DSA image recognition method, device and storage medium

Abstract

This invention provides a DSA image recognition method, apparatus, and storage medium, comprising: receiving 2D-DSA image data, the 2D-DSA image data including one or more 2D-DSA images; performing a first-stage region recognition on the 2D-DSA images based on a CNN model to locate a specific region in the 2D-DSA images; after locating the specific region, performing a second-stage aneurysm recognition on the 2D-DSA images based on the CNN model to obtain recognition result data; and outputting the recognition result data, wherein the recognition result data includes an aneurysm label marked on the 2D-DSA images. This invention, through a CNN network algorithm, can perform two-stage processing on 2D-DSA images, progressively locating aneurysms, effectively reducing interference from other factors and improving detection accuracy.

Description

Technical Field

[0001] This invention relates to DSA image recognition technology, and more particularly to a DSA image recognition method, apparatus and storage medium. Background Technology

[0002] Intracranial aneurysms are cerebrovascular diseases caused by abnormal localized dilation of the walls of cerebral arteries. They are often caused by thinning of the intima and appear to be associated with polycystic kidney disease, fibrosis, and aortic coarctation. If not detected and treated promptly, rupture of an intracranial aneurysm can lead to subarachnoid hemorrhage, resulting in severe neurological sequelae and a high mortality rate.

[0003] Intracranial aneurysms are relatively common, with an incidence rate of 3% worldwide and 7% in China. Many people have aneurysms without knowing it, making screening for intracranial aneurysms crucial. Digital subtraction angiography (DSA) has long been considered the gold standard for diagnosing intracranial aneurysms. DSA is recommended for the identification and evaluation of intracranial aneurysms when considering surgical or endovascular treatment. Compared to CT angiography (CTA) and magnetic resonance angiography (MRA), DSA offers higher resolution and detection sensitivity, especially for intracranial aneurysms smaller than 3 mm.

[0004] DSA records the acquired two-dimensional DSA angiographic images in a standard Digital Imaging and Communications in Medicine (DICOM) format. DSA images are sampled sequentially as the contrast agent flows through the entire artery at a sampling rate of 3 to 5 frames per second, with the sampling process typically lasting 3 to 15 seconds. During imaging, the patient's head should be in a stable position to obtain a clear imaging sequence. The physician should examine the entire two-dimensional DSA image sequence to identify the intracranial aneurysm. The examination process may take longer than expected, and small intracranial aneurysms are sometimes misdiagnosed due to the potential for errors in the location and shape of the aneurysm. Traditional diagnostic methods are laborious, requiring the involvement of skilled and experienced physicians. Nevertheless, cases where intracranial aneurysms cannot be identified still occur from time to time. Therefore, automated systems for diagnosing intracranial aneurysms can alleviate the burden on physicians to some extent.

[0005] For patients with subarachnoid hemorrhage, physicians need to obtain anteroposterior two-dimensional (2D-DSA) and lateral views of the four intracranial arteries (left and right common carotid arteries and left and right vertebral arteries). If an intracranial aneurysm is found in the artery, three-dimensional (3D-DSA) reconstruction of the entire artery should be performed immediately. The reconstructed 3D DSA provides essential information about the vascular morphology and pathology. Physicians can use it to analyze the geometry of the vessel in detail, i.e., the vessel diameter, the location and size of the aneurysm, in order to make clinical decisions. However, the quality of 3D-DSA images can vary from patient to patient. In clinical practice, many factors affect the quality of 3D-DSA images, including the amount of contrast agent injected, the timing of injection, the hemodynamic parameters of the contrast agent, and differences in factors such as blood flow and cardiac output. In addition, the reconstruction parameters may also differ between patients. These factors can lead to biases in physicians' selection of treatment options for patients.

[0006] On the other hand, compared with 2D-DSA mode, 3D-DSA mode has more information and can easily identify aneurysms. However, most hospitals in developing countries can only purchase 2D angiography equipment at the price of expensive 3D equipment, and the 3D-DSA data for invasive examinations is quite limited.

[0007] Current research on automated computer-aided diagnosis using DSA mode is based on classical digital image processing methods, employing 2D-DSA images. However, classical digital image processing methods cannot solve the problem of automatic aneurysm localization. Current feature extraction methods are severely hampered by the complex distribution of blood vessels, and the sliding window method is extremely time-consuming during search and feature extraction. Therefore, classical digital image processing methods are not the optimal approach.

[0008] Compared to classical digital image processing methods, Convolutional Neural Networks (CNNs) have proven to be more effective in feature extraction. In recent years, CNN architectures have been widely applied in object detection with good results. Furthermore, CNN methods have also performed well in some medical image detection tasks. However, currently, there is no technical solution that combines CNN architecture with aneurysm recognition. Summary of the Invention

[0009] This invention provides a DSA image recognition method, device, and storage medium that combines a CNN architecture with aneurysm recognition, thereby improving the accuracy of aneurysm recognition and detection.

[0010] A first aspect of the present invention provides a DSA image recognition method, comprising:

[0011] Receive 2D-DSA image data, wherein the 2D-DSA image data includes one or more 2D-DSA images;

[0012] The first stage of region recognition is performed on the 2D-DSA image based on the CNN model to locate specific regions in the 2D-DSA image. After locating the specific regions, the second stage of aneurysm recognition is performed on the 2D-DSA image based on the CNN model to obtain the recognition result data.

[0013] The identification result data is output, wherein the identification result data includes aneurysm labels marked on the 2D-DSA image.

[0014] Optionally, in one possible implementation of the first aspect, the first-stage region identification based on the CNN model of the 2D-DSA image, locating a specific region in the 2D-DSA image, includes:

[0015] The framework of the CNN model is built on a feature pyramid network, with a residual neural network as its backbone.

[0016] The 2D-DSA image includes a frontal image and a side image. This represents the j-th image in the sequence of frontal images of the i-th patient. This represents the j-th image in the lateral image sequence of the i-th patient. and Correspondingly;

[0017] The frontal image and the side image are respectively input into the feature pyramid network. The target region is determined in the frontal image, and the corresponding region corresponding to the target region is determined in the side image.

[0018] Optionally, in one possible implementation of the first aspect, the second-stage aneurysm identification based on a CNN model after locating a specific region to obtain identification data includes:

[0019] Obtain the target region in the frontal image and the corresponding region in the side image;

[0020] The target region and the corresponding region are connected and fed into the feature pyramid network in a dual input manner to obtain the recognition result data;

[0021] The identification result data includes aneurysm labels in the images, which are used to indicate the confidence level that a certain location in the image is an aneurysm.

[0022] Optionally, in one possible implementation of the first aspect, after obtaining the recognition result data, the recognition result data is subjected to region average grayscale suppression processing, including:

[0023] The confidence level of the aneurysm is compared with the standard threshold. If the confidence level of the aneurysm is lower than the standard, the region average gray level suppression algorithm is applied.

[0024] An adaptive grayscale threshold selection method is used to coarsely extract the vascular model from the original target aneurysm region image;

[0025] Based on the vascular model and the original aneurysm region image, the blood vessels in the target aneurysm region image are calculated logically.

[0026] Among them, the gray value of the aneurysm is G 动脉瘤 The pixel value is P 动脉瘤 The average gray value of the aneurysm is:

[0027] AG 动脉瘤 = 动脉瘤 / 动脉瘤

[0028] The average gray value of the vessels surrounding the aneurysm is:

[0029] AG 动脉瘤周边 =(G 扩大区域 -G 动脉瘤 ) / (P 扩大区域 -P 动脉瘤 )

[0030] If, AG 动脉瘤 Less than AG 动脉瘤周边 If the image of that region is an overlapping blood vessel, then if AG 动脉瘤 Greater than AG 动脉瘤周边 If so, the region image is an aneurysm.

[0031] Alternatively, in one possible implementation of the first aspect, the CNN model is trained by the following steps:

[0032] Acquire the patient's 2D-DSA image data, which includes one or more 2D-DSA images;

[0033] Images below a preset value are removed from the 2D-DSA images to obtain 2D-DSA images that meet the preset value.

[0034] The 2D-DSA images that meet the preset values ​​are divided into training datasets and test datasets based on a preset ratio.

[0035] Optionally, in one possible implementation of the first aspect, after dividing the 2D-DSA images that satisfy the preset values ​​into training and test datasets based on a preset ratio, the method further includes:

[0036] The multiple 2D-DSA images constitute a DSA image sequence;

[0037] The DSA image is locally normalized so that the local mean value of the DSA image is 0 and the SD value is 1.

[0038] A second aspect of the present invention provides a DSA image recognition device, comprising:

[0039] A 2D-DSA image receiving module is used to receive 2D-DSA image data, wherein the 2D-DSA image data includes one or more 2D-DSA images;

[0040] The recognition module is used to perform a first-stage region recognition on the 2D-DSA image based on a CNN model, locate a specific region in the 2D-DSA image, and after locating the specific region, perform a second-stage aneurysm recognition on the 2D-DSA image based on a CNN model to obtain recognition result data.

[0041] An output module is used to output the recognition result data, wherein the recognition result data includes aneurysm labels that are marked on the 2D-DSA image.

[0042] Optionally, in one possible implementation of the second aspect, the identification module includes a one-stage identification unit, which is used to perform the following steps:

[0043] The framework of the CNN model is built on a feature pyramid network, with a residual neural network as its backbone.

[0044] The 2D-DSA image includes a frontal image and a side image. This represents the j-th image in the sequence of frontal images of the i-th patient. This represents the j-th image in the lateral image sequence of the i-th patient. and Correspondingly;

[0045] The frontal image and the side image are respectively input into the feature pyramid network. The target region is determined in the frontal image, and the corresponding region corresponding to the target region is determined in the side image.

[0046] Optionally, in one possible implementation of the second aspect, the identification module includes a two-stage identification unit, which is used to perform the following steps:

[0047] Obtain the target region in the frontal image and the corresponding region in the side image;

[0048] The target region and the corresponding region are connected and fed into the feature pyramid network in a dual input manner to obtain the recognition result data;

[0049] The identification result data includes aneurysm labels in the images, which are used to indicate the confidence level that a certain location in the image is an aneurysm.

[0050] A third aspect of the present invention provides a readable storage medium storing a computer program, which, when executed by a processor, is used to implement the methods of the first aspect of the present invention and various possible designs of the first aspect.

[0051] This invention provides a DSA image recognition method, device, and storage medium. Through a CNN network algorithm, it can perform two-stage processing on 2D-DSA images to progressively locate aneurysms, effectively reducing interference from other factors and improving detection accuracy. The network algorithm for the first-stage region identification has an accuracy of 93%, meaning that it can effectively detect aneurysm regions in 93 out of 100 test images. This effectively solves the problem of background information interference and improves detection accuracy.

[0052] Furthermore, comparing the results of the present invention with those of the single-input model, the convergence curve of the dual-input model in the five-fold cross-validation results fluctuated between 0.59 and 0.72, while the curve convergence of the single-input model was between 0.56 and 0.82. The dual-input model, which combines frontal and lateral information, showed more stable performance. In addition, the region-average grayscale suppression algorithm effectively suppressed false positives and could better distinguish between intracranial aneurysms and overlapping blood vessels. The detection results of this invention have higher sensitivity, specificity, and accuracy than those of senior physicians. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the first embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of a second embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of a model according to an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the network architecture for the region identification stage in an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of the network architecture for the aneurysm recognition stage according to an embodiment of the present invention;

[0058] Figure 6 This is a schematic diagram illustrating the training of the recognition algorithm according to an embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of a third embodiment of the present invention;

[0060] Figure 8 This is a schematic diagram of the fourth embodiment of the present invention;

[0061] Figure 9 This is a schematic diagram of the blood vessel centerline according to an embodiment of the present invention;

[0062] Figure 10 This is a schematic diagram of the fifth embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0065] It should be understood that in the various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0066] It should be understood that in this invention, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0067] It should be understood that in this invention, "multiple" refers to two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, "and / or B" can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "Contains A, B, and C", "Contains A, B, and C" means that all three A, B, and C are contained; "Contains A, B, or C" means that one of A, B, and C is contained; "Contains A, B, and / or C" means that any one, two, or three of A, B, and C are contained.

[0068] It should be understood that in this invention, "B corresponding to A", "B corresponding to A", "A and B correspond", or "B and A correspond" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Matching A and B is defined as a similarity between A and B that is greater than or equal to a preset threshold.

[0069] Depending on the context, "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection."

[0070] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0071] Digital subtraction angiography (DSA) is commonly used to assess the structural basis of neurovascular diseases such as stenosis, arteriovenous malformations, and intracranial aneurysms. In addition to the geometry and connectivity displayed by standard 2D-DSA, parametric angiography helps clinicians assess the functional properties of these lesions. Parametric angiography is an imaging method that uses digital subtraction angiography to describe the dynamic properties of contrast agents throughout the vascular system. It uses convolutional neural networks (CNNs) to identify and segment intracranial aneurysms in DSA and extracts aneurysm imaging features with minimal error. This may help in understanding the natural history of the disease, predicting disease progression, or planning interventions.

[0072] 2D-DSA is considered the gold standard for diagnosing aneurysms. However, for patients screened for subarachnoid hemorrhage, doctors need to obtain anteroposterior and lateral 2D-DSA images of the four intracranial arteries. If an intracranial aneurysm is found in an artery, doctors should immediately perform 3D-DSA to reconstruct the entire artery. The reconstructed 3D-DSA provides more information to help select treatment. However, this 3D-DSA needs to be compared with the 2D-DSA to avoid errors caused by differences in patient image quality.

[0073] Embodiments of the present invention provide a DSA image recognition method, such as... Figure 1 and Figure 2 As shown, it includes:

[0074] Step S110: Receive 2D-DSA image data, wherein the 2D-DSA image data includes one or more 2D-DSA images;

[0075] Step S120: Perform a first-stage region recognition on the 2D-DSA image based on a CNN model to locate a specific region in the 2D-DSA image. After locating the specific region, perform a second-stage aneurysm recognition on the 2D-DSA image based on a CNN model to obtain recognition result data.

[0076] Step S130: Output the recognition result data, wherein the recognition result data includes aneurysm tags that are marked on the 2D-DSA image.

[0077] In this embodiment, there are two main steps. First, a 2D DSA image in DICOM format is uploaded to the aneurysm recognition system via a DSA device and server. A unique two-stage CNN algorithm model automatically identifies and measures the aneurysm, directly outputting the recognition results, including aneurysm size, aspect ratio, and other morphological parameters. Second, the two-stage CNN aneurysm detection algorithm outputs a standard 2D-DSA segmentation result, which serves as the gold standard for correcting the 3D-DSA analysis results. Upon aneurysm detection, 3D-DSA data is generated using 3D DSA. A centerline is generated using the algorithm, and 3D vessel segmentation is performed based on an ellipse. Future results are predicted, compared with the standard 2D-DSA segmentation result, and the optimized 2D / 3D image results are integrated and returned to the 3D vessel segmentation system, ultimately outputting the verified 3D-DSA result.

[0078] For aneurysm identification, 2D-DSA image data was acquired. Considering that the aneurysm contour is generally between 1mm and 40mm, the average number of frames in the DSA sequence was 25. All frames were rescaled to 512×512 pixels with a pixel pitch of 0.391mm, forming a 200×200mm² field of view. Since it is unlikely to find the most prominent aneurysm frame at the beginning or end of the DSA sequence, only 16 frames surrounding the central frame were used. To reduce background noise and enhance vascular signals, the frames in the DSA sequence were averaged together to form a single frame when the artery and aorta were fully fused. The averaged frame was adjusted for background effects caused by motion artifacts using a 3×3 pixel two-dimensional median filter. An intensity threshold was then applied to separate the background from the vascular system. Background pixels were assigned a value of 0, pixels within the vascular system category were assigned a value of 1, and pixels within the aneurysm sac were assigned a value of 2. The entire process of processing each DSA sequence took approximately 2 minutes.

[0079] The CNN architecture utilizes a Visual Geometric Group (VGG-16) encoder, two fully convolutional layers (with rectified linear unit activators), and a decoder consisting of transposed convolutional layers to upsample the image data back to the original input image size. The system assigns this network a pixel-by-pixel semantic segmentation problem, with each pixel divided into three groups: background, blood vessel, or aneurysm. Modeling is performed on an NVIDIA GPU. The network employs an optimizer, an adaptive gradient descent method that adjusts the learning rate over time, ensuring continuous learning even after many epochs. The loss between the network's predicted mask and the benchmark ground truth mask in the training queue after each training epoch is calculated using the sum of dice (DICE) losses (dice similarity coefficients and the complement of binary cross-entropy), guiding gradient descent during training. After network weight optimization, the model is evaluated using a test queue. Network training relies on the accuracy of the benchmark ground truth labels for appropriate weight optimization. Compared to shallower network architectures, the VGG-16 architecture for the network encoder comes at the cost of computational time, requiring approximately 24 hours of training time to achieve model weight convergence. Compared to other segmentation structures such as U-Net, which has more than 31 million adjustable parameters, the network used in this paper has fewer parameters, allowing for faster and less data optimization of network weights.

[0080] To address the technical challenges mentioned above, the aneurysm identification process utilizes a two-stage convolutional neural network (CNN) architecture to automatically detect intracranial aneurysms in 2D-DSA images. The two stages are region identification and aneurysm detection. In the first stage, region identification, the system can locate specific regions to reduce interference from other areas. In the second stage, aneurysm detection, the detection algorithm leverages a false positive suppression algorithm, combining information from anterior and lateral angiography to identify intracranial aneurysms. In this aneurysm identification system, a CNN-based two-stage detection network was developed to automatically detect intracranial aneurysms in 2D-DSA images, such as... Figure 3 As shown, the region identification stage locates the aneurysm region from the 2D-DSA image, and the subsequent aneurysm detection stage identifies the intracranial aneurysm from the output of the region identification.

[0081] In one embodiment, the first-stage region identification based on the CNN model of the 2D-DSA image, locating a specific region in the 2D-DSA image, includes:

[0082] The framework of the CNN model is built on a feature pyramid network, with a residual neural network as its backbone.

[0083] The 2D-DSA image includes a frontal image and a side image. This represents the j-th image in the sequence of frontal images of the i-th patient. This represents the j-th image in the lateral image sequence of the i-th patient. and Correspondingly;

[0084] The frontal image and the side image are respectively input into the feature pyramid network. The target region is determined in the frontal image, and the corresponding region corresponding to the target region is determined in the side image.

[0085] The two-stage CNN framework is based on a feature pyramid network with a residual neural network (ResNet50 network) as its backbone. The 2D-DSA device simultaneously generates two matched sequences for each patient, including frontal and side views of the head. Therefore, the system introduces... Let j be the j-th image in the frontal sequence of the i-th patient, and introduce... To represent the corresponding outer image. The system uses... This represents the target image that the system needs to detect. It is the corresponding image of the same location in another angiographic view. In the region recognition stage, the system inputs the target image and its corresponding image. Then, the region recognition model is applied to obtain the corresponding region image. In the aneurysm detection stage, the images of the two target regions are concatenated through dual inputs and sent to the feature pyramid network for aneurysm detection to detect intracranial aneurysms. Finally, the system employs a region average grayscale suppression algorithm to suppress false positives.

[0086] The high resolution of raw 2D-DSA images introduces additional image interference and processing time, therefore it is essential to minimize the impact of irrelevant structures on image recognition. The system automatically locates specific regions, and the network architecture for the region recognition stage is as follows: Figure 4 As shown in the figure, "Convf×f,c, / s" indicates that the size of the convolutional layer is the filter size f, the number of channels is c, and the stride is s (the default stride is 1). Note that each convolutional layer (Conv layer) is followed by a batch normalization (BN) layer and a Rectified Linear Unit (ReLU) activation layer. "Poolf×f, / s" refers to a max pooling layer with a filter size of f and a stride of s. "Up-sanple, / r" indicates the most recently sampled peripheral upsampling with an upsampling rate of r. "Anchor" is the anchor box used to predict the aneurysm region. The first part receives the input of the three-channel image, and the following feature extraction module is a feature pyramid network with a ResNet50 network backbone. Finally, the anchor box outputs the aneurysm region of the original input image.

[0087] Features are extracted from the input image using a feature pyramid network, which can adapt to various resolutions and extract multi-scale features. ResNet50 is used as the backbone of the feature pyramid network, where each convolutional layer is followed by a batch normalization (BN) layer and a ReLU activation layer. The algorithm then sends the extracted features to anchor boxes to predict the aneurysm region in the original input image. Specifically, the detector may predict a total of 6k parameters. These features are extracted by the feature pyramid network, which is sensitive to objects at various scales. k represents the number of objects, 6 represents a classification label, four spatial boxes (x, y, w, h represent the horizontal and vertical coordinates, width and height, respectively), and a classification confidence. Depending on the algorithm's network architecture, other aneurysm regions can be easily extended and supplemented with labeled data. Finally, based on the resulting bounding boxes, the system's model outputs an image of the aneurysm region. The system then resizes the image to a 288×288 unified interface, easily applying the aneurysm recognition detector to different internal carotid artery regions.

[0088] In one embodiment, after locating a specific region, the identification data obtained by performing a second-stage aneurysm identification on the 2D-DSA image based on a CNN model includes:

[0089] Obtain the target region in the frontal image and the corresponding region in the side image;

[0090] The target region and the corresponding region are connected and fed into the feature pyramid network in a dual input manner to obtain the recognition result data;

[0091] The identification result data includes aneurysm labels in the images, which are used to indicate the confidence level that a certain location in the image is an aneurysm.

[0092] The second stage is the aneurysm identification stage, which is a method to distinguish intracranial aneurysms from overlapping blood vessels. The overall framework for aneurysm identification is as follows: Figure 5 As shown, the system architecture mainly consists of four steps: dual input, feature extraction, result output, and region-averaged gray-scale suppression. Dual input involves concatenating the target aneurysm region image and its corresponding aneurysm region image as input variables. The features are extracted by a feature pyramid network, which is sensitive to objects of various scales. Then, the anchor box outputs the detector's results. Similar to the first-stage region recognition, the detector may predict a total of 6k parameters, and the output step retains objects with a confidence level higher than 0.6. Finally, because overlapping blood vessels can affect the detection results, a region-averaged gray-scale suppression algorithm is applied to predict objects labeled as aneurysms, but with a confidence level below a preset threshold.

[0093] exist Figure 5 In the aneurysm recognition algorithm, "t×" in curly braces indicates that these layers are repeated t times. A dual-input layer concatenates the target aneurysm region image and the corresponding aneurysm region image, followed by a feature pyramid network applied to aneurysms of various scales. Then, the anchor box outputs the prediction result, following a region-averaged grayscale suppression algorithm to mitigate false positives.

[0094] The dual-input design primarily addresses the issues of information loss and image overlap recognition. Clinically, even experienced neurosurgeons struggle to distinguish between intracranial aneurysms and overlapping vascular regions. By providing both frontal and lateral DSA images, the system can correctly infer whether it's an aneurysm or a vascular overlap based on vascular trends. The basic input architecture of this recognition algorithm combines information from the frontal and lateral DSA views, concatenating the target aneurysm image with the corresponding aneurysm image output from the region recognition stage as input variables for the aneurysm recognition stage. Through this concatenation, the system places the target image first, followed by the corresponding image. Therefore, the neural network can utilize the dual-input information to partially compensate for the loss of 2D-DSA modality information during feature extraction.

[0095] The dual-input model improves accuracy and stability to some extent, but overlapping blood vessels can still easily be mistaken for aneurysms. Therefore, the system needs an algorithm to suppress false positives. According to DSA principles, the grayscale of an image is determined by the density of the contrast agent. The grayscale of a pixel in a 2D-DSA image depends on the density of the radioactive contrast agent in the blood within the corresponding 3D space of the pixel. Compared to the surrounding area, areas of overlapping blood vessels correspond to more blood, which may result in a darker grayscale. Large aneurysms typically have higher grayscale because they contain a large amount of contrast agent, but they are easily identified using computer-aided methods with a confidence level close to 1.0. In contrast, small aneurysms, such as hemangiomas, are typically smaller than the diameter of the surrounding veins, and therefore have a lighter grayscale.

[0096] In one embodiment, after obtaining the recognition result data, performing region average grayscale suppression processing on the recognition result data includes:

[0097] The confidence level of the aneurysm is compared with the standard threshold. If the confidence level of the aneurysm is lower than the standard, the region average gray level suppression algorithm is applied.

[0098] An adaptive grayscale threshold selection method is used to coarsely extract the vascular model from the original target aneurysm region image;

[0099] Based on the vascular model and the original aneurysm region image, the blood vessels in the target aneurysm region image are calculated logically.

[0100] Among them, the gray value of the aneurysm is G 动脉瘤 The pixel value is P 动脉瘤 The average gray value of the aneurysm is:

[0101] AG 动脉瘤 =G 动脉瘤 / P 动脉瘤

[0102] The average gray value of the vessels surrounding the aneurysm is:

[0103] AG 动脉瘤周边 =(G 扩大区域 -G 动脉瘤 ) / (P 扩大区域 -P 动脉瘤 )

[0104] If, AG 动脉瘤 Less than AG 动脉瘤周边 If the image of that region is an overlapping blood vessel, then if AG 动脉瘤 Greater than AG 动脉瘤周边 If so, the region image is an aneurysm.

[0105] A region-averaged gray-level suppression algorithm is used to evaluate outputs with aneurysm classification labels but confidence levels below a standard. The system compares the predicted aneurysm confidence level with a standard threshold to determine whether to retain the aneurysm label. If the aneurysm confidence level is below the standard, the region-averaged gray-level suppression algorithm is applied. First, an adaptive gray-level threshold selection method is used to coarsely extract the vascular model of the original target aneurysm region image. Then, the vascular model and the original aneurysm region image are used to logically compute the vessels of the target aneurysm region image. By defining the region, the system defines the aneurysm gray-level value as G. 动脉瘤 The pixel value is defined as P. 动脉瘤 The average gray value of the aneurysm is:

[0106] AG 动脉瘤 =G 动脉瘤 / P 动脉瘤

[0107] Meanwhile, the average grayscale value of the peripheral blood vessels is:

[0108] AG 动脉瘤周边 =(G 扩大区域 -G 动脉瘤 ) / (P 扩大区域 -P 动脉瘤 )

[0109] Finally, compare AG 动脉瘤 and AG 动脉瘤周边 Numerical cases. If AG 动脉瘤 Compared to AG 动脉瘤周边 The aneurysm should be small, meaning its grayscale should be darker than surrounding vessels. This is because it allows the system to identify that the object is an overlapping vessel. Otherwise, the system should retain the object as an aneurysm.

[0110] The loss formula for object classification is defined as follows:

[0111] FL(p t )=-α(1-p t )γlog(p t )

[0112] In the formula, α is a balancing parameter, and γ is the weight decrease rate. Additionally, p... t According to the following definition:

[0113]

[0114] Where p represents the prediction confidence level, and y∈{±1} is defined as the basic fact class. For bounding box regression, the system uses the following loss formula:

[0115] SL(t,v)=∑ i∈{x,y,w,h} Smooth L 1(ti -v i ),

[0116] in:

[0117]

[0118] Where t is the bounding box of the predicted object, v represents the bounding box of the basic facts, and σ is a tradeoff factor. The first metric in the region identification stage is accuracy, which can be defined as the number of images from which the network finds aneurysm regions in the test data. Typically, aneurysm identification focuses on image intersections, and slight movements or size changes of the bounding boxes have a relatively small impact on actual detection. Therefore, a regularized offset relative to manually labeled data is proposed to evaluate the performance of region identification, defined as follows:

[0119]

[0120]

[0121] f x Represents the regular offset along the X-axis. annotation The x-coordinate representing the manually labeled centroid, x prdiction The X-coordinate represents the centroid predicted by the region recognition algorithm. The same applies to the Y-axis. W and H represent the width and height of the bounding box, respectively.

[0122] Furthermore, a five-fold cross-validation scheme was employed in the aneurysm recognition algorithm. To compare performance, five dual-input and unidirectional-input models were trained using the same dataset. The aneurysm recognition system used average precision to evaluate the classification accuracy of aneurysms and overlapping blood vessels. For bounding box regression, Smooth L1 loss was used for analysis and evaluation. Moreover, the clinical purpose of the aneurysm recognition system is to provide physicians with accurate indications of whether a patient has an aneurysm. To avoid missed diagnoses, strict evaluation metrics were defined as follows:

[0123]

[0124] Among them, Diag i This represents whether the i-th patient has an aneurysm. When the j-th image of the i-th patient is predicted to have an aneurysm from the V angle (F or L), This is equivalent to a true positive. In other words, the formula means that even if a single frontal or lateral image with an aneurysm is detected in the entire DSA sequence, the patient will be defined as having an aneurysm. Furthermore, a confusion matrix and receiver operating characteristic curves were used in the assessment. The system calculated true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) to calculate sensitivity, specificity, and accuracy, defined as follows:

[0125]

[0126]

[0127]

[0128] In one embodiment, the CNN model is trained through the following steps:

[0129] Acquire the patient's 2D-DSA image data, which includes one or more 2D-DSA images;

[0130] Images below a preset value are removed from the 2D-DSA images to obtain 2D-DSA images that meet the preset value.

[0131] The 2D-DSA images that meet the preset values ​​are divided into training datasets and test datasets based on a preset ratio.

[0132] CNN algorithm models require large datasets to optimize the weights of each network layer. The system uses a data augmentation scheme created in Python to increase the size of the training dataset. A combination of rotation and scaling settings is used to expand both the training and test sets. Each frame of the DSA sequence is rotated by an average of 90°, 180°, and 270°, and enlarged to 200% of its original size. The augmentation process expands the training set to 1500 examples. Enlarging the image data regions during augmentation generates new true negative examples, improving the generalization ability of the CNN algorithm network in the absence of aneurysms.

[0133] In one embodiment, after dividing the 2D-DSA images that meet the preset values ​​into training and test datasets based on a preset ratio, the method further includes:

[0134] The multiple 2D-DSA images constitute a DSA image sequence;

[0135] The DSA image is locally normalized so that the local mean value of the DSA image is 0 and the SD value is 1.

[0136] A large amount of 2D DSA image sequence data was used to train a two-stage CNN aneurysm recognition algorithm. The entire training process is divided into two steps, such as... Figure 6As shown. The first step is image quality control, removing DSA image sequences with poor image quality; the second step is to divide the data into training and testing datasets proportionally. Before inputting into the CNN network, the pixel intensity of the DSA sequences is locally normalized by setting the local mean to 0 and the SD to 1. Next, the proposed network is trained and validated using patients and DSA sequences, and its accuracy is tested on patients and DSA sequences. During network training, the input frames and corresponding ground truth masks are cropped to 256×256. When testing on the test set, the system inputs the original 512×512 frames, and the network runs in fully convolutional mode.

[0137] This invention also provides a DSA image recognition device, such as... Figure 7 As shown, it includes:

[0138] A 2D-DSA image receiving module is used to receive 2D-DSA image data, wherein the 2D-DSA image data includes one or more 2D-DSA images;

[0139] The recognition module is used to perform a first-stage region recognition on the 2D-DSA image based on a CNN model, locate a specific region in the 2D-DSA image, and after locating the specific region, perform a second-stage aneurysm recognition on the 2D-DSA image based on a CNN model to obtain recognition result data.

[0140] An output module is used to output the recognition result data, wherein the recognition result data includes aneurysm labels that are marked on the 2D-DSA image.

[0141] In one embodiment, the identification module includes a one-stage identification unit, which is used to perform the following steps:

[0142] The framework of the CNN model is built on a feature pyramid network, with a residual neural network as its backbone.

[0143] The 2D-DSA image includes a frontal image and a side image. This represents the j-th image in the sequence of frontal images of the i-th patient. This represents the j-th image in the lateral image sequence of the i-th patient. and Correspondingly;

[0144] The frontal image and the side image are respectively input into the feature pyramid network. The target region is determined in the frontal image, and the corresponding region corresponding to the target region is determined in the side image.

[0145] In one embodiment, the identification module includes a two-stage identification unit, which is used to perform the following steps:

[0146] Obtain the target region in the frontal image and the corresponding region in the side image;

[0147] The target region and the corresponding region are connected and fed into the feature pyramid network in a dual input manner to obtain the recognition result data;

[0148] The identification result data includes aneurysm labels in the images, which are used to indicate the confidence level that a certain location in the image is an aneurysm.

[0149] The identification results data include 2D-DSA segmentation results, which is considered the gold standard. After outputting the identification results data, physicians need to obtain more information through 3D-DSA. In modern neuroradiology, 3D rotational angiography (3D-DSA) is a technique that can both visualize complex cerebrovascular lesions and guide interventional treatment. Increasingly, 3D-DSA data is also being used for further quantitative analysis to support treatment planning and procedures for patients with cerebrovascular diseases. This analysis requires reliable three-dimensional vessel segmentation methods to accurately depict vessel boundaries as precisely as possible to obtain accurate vessel measurements, such as vessel diameter, bifurcation angles, and aneurysm dome size. The segmentation results and their corresponding mesh representations form the basis for more reliable analysis. Small variations in the segmentation results can lead to entirely different flow patterns or wall shear stress distributions. Therefore, vessel segmentation results become increasingly important for post-processing applications in clinical settings. Accurate vessel segmentation based on 3D angiography is challenging because it depends on the quality of the 3D-DSA images, which can vary from patient to patient. Many factors can affect image quality, including the amount of contrast agent injected, the timing of injection, the hemodynamic mixing of the contrast agent, and factors such as blood flow and cardiac output. Reconstruction parameters may also differ between patients. Edge gradations between the background and vessel intensity vary depending on the applied reconstruction. These contours provide an opportunity for different segmentation methods to interpret these edge gradations, which may lead to variations in vessel boundary positions, in different ways. Contrast agent may leak out under different injection conditions, resulting in localized blurring of vessel boundaries. Even with identical acquisition parameters, the final 3D-DSA image volume will vary slightly due to inhomogeneous hemodynamic mixing of the contrast agent and blood. Therefore, validating 3D-DSA vessel segmentation results becomes quite difficult, and 2D-DSA imaging remains considered the gold standard for quantitative assessment and measurement.

[0150] This invention also provides a DSA image recognition method, such as... Figure 8 As shown, it includes:

[0151] Step S210: Receive 3D-DSA image data, wherein the 3D-DSA image data includes one or more 3D-DSA images;

[0152] Step S220: Obtain the center point of the 3D-DSA image, and generate the center line of the 3D-DSA image based on the center point;

[0153] Step S230: Segment the 3D-DSA image based on the centerline using an ellipsoidal blood vessel model to obtain a segmented 3D-DSA image;

[0154] Step S240: Receive 2D-DSA gold standard segmentation result data, and verify the 3D-DSA image based on the 2D-DSA gold standard segmentation result data to obtain verified 3D-DSA image data.

[0155] To calibrate 3D-DSA images, 3D-DSA image information is first collected, then centerlines are generated, and ellipsoidal vessel models are used for 3D segmentation. To evaluate the accuracy of 3D vessel segmentation, its forward projection is iteratively superimposed with the corresponding 2D DSA projection. Local vessel differences are modeled using a global 2D / 3D optimization function to adjust the correspondence between the 3D vessel segmentation results and the 2D vessel contours. Validating any 3D vessel segmentation method is challenging; to address this, this invention employs a novel 3D vessel segmentation method based on additional 2D vessel information. 2D-DSA acquisition serves as the validation basis for 3D segmentation and is also the driving force for adapting the initial 3D segmentation. Each 3D-DSA dataset is automatically accompanied by 133 2D-DSA projections, displaying the patient's organs from different perspectives. Furthermore, these projections are registered and calibrated using the 3D-DSA dataset. The advantage of displaying selected 2D-DSA images is that vessel boundaries are clearly visible, rather than blurred as often occurs in 3D-DSA images. The innovation of this invention lies in the fact that an automatic segmentation algorithm formalizes a new global 2D / 3D optimization function that smoothly combines 2D vascular information with 3D images.

[0156] In one embodiment, obtaining the center point of the 3D-DSA image and generating the center line of the 3D-DSA image based on the center point includes:

[0157] Select the bottom slice of the 3D-DSA image and perform a circular Hough transform on the bottom slice to locate the blood vessel with the largest diameter. The center of the circle with the largest diameter is the center point.

[0158] After the 3D DSA image data is input into the system, the main purpose of the system's 3D-DSA segmentation method is to detect and segment large and medium-sized blood vessels in the 3D-DSA dataset, providing input for subsequent adaptive 2D-DSA information. 3D-DSA is based on direct intra-arterial injection of contrast agent. Due to high contrast and subtraction, it can be assumed a priori that blood vessels appear in the highest intensity bin of the histogram. Therefore, the 3D-DSA dataset consists of three different types of intensity levels: foreground (blood vessels), artifacts, and background. Blood vessels in the 3D-DSA dataset enter the image volume from the bottom, due to the prior known acquisition settings, the anatomical structure of the cerebral vascular tree, and the patient's position on the C-arm system workbench. Therefore, the contrast agent flows from bottom to top. The location of the largest blood vessel entering the volume is automatically found. A circular Hough transform is applied to the bottom slice of the 3D-DSA image to locate the blood vessel with the largest diameter, and this location is then used as the seed point for the centerline calculation. Figure 9 The example shown illustrates the centerline calculation results for blood vessels. This seed point initialization is limited to a selected specific volume used for secondary reconstruction. After centerline calculation, each blood vessel branch is associated with a centerline. All centerline sections are stored in a tree structure, with the centerline of the blood vessel lumen entering the volume serving as the root in the tree structure. This structure readily reveals the parent-child relationships between different blood vessel branches. The blood vessel centerline tree is used as the initialization for ellipsoid-based blood vessel segmentation.

[0159] In one embodiment, segmenting the 3D-DSA image based on the centerline using an ellipsoidal blood vessel model to obtain the segmented 3D-DSA image includes:

[0160] An ellipsoidal blood vessel model is constructed, represented by the following formula:

[0161]

[0162] Where, x∈R 3 Represents a point on the surface of a sphere;

[0163] Nine transformation parameters are set, including three rotation parameters, three scaling parameters, and three translation parameters. The transformation function is expressed by the following formula:

[0164] T(m,x)=R(r)S(s)x+t

[0165] Where R represents a rotation parameter, r∈R 3 A 3×3 rotation matrix, with scaling parameter s∈R 3 In a 3×3 diagonal matrix S(s), t∈R 3 Represents the translation vector;

[0166] All transformation parameters are in vector m = (r, s, t) ∈ R 9 The inner summation and intrinsic shape model can be rewritten as:

[0167] f T (m,x):=f(T(m,x))=f(R(r)S(s)x+t)

[0168] The space is divided into three distinct areas:

[0169]

[0170] Where S(m) and I(m) represent the surface and interior of the ellipsoid, respectively, and O represents the image domain.

[0171] Hyperellipsoids, ellipsoids, or spheres are well-suited for describing the geometry of local vascular segments. The coordinates of a unit sphere located at its coordinate center are defined as follows:

[0172]

[0173] Where, x∈R 3 Let represent a point on the sphere. To approximate a local blood vessel segment, a coordinate transformation must be applied to allow for arbitrary rotation, scaling, and translation; that is, nine transformation parameters must be estimated (three for rotation, three for scaling, and three for translation). The entire transformation function is expressed as:

[0174] T(m,x)=R(r)S(s)x+t

[0175] Where R represents a rotation parameter, r∈R 3 A 3×3 rotation matrix. Scale parameter s∈R 3 In a 3×3 diagonal matrix S(s), t∈R 3 Let m represent the translation vector. All transformation parameters are in the vector m = (r, s, t) ∈ R. 9 Inner summation. Therefore, the intrinsic shape model can be rewritten as:

[0176] f T (m,x):=f(T(m,x))=f(R(r)S(s)x+t)

[0177] The space is divided into three distinct areas:

[0178]

[0179] Where S(m) and I(m) represent the surface and interior of the ellipsoid, respectively, and O represents the image domain. These parameters are used as extrinsic parameters in the global objective function to accurately estimate the parameters. The previously calculated centerline is used to initialize the parameters, i.e., the center of the ellipsoid is placed on the centerline. The centerline is rotated so that its tangent vector coincides with the local Z-axis of the ellipsoid's interior coordinate system. The curvature of the centerline is used to determine the number of ellipsoids required to approximate vascular branches. In a high-curvature centerline section, the number of ellipsoids increases, while the proportion of ellipsoids in the Z-direction decreases simultaneously.

[0180] In one embodiment, segmenting the 3D-DSA image based on the centerline using an ellipsoidal blood vessel model to obtain the segmented 3D-DSA image includes:

[0181] Rotate the centerline so that its tangent vector coincides with the local Z-axis of the coordinate system inside the ellipsoid;

[0182] The parameter vector of the ellipsoidal blood vessel model is k = (m0, m1, ..., m M-1 )∈R 9×M This represents all the ellipsoidal parameters of the entire model, where M is the total number of ellipsoids used.

[0183] Each ellipsoidal local image domain is divided into foreground intensity and background intensity, and a specific local intensity threshold is obtained. The threshold is stored in the vector h = (h0, h1, ..., h...). M-1 ) T ∈R M middle.

[0184] The ellipsoidal scale perpendicular to the blood vessel direction is initialized using the cross-sectional radius intensity profile. The parameter vector k = (m0, m1, ..., m M-1 )∈R 9×M This represents all ellipsoidal parameters for the entire model, and M represents the total number of ellipsoids used. For each ellipsoid, the local image domain is divided into foreground and background intensities, and a specific local intensity threshold is estimated. The threshold is stored in the vector h = (h0, h1, ..., h...). M-1 ) T ∈R M middle.

[0185] In one embodiment, the initialized ellipsoidal tube model is optimized so that it evolves towards the vessel boundary according to a predefined energy, with the total energy formula G. tot (k,h) consists of two terms:

[0186] G tot (k,h):R 9×M ×R M →R

[0187] Among them G ext(k,h) and G int (k) represent the external energy term and the internal energy term, respectively, and α defines the weighting factor.

[0188] 6. The DSA image recognition method according to claim 5, characterized in that,

[0189] The inherent properties of the internal energy term are related to the ellipsoid. The internal energy term consists of three cubic curves to ensure the smoothness of rotation, scaling, and translation between adjacent ellipsoids, as expressed by the following formula:

[0190]

[0191] In the formula, B i (l,k)′ and B i (l,k)′′ represent B respectively i First and second derivatives, L i is the total length of the B-spline curve, and l is the position on the spline curve.

[0192] The initial ellipsoidal tube model has now been optimized, causing it to evolve toward the vessel boundary based on a predefined energy. The total energy formula is G. tot (k,h) consists of two terms:

[0193] G tot (k,h):R 9×M ×R M →R

[0194] Among them G ext (k,h) and G int (k) represent the external energy term and the internal energy term, respectively, and α defines the weighting factor.

[0195] The inherent properties of the internal energy term are related to the ellipsoid, and the parameters between two subsequent ellipsoids must be modeled in a smooth manner. Therefore, the internal energy term consists of three cubic curves to ensure smoothness in rotation, scaling, and translation between adjacent ellipsoids:

[0196]

[0197] In the formula, B i (l,k)′ and B i (l,k)′′ represent B respectively i First and second derivatives, L i is the total length of the B-spline curve, and l is the position on the spline curve.

[0198] In one embodiment, the external energy term drives the model of the surrounding vascular structure, and the ellipsoidal vascular shape model divides the image domain into foreground voxels and background voxels. The foreground voxels and background voxels are calculated by the set S(k) i) and I(k i The number of foreground and background voxel intensities is obtained, where the formula for the external energy term is:

[0199]

[0200] The initialization of the ellipsoid parameters and the local density ensure that the number of foreground voxels is always greater than the number of background voxels.

[0201] The external energy term drives the model of the surrounding vascular structure. The system's ellipsoidal vascular shape model attempts to segment the image domain into foreground and background voxels as best as possible. These two parts are determined by calculating the set S(k i ) and I(k i The external energy term is evaluated based on the number of foreground and background voxel intensities. The formula for the external energy term is:

[0202]

[0203] The initialization of the ellipsoid parameters and local density ensures that the number of foreground voxels is always greater than the number of background voxels. A local intensity threshold, centered at the current ellipsoid center, is set within a box with dimensions between 32×32×32 and 64×64×64, depending on the scale of the current blood vessel. The objective is to minimize an equation that perfectly matches the ellipsoid parameters to the local vascular structure. This local foreground / background estimation ensures that the external energy adapts to vascular structures exhibiting even high intensity variations across the entire dataset. This is achieved by minimizing the total energy formula G. tot The optimal ellipsoidal parameters are obtained by using (k,h) and local intensity thresholds. Since the system model is initialized by pre-calculated centerline and cross-sectional radius profile estimates, the initial parameter set is at the global optimum.

[0204] In one embodiment, receiving 2D-DSA gold standard segmentation result data and verifying the 3D-DSA image based on the 2D-DSA gold standard segmentation result data to obtain verified 3D-DSA image data includes:

[0205] The difference between 2D-DSA gold standard segmentation results and 3D-DSA segmentation images is measured using the 2D / 3D external energy function Gext2D / 3D(k). The formula for the 2D / 3D external energy function Gext2D / 3D(k) is as follows:

[0206]

[0207] In the formula, Q represents the number of projected images used to adjust the 3D segmentation result, w(i, u) represents the intensity value of the two-dimensional fundamental truth of the projected image at position u = (x1, x2) ∈ R, and the intensity value of the forward projection v(i, u, k) is obtained from the maximum intensity projection.

[0208] The overlap between the forward projection of the 3D-DSA segmentation image and the gold standard of the 2D DSA image is evaluated using two metrics: the Dice coefficient and accuracy. The Dice coefficient is:

[0209]

[0210] Where X represents all pixels segmented by the gold standard, and Y represents all pixels projected in orthographic projection;

[0211] The formula for defining accuracy is as follows:

[0212]

[0213] Where TP represents the number of true positive cases and FP represents the number of false positive cases.

[0214] The emergence of 3D-DSA volumetric data depends on four influencing factors: contrast agent dosage, injection time, hemodynamic mixture of contrast agent, and the applied reconstructive kernel. 3D-DSA segmentation results are compared with 2D-DSA segmentation results. The 3D segmentation results are validated by overlay matching using 2D-DSA image analysis results within the system; this comparison is achieved by using the orthographic projection of the 3D vascular segmentation using ray projection. The local differences between 2D vascular information and the orthographic projection are now formalized into a new objective function that utilizes 3D parametric segmentation. This derives a new 2D / 3D external energy function, Gext2D / 3D(k), which measures the difference between 2D and 3D segmentations. It is defined as follows:

[0215]

[0216] In the formula, Q represents the number of projected images used to adjust the 3D segmentation result, and w(i, u) represents the intensity value of the 2D fundamental truth of the projected image at position u = (x1, x2) ∈ R. The intensity value of the forward projection v(i, u, k) is given by the maximum intensity projection. The new 2D / 3D external energy term forms a new total energy. Therefore, the internal energy term ensures that the 3D ellipsoidal model will not be distorted into non-compliant blood vessel distortion. The 3D blood vessel segmentation method is validated through systematic 2D / 3D image optimization, and finally, the validated 3D-DSA image result is output.

[0217] The overlap between the forward projection of the 3D-DSA segmentation result and the gold standard 2D DSA image is evaluated using two metrics: the Dice coefficient and accuracy. The Dice coefficient is defined as:

[0218]

[0219] Where X represents all pixels segmented by the gold standard, and Y represents all pixels projected in orthographic projection.

[0220] The formula for defining accuracy is as follows:

[0221]

[0222] Where TP represents the number of true positive cases and FP represents the number of false positive cases.

[0223] Embodiments of the present invention also provide a DSA image recognition device, such as... Figure 10 As shown, it includes:

[0224] A 3D-DSA image receiving module is used to receive 3D-DSA image data, wherein the 3D-DSA image data includes one or more 3D-DSA images;

[0225] The centerline acquisition module is used to acquire the center point of the 3D-DSA image and generate the centerline of the 3D-DSA image based on the center point;

[0226] The segmentation module is used to segment the 3D-DSA image based on the centerline using an ellipsoidal blood vessel model to obtain a segmented 3D-DSA image;

[0227] The verification module is used to receive 2D-DSA gold standard segmentation result data, verify the 3D-DSA image based on the 2D-DSA gold standard segmentation result data, and obtain verified 3D-DSA image data.

[0228] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0229] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0230] In the above-described terminal or server embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0231] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for DSA image recognition, characterized in that, include: Receive 2D-DSA image data, wherein the 2D-DSA image data includes one or more 2D-DSA images; The first stage of region recognition is performed on the 2D-DSA image based on the CNN model to locate specific regions in the 2D-DSA image. After locating the specific regions, the second stage of aneurysm recognition is performed on the 2D-DSA image based on the CNN model to obtain the recognition result data. The identification result data is output, wherein the identification result data includes aneurysm tags labeled on the 2D-DSA image; The recognition result data is subjected to region average grayscale suppression processing, including: The confidence level of the aneurysm is compared with the standard threshold. If the confidence level of the aneurysm is lower than the standard, the region average gray level suppression algorithm is applied. An adaptive grayscale threshold selection method is used to coarsely extract the vascular model from the original target aneurysm region image; Based on the vascular model and the original aneurysm region image, the blood vessels in the target aneurysm region image are calculated logically. Among them, the gray value of aneurysm is Pixel value The average gray value of the aneurysm is: The average gray value of the vessels surrounding the aneurysm is: like, Less than If so, the image of that region consists of overlapping blood vessels. Greater than If so, the region image is an aneurysm.

2. The DSA image recognition method according to claim 1, characterized in that, The first-stage region recognition based on the CNN model of the 2D-DSA image, locating specific regions in the 2D-DSA image, includes: The framework of the CNN model is built on a feature pyramid network, with a residual neural network as its backbone. The 2D-DSA image includes a frontal image and a side image. Indicates the first For example, the j-th image in a sequence of frontal images of a patient. Indicates the first For example, the j-th image in the patient's lateral image sequence, the and Correspondingly; The frontal image and the side image are respectively input into the feature pyramid network. The target region is determined in the frontal image, and the corresponding region corresponding to the target region is determined in the side image.

3. The DSA image recognition method according to claim 2, characterized in that, The second-stage aneurysm identification data obtained by performing a CNN model on the 2D-DSA image after locating a specific region includes: Obtain the target region in the frontal image and the corresponding region in the side image; The target region and the corresponding region are connected and fed into the feature pyramid network in a dual input manner to obtain the recognition result data; The identification result data includes aneurysm tags that label the 2D-DSA image. The aneurysm tags are used to indicate the confidence level that a certain location in the image is an aneurysm.

4. The DSA image recognition method according to claim 1, characterized in that, The CNN model is trained through the following steps: Acquire the patient's 2D-DSA image data, which includes one or more 2D-DSA images; Images below a preset value are removed from the 2D-DSA images to obtain 2D-DSA images that meet the preset value. The 2D-DSA images that meet the preset values ​​are divided into training datasets and test datasets based on a preset ratio.

5. The DSA image recognition method according to claim 4, characterized in that, After dividing the 2D-DSA images that meet the preset values ​​into training and test datasets based on a preset ratio, the process further includes: The multiple 2D-DSA images constitute a DSA image sequence; The DSA image is locally normalized so that the local mean value of the DSA image is 0 and the SD value is 1.

6. A DSA image recognition device, characterized in that, include: A 2D-DSA image receiving module is used to receive 2D-DSA image data, wherein the 2D-DSA image data includes one or more 2D-DSA images; The recognition module is used to perform a first-stage region recognition on the 2D-DSA image based on a CNN model, locate a specific region in the 2D-DSA image, and after locating the specific region, perform a second-stage aneurysm recognition on the 2D-DSA image based on a CNN model to obtain recognition result data. The output module is used to output the recognition result data, wherein the recognition result data includes aneurysm labels marked on the 2D-DSA image; The module for performing region-average grayscale suppression processing on the recognition result data is used for: The confidence level of the aneurysm is compared with the standard threshold. If the confidence level of the aneurysm is lower than the standard, the region average gray level suppression algorithm is applied. An adaptive grayscale threshold selection method is used to coarsely extract the vascular model from the original target aneurysm region image; Based on the vascular model and the original aneurysm region image, the blood vessels in the target aneurysm region image are calculated logically. Among them, the gray value of aneurysm is Pixel value The average gray value of the aneurysm is: The average gray value of the vessels surrounding the aneurysm is: like, Less than If so, the image of that region consists of overlapping blood vessels. Greater than If so, the region image is an aneurysm.

7. The DSA image recognition device according to claim 6, characterized in that, The identification module includes a one-stage identification unit, which is used to perform the following steps: The framework of the CNN model is built on a feature pyramid network, with a residual neural network as its backbone. The 2D-DSA image includes a frontal image and a side image. Indicates the first For example, the j-th image in a sequence of frontal images of a patient. Indicates the first For example, the j-th image in the patient's lateral image sequence, the and Correspondingly; The frontal image and the side image are respectively input into the feature pyramid network. The target region is determined in the frontal image, and the corresponding region corresponding to the target region is determined in the side image.

8. The DSA image recognition device according to claim 7, characterized in that, The identification module includes a two-stage identification unit, which is used to perform the following steps: Obtain the target region in the frontal image and the corresponding region in the side image; The target region and the corresponding region are connected and fed into the feature pyramid network in a dual input manner to obtain the recognition result data; The identification result data includes aneurysm tags that label the 2D-DSA image. The aneurysm tags are used to indicate the confidence level that a certain location in the image is an aneurysm.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, is used to implement the method described in any one of claims 1 to 5.

Images