Intracranial aneurysm detection and segmentation method, device and equipment and storage medium

Abstract

The application relates to an intracranial aneurysm detection and segmentation method, device, equipment and storage medium. The method comprises the following steps: performing forward propagation on an intracranial blood vessel image to be processed based on a trained aneurysm detection model, mapping a detection probability image obtained based on a probability threshold, and obtaining center point coordinates and three-dimensional dimensions of an aneurysm in a target connected domain; performing block cutting based on a segmentation block cutting strategy according to the center point coordinates and the three-dimensional dimensions of the aneurysm, and obtaining an aneurysm first block cutting image; performing forward propagation and aneurysm classification identification on each aneurysm first block cutting image based on an aneurysm segmentation model which is trained and has a classification branch network inserted, and obtaining a classification probability and a segmentation probability image; and integrating the segmentation probability image and the classification probability of each aneurysm first block cutting image, and obtaining an aneurysm segmentation result. The application can filter false positive interference introduced in the detection stage without loss of recall rate, and realize aneurysm detection and segmentation with high detection rate and high segmentation accuracy.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to methods, apparatus, devices and storage media for the detection and segmentation of intracranial aneurysms. Background Technology

[0002] Intracranial aneurysms are high-risk vascular lesions formed by abnormal local dilation of intracranial arteries. Rupture can lead to fatal subarachnoid hemorrhage, seriously threatening the patient's life. Current clinical diagnosis mainly relies on multimodal vascular imaging such as digital subtraction angiography (DSA), CT angiography (CTA), and magnetic resonance angiography (MRA), providing crucial data for aneurysm detection and segmentation.

[0003] In existing technologies, aneurysm detection methods are mainly divided into traditional methods and deep learning methods. Traditional methods rely on manually designed features for identification, such as morphological features, gradient features, or geometric features. Deep learning methods automatically extract image features through convolutional neural networks. However, in the case of a single task of aneurysm detection, to ensure a high recall rate for aneurysms, the threshold setting must be low. A low threshold will include some non-aneurysm predictions, ultimately leading to false positives for aneurysms. Secondly, aneurysm segmentation algorithms based on vascular masks include traditional image segmentation algorithms and deep learning segmentation algorithms. Among them, deep learning methods mostly improve segmentation accuracy by modifying the network structure, such as adding multiple feature enhancement fusions and attention mechanisms. Although the segmentation accuracy is improved, it is constrained by the size of the aneurysm, making it difficult to achieve good segmentation results for both very large and very small aneurysms simultaneously.

[0004] It is evident that current technologies for the detection and segmentation of aneurysms cannot simultaneously achieve both high recall and low false positive rates. Summary of the Invention

[0005] This application provides a method, device, equipment, and storage medium for the detection and segmentation of intracranial aneurysms, in order to solve the technical problem in the prior art that the detection and segmentation of aneurysms cannot simultaneously achieve high recall and low false positives.

[0006] According to one aspect of the embodiments of this application, this application provides a method for detecting and segmenting intracranial aneurysms. The method includes: performing forward propagation on an intracranial vascular image to be processed based on a trained aneurysm detection model; mapping the obtained detection probability image based on a probability threshold to obtain the coordinates of the center point and the three-dimensional dimensions of the aneurysm in at least one target connected region; performing segmentation based on the coordinates of the center point and the three-dimensional dimensions of the aneurysm in each target connected region, and obtaining a first segmented image of the aneurysm corresponding to the target connected region; performing forward propagation and aneurysm classification recognition on each of the first segmented images of the aneurysm based on a trained aneurysm segmentation model with inserted classification branch networks, and obtaining a segmentation probability image corresponding to the classification probability and the first segmented image of the aneurysm; and integrating the segmentation probability image and the classification probability of each first segmented image of the aneurysm to obtain an aneurysm segmentation result.

[0007] Optionally, before performing forward propagation on the intracranial vascular image to be processed based on the trained aneurysm detection model, mapping the obtained detection probability image based on a probability threshold, and obtaining the center point coordinates and three-dimensional dimensions of the aneurysm in at least one target connected region, the method further includes: performing data preprocessing on the aneurysm sample data to obtain multiple aneurysm second block images, the data preprocessing including aneurysm size adjustment and block segmentation according to the detection block segmentation strategy, including segmenting positive sub-image blocks containing aneurysms and negative sub-image blocks not containing aneurysms; iteratively optimizing the aneurysm detection model, performing forward propagation based on each of the aneurysm second block images to obtain the detection probability image, and performing model convergence based on the detection loss function using gradient descent to obtain the trained aneurysm detection model.

[0008] Optionally, the process of forward propagating the trained aneurysm detection model to the intracranial vascular image to be processed, mapping the obtained detection probability image based on a probability threshold, and obtaining the center point coordinates and three-dimensional dimensions of the aneurysm in at least one target connected region includes: cropping the intracranial vascular image to be processed into multiple sub-images with overlapping regions, and inputting the multiple sub-images to be processed into the trained aneurysm detection model for forward propagation to obtain the detection probability sub-images corresponding to each sub-image; restoring the pre-cropping position of each detection probability sub-image to obtain the corresponding detection probability image; performing binary mapping on the detection probability images based on the probability threshold, and selecting at least one target connected region based on the mapped binary image; numbering the selected target connected regions, and calculating the center point coordinates and three-dimensional dimensions of the aneurysm in each target connected region.

[0009] Optionally, before performing forward propagation and aneurysm classification recognition on each of the aneurysm first block images based on the trained aneurysm segmentation model with inserted classification branch network, and obtaining the classification probability and the segmentation probability image corresponding to each of the aneurysm first block images, the method further includes: preprocessing the aneurysm sample data to obtain multiple aneurysm first block images; the preprocessing includes segmenting blocks according to the center point coordinates and three-dimensional dimensions of the aneurysm in the sample data using the segmentation block strategy, including segmenting blocks for positive sub-images containing aneurysms and negative sub-images not containing aneurysms; adding the classification branch network predicting whether an aneurysm exists in the aneurysm first block image to the highest-dimensional feature in the aneurysm segmentation model; iteratively optimizing the aneurysm segmentation model, performing forward propagation and aneurysm classification recognition based on each of the aneurysm first block images, obtaining the segmentation probability image and the classification probability, and performing model convergence based on the segmentation loss function and the classification loss function using the gradient descent method to obtain the trained aneurysm detection model.

[0010] Optionally, the step of performing block segmentation based on the center point coordinates and three-dimensional dimensions of the aneurysm in each of the target connected regions, and obtaining a first segmented aneurysm image corresponding to the target connected region, includes: calculating the size of the corresponding first segmented aneurysm image based on the size of the aneurysm in each dimension of the three-dimensional space according to the segmentation block segmentation strategy; and extracting the first segmented aneurysm image corresponding to each of the target connected regions based on the size of the first segmented aneurysm image, with the center point coordinates of the aneurysm as the center.

[0011] Optionally, the aneurysm segmentation model based on the trained model with inserted classification branch network performs forward propagation and aneurysm classification recognition on each of the first segmented images of the aneurysm to obtain the classification probability and the segmentation probability image corresponding to each of the first segmented images of the aneurysm, including: inputting each of the first segmented images of the aneurysm into the trained aneurysm segmentation model for forward propagation and aneurysm classification recognition; performing forward propagation based on the segmentation branch to obtain the segmentation probability image corresponding to each of the first segmented images of the aneurysm; and performing aneurysm recognition on each of the first segmented images of the aneurysm based on the classification branch network to obtain the classification probability of the aneurysm.

[0012] Optionally, the step of integrating the segmentation probability image and the classification probability of each aneurysm first block image to obtain the aneurysm segmentation result includes: calculating the product of the segmentation probability image and the classification probability of each aneurysm first block image; performing binary map mapping based on the product of the segmentation probability image and the classification probability to obtain the segmentation binary map corresponding to each aneurysm first block image, and performing spatial restoration on the segmentation binary map to obtain the binary map restoration result corresponding to each aneurysm first block image; and performing binary map integration and connected component analysis and numbering on the binary map restoration results of all aneurysm first block images to obtain the aneurysm segmentation result.

[0013] According to another aspect of the embodiments of this application, this application provides an apparatus for detecting and segmenting intracranial aneurysms. The apparatus includes: a detection module, used to perform forward propagation on an intracranial vascular image to be processed based on a trained aneurysm detection model, map the obtained detection probability image based on a probability threshold, and obtain the coordinates of the center point and the three-dimensional dimensions of the aneurysm in at least one target connected region; a block segmentation module, used to perform block segmentation based on the coordinates of the center point and the three-dimensional dimensions of the aneurysm in each target connected region according to a segmentation block segmentation strategy, to obtain a first block image of the aneurysm corresponding to the target connected region; a segmentation module, used to perform forward propagation and aneurysm classification recognition on each of the first block images of the aneurysm based on a trained aneurysm segmentation model with inserted classification branch networks, to obtain the classification probability and the segmentation probability image corresponding to each of the first block images of the aneurysm; and a result calculation module, used to integrate the segmentation probability image and the classification probability of each of the first block images of the aneurysm to obtain the aneurysm segmentation result.

[0014] According to another aspect of the embodiments of this application, this application provides an electronic device, including: a processor, a memory, and a network interface. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory through the network interface, and the processor executes the machine-readable instructions to perform the steps of the method for detecting and segmenting intracranial aneurysms.

[0015] According to another aspect of the embodiments of this application, this application provides a storage medium having processor-executable non-volatile program code, the program code causing the processor to perform the steps of the intracranial aneurysm detection and segmentation method.

[0016] Compared with related technologies, the technical solutions provided in this application have the following advantages: The intracranial aneurysm detection and segmentation method provided in this application, through a trained aneurysm detection model, can ensure high recall by employing a probability threshold strategy in the aneurysm detection stage. It simultaneously extracts the center point coordinates and three-dimensional dimensions of the aneurysm on each target connected region, providing a reliable localization basis for subsequent accurate block segmentation. This avoids missing small aneurysms and provides the aneurysm segmentation model with input images of a more suitable scale. Furthermore, based on an aneurysm segmentation model embedded with a classification branch network, it simultaneously completes segmentation and category determination for the first block image of the aneurysm corresponding to each target connected region. Finally, it fuses the two types of output results to complete the output of all aneurysms. Without losing the initial high recall rate throughout the process, it significantly filters out false positive interference introduced in the detection stage, ultimately achieving fully automatic intracranial aneurysm detection and segmentation that balances high detection rate and high segmentation accuracy. This effectively solves the problem that it is difficult to simultaneously achieve high recall and low false positives in traditional solutions. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0018] Figure 1 This is a flowchart of an optional method for detecting and segmenting intracranial aneurysms according to an embodiment of this application; Figure 2 This is an optional aneurysm size adjustment flowchart provided according to an embodiment of this application; Figure 3 This is a flowchart of an optional aneurysm detection model training process provided according to an embodiment of this application; Figure 4 This is a flowchart illustrating an optional aneurysm segmentation model training process according to an embodiment of this application. Figure 5 This is a flowchart of another optional method for detecting and segmenting intracranial aneurysms according to an embodiment of this application; Figure 6 This is a structural diagram of an optional intracranial aneurysm detection and segmentation device according to an embodiment of this application; Figure 7 This is a schematic diagram of an optional electronic device structure provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To address the problems mentioned in the background art, according to one aspect of the embodiments of this application, an embodiment of a method for detecting and segmenting intracranial aneurysms is provided.

[0021] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for detecting and segmenting intracranial aneurysms according to an embodiment of the present invention. Taking the method for detecting and segmenting intracranial aneurysms executed by a server as an example, the method includes the following steps: Step S102: Based on the trained aneurysm detection model, forward propagation is performed on the intracranial vascular image to be processed. The obtained detection probability image is mapped based on the probability threshold to obtain the center point coordinates and three-dimensional dimensions of the aneurysm in at least one target connected region.

[0022] In this embodiment, the intracranial aneurysm detection and segmentation method provided by the present invention is applicable to data scenarios of different modalities such as CT angiography, magnetic resonance imaging (MRA), and digital subtraction angiography (DSA).

[0023] The intracranial vascular images to be processed serve as input data for the aneurysm detection model's inference. Before inference, the aneurysm detection model is pre-trained. This can be done using pre-labeled aneurysm sample data, trained through a deep learning network model with an encoder-decoder structure, including the Unet model, such as ResUnet or an expanded version of Unet. Through forward propagation of the images, the encoder extracts vascular and aneurysm features from the intracranial vascular images to be processed, and the decoder restores the spatial resolution based on the extracted features. After completing the forward computation of all network layers, the detection probability image of the output layer is obtained through an interface.

[0024] To ensure high recall of aneurysms, the probability threshold is set to a low value, for example, between 0.1 and 0.3. The probability threshold can be preset manually, for example, fixed at 0.2. Alternatively, it can be automatically set based on historical probability thresholds, for example, by taking the average of historical probability thresholds. The detection probability image is an image formed based on probability representation; therefore, the detection probability image can be compared with the probability threshold to map it into a binary image. Analysis based on the binary image allows for the acquisition of at least one target connected region, which represents the area where the aneurysm is located. Then, the coordinates of the aneurysm's center point and its three-dimensional dimensions are calculated for each target connected region. The three-dimensional dimensions refer to the size of the aneurysm in the X, Y, and Z dimensions, i.e., the lengths of the three axes of the aneurysm's minimum bounding box, and the center point coordinates are the center point of the minimum bounding box.

[0025] Step S104: Based on the center point coordinates and three-dimensional dimensions of the aneurysm in each target connected region, perform segmentation and block segmentation based on the segmentation and block segmentation strategy to obtain the first segmented block image of the aneurysm corresponding to the target connected region.

[0026] The segmentation block selection strategy is a pre-defined strategy for segmenting aneurysms during the segmentation process. To improve the performance of the aneurysm detection model, a segmentation block selection strategy can be preset. Before segmenting the aneurysm, the coordinates of the center point and the three-dimensional dimensions of the aneurysm in the target connected region are calculated. Blocks are then selected according to the segmentation block selection strategy, resulting in the first segmented image of the aneurysm corresponding to aneurysms of different sizes in each target connected region. Since different aneurysms vary greatly in size, adaptive block selection using size information ensures that aneurysms of different sizes have a similar proportion in the first segmented image, thereby improving segmentation accuracy.

[0027] Step S106: Based on the aneurysm segmentation model trained and with inserted classification branch network, forward propagation and aneurysm classification recognition are performed on each of the first segmented block images of the aneurysm to obtain the classification probability and the segmentation probability image corresponding to each of the first segmented block images of the aneurysm.

[0028] Before aneurysm segmentation, an aneurysm segmentation model is pre-trained. Similarly, the network core of the aneurysm segmentation model can also be an encoder-decoder structure, such as the UNet structure. To further improve the segmentation accuracy of aneurysms and reduce the impact of false positives in aneurysm detection models aimed at ensuring high recall, this embodiment adds a classification branch network to the highest-dimensional features of the aneurysm segmentation model. For example, the fully connected layer is mapped to one dimension, and then a sigmoid activation function is used to generate probability values. This classification branch network can be used to predict whether an aneurysm exists in the first segmented image.

[0029] Furthermore, based on the aneurysm segmentation model trained and with inserted classification branch network, forward propagation is performed on the first block image of each aneurysm. Based on the segmentation branch in the aneurysm segmentation model, the aneurysm segmentation region is identified to obtain the segmentation probability image corresponding to the first block image of each aneurysm. Based on the classification branch in the aneurysm segmentation model, the aneurysm binary output is performed to obtain the classification probability corresponding to the first block image of each aneurysm.

[0030] Step S108: Integrate the segmentation probability image and the classification probability of each aneurysm first block image to obtain the aneurysm segmentation result.

[0031] Specifically, the segmentation probability image and classification probability of a single aneurysm first block image are fused and calculated, and mapped to a binary image. The binary images corresponding to all aneurysm first block images are then integrated to obtain a complete binary image. Connectivity analysis is performed on the complete binary image to finally output the aneurysm segmentation result.

[0032] In this embodiment of the invention, the trained aneurysm detection model can ensure high recall by employing a probability threshold strategy in the aneurysm detection stage. It simultaneously extracts the center point coordinates and three-dimensional dimensions of the aneurysm on each target connected domain, providing a reliable localization basis for subsequent accurate block segmentation. This avoids missing small aneurysms and provides the aneurysm segmentation model with input images of a more suitable scale. Furthermore, based on the aneurysm segmentation model with embedded classification branch network, the first block image of the aneurysm corresponding to each target connected domain is segmented and classified simultaneously. Finally, the two types of output results are fused to complete the full aneurysm output. Without losing the initial high recall rate, the false positive interference introduced in the detection stage is greatly filtered out. Ultimately, a fully automatic intracranial aneurysm detection and segmentation that balances high detection rate and high segmentation accuracy is achieved, effectively solving the problem that it is difficult to simultaneously achieve high recall and low false positives in traditional solutions.

[0033] In some alternative embodiments, prior to step S102 described above, the method further includes: S11, perform data preprocessing on the aneurysm sample data to obtain multiple second aneurysm block images. The data preprocessing includes aneurysm size adjustment and block segmentation according to the detection block segmentation strategy, including segmenting positive sub-image blocks containing aneurysms and negative sub-images that do not contain aneurysms. S12, the aneurysm detection model is iteratively optimized by forward propagation based on the second block image of each aneurysm to obtain the detection probability image, and the model is converged by gradient descent based on the detection loss function to obtain the trained aneurysm detection model.

[0034] Before inference based on the aneurysm detection model, the process includes data preprocessing and aneurysm detection model training. The aneurysm sample data used for training is too large. To save resources and improve training speed, actual training is performed using image patching. That is, in each iteration of training, image patches are randomly extracted from the intracranial vascular image, including both positive and negative image patches, with one type extracted per iteration.

[0035] In some possible embodiments, during the data preprocessing of aneurysm sample data, random block segmentation can be performed according to a detection segmentation strategy. This strategy includes: to ensure a balance between positive and negative samples, proportionally segmenting positive sub-image blocks containing aneurysms and negative sub-image blocks not containing aneurysms as the second segmented image for aneurysms. For example, the ratio of positive to negative image blocks is approximately 1.2. The negative sub-image blocks are then directly input into the aneurysm detection model for training, while the positive sub-image blocks undergo data augmentation. Specifically, when segmenting positive sub-image blocks, a center point is randomly selected within the aneurysm-marked region of the labeled aneurysm sample data, and then a positive sub-image block of the target size is segmented. When segmenting negative sub-image blocks, a center point is randomly selected outside the aneurysm-marked region, and then a negative sub-image block of the target size is segmented, ensuring that the segmented negative sub-image block of the target size does not contain any pixels of the aneurysm.

[0036] In other possible embodiments, the detection block segmentation strategy further includes: to ensure the detection performance of image edge regions, sampling is performed proportionally in the image edge regions. When the random probability is less than 0.2, the block segmentation will directly sample the edges of the intracranial blood vessel images in the sample data. In this case, the resulting second block image of the aneurysm can still be divided into positive and negative sub-image blocks. If the random probability is greater than 0.2, a block segmentation strategy that balances positive and negative samples is adopted. The random probability can be adaptively adjusted.

[0037] Combination Figure 2 As shown, Figure 2 This embodiment provides an optional aneurysm size adjustment flowchart. Because the number of aneurysms of different sizes in the training data is uneven, the size of aneurysms in positive sub-image patches needs to be adjusted in real time during training. This includes proportionally performing morphological operations such as dilation, erosion, and smoothing on existing aneurysms to dynamically adjust their size. The dilation coefficient is randomly generated, referring to the number of times the aneurysm is dilated or eroded; specifically, it is randomly generated within a certain range based on the current aneurysm voxel value. By adjusting the aneurysm size, the diversity of aneurysm morphology can be improved, thereby enhancing the performance of the aneurysm detection model.

[0038] In other possible embodiments, data preprocessing also includes: aneurysm sample data augmentation and data normalization. Data augmentation involves expanding the data by translating, rotating, or flipping the samples, making the sample set richer and more diverse. To improve the performance of the detection model, data of different dimensions and ranges need to be converted to a unified standard, i.e., data normalization is performed. The specific calculation formula is shown in equation (1) below: (1) Among them, I Let be the pixel value of the aneurysm image at coordinates (x, y) in the sample data. Let I(x,y) be the pixel value mapped from the pixel at coordinate (x,y), and Min(I(x,y)) and Max(I(x,y)) be the minimum and maximum pixel values ​​in the aneurysm image in the sample data, respectively.

[0039] Furthermore, combined Figure 3 As shown, after data preprocessing, the aneurysm detection model is iteratively optimized. Forward propagation is performed based on the preprocessed second-block images of each aneurysm, and model convergence is achieved using gradient descent to obtain the trained aneurysm detection model. The model output is a detection probability image. During forward propagation, the second-block images of the aneurysm are used as input data, sequentially passing through the convolution and pooling layers of the encoder to extract multi-scale vascular and aneurysm features. Then, through the upsampling and skip connection layers of the decoder, the spatial resolution of the feature map is gradually restored. After completing the forward computation of all network layers, the prediction results of the output layer are obtained through an interface. During training, if the second-block image of the aneurysm contains an aneurysm, the aneurysm segmentation region is identified; if the second-block image of the aneurysm does not contain an aneurysm, a zero-image is obtained. The model training uses gradient descent until the aneurysm detection model converges. The Dice loss function is used as the detection loss function, and loss calculation is performed based on manually labeled aneurysm masks, and the detection model is trained and optimized based on the loss. The specific calculation formula for the detection loss function is as follows: (2) in, This is a detection probability image output after training the aneurysm detection model. is the aneurysm mask or all-zero image, and i is the sequence number of the second block image of the aneurysm.

[0040] In this embodiment, adjusting the aneurysm size enhances the diversity of aneurysm morphology, thereby improving the performance of the aneurysm detection model. By using a detection-based block-segmentation strategy to segment aneurysms of different sizes, the proportion of aneurysms in the overall image input to the model is kept similar for all sizes. Data augmentation ensures data diversity. Transforming data of different dimensions and ranges to a unified standard improves the segmentation model's performance. Through block segmentation, the model is trained using forward propagation based on the second segmented aneurysm image, coupled with gradient descent driven by the Dice loss function for iterative training until convergence. This approach better suits the class-imbalanced characteristics of small aneurysms in medical imaging, effectively avoiding the bias towards background inherent in traditional cross-entropy loss, and significantly improving the segmentation accuracy of small aneurysms, thus better meeting the high-precision requirements for intracranial aneurysm segmentation in clinical practice.

[0041] In some optional embodiments, step S102 above includes: S1021, the intracranial vascular image to be processed is cropped into multiple sub-images to be processed with overlapping areas, and the multiple sub-images to be processed are input into the trained aneurysm detection model for forward propagation to obtain the detection probability sub-images corresponding to each sub-image to be processed. S1022, restore the pre-cropping position of each of the detection probability sub-images to obtain the corresponding detection probability image; S1023, perform binary mapping on the detection probability image based on the probability threshold, and select at least one target connected component based on the mapped binary image; S1024, Number the selected target connected regions, and calculate the coordinates of the center point and the three-dimensional dimensions of the aneurysm in each target connected region.

[0042] In this process, the intracranial vascular image to be processed is pre-normalized and then cropped into several sub-images with overlapping regions according to the required training size. These sub-images are then input into the aneurysm detection model for forward propagation inference, resulting in detection probability sub-images for each sub-image. The higher the value of the detection probability sub-image, the greater the probability that the corresponding sub-image is an aneurysm region. Cropping the image with overlapping regions allows for more natural transitions between image blocks. The detection probability sub-images are then stitched together to restore the original positions before cropping, and the average value of the overlapping regions is taken to obtain the corresponding detection probability images.

[0043] In some possible embodiments, given that the original image size, cropping size, and overlapping area size are known, the Z, Y, and X axes are cropped sequentially. If the cropping is not divisible, the Z, Y, and X axes can be padded with 0 values.

[0044] Furthermore, a probability threshold is preset; to ensure a higher recall rate, a relatively low probability threshold is set, for example, 0.2. The stitched detection probability image is then... Mapped to the corresponding binary image The specific calculation formula is as follows: (3) Furthermore, after performing binary graph mapping, based on the binary graph... =1, which yields at least one connected component, representing the location of the aneurysm. Further, based on the connected component selection criteria, at least one target connected component can be selected and numbered. The coordinates of the aneurysm's center point and its dimensions (three-dimensional dimensions) in the X, Y, and Z dimensions are then calculated for each target connected component.

[0045] In some examples, the selection of target connected components based on the above-mentioned connected component selection criteria includes: performing connected component analysis on the binary graph and retaining connected components with more than k pixels, for example, retaining connected components with more than 10 pixels. Further, the k target connected components are numbered, for example, 1-1, 1-2, 1-3, ..., 1-k.

[0046] In this embodiment, cropping the intracranial vascular image to be processed into sub-images with overlapping regions allows for more natural transitions between image blocks. Forward propagation of each sub-image yields a detection probability sub-image for each. Further position restoration and binary mapping of these probability sub-images enable classification, allowing for rapid identification of aneurysm regions based on the mapped binary images. Selecting and numbering target connected components filters out scattered noise and minor artifacts, preventing invalid small regions from consuming computational resources. A unique identifier is assigned to each aneurysm for accurate differentiation, facilitating subsequent ordered segmentation. Calculating the center point coordinates and 3D dimensions of the aneurysm corresponding to each target connected component provides structured basic data for subsequent spatial localization and morphological quantification analysis of aneurysms, significantly reducing the probability of missed or false detections and improving the overall accuracy and computational efficiency of aneurysm segmentation and measurement in 3D medical imaging.

[0047] In some optional embodiments, prior to step S106 described above, the following is included: S21, perform data preprocessing on the aneurysm sample data to obtain multiple aneurysm first block images. The preprocessing includes block segmentation based on the center point coordinates and three-dimensional dimensions of the aneurysm in the sample data using the segmentation block segmentation strategy. This includes segmenting positive sub-images containing aneurysms and negative sub-images not containing aneurysms. S22, add the classification branch network that predicts whether an aneurysm exists in the first block image of the aneurysm to the highest dimension feature of the aneurysm segmentation model; S23, the aneurysm segmentation model is iteratively optimized by performing forward propagation and aneurysm classification based on the first block image of each aneurysm, obtaining the segmentation probability image and the classification probability, and performing model convergence based on the segmentation loss function and the classification loss function using the gradient descent method to obtain the trained aneurysm detection model.

[0048] Combination Figure 4 As shown, the sample data for training the aneurysm segmentation model is also intracranial vascular images. The data preprocessing includes segmenting the aneurysm into blocks based on the center point coordinates and three-dimensional dimensions of the aneurysm in the sample data. The aneurysm mask is manually labeled and used to calculate the loss of the segmentation probability image during the segmentation model training. The model is then optimized based on the loss.

[0049] In some possible embodiments, the segmentation strategy includes: size-adaptive segmentation: First, based on the center point coordinates and the aneurysm's dimensions in the X, Y, and Z dimensions, the starting and ending coordinates of the segmentation block are calculated, and the segmentation block size is estimated based on these two coordinates. Then, a sub-image of the estimated segmentation block size is extracted, centered on the center point coordinates. If the segmentation block size of the sub-image is inconsistent with the target size, the size of the sub-image can be adjusted to the target size through interpolation. The segmented sub-image in the size-adaptive segmentation contains the positive sub-image of the aneurysm.

[0050] In some other possible embodiments, since the aneurysm detection model will include a certain number of false positive aneurysm detection points in its forward propagation output to ensure high recall, in order to improve the aneurysm segmentation performance, the segmentation block-cutting strategy in the aneurysm segmentation model training also includes: randomly cutting out negative sub-images that do not contain aneurysms at a preset ratio.

[0051] In other possible embodiments, data preprocessing also includes: aneurysm sample data augmentation and data normalization. Data augmentation involves expanding the data by translating, rotating, or flipping the samples, making the sample set richer and more diverse. To improve the performance of the detection model, data of different dimensions and ranges need to be converted to a unified standard, i.e., data normalization is performed. The calculation formula for data normalization is as shown in equation (1) above.

[0052] In some examples, the aneurysm segmentation model employs a 3D U-Net architecture, which includes a 4-layer downsampling encoding module and a 4-layer upsampling decoding module. Each convolutional kernel in the encoding module extracts features, progressively increasing the feature dimension of the input aneurysm first patch image from an initial 64 channels to a maximum of 512 channels. Following the 512-channel feature layer, a binary classification branch network is connected in parallel. This classification branch network can be composed of two global average pooling layers, two fully connected layers, and one sigmoid activation layer connected in series. The input of this classification branch network is directly connected to the output of the highest-dimensional feature layer of the aneurysm segmentation model, outputting the classification probability of whether an aneurysm exists in the first patch image. The classification branch network shares the highest-dimensional features with the decoding branch of the aneurysm segmentation model, eliminating the need for numerous additional convolutional operations and achieving multi-task joint training of segmentation and classification tasks with almost no increase in the overall model parameter count.

[0053] Furthermore, the training parameters for model iterative optimization are set. The Adam gradient descent optimizer is used during model training, and the total loss function of the model is obtained from two parts: segmentation loss and classification loss. Preferably, the Dice loss function with added boundary constraints is used as the segmentation loss function, and the binary cross-entropy loss function is used as the classification loss function. The total loss function is a weighted sum of the two, which can simultaneously consider the pixel-level accuracy of the segmented region and the global discriminative ability of the classification task. The weights are not uniquely defined. At the start of training, the first block images of each aneurysm are batch-input into the aneurysm segmentation model for forward propagation. For the classification network, if the first block image of the aneurysm contains the aneurysm, the segmented region of the aneurysm is obtained; otherwise, an image with all zeros is obtained. For the classification branch network, if the first block image of the aneurysm contains the aneurysm, a 1 is obtained; otherwise, a 0 is obtained. Finally, the segmentation network outputs a segmentation probability image with the same size as the input patch image, while the classification branch network simultaneously outputs the classification probability of the presence of an aneurysm in the first patch image. Then, based on the segmentation probability image obtained from forward propagation, the classification probability, and the corresponding manually labeled ground truth, the total loss function value for the current batch is calculated. Following the gradient backpropagation rule, the loss gradient is propagated layer by layer back to all parameter layers of the aneurysm segmentation model and the classification branch network, updating the convolutional weight parameters of each layer using gradient descent. Further model convergence is then determined. Pre-defined convergence conditions can be used, such as when the total loss value of the validation set no longer decreases and the fluctuation range is less than 1e-5 for 15 consecutive epochs, and simultaneously the Dice coefficient for aneurysm segmentation on the validation set is consistently higher than 0.92 and the classification accuracy is consistently higher than 0.97, the model is considered to have converged, the iterative training process is terminated, all weight parameters of the current model are saved, and the trained aneurysm detection model is obtained.

[0054] The segmentation loss function for the segmentation branch is as follows: (4) in, This is a segmentation probability image obtained through segmentation branch inference calculations using an aneurysm segmentation model. The image is either a manually annotated aneurysm mask or an all-zero image.

[0055] The classification loss function of the classification branch network is as follows: (5) in, y represents the classification probability calculated through the classification branch inference of the aneurysm segmentation model, where y is the classification label.

[0056] The total loss function is shown in the following formula: (6) In this embodiment, by adding a classification branch to the highest-dimensional feature layer of the aneurysm segmentation model, a multi-task joint training mode can be implemented to simultaneously complete pixel-level accurate segmentation and classification of each aneurysm first block image. Based on the segmentation probability image, the voxel-level contour information of the aneurysm is output. At the same time, the authenticity of the aneurysm in the first block image is verified from the global semantic dimension through the classification probability. While fully retaining the original high recall rate, the classification branch filters out a large number of false positive interference areas such as vascular bifurcation, arterial infundibulum, and calcification artifacts that are difficult to distinguish by the segmentation task alone. Finally, without reducing the aneurysm recall rate, the false positive rate of the aneurysm detection results is significantly reduced, and the segmentation accuracy of the aneurysm is significantly improved.

[0057] In some optional embodiments, step S104 above includes: S1041, Based on the segmentation and slicing strategy, calculate the size of the corresponding first slicing block image of the aneurysm according to the size of the aneurysm in each dimension of the three-dimensional space in each of the target connected domains; S1042, using the coordinates of the center point of the aneurysm as the center, extract the first block image of the aneurysm corresponding to each of the target connected regions based on the size of the first block image of the aneurysm.

[0058] In some examples, before performing segmentation inference on the intracranial vascular image to be processed, aneurysm segmentation is performed on each target connected component of the intracranial vascular image to be processed according to the segmentation segmentation strategy. Specifically, all target connected components are traversed, and the calculated pixel dimensions W, H, and D of each aneurysm in the X, Y, and Z dimensions are read. Optionally, a safety expansion coefficient of 1.5 can be pre-configured, and the dimensions W, H, and D in each dimension are multiplied by the safety expansion coefficient and rounded up to an integer multiple of 8 to obtain the three-dimensional target dimensions W', H', and D' of the first segmented image, which adapts to the downsampling alignment requirements of the subsequent U-Net network. Based on the three-dimensional center point coordinates of the aneurysm obtained from the connected component analysis, the calculated segmented size is combined with bidirectional expansion in three dimensions to verify that the segmented boundary does not exceed the spatial range of the original three-dimensional image. If it exceeds the boundary, the boundary is anchored to the edge position of the original image or filled with 0 values ​​to avoid invalid empty pixels. Furthermore, using the center coordinates of the verified aneurysm as the center, three-dimensional voxel regions of corresponding W', H', and D' sizes are extracted along the X, Y, and Z dimensions respectively to generate the first segmented image of each aneurysm. After completion, all the first segmented images of aneurysms are indexed and stored according to the aneurysm number, and after data normalization, they are input into the subsequent aneurysm segmentation model to complete the inference.

[0059] In this embodiment, the three-dimensional segmentation of each target connected region is performed using the coordinates of the center point of the aneurysm as the anchor point. This can generate a standardized independent first segmented image of the aneurysm without destroying the spatial morphology of the lesion. This not only significantly reduces the computational load of the full large volume image, but also provides a high signal-to-noise ratio input for lesion focusing for subsequent model inference, effectively reduces the interference of background redundant information, and significantly improves the segmentation accuracy and computational efficiency of small aneurysms.

[0060] In some optional embodiments, step S106 above includes: S1061, Input the first block image of each aneurysm into the trained aneurysm segmentation model for forward propagation and aneurysm classification and recognition; S1062, perform forward propagation based on segmentation branches to obtain the segmentation probability image corresponding to each of the first block images of the aneurysm; S1063, perform aneurysm identification on each of the first block images of the aneurysm based on the classification branch network, and obtain the classification probability of the aneurysm.

[0061] In this embodiment, the preprocessed first block images of each aneurysm are input into the trained aneurysm segmentation model for forward propagation inference, resulting in an aneurysm segmentation probability image corresponding to each first block image of the aneurysm. And the classification probability of aneurysms obtained by classifying and recognizing all aneurysm first block images based on a classification branch network. .

[0062] In some examples, during inference, the encoder uses a 4-layer 3D residual convolutional module to perform feature extraction and downsampling layer by layer, outputting multi-scale deep semantic feature maps after multi-layer downsampling. During forward propagation based on the segmentation branch to generate the segmentation probability image, the segmentation branch can employ a 4-layer 3D deconvolutional decoder symmetrical to the encoder, combined with skip connections to fuse shallow detail features. Finally, a single-channel segmentation probability image is output via a sigmoid activation function, with dimensions identical to the input block. Each voxel in the image ranges from 0 to 1, with higher values ​​indicating a greater probability that the voxel belongs to an aneurysm. The classification branch can directly reuse the 128-channel deep semantic feature map output by the encoder, eliminating the need for separate feature extraction and significantly reducing the number of model parameters. The classification branch can perform 3D global average pooling on the feature map using a global pooling layer, perform dimensional mapping through a fully connected layer, and output a binary classification probability distribution via a Softmax activation function. Here, the 0th dimension represents the probability of a false positive ("no aneurysm"), and the 1st dimension represents the probability of "aneurysm present."

[0063] It is important to understand that the aforementioned first and second aneurysm block images are used to distinguish blocks during the training and inference processes of different models. For the same model, although the objects of block segmentation are different during the training and inference processes, they are still defined as the block segmentation results of data preprocessing under that model. For example, in the aneurysm detection model, the first aneurysm block image is used as the block segmentation result of sample data during training and as the block segmentation result of intracranial blood vessel image to be processed during inference.

[0064] In this embodiment, by adding a classification branch to the highest-dimensional feature layer of the aneurysm segmentation model for task inference, pixel-level accurate segmentation and classification can be achieved simultaneously for each aneurysm first block image. Not only is the voxel-level contour information of the aneurysm output based on the segmentation probability image, but the authenticity of the aneurysm in the first block image is also verified by the classification probability. While fully preserving the original high recall rate, the classification branch is used to distinguish false positive interference areas. Finally, without reducing the aneurysm recall rate, the false positive rate of the aneurysm detection results is greatly reduced, and the segmentation accuracy of the aneurysm is significantly improved.

[0065] In some optional embodiments, step S108 above includes: S1081, Calculate the product of the segmentation probability image and the classification probability for each of the first block images of the aneurysm; S1082, perform binary image mapping based on the product of the segmentation probability image and the classification probability to obtain the segmentation binary image corresponding to each of the aneurysm first block images, and perform spatial restoration on the segmentation binary image to obtain the binary image restoration result corresponding to each of the aneurysm first block images; S1083, perform binary graph integration and connected component analysis and numbering on the binary graph restoration results of all the first block images of the aneurysm to obtain the aneurysm segmentation result.

[0066] In this embodiment, the segmentation probability images and classification probabilities of each aneurysm obtained from the aneurysm segmentation model are fused. The product of each segmentation probability image and the classification probability is calculated, and the result is mapped to a segmentation binary image. The specific calculation formula is as follows: (7) Furthermore, the spatial restoration described above includes interpolating the obtained segmented binary image back to the original block size and restoring it to the block position in the intracranial vessel image to be processed before block removal, thus obtaining the binary image restoration result corresponding to the first block image of each aneurysm. Spatial restoration is the inverse operation of the block removal step, and will not be elaborated further here.

[0067] Further, aneurysm structure post-processing is performed, integrating all the binary image reconstruction results into a complete segmentation binary image. Then, connected component analysis is performed on the complete segmentation binary image, removing smaller connected components, such as those with fewer than 15 pixels. Finally, the remaining connected components are renumbered to obtain the final aneurysm segmentation result of the intracranial vessel image to be processed.

[0068] like Figure 5 As shown, for data post-processing, during the inference process, an intracranial vascular image to be processed is input. Based on the aneurysm detection model, n aneurysm detection results are obtained. The aneurysm detection results include the coordinates of the center point and the three-dimensional dimensions of the aneurysm in the n target connected regions obtained from the detection probability image. The output of the aneurysm detection model is used as the input of the aneurysm segmentation model. Based on the aneurysm segmentation model, the aneurysm segmentation results corresponding to the n target connected regions are obtained and presented in the form of a binary map, thus obtaining n aneurysm segmentation masks. Further data post-processing is performed to restore the position of the n aneurysm segmentation masks and integrate them globally. The final connected region is selected as the determined aneurysm region for segmentation, thereby obtaining the aneurysm segmentation result.

[0069] In this embodiment of the invention, by fusing the segmentation probability image and the classification probability during the segmentation process, the false positive rate of aneurysm detection results can be reduced without decreasing the aneurysm recall rate, thereby improving segmentation accuracy. Interpolating the binary image of the segmented block back to the original segmented block size and accurately mapping it back to the corresponding segmented block position in the original image allows for pixel-level coordinate alignment to restore the true spatial distribution of the aneurysm segmentation results in the full-frame medical image, avoiding edge misalignment and detail loss during segmentation scaling. Connected component analysis, while preserving the large connected components of the aneurysm body, eliminates scattered noise introduced during interpolation and tiny false positive regions mistakenly generated by segmentation edges. This approach not only preserves the true morphology and size accuracy of the aneurysm but also significantly reduces the false positive rate of the segmentation results, making it more suitable for the high sensitivity and low false detection requirements of intracranial aneurysm segmentation scenarios.

[0070] According to another aspect of the embodiments of this application, such as Figure 6 As shown, corresponding to the intracranial aneurysm detection and segmentation method in the above embodiments, this embodiment provides an intracranial aneurysm detection and segmentation device, the device comprising: The detection module 601 is used to perform forward propagation on the intracranial vascular image to be processed based on the trained aneurysm detection model, map the obtained detection probability image based on the probability threshold, and obtain the center point coordinates and three-dimensional dimensions of the aneurysm in at least one target connected region. The block segmentation module 603 is used to perform block segmentation based on the center point coordinates and three-dimensional dimensions of the aneurysm in each of the target connected regions, and to obtain the first block image of the aneurysm corresponding to the target connected region. The segmentation module 605 is used to perform forward propagation and aneurysm classification recognition on each of the first block images of the aneurysm based on the aneurysm segmentation model after training and inserting a classification branch network, and to obtain the classification probability and the segmentation probability image corresponding to each of the first block images of the aneurysm. The result calculation module 607 is used to integrate the segmentation probability image and the classification probability of each aneurysm first block image to obtain the aneurysm segmentation result.

[0071] It should be noted that in this embodiment, the detection module 601 can be used to execute step S102 in this application embodiment, the block-cutting module 603 in this embodiment can be used to execute step S104 in this application embodiment, the segmentation module 605 in this embodiment can be used to execute step S106 in this application embodiment, and the result calculation module 507 in this embodiment can be used to execute step S106 in this application embodiment.

[0072] It should be noted that the examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. The suffixes such as module, submodule, and subunit used to represent elements in the above device are only for the purpose of illustration in this application and have no specific meaning in themselves. Therefore, they can be used in combination.

[0073] According to another aspect of the embodiments of this application, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the intracranial aneurysm detection and segmentation method in any of the above embodiments.

[0074] According to another aspect of the embodiments of this application, this application also provides an electronic device, such as... Figure 7 As shown, the device includes a memory 701, a processor 703, and a network interface 705. The memory 701 stores a computer program that can run on the processor 703. The memory 701 and the processor 703 communicate through the network interface 705 and a communication bus 707. When the electronic device is running, the processor 703 and the memory 701 communicate through the network interface 705. When the processor 703 executes the computer program, it implements the steps of the above-described method for detecting and segmenting intracranial aneurysms.

[0075] The memory and processor in the aforementioned electronic device communicate with each other via a communication bus and a communication interface. The communication bus can be a peripheral component interconnect standard (PCI) bus or an extended industry standard structure (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. The memory can include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor. The aforementioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0076] It is understood that the embodiments described herein can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits, digital signal processors, digital signal processing devices, microprocessors, and other electronic units or combinations thereof for performing the functions described herein. For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. Software code can be stored in memory and executed by a processor.

[0077] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0078] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, device or unit, and may be electrical, mechanical or other forms.

[0079] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the various embodiments of this application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0080] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, essentially, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0081] It should be noted that, in this document, relational terms such as first, second, etc., are used only to distinguish one entity or operation from another entity or operation. The terms include, encompass, or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for detecting and segmenting intracranial aneurysms, characterized in that, The method includes: The trained aneurysm detection model is used to perform forward propagation on the intracranial vascular image to be processed. The obtained detection probability image is mapped based on the probability threshold to obtain the coordinates of the center point and the three-dimensional size of the aneurysm in at least one target connected region. Based on the coordinates of the center point and the three-dimensional size of the aneurysm in each target connected region, blocks are cut out according to the segmentation block cutting strategy to obtain the first block image of the aneurysm corresponding to the target connected region. Based on the aneurysm segmentation model trained and with a classification branch network inserted, forward propagation and aneurysm classification recognition are performed on the first block image of each aneurysm to obtain the classification probability and the segmentation probability image corresponding to the first block image of each aneurysm. The segmentation result of the aneurysm is obtained by integrating the segmentation probability image and the classification probability of each aneurysm's first block image.

2. The method of claim 1, wherein, Before performing forward propagation on the intracranial vascular image to be processed based on the trained aneurysm detection model, mapping the obtained detection probability image based on a probability threshold, and obtaining the center point coordinates and three-dimensional dimensions of at least one aneurysm in the target connected region, the method further includes: Data preprocessing is performed on aneurysm sample data to obtain multiple second aneurysm block images. The data preprocessing includes aneurysm size adjustment and block segmentation according to the detection block segmentation strategy, including segmenting positive sub-image blocks containing aneurysms and negative sub-image blocks not containing aneurysms. The aneurysm detection model is iteratively optimized by forward propagation based on the second block image of each aneurysm to obtain the detection probability image, and by using gradient descent based on the detection loss function to achieve model convergence, thereby obtaining the trained aneurysm detection model.

3. The method according to claim 1, characterized in that, The trained aneurysm detection model performs forward propagation on the intracranial vascular image to be processed, maps the obtained detection probability image based on a probability threshold, and obtains the coordinates and three-dimensional dimensions of the center point of the aneurysm in at least one target connected region, including: The intracranial vascular image to be processed is cropped into multiple sub-images with overlapping regions, and the multiple sub-images to be processed are input into the trained aneurysm detection model for forward propagation to obtain the detection probability sub-images corresponding to each sub-image to be processed. The positions of each detection probability sub-image are restored before cropping to obtain the corresponding detection probability image; Based on the probability threshold, the detection probability image is mapped to a binary image, and at least one target connected component is selected based on the mapped binary image. The selected target connected regions are numbered, and the coordinates of the center point and the three-dimensional dimensions of the aneurysm in each target connected region are calculated.

4. The method according to claim 1, characterized in that, Before the aneurysm segmentation model based on the trained network with inserted classification branches performs forward propagation and aneurysm classification recognition on each of the first segmented images of the aneurysm to obtain the classification probability and the segmentation probability image corresponding to each of the first segmented images of the aneurysm, the method further includes: Data preprocessing is performed on aneurysm sample data to obtain multiple aneurysm first segmented images. The preprocessing includes segmenting the aneurysm into blocks based on the center point coordinates and three-dimensional dimensions of the aneurysm in the sample data using the segmentation and segmentation strategy. This includes segmenting positive sub-images containing aneurysms and negative sub-images not containing aneurysms. The classification branch network that predicts whether an aneurysm exists in the first block image of the aneurysm segmentation model is added to the highest dimension feature of the aneurysm segmentation model. The aneurysm segmentation model is iteratively optimized by performing forward propagation and aneurysm classification based on the first block image of each aneurysm, obtaining the segmentation probability image and the classification probability, and then converging the model according to the gradient descent method based on the segmentation loss function and the classification loss function to obtain the trained aneurysm detection model.

5. The method according to claim 4, characterized in that, The step of segmenting aneurysms based on their center point coordinates and three-dimensional dimensions within each target connected region, and obtaining a first segmented image of the aneurysm corresponding to the target connected region, includes: Based on the segmentation and slicing strategy, the size of the corresponding first slicing block image of the aneurysm is calculated according to the size of the aneurysm in each dimension of the three-dimensional space in each of the target connected regions; Centered on the coordinates of the center point of the aneurysm, the first block image of the aneurysm corresponding to each of the target connected regions is extracted based on the size of the first block image of the aneurysm.

6. The method according to claim 1, characterized in that, The aneurysm segmentation model based on the trained network with inserted classification branches performs forward propagation and aneurysm classification recognition on each of the first segmented images of the aneurysm, and obtains a segmentation probability image corresponding to the classification probability and the first segmented image of the aneurysm, including: The first block image of each aneurysm is input into the trained aneurysm segmentation model for forward propagation and aneurysm classification and recognition. Forward propagation is performed based on the segmentation branch to obtain the segmentation probability image corresponding to the first block image of each aneurysm; Based on the classification branch network, aneurysm identification is performed on the first block image of each aneurysm to obtain the classification probability of the aneurysm.

7. The method according to any one of claims 1 to 6, characterized in that, The segmentation probability image and classification probability based on the first block image of each aneurysm are integrated to obtain the aneurysm segmentation result, including: Calculate the product of the segmentation probability image and the classification probability for each of the first block images of the aneurysm; Binary image mapping is performed based on the product of the segmentation probability image and the classification probability to obtain the segmentation binary image corresponding to each aneurysm first block image, and spatial restoration is performed on the segmentation binary image to obtain the binary image restoration result corresponding to each aneurysm first block image; The binary graph restoration results of all the first segmented block images of the aneurysm are integrated and connected component analysis and numbered to obtain the aneurysm segmentation result.

8. A device for detecting and segmenting intracranial aneurysms, characterized in that, The device includes: The detection module is used to perform forward propagation on the intracranial vascular image to be processed based on the trained aneurysm detection model, map the obtained detection probability image based on the probability threshold, and obtain the center point coordinates and three-dimensional dimensions of the aneurysm in at least one target connected region. The block segmentation module is used to perform block segmentation based on the center point coordinates and three-dimensional dimensions of the aneurysm in each of the target connected regions, and to obtain the first block image of the aneurysm corresponding to the target connected region. The segmentation module is used to perform forward propagation and aneurysm classification recognition on each of the first block images of the aneurysm based on the trained aneurysm segmentation model with inserted classification branch network, and to obtain the classification probability and the segmentation probability image corresponding to each of the first block images of the aneurysm. The result calculation module is used to integrate the segmentation probability image and the classification probability of each aneurysm first block image to obtain the aneurysm segmentation result.

9. An electronic device, comprising: The electronic device comprises a processor, a memory, and a network interface, wherein the memory stores machine-readable instructions executable by the processor, characterized in that: when the electronic device is in operation, the processor communicates with the memory via the network interface, and the processor executes the machine-readable instructions to perform the steps of the method for detecting and segmenting intracranial aneurysms as described in any one of claims 1 to 7.

10. A storage medium having processor-executable non-volatile program code, characterized in that, The program code causes the processor to execute the steps of the method for detecting and segmenting intracranial aneurysms according to any one of claims 1 to 7.

Images