Method, device, computer equipment and storage medium for identifying and detecting intracranial aneurysms

Abstract

The present application relates to a method, apparatus, computer equipment and storage medium for identifying and detecting intracranial aneurysms, the method comprising: obtaining positive and negative sample image blocks of the original intracranial image, the positive and negative sample image blocks being three-dimensional image blocks; processing the positive and negative sample image blocks using a three-dimensional semantic segmentation model, and outputting corresponding detection results, the three-dimensional semantic segmentation model being based on a convolutional neural network and constructed through a Transformer module, an encoder and a decoder, the processing process of the three-dimensional semantic segmentation model being divided into multiple stages according to the specifications of the positive and negative sample image blocks, the encoder and decoder being coupled and connected in the same stage through a first type of path, the first type of path being multiple paths corresponding to a single-dimensional vector of the three-dimensional image block. The present application can better perform feature extraction on the positive and negative sample image blocks, improve the detection rate of microaneurysms, maintain the coherence between the original intracranial images, and improve the continuity between layers.

Description

Technical Field

[0001] The present application relates to the intersection of the medical field and image processing, and in particular to a method, apparatus, computer equipment, and storage medium for identifying and detecting intracranial aneurysms. Background Art

[0002] Intracranial aneurysms are one of the most common cerebrovascular diseases. Although they contain the word "tumor," intracranial aneurysms are not tumors. Instead, they are localized damage to the intracranial arterial wall caused by factors such as congenital developmental abnormalities or acquired injury. Under the combined influence of hemodynamics and other factors, they gradually expand and form an abnormal bulge, a cerebrovascular disease. The incidence of intracranial aneurysms in the population is approximately 2% to 7%, and they can occur in any age group, with a peak incidence between the ages of 40 and 60.

[0003] Subarachnoid hemorrhage (SAH) caused by ruptured intracranial aneurysms has extremely high rates of disability and mortality, accounting for 5% to 10% of all strokes, with 80% of SAHs caused by ruptured intracranial aneurysms. In a sense, aneurysms are like a "ticking time bomb" buried within the skull. Therefore, rapid and accurate early screening for intracranial aneurysms is crucial.

[0004] Conventional examinations for intracranial aneurysms include intracranial CT angiography (CTA), intracranial digital subtraction angiography (DSA), and intracranial magnetic resonance angiography (MRA). CTA, as part of a comprehensive intracranial CT scan, offers fast imaging speed and high resolution, making it a commonly used aneurysm screening method. Traditional manual image reading and screening techniques are inefficient and have a high rate of missed detections. In recent years, research on automated aneurysm screening and detection based on artificial intelligence and deep learning technologies has become increasingly widespread.

[0005] In 2019, the medical journal JAMA Network Open published the paper "Deep Learning-Assisted Diagnosis of Cerebral Aneurysms Using the HeadXNet Model," which proposed the CTA-based aneurysm detection model, HeadXNet. Experiments showed that doctors equipped with the HeadXNet aneurysm-assisted diagnosis system had significantly higher accuracy in diagnosing aneurysms than those without algorithmic assistance.

[0006] In November 2020, Zhang Longjiang et al. published "A clinically applicable deep-learning model for detecting intracranial aneurysm incomputed tomography angiography images" in the journal Nature Communications, proposing an aneurysm detection model DAResUNet based on CTA images. The model is designed with multiple convolutional neural network (CNN) structures and uses DSA images as the gold standard for aneurysm diagnostic labels. Multiple comparative clinical experiments were designed. The final model had a sensitivity of up to 0.97 and a sensitivity of 0.85 on an independent external validation set. It also had a high detection rate for small aneurysms less than 5 mm.

[0007] In terms of MR imaging aneurysm diagnosis, Daiju et al. published "Deep Learning for MR Angiography: Automated Detection of Cerebral Aneurysms." in June 2019, proposing the AneuDect model for MRA images, which had sensitivities of 0.91 and 0.93 on internal and external validation sets, respectively.

[0008] Among the existing technical solutions mentioned above, although there has been some research on intracranial aneurysm detection based on deep learning, the following technical issues exist: First, the area occupied by aneurysms on the entire CTA image is very small, and the ratio between the background area and the aneurysm target area is seriously unbalanced during actual detection, resulting in a low detection rate for some small or tiny aneurysms; Second, the current design based on the classic 3D UNet network structure has a poor long-term reliance on capturing aneurysms in CTA images, and the continuity between layers needs to be further improved. Summary of the Invention

[0009] Based on this, it is necessary to provide a method for identifying and detecting intracranial aneurysms to address the above technical issues.

[0010] The method for identifying and detecting intracranial aneurysms in this application includes:

[0011] Obtaining positive and negative sample image blocks of the original intracranial image, wherein the positive and negative sample image blocks are three-dimensional image blocks;

[0012] A three-dimensional semantic segmentation model is used to process the positive and negative sample image blocks and output corresponding detection results. The three-dimensional semantic segmentation model is based on a convolutional neural network and is constructed through a Transformer module, an encoder and a decoder. The processing process of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and decoder are coupled and connected through a first-class path in the same stage. The first-class path is a plurality of paths corresponding to the single-dimensional vector of the three-dimensional image block.

[0013] Optionally, the encoder and decoder respectively include multiple encoding blocks and multiple decoding blocks, and at least a part of the encoding blocks and the decoding blocks in the same stage are coupled and connected through a first type of path, and at least another part of the encoding blocks and the decoding blocks are coupled and connected through a second type of path, and the second type of path is a long-range jump connection.

[0014] Optionally, the first type of path includes a Transformer module; the second type of path includes a Transformer module.

[0015] Optionally, at each stage, the decoding block receives single-dimensional vectors of positive and negative sample image blocks through the first type of path, and concatenates and fuses the single-dimensional vectors into a group. The decoder outputs the detection result in the form of a three-dimensional feature map.

[0016] Optionally, data enhancement and bone removal processing are also performed on the positive and negative sample image blocks.

[0017] The data enhancement process includes a random affine transformation process, wherein the random affine transformation process includes rotating the three-dimensional image block by an arbitrary angle along one or more dimensions;

[0018] The three-dimensional semantic segmentation model receives positive and negative sample image blocks that have undergone data enhancement and bone removal processing, and outputs detection results.

[0019] Optionally, the data processing also includes random redirection axis direction processing.

[0020] Optionally, obtaining positive and negative sample image blocks of the original intracranial image specifically includes:

[0021] sequentially receiving and annotating original intracranial images to obtain annotated image data, and performing image matrix processing on the annotated image data to obtain three-dimensional image data;

[0022] The critical line of the aneurysm in the three-dimensional space is extracted using the three-dimensional image data, and positive and negative sample image blocks of the three-dimensional image data are distinguished and obtained according to the critical line. The positive sample image block is in the area where the aneurysm is located, and the negative sample image block is in the area where the aneurysm is not located.

[0023] The present application also provides a device for identifying and detecting intracranial aneurysms, the device comprising:

[0024] A receiving module, configured to obtain positive and negative sample image blocks of the original intracranial image, wherein the positive and negative sample image blocks are three-dimensional image blocks;

[0025] The detection module uses a three-dimensional semantic segmentation model to process the positive and negative sample image blocks and output corresponding detection results. The three-dimensional semantic segmentation model is based on a convolutional neural network and is constructed through a Transformer module, an encoder and a decoder. The processing process of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and decoder are coupled and connected through a first-class path in the same stage. The first-class path is a plurality of paths corresponding to the single-dimensional vector of the three-dimensional image block.

[0026] The present application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the method for identifying and detecting intracranial aneurysms described in the present application when executing the computer program.

[0027] The present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for identifying and detecting intracranial aneurysms described in the present application.

[0028] The method for identifying and detecting intracranial aneurysms in this application has at least one of the following effects:

[0029] This application is based on a convolutional neural network and innovates in the selection of the backbone network. By rewriting the 2D CNN network into a 3D CNN network, the positive and negative sample image blocks of the original intracranial image are rewritten into single-dimensional vectors and the output results are restored into three-dimensional feature maps. This can better extract features from the positive and negative sample image blocks, improve the detection rate of small aneurysms, maintain the coherence between the original intracranial CTA images, and improve the continuity between layers.

[0030] This application uses data enhancement processing and bone removal processing (data enhancement processing includes random affine transformation processing and / or random redirection axis direction processing) to effectively improve the performance of the three-dimensional semantic segmentation model, enhance robustness, and further improve the accuracy of optimized detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the process of obtaining positive and negative sample image blocks in the intracranial aneurysm identification and detection method in one embodiment of the present application;

[0032] Figure 2 This is a flow chart of a method for identifying and detecting intracranial aneurysms in one embodiment of the present application;

[0033] Figure 3 for Figure 1 Schematic diagram of the original intracranial images during the process shown;

[0034] Figure 4 shows Figure 1 Schematic diagram of the aneurysm annotation results in the workflow shown;

[0035] FIG5 is a schematic diagram of an intracranial aneurysm prediction output result according to an embodiment of the present application;

[0036] FIG6( a ) is a Nifti image to be processed for data augmentation obtained by the method for identifying and detecting intracranial aneurysms according to an embodiment of the present application;

[0037] Figure 6(b) to Figure 6(d) This is a schematic diagram of an original intracranial image of an aneurysm obtained during the data augmentation process of the method for identifying and detecting an intracranial aneurysm in one embodiment of the present application;

[0038] Figure 7 for Figure 2 Schematic diagram of the network structure of the 3D semantic segmentation model in the process shown;

[0039] Figure 8 for Figure 7 Schematic diagram of the Transformer structure added;

[0040] Figure 9 This is a structural block diagram of an intracranial aneurysm identification and detection device in one embodiment of the present application. DETAILED DESCRIPTION

[0041] The purpose of intracranial imaging detection is to efficiently and accurately identify and detect intracranial aneurysms. However, existing technologies have low detection rates for small or tiny aneurysms, poor continuity between different layers, and low robustness.

[0042] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0043] To solve the above technical problems, see Figure 1 and Figure 2 In one embodiment of the present application, a method for identifying and detecting intracranial aneurysms is provided, which specifically includes the following steps:

[0044] Step S10, obtaining positive and negative sample image blocks of the original intracranial image, where the positive and negative sample image blocks are three-dimensional image blocks;

[0045] In one embodiment, step S10 specifically includes steps S11 to S13;

[0046] Step S11, obtaining original intracranial CTA images;

[0047] Step S12, marking after removing skull tissue;

[0048] Step S13: obtaining positive and negative sample image blocks.

[0049] Specifically, in step S10, original intracranial images are received and annotated in sequence to obtain annotated image data, and the annotated image data is subjected to image matrix processing to obtain three-dimensional image data; the three-dimensional image data is used to extract the critical line of the aneurysm in three-dimensional space, and the positive and negative sample image blocks of the three-dimensional image data are distinguished and obtained based on the critical line, where the positive sample image block is in the area where the aneurysm is located, and the negative sample image block is in the area where the aneurysm is not located.

[0050] Furthermore, step S10 includes obtaining a number of CTA images and bone-removed CTA images as original training images, and constructing an initial image set together with the annotated corresponding labels; and obtaining equal-sized image blocks with a positive-to-negative sample ratio of 1:3 according to the aneurysm ROI.

[0051] It can be understood that the format of the original intracranial images received and annotated in sequence can be Dicom or Nifti, and the image matrix processing is targeted at the image format Nifti; if the format of the original intracranial image is Nifti, no conversion is required. If the format of the original intracranial image is Dicom, it is necessary to convert the image Dicom sequence into Nifti format, and then resample to obtain image blocks of the same size and read the image matrix. The obtained matrix size is (1, 64, 64, 64), and the size unit is pixel. In this embodiment, the format of the original intracranial image is preferably Dicom format, such as Figure 3 The original CTA image is shown.

[0052] Sequentially receiving and annotating original intracranial images can be obtaining CTA images and corresponding labels, specifically including: first manually selecting CTA images that meet the requirements in the original sequence, then obtaining bone-free CTA through image silhouette technology and feeding the results back to the expert group for data annotation, and manually outlining the contour of the aneurysm area based on image features and diagnostic experience to obtain the label map (ground truth of image annotation) shown in Figure 4, forming a one-to-one corresponding usable data set.

[0053] It can be understood that the method for obtaining positive and negative sample image blocks is to extract positive and negative sample data according to the aneurysm ROI (region of interest), specifically including: first, dividing the positive and negative sample areas of the image according to the label matrix, that is, the aneurysm and other intracranial tissues are distinguished by different pixel values, and the location of the aneurysm is regarded as a positive sample. Then the positive sample ROI is calculated, and the coordinates of the critical points are recorded from three different dimensions of x, y, and z in three-dimensional space to perform image block resampling. In one embodiment, image blocks with a positive and negative sample ratio of 1:3 are obtained through sequential sampling and random sampling, and the size is 64×64×64. The image block specifications of the positive and negative samples are in pixels. Unless otherwise specified below, the unit of specification size shall be pixel.

[0054] To enhance the robustness of the detection process, in one embodiment, step S10 also includes data enhancement processing and bone removal processing on the positive and negative sample image blocks, where the data enhancement processing includes random affine transformation processing, and the random affine transformation processing includes rotating the three-dimensional image block by any angle along one or more dimensions; the three-dimensional semantic segmentation model receives the positive and negative sample image blocks that have undergone data enhancement processing and bone removal processing, and outputs the detection results.

[0055] In one embodiment, data processing also includes randomized reorientation axis processing. It is understood that randomized affine transformations and randomized reorientation axes belong to randomized transformation methods for 3D data. Specifically, the randomized 3D data transformation method is applied by selecting an online data augmentation strategy for image preprocessing. First, the image blocks are randomly reorientated axiswise, with the default axis labels being left (L), right (R), posterior (P), anterior (A), inferior (I), and superior (S). For example, the initial "RPI" is reorientated to the "PRI" direction. Next, the image blocks are randomly affine transformed, specifically by randomized angle rotations. Each rotation can be considered as a random rotation of α, β, and γ along the x, y, and z axes, respectively. As shown in Figure 6, Figure 6(a) shows the obtained Nifti image, Figure 6(b) shows the image after the reorientation axis, and Figures 6(c) and 6(d) show the images rotated 45 and 90 degrees, respectively. Finally, bone-removed CTA images obtained using the silhouette technique are used as supplementary data, i.e., a portion of the bone-removed image is randomly selected as the original input to enhance the model's expressiveness.

[0056] In one embodiment, the 3D data augmentation process includes sequentially performing random affine transformations and random reorientation axis processing. To increase the reliability of the positive and negative sample image blocks to be detected, the positive and negative sample image blocks undergo image normalization and vectorization after the 3D data augmentation process to complete the preliminary data processing.

[0057] In existing technologies, data augmentation methods applicable to 3D medical images are relatively limited, yet their impact on model results is crucial. Therefore, various embodiments employing data augmentation effectively improve model performance and robustness through axis reorientation, random affine transformation, and random selection of bone-removing CTAs.

[0058] Step S20, constructing a 3D semantic segmentation model based on a convolutional neural network through a Transformer module, an encoder, and a decoder;

[0059] In this embodiment, the three-dimensional semantic segmentation model is specifically a Res-Trans U-Net model. Both the encoder and decoder can change the specifications of the positive and negative sample image blocks. The encoder and decoder respectively include multiple encoding blocks and multiple decoding blocks. The Res-Trans U-Net model follows the U-shaped design concept and adds a Transformer module to the symmetrical encoding and decoding structure to coordinate the features of each stage.

[0060] The encoder is based on the 2D ResNet-50 and built using a 3D CNN. The decoder is composed of convolution and deconvolution layers, gradually restoring the feature map size. The Res-Trans U-Net architecture described above combines the CNN and Transformer architectures. Convolutional neural networks (CNNs), 2D ResNet-50, 3D CNN, and Transformer architectures are all commonly used network models and modules in the art and will not be explained in detail here.

[0061] In this embodiment, the encoder uses 2D ResNet-50 as a prototype and contains a total of four encoding blocks. Specifically, a 3DCNN is used to build a residual bottleneck block. The bottleneck block consists of three convolution blocks and a short-range skip connection. The structure of the first convolution block is formed by a 1×1 convolution, a batch normalization layer, and a ReLU activation layer connected in series; the structure of the second convolution block is formed by a 3×3 convolution, a batch normalization layer, and a ReLU activation layer connected in series; the third convolution block has the same structure as the first convolution block; the short-range skip connection is obtained by adding the input feature map of the first convolution block to the output feature map of the third convolution block. The decoder contains a total of four decoding blocks, each of which is composed of two 3×3 convolution layers and a deconvolution layer connected in series, and a BN batch normalization layer and a ReLU activation layer are added after each convolution layer.

[0062] In step S30, the processing of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and the decoder are coupled and connected through the first type of path in the same stage. The first type of path is a plurality of paths corresponding to the single-dimensional vectors of the three-dimensional image blocks.

[0063] In this embodiment, the encoder divides the positive and negative sample image blocks into single-dimensional vectors of multiple dimensions based on the different dimensions of the 3D image blocks in 3D space. Specifically, the 3D feature maps of the positive and negative sample image blocks are divided into one-dimensional vectors. The number of single-dimensional vectors can be, for example, three types of single-dimensional vectors based on the three dimensions of the positive and negative sample image blocks.

[0064] It can be understood that since both the encoder and the decoder can change the size of the positive and negative sample image blocks; taking the encoder as an example, the feature map size is reduced by half after each encoding block; the corresponding feature map size is doubled after each decoding block. Figure 7 As shown, in the same stage, the specifications of the positive and negative sample image blocks are the same. In this embodiment, since the number of encoding blocks and decoding blocks is the same and both are four, it is divided into four stages. In the same stage, the encoding block and the decoding block are coupled and connected through the first type of path. The coupling connection can be a direct connection or a systematic connection. The systematic connection includes an indirect connection and a connection that introduces other parameters. The first type of path includes multiple paths, each of which corresponds to one of the above-mentioned one-dimensional vectors; to match the above-mentioned three types of one-dimensional vectors, there can be three first type paths.

[0065] This application innovates in its choice of backbone network. While existing techniques typically use CNN models to process two-dimensional images, this application rewrites the 2D CNN network into a 3D form to maintain coherence between three-dimensional CTA images, enabling better feature extraction for three-dimensional image blocks. Specifically, by fusing features of single-dimensional vectors at the same stage, it enables better feature extraction for positive and negative sample image blocks, improving the detection rate of microaneurysms, maintaining coherence between original intracranial CTA images, and improving continuity between layers.

[0066] As a preferred approach, the encoding block of the residual bottleneck block constructed using 3D CNN is repeated 3, 4, 6, and 3 times in each stage respectively.

[0067] In one embodiment, at least a portion of the encoding blocks and the decoding blocks in the same stage are coupled via a first type of path, and at least another portion of the encoding blocks and the decoding blocks are coupled via a second type of path, and the second type of path is a long-range jump connection.

[0068] This embodiment further maintains the connection between features in the same stage through the second type of path to construct a three-dimensional semantic segmentation model to achieve the detection task of intracranial aneurysms. The input of the Transformer module and the second type of path comes from the output of each stage of the encoder (each stage corresponds to an encoding block), that is, the feature map of each stage of the encoder can also be fused with the feature map of the decoding block of the same stage (each stage corresponds to a decoding block) through the second type of path to optimize image details. The second type of path is specifically a long-range skip connection.

[0069] In this embodiment, the first type of path is enhanced by assisting the second type of path, relying on the coupling effect of the three-dimensional image (the second type of path) and the one-dimensional vector (the first type of path), the coherence between the original intracranial CTA images is further maintained, and the continuity between layers is improved. It can be understood that if the algorithm relies only on the first type of path, the information lost in the input image will not be well restored, which will greatly reduce the segmentation accuracy for three-dimensional images, so the second type of path is needed to assist; if the algorithm relies only on the second type of path, a large amount of global information will be lost, which will also cause the problem of inaccurate segmentation effect. In this embodiment, the Res-Trans U-Net model includes an encoder, a decoder, a long-range skip connection and a Transformer module.

[0070] Specifically, at the same stage, the encoder and decoder are coupled through the Transformer module, specifically including: the first type of path includes the Transformer module; the second type of path includes the Transformer module. Figure 7 As shown in the figure, the first type of path connects the corresponding encoding block and decoding block through the Transformer module; the second type of path connects the corresponding encoding block and decoding block through the Transformer module.

[0071] Step S40: Process the positive and negative sample image blocks using a three-dimensional semantic segmentation model and output corresponding detection results.

[0072] Specifically, the three-dimensional semantic segmentation model receives positive and negative sample image blocks, and outputs detection results through multiple serial encoding blocks and multiple serial decoding blocks in sequence.

[0073] It can be understood that in the entire workflow of each encoding block and each decoding block, four times of image size reduction (implemented by the encoding block) and four times of image size enlargement (implemented by the decoding block) are performed in sequence; in terms of the division of specific stages, such as Figure 7As shown, the process of four successive size reductions and size enlargements is divided into four stages. It can be understood that the first size reduction stage is the same as the fourth size enlargement stage, both belonging to the first stage; the second size reduction stage is the same as the third size enlargement stage, both belonging to the second stage; the third size reduction stage is the same as the second size enlargement stage, both belonging to the third stage; and the fourth size reduction stage is the same as the first size enlargement stage, both belonging to the fourth stage.

[0074] In this embodiment, the positive and negative sample image blocks received by the three-dimensional semantic segmentation model are preferably image blocks after data enhancement; specifically, the size of the positive and negative sample images is 1×64×64×64, and the output is a segmentation prediction map of the same size, as shown in Figure 5, and the number of categories is two.

[0075] It can be understood that at each stage, the decoding block receives the single-dimensional vectors of the positive and negative sample image blocks through the first type of path, and concatenates and fuses the single-dimensional vectors into a group. The decoder outputs the detection results in the form of a three-dimensional feature map.

[0076] Specifically, the second path converts the feature map into a one-dimensional vector first, and then inputs the feature vectors from different stages into the Transformer module. The second path includes a total of 6 repeated Transformer structures. The vectors are first fused into a set of vectors by splicing at input, and then restored to a three-dimensional feature map at output. The Transformer structure is as follows Figure 8 As shown, the complete Multi-Trans 3D-UNet network structure diagram (the three-dimensional semantic segmentation model described in each embodiment of this application) is as follows Figure 7 As shown in Figure 2, the encoder inputs positive and negative sample image blocks (1, D, H, W), and the decoder block outputs the detection results (2, D, H, W).

[0077] In one embodiment, the accuracy of the intracranial aneurysm detection method is further tested using the aforementioned 3D semantic segmentation model. This includes using the trained 3D semantic segmentation model to detect intracranial aneurysms on CTA images, using the intersection-over-union ratio between the predicted results (as shown in FIG5 ) and the true labels (as shown in FIG4 ) as the primary metric, and simultaneously testing the model's sensitivity from two different perspectives: the complete case and the specific lesion.

[0078] It can be understood that by testing the three-dimensional semantic segmentation model, the changes in the order of steps, number of steps and specific details of the steps in the above steps S10 to S50 can be comprehensively obtained, and optimized based on the inventive concept of this application to obtain the optimal technical solution.

[0079] Steps S20 to S40 can be summarized as: using a three-dimensional semantic segmentation model to process the positive and negative sample image blocks and output corresponding detection results. The three-dimensional semantic segmentation model is based on a convolutional neural network and is constructed through a Transformer module, an encoder and a decoder. The processing process of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and the decoder are coupled and connected through a first-class path in the same stage. The first-class path is a plurality of paths corresponding to the single-dimensional vector of the three-dimensional image block.

[0080] In summary, this application builds a deep learning algorithm for automatic detection of intracranial aneurysms based on the attention mechanism network. Specifically, the purpose is to improve the accuracy of the aneurysm detection algorithm by designing a three-dimensional image augmentation method. The algorithm targets the original intracranial three-dimensional images and introduces a variety of three-dimensional image augmentation methods in the data processing stage to improve the robustness of the algorithm, so as to achieve the purpose of optimizing the detection accuracy. This application makes innovations in the model structure and proposes a Res-Trans U-Net network model. The encoder uses 3D CNN to build ResNet as the main path. The network has a strong ability in image feature extraction; using long-range jumps and Transformer modules as a means of feature fusion, this application simultaneously applies three-dimensional image feature fusion, single-dimensional image feature fusion, and three-dimensional data augmentation processing; and effectively utilizes the output results of each stage of the encoder. Experiments have shown that in the detection task of intracranial aneurysms, the Res-Trans U-Net model (three-dimensional semantic segmentation model) established by this application has good performance and can achieve the technical effects described in the invention content of this application.

[0081] It should be understood that although Figure 1-2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1-2 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0082] In one embodiment, a device for identifying and detecting intracranial aneurysms is provided, comprising: a receiving module, a dividing module, a building module, a connecting module, and a detecting module, wherein:

[0083] A receiving module, configured to obtain positive and negative sample image blocks of the original intracranial image, wherein the positive and negative sample image blocks are three-dimensional image blocks;

[0084] The detection module uses a three-dimensional semantic segmentation model to process the positive and negative sample image blocks and output corresponding detection results. The three-dimensional semantic segmentation model is based on a convolutional neural network and is constructed through a Transformer module, an encoder and a decoder. The processing process of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and decoder are coupled and connected through a first-class path in the same stage. The first-class path is a plurality of paths corresponding to the single-dimensional vector of the three-dimensional image block.

[0085] The specific definition of the device for identifying and detecting intracranial aneurysms can be found in the definition of the method for identifying and detecting intracranial aneurysms above, and will not be repeated here. The various modules in the above-mentioned device for identifying and detecting intracranial aneurysms can be implemented in whole or in part by software, hardware, or a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0086] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for identifying and detecting intracranial aneurysms is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0087] Those skilled in the art will understand that Figure 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0088] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the method for identifying and detecting intracranial aneurysms described in each embodiment of the present application is implemented.

[0089] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0090] Step S10, obtaining positive and negative sample image blocks of the original intracranial image, where the positive and negative sample image blocks are three-dimensional image blocks;

[0091] Step S20, constructing a 3D semantic segmentation model based on a convolutional neural network through a Transformer module, an encoder, and a decoder;

[0092] In step S30, the processing of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and the decoder are coupled and connected through the first type of path in the same stage. The first type of path is a plurality of paths corresponding to the single-dimensional vectors of the three-dimensional image blocks.

[0093] Step S40: Process the positive and negative sample image blocks using a three-dimensional semantic segmentation model and output corresponding detection results.

[0094] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for identifying and detecting intracranial aneurysms described in each embodiment of the present application is implemented.

[0095] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0096] Step S10, obtaining positive and negative sample image blocks of the original intracranial image, where the positive and negative sample image blocks are three-dimensional image blocks;

[0097] Step S20, constructing a 3D semantic segmentation model based on a convolutional neural network through a Transformer module, an encoder, and a decoder;

[0098] In step S30, the processing of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and the decoder are coupled and connected through the first type of path in the same stage. The first type of path is a plurality of paths corresponding to the single-dimensional vectors of the three-dimensional image blocks.

[0099] Step S40: Process the positive and negative sample image blocks using a three-dimensional semantic segmentation model and output corresponding detection results.

[0100] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0101] The technical features of the above embodiments may be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there are no conflicts in the combination of these technical features, they should be considered to be within the scope of this specification. When technical features in different embodiments are reflected in the same figure, it can be regarded as that figure also discloses the combination examples of the various embodiments involved.

[0102] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for identifying and detecting intracranial aneurysms, characterized in that: include: Obtaining positive and negative sample image blocks of the original intracranial image, wherein the positive and negative sample image blocks are three-dimensional image blocks; Processing the positive and negative sample image blocks using a three-dimensional semantic segmentation model and outputting corresponding detection results. The three-dimensional semantic segmentation model is based on a convolutional neural network and is constructed using a Transformer module, an encoder, and a decoder. The processing of the three-dimensional semantic segmentation model is divided into multiple stages based on the specifications of the positive and negative sample image blocks. The encoder and decoder respectively include multiple encoding blocks and multiple decoding blocks. The specifications of the positive and negative sample image blocks in the same stage are the same. The encoder and decoder are coupled and connected through a first type of path and a second type of path in the same stage. The first type of path includes a Transformer module, and the second type of path includes a Transformer module. The first type of path is a plurality of paths corresponding to the single-dimensional vectors of the three-dimensional image blocks. The number of single-dimensional vectors is divided into three types of single-dimensional vectors according to the three dimensions of the positive and negative sample image blocks, and the second type of path is a long-range jump connection. The input of the second type of path comes from the output of each stage of the encoder, and the feature vectors from different stages are input together into the Transformer module included in the second type of path. The feature vector is fused into a group of vectors by splicing when input, and is restored to a three-dimensional feature map when output.

2. The method for identifying and detecting intracranial aneurysms according to claim 1, wherein: At each stage, the decoding block receives the single-dimensional vectors of the positive and negative sample image blocks through the first type of path, and concatenates and fuses the single-dimensional vectors into a group. The decoder outputs the detection results in the form of a three-dimensional feature map.

3. The method for identifying and detecting intracranial aneurysms according to claim 1, wherein: It also includes data enhancement and bone removal processing for positive and negative sample image blocks. The data enhancement process includes a random affine transformation process, wherein the random affine transformation process includes rotating the three-dimensional image block by an arbitrary angle along one or more dimensions; The three-dimensional semantic segmentation model receives positive and negative sample image blocks that have undergone data enhancement and bone removal processing, and outputs detection results.

4. The method for identifying and detecting intracranial aneurysms according to claim 3, wherein: The data enhancement process also includes random redirection axis direction processing.

5. The method for identifying and detecting intracranial aneurysms according to claim 1, wherein: The obtaining of positive and negative sample image blocks of the original intracranial image specifically includes: sequentially receiving and annotating original intracranial images to obtain annotated image data, and performing image matrix processing on the annotated image data to obtain three-dimensional image data; The critical line of the aneurysm in the three-dimensional space is extracted using the three-dimensional image data, and positive and negative sample image blocks of the three-dimensional image data are distinguished and obtained according to the critical line. The positive sample image block is in the area where the aneurysm is located, and the negative sample image block is in the area where the aneurysm is not located.

6. A device for identifying and detecting intracranial aneurysms, characterized in that: The device comprises: A receiving module, configured to obtain positive and negative sample image blocks of the original intracranial image, wherein the positive and negative sample image blocks are three-dimensional image blocks; The detection module uses a three-dimensional semantic segmentation model to process the positive and negative sample image blocks and output corresponding detection results. The three-dimensional semantic segmentation model is based on a convolutional neural network and is constructed by a Transformer module, an encoder and a decoder. The processing process of the three-dimensional semantic segmentation model is divided into multiple stages according to the specifications of the positive and negative sample image blocks. The encoder and decoder respectively include multiple encoding blocks and multiple decoding blocks. The specifications of the positive and negative sample image blocks in the same stage are the same. The encoder and decoder are coupled and connected through a first type of path and a second type of path in the same stage. The first type of path includes a Transformer module, and the second type of path includes a Transformer module. The first type of path is multiple paths corresponding to the single-dimensional vectors of the three-dimensional image blocks. The number of single-dimensional vectors is divided into three types according to the three dimensions of the positive and negative sample image blocks, and the second type of path is a long-range skip connection. The input of the second type of path comes from the output of each stage of the encoder. The feature vectors from different stages are input together into the Transformer module included in the second type of path. The feature vectors are fused into a group of vectors by splicing when input and restored to a three-dimensional feature map when output.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images