voxel at coordinate is predicted to be the category .

[0144] Indicates that the voxel at coordinate is predicted to be the category .

[0145] Indicates that the true class label of the voxel at coordinate is .

[0146] S450. Multiply the false positive adjacent error voxel set, the false negative adjacent error voxel set, and the pre-determined second encoded image with the prediction probability map and the weight map respectively to obtain the probability weighted map of the false positive adjacent error voxel set, the probability weighted map of the false negative adjacent error voxel set, and the probability weighted map of the true positive voxel set; wherein, the second encoded image is the one-hot encoded image of the true annotation image, and the determination method of the second encoded image can refer to the description of S310 - S320 above;

[0147] Multiply , , with the prediction probability map and the weight map calculated in step S210 above Multiplication:

[0148]

[0149]

[0150]

[0151] wherein, represents the probability-weighted graph of the false positive adjacent error voxel set, represents the probability-weighted graph of the false negative adjacent error voxel set, represents the probability-weighted graph of the true positive voxel set.

[0152] S460. 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, the adjacent relationship perception loss is calculated using the Dice coefficient formula. It is used to evaluate the consistency of the predicted category image and the true annotation image in the anatomical adjacent relationship.

[0153] Calculate the final loss function using the Dice coefficient formula, that is, the above adjacent relationship perception loss :

[0154]

[0155] wherein, , , respectively represent the sum of the voxel values in the probability-weighted graph of the true positive voxel set, the probability-weighted graph of the false positive adjacent error voxel set, and the probability-weighted graph of the false negative adjacent error voxel set.

[0156] Overall segmentation loss The calculation formula is as follows:

[0157]

[0158] wherein, and are hyperparameters that can be adjusted according to the effect to adapt to different datasets.

[0159] By introducing a list of prior knowledge and comparing it with the adjacent category list, the present invention can accurately identify false positive categories and false negative categories, thereby correcting the incorrect adjacency relationships in the predicted category images. This helps to improve the accuracy of anatomical relationships and provides a reliable basis for subsequent medical analysis and diagnosis. By considering the constraints of anatomical adjacency relationships, this 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 and enables it to 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 this loss function, the model can gradually learn more accurate anatomical adjacency relationships, thereby improving the training efficiency and convergence speed.

[0160] In an optional embodiment, after obtaining the segmentation result of the multi-vessel segments of the Circle of Willis on the target region of interest block, referring to Figure 1 as shown, the method further includes:

[0161] S160. Extract the coordinates of the target region of interest block on the target whole-brain image, and restore the segmentation result of the multi-vessel segments of the Circle of Willis from the target region of interest block to the whole-brain space to obtain the segmentation result of the multi-vessel segments of the Circle of Willis on the whole brain.

[0162] In the model inference stage, for the segmentation task of the multi-vessel segments of the Circle of Willis in the whole-brain time-of-flight magnetic resonance angiography (TOF-MRA) image, first, use registration to locate the region of interest (ROI) where the Circle of Willis is located in the whole-brain TOF-MRA image. This step aims to quickly identify and extract a small part of the image region containing the key vascular structures of the Circle of Willis to reduce the computational amount and complexity of subsequent processing. Then, according to the positioning result, crop the whole-brain image to obtain an image block containing only the ROI of the Circle of Willis, that is, the above-mentioned target region of interest block. This cropping process needs to accurately record the coordinate information of the ROI on the original whole-brain image for subsequent restoration of the segmentation result to the whole-brain space.

[0163] Input the cropped target region of interest (ROI) block into the pre-trained intracranial artery multi-class segmentation model (using the above-mentioned three-dimensional U-shaped network architecture). This model is trained with a large amount of labeled data and can accurately identify and segment multiple vascular segments of the circle of Willis. The output of the model is the segmentation result of multiple vascular segments of the circle of Willis on the ROI, presented in the form of a multi-channel probability map or a multi-value segmentation map. After obtaining the segmentation result on the ROI, use the previously saved coordinate information of the ROI on the original whole-brain image to accurately restore the segmentation result 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 can accurately reflect the position and morphology of the vascular segments in the whole brain.

[0164] The present invention first performs ROI localization and cropping on the whole-brain image, and then performs segmentation prediction on the cropped ROI. This can significantly reduce the computational amount and processing time, which is particularly important for the processing of large-scale image data. Focusing on the segmentation task of the ROI can enable the model to focus more on the feature learning of the vascular structure of the circle of Willis, thereby improving the accuracy and robustness of the segmentation.

[0165] Next, the intracranial artery multi-class segmentation device based on anatomical and topological perception provided by the present invention will be described. The intracranial artery multi-class segmentation device based on anatomical and topological perception described below can be mutually referred to with the intracranial artery multi-class segmentation method based on anatomical and topological perception described above.

[0166] The intracranial artery multi-class segmentation device based on anatomical and topological perception provided by the present invention, referring to Figure 4 as shown, includes:

[0167] A weight determination module 510, configured to extract a region of interest block containing the circle of Willis from the obtained whole-brain image to obtain an input image and a true annotation image, and determine the weight of each vascular voxel based on the true annotation image to obtain a weight map;

[0168] A first prediction module 520, configured to input the input image into a pre-constructed intracranial artery multi-class segmentation model to obtain a multi-class prediction probability map;

[0169] A loss determination module 530, configured to calculate multiple segmentation losses based on the prediction probability map, the true annotation image, and the weight map, where the multiple segmentation losses include a Dice loss based on radius correction, a vascular break perception loss, an adjacency relationship perception loss, and a multi-class cross-entropy loss;

[0170] A model training module 540, configured to determine an overall segmentation loss according to the multiple segmentation losses, and train the intracranial artery multi-class segmentation model based on the overall segmentation loss to obtain a trained intracranial artery multi-class segmentation model;

[0171] A second prediction module 550, configured 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-class segmentation model for prediction to obtain a segmentation result of multiple vascular segments of the circle of Willis on the target region of interest block.

[0172] Figure 5 FIG. illustrates a schematic physical structure diagram of an electronic device, as Figure 5 shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communication interface 620, and the memory 630 complete communication with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the intracranial artery multi-class segmentation method based on anatomy and topology awareness.

[0173] In addition, when the logic instructions in the above-mentioned memory 630 are implemented in the form of software functional units and sold or used as an independent product, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0174] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that 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 intracranial artery multi-class segmentation method provided by the above-mentioned various methods.

[0175] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method for multi-class segmentation of intracranial arteries based on anatomy and topology perception provided by the above-mentioned various methods.

[0176] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0177] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-class segmentation method for intracranial arteries based on anatomical and topological perception, characterized in that, Including: Extract the region of interest block containing the circle of Willis from the acquired whole-brain image to obtain the input image and the ground truth image, and determine the weight of each vascular voxel based on the ground truth image to obtain the weight map; the weight is determined based on the thickness of the blood vessel; Input the input image into the pre-constructed intracranial artery multi-class segmentation model to obtain a multi-class prediction probability map; Calculate multiple segmentation losses based on the prediction probability map, the ground truth image, and the weight map. The multiple segmentation losses include the Dice loss based on radius correction, the blood vessel breakage perception loss, the adjacency relationship perception loss, and the multi-class cross-entropy loss; among them, the multi-class cross-entropy loss and the Dice loss based on radius correction are calculated based on the multi-class prediction probability map and the ground truth image respectively; perform convolution operations on the multi-class prediction probability map and the ground truth image respectively to obtain the predicted neighborhood fusion result and the ground truth neighborhood result; calculate the blood vessel breakage perception loss based on the obtained predicted neighborhood fusion result and the ground truth neighborhood result; calculate the adjacency relationship perception loss based on the multi-class prediction results in the multi-class prediction probability map and the ground truth image; Determine the overall segmentation loss according to the multiple segmentation losses, and train the intracranial artery multi-class segmentation model based on the overall segmentation loss to obtain the trained intracranial artery multi-class segmentation model; Extract the target region of interest block containing 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-class segmentation model to predict the segmentation result of multiple vascular segments of the circle of Willis on the target region of interest block.

2. The intracranial artery multi-class segmentation method based on anatomy and topology awareness according to claim 1, wherein, The method further includes: Extract the coordinates of the target region of interest block on the target whole-brain image, and restore the segmentation result of multiple vascular segments 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 multiple vascular segments of the circle of Willis on the whole brain.

3. The intracranial artery multi-class segmentation method based on anatomical and topological awareness according to claim 1, wherein The intracranial artery multi-class 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 multiple downsampling modules, and the multiple downsampling modules are connected in sequence and gradually reduce the resolution; The decoder includes multiple upsampling modules, and the multiple upsampling modules are connected in sequence and gradually increase the resolution; The number of upsampling modules is the same as the number of downsampling modules, and the upsampling module is connected to the downsampling module at the corresponding position.

4. The intracranial artery multi-class segmentation method based on anatomical and topological perception according to claim 1, wherein, The Dice loss based on radius correction is determined by the following method: Multiply the weight map with the ground truth image and the prediction probability map respectively to obtain the weighted ground truth image and the weighted prediction probability map; Based on the weighted ground truth image and the weighted prediction probability map, use the Dice coefficient formula to calculate the Dice loss based on radius correction.

5. The intracranial artery multi-class segmentation method based on anatomical and topological awareness according to claim 1, characterized in that The blood vessel breakage perception loss is determined by the following method: For each voxel, determine the class with the maximum probability of the voxel in the prediction probability map to obtain the target class; Create one-hot encodings for the predicted probability map and the ground truth annotation image respectively according to the target category, obtaining a first encoded image and a second encoded image; Use 3D grouped convolution operations to perform local summation on the first encoded image and the second encoded image for each category, obtaining a first convolution map and a second convolution map for each category; Determine the weighted mean square error between the first convolution map and the second convolution map for each category according to the weight map, obtaining the vascular rupture perception loss.

6. The intracranial artery multi-class segmentation method based on anatomical and topological perception according to claim 1, wherein The adjacency relationship perception loss is determined in the following manner: For each voxel, determine the category with the highest probability in the predicted probability map of the voxel, obtaining 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 by the predicted category image to obtain a dilated neighborhood category superposition map; Extract the vascular segment categories existing in the dilated neighborhood category superposition map to obtain an adjacency category list, compare the adjacency category list with a pre-established prior knowledge list, determine the false positive categories and the false negative categories, obtaining a false positive category list and a false negative category list; wherein, the false positive categories are the neighbor categories that appear in the adjacency category list but do not appear in the prior knowledge list, and the false negative categories are the neighbor categories that appear in the prior knowledge list but do not appear in the adjacency category list; Determine a false positive adjacency error voxel set according to the false positive category list, the predicted category image and the ground truth annotation image, and determine a false negative adjacency error voxel set according to the false negative category list, the predicted category image and the ground truth annotation image; Multiply the false positive adjacency error voxel set, the false negative adjacency error voxel set and a pre-determined second encoded image by the predicted probability map and the weight map respectively to obtain a probability weighted map of the false positive adjacency error voxel set, a probability weighted map of the false negative adjacency error voxel set and a probability weighted map of the true positive voxel set; wherein, the second encoded image is the one-hot encoded image of the ground truth annotation image; According to the probability weighted map of the false positive adjacency error voxel set, the probability weighted map of the false negative adjacency error voxel set and the probability weighted map of the true positive voxel set, calculate the adjacency relationship perception loss using the Dice coefficient formula.

7. An intracranial artery multi-class segmentation device based on anatomical and topological awareness, characterized in that, Includes: A weight determination module, configured to extract an interested region block containing the circle of Willis from the acquired whole-brain image to obtain an input image and a ground truth annotation image, and determine the weight of each vascular voxel based on the ground truth annotation image to obtain a weight map; the weight is determined based on the thickness of the blood vessel; A first prediction module, configured to input the input image into a pre-constructed intracranial artery multi-class segmentation model to obtain a multi-class predicted probability map; The loss determination module is used to calculate multiple segmentation losses based on the predicted probability map, the ground truth annotation image, and the weight map. The multiple segmentation losses include the Dice loss based on radius correction, the blood vessel breakage perception loss, the adjacency relationship perception loss, and the multi-class cross-entropy loss. Among them, the multi-class cross-entropy loss and the Dice loss based on radius correction are calculated respectively based on the multi-class predicted probability map and the ground truth annotation image. Convolution operations are respectively performed on the multi-class predicted probability map and the ground truth annotation image to obtain the predicted neighborhood fusion result and the ground truth annotation neighborhood result. The blood vessel breakage perception loss is calculated based on the obtained predicted neighborhood fusion result and the ground truth annotation neighborhood result. The adjacency relationship perception loss is calculated based on the multi-class prediction results in the multi-class predicted probability map and the ground truth annotation image. The model training module is used to determine the overall segmentation loss according to the multiple segmentation losses, and train the intracranial artery multi-class segmentation model based on the overall segmentation loss to obtain the trained intracranial artery multi-class segmentation model. The second prediction module is used to extract the target region of interest block containing the circle of Willis from the obtained target whole brain image, and input the target region of interest block into the trained intracranial artery multi-class segmentation model to predict the segmentation result of the multi-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 on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the intracranial artery multi-class segmentation method based on anatomical and topological perception according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the intracranial artery multi-class segmentation method based on anatomical and topological perception according to 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 the processor, it implements the intracranial artery multi-class segmentation method based on anatomical and topological perception according to any one of claims 1 to 6.

Images