Deep learning-based vascular embolization recognition method and related device

Abstract

The application relates to a deep learning-based vascular embolism identification method, which comprises the following steps: obtaining seed points of a to-be-identified abdominal medical image according to an abdominal aorta blood vessel image and a trained first model; performing semi-automatic segmentation on a mesenteric artery region in the image by using a GrowCut algorithm on the seed points; extracting the image of the mesenteric artery region; inputting the image of the mesenteric artery region into a VTK platform; demarcating an arterial blood vessel center line of the image of the mesenteric artery region; inputting the image of the mesenteric artery region with the demarcated arterial blood vessel center line into a second model; and determining whether a thrombus exists in the to-be-identified abdominal medical image and the position of the thrombus. The application can more comprehensively capture influencing factors of acute superior mesenteric artery thrombosis, screen out relevant strong predictors, accurately identify the superior mesenteric artery thrombosis on a CT image, and enhance the prediction ability of the model. The application also relates to an equipment and a storage medium.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a vascular embolism recognition method based on deep learning and related equipment. Background Art

[0002] Existing methods for CT images of single-component thrombi often use grayscale characteristics as a starting point. By traversing the area surrounding the pixel, the researchers search for patterns in pixel value distribution to identify the lesion's edge. This is a binary classification task. Current image classification methods based on statistical learning are unable to fully capture the factors influencing superior mesenteric artery thrombosis in CT images, making it difficult to accurately identify superior mesenteric artery thrombosis in CT images. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a vascular embolism identification method and related equipment based on deep learning.

[0004] In a first aspect, the present invention provides a method for identifying vascular embolism based on deep learning, the method comprising:

[0005] Extracting a region image from an abdominal medical image to be identified, wherein the region image is an abdominal aorta blood vessel image;

[0006] Obtaining seed points of the abdominal medical image to be identified based on the abdominal aorta vascular image and a trained first model, wherein the first model is constructed based on a Unet network and trained using a set of abdominal medical sample images, and the seed points are pixels of the abdominal aorta in the medical image to be identified;

[0007] Semi-automatically segmenting the mesenteric artery region in the regional image using the GrowCut algorithm using the seed points to extract an image of the mesenteric artery region;

[0008] Delineating an arterial centerline of the image of the mesenteric artery region using a VTK platform, wherein the VTK platform is an application for three-dimensional computer graphics;

[0009] The image of the mesenteric artery region with the arterial centerline delineated is input into the second model to determine whether there is a thrombus and the location of the thrombus in the abdominal medical image to be identified. The second model is established based on the SVN classifier and is used to determine whether there is a thrombus based on the normal vascular image characteristics and the thrombus vascular image characteristics.

[0010] Based on the above embodiment, obtaining the seed point of the to-be-identified abdominal medical image according to the abdominal aorta vascular image and the trained first model specifically includes:

[0011] inputting the abdominal aorta blood vessel image into the trained first model to obtain an abdominal aorta corresponding pixel point in the abdominal aorta blood vessel image;

[0012] determining a seed point of the to-be-identified abdominal medical image based on the abdominal aorta corresponding pixel point in the abdominal aorta blood vessel image.

[0013] Based on the above embodiments, the seed point of the to-be-identified abdominal medical image is determined based on the abdominal aorta corresponding pixel point in the abdominal aorta blood vessel image, specifically including:

[0014] determining a sliding window range according to the positions of the pixel points in the abdominal aorta blood vessel image;

[0015] traversing the pixel points in the sliding window range to obtain an average value of the pixel values in the sliding window range;

[0016] determining an original seed point in the abdominal aorta blood vessel image according to the average values of the pixel values in all the sliding window ranges;

[0017] calculating the weighted distances between the pixel points in the abdominal aorta blood vessel image and the original seed point;

[0018] determining the seed point of the to-be-identified abdominal medical image according to the pixel values of the pixel points in the abdominal aorta blood vessel image, and the weighted distances, the maximum weighted distance and the minimum weighted distance between the pixel points and the original seed point.

[0019] Based on the above embodiments, the seed point is used to semi-automatically segment the mesenteric artery region in the region image by using the GrowCut algorithm, and the image of the mesenteric artery region is extracted, specifically including:

[0020] S1, marking the seed point of the abdominal aorta blood vessel image as a foreground point, and marking other pixel points as background points;

[0021] S2, obtaining the neighborhood points of each foreground point, and calculating the similarity values between the foreground points and the neighborhood points, wherein the similarity values include calculating the color similarity, gray value similarity and spatial position similarity between the foreground points and the neighborhood points;

[0022] S3, expanding the neighborhood points with a similarity value greater than a preset threshold value to the foreground points, and returning to step S2;

[0023] S4, if no neighborhood point is expanded to a foreground point in this round of iteration, extracting all the foreground points in the abdominal aorta blood vessel image to obtain a foreground image, and the remaining pixel points are background images, wherein the foreground image is the image of the mesenteric artery region.

[0024] Based on the above embodiment, the image of the mesenteric artery region is extracted using the VTK platform, and the arterial vessel centerline of the image of the mesenteric artery region is drawn, wherein the VTK platform is an application program for three-dimensional computer graphics, specifically comprising:

[0025] The CSV file of the image of the mesenteric artery region is extracted, and the RAS coordinates of the termination point of the superior mesenteric artery blood vessel in the CSV file are converted into data objects in the VTK platform, wherein the RAS coordinates are right upper front coordinates used in the medical field;

[0026] The RAS coordinates of the termination point of the superior mesenteric artery blood vessel are converted into data objects in the VTK platform;

[0027] The polygon data object of the superior mesenteric artery blood vessel in the VTK platform and the data object of the termination point of the superior mesenteric artery blood vessel in the VTK platform are used to calculate the arterial vessel centerline.

[0028] Based on the above embodiment, the method further comprises:

[0029] Obtaining the sample image of the mesenteric artery region in which the arterial vessel centerline is drawn from the abdominal medical sample image set;

[0030] Extracting the local binary pattern feature vector of the sample image of the mesenteric artery region in which the arterial vessel centerline is drawn;

[0031] Extracting the gray level co-occurrence matrix feature vector of the sample image of the mesenteric artery region in which the arterial vessel centerline is drawn;

[0032] Combining the local binary pattern feature vector and the gray level co-occurrence matrix feature vector to obtain a comprehensive feature vector;

[0033] Training the SVN classifier using the comprehensive feature vector to obtain the second model.

[0034] Based on the above embodiment, the method further comprises:

[0035] Obtaining an abdominal medical sample image original set, and screening out images with a layer thickness of 5mm from the abdominal medical sample image original set;

[0036] Reducing noise interference of the screened out images using Gaussian filtering, and adjusting the screened out images by setting a window width of 3200 and a window level of -1400 to obtain the abdominal medical sample image set.

[0037] In a second aspect, the present application also provides a blood vessel embolism recognition device based on deep learning, which comprises:

[0038] A first processing module is used to extract a regional image from the abdominal medical image to be identified, wherein the regional image is an abdominal aorta blood vessel image;

[0039] a second processing module, configured to obtain seed points of the abdominal medical image to be identified based on the abdominal aorta vascular image and a trained first model, wherein the first model is constructed based on a Unet network and trained using a set of abdominal medical sample images, and the seed points are pixels of the abdominal aorta in the medical image to be identified;

[0040] a third processing module, configured to semi-automatically segment the mesenteric artery region in the regional image using the seed points using a GrowCut algorithm, and extract an image of the mesenteric artery region;

[0041] a fourth processing module, configured to delineate an arterial centerline of the image of the mesenteric artery region using a VTK platform, wherein the VTK platform is an application for three-dimensional computer graphics;

[0042] The fifth processing module is used to input the image of the mesenteric artery area with the centerline of the arterial blood vessel into the second model to determine whether there is a thrombus and the location of the thrombus in the abdominal medical image to be identified. The second model is established based on the SVN classifier and is used to determine whether there is a thrombus based on the normal blood vessel image characteristics and the thrombus blood vessel image characteristics.

[0043] Based on the above technical solution, further based on the above embodiment, the second processing module is specifically used to input the abdominal aorta blood vessel image into the trained first model to obtain the corresponding pixel points of the abdominal aorta in the abdominal aorta blood vessel image;

[0044] The seed point of the abdominal medical image to be identified is determined based on the pixel points corresponding to the abdominal aorta in the abdominal aorta blood vessel image.

[0045] Based on the above embodiment, the second processing module is specifically configured to determine a sliding window range according to the position of each pixel point in the abdominal aorta blood vessel image;

[0046] Traversing the pixel points within the sliding window to obtain the average value of the pixel values ​​within the sliding window;

[0047] determining an original seed point in the abdominal aorta image according to an average value of all pixel values ​​within the sliding window;

[0048] Calculating the weighted distance between each pixel point in the abdominal aorta blood vessel image and the original seed point;

[0049] The seed point of the abdominal medical image to be identified is determined according to the pixel value of each pixel point in the abdominal aorta blood vessel image, and the weighted distance, maximum weighted distance and minimum weighted distance between each pixel point and the original seed point.

[0050] Based on the above embodiment, the third processing module includes a first subunit, a second subunit, a third subunit, and a fourth subunit;

[0051] A first subunit is configured to mark the seed point of the abdominal aorta blood vessel image as a foreground point and mark other pixel points as background points;

[0052] A second subunit is configured to obtain neighboring points of each foreground point and calculate similarity values ​​between the foreground point and each neighboring point, wherein the similarity values ​​include color similarity, gray value similarity, and spatial position similarity between the foreground point and each neighboring point;

[0053] A third subunit is configured to expand the neighborhood points having similarity values ​​greater than a preset threshold into the foreground points, and return the result to the second subunit;

[0054] The fourth subunit is used to extract all the foreground points in the abdominal aorta vascular image when no neighboring points are expanded as foreground points in this iteration to obtain a foreground image, and the remaining pixels are the background image, wherein the foreground image is an image of the mesenteric artery area.

[0055] Based on the above embodiment, the fourth processing module is specifically used to extract the CSV file of the image of the mesenteric artery region, and the RAS coordinates of the end point of the superior mesenteric artery in the CSV file, where the RSA coordinates are right upper anterior coordinates used in the medical field;

[0056] Convert the RAS coordinates of the end point of the superior mesenteric artery into a data object in the VTK platform;

[0057] The centerline of the artery is calculated using the polygonal data object of the superior mesenteric artery in the VTK platform and the data object of the end point of the superior mesenteric artery in the VTK platform.

[0058] Based on the above embodiment, a sixth processing module is further included, the sixth processing module being configured to obtain a sample image of the mesenteric artery region where the arterial centerline is delineated in the abdominal medical sample image set;

[0059] Extracting a local binary pattern feature vector of the sample image of the mesenteric artery region delineating the arterial vessel centerline;

[0060] Extracting a gray level co-occurrence matrix eigenvector of a sample image of the mesenteric artery region delineating the arterial vessel centerline;

[0061] Combining the local binary pattern eigenvector and the gray level co-occurrence matrix eigenvector to obtain a comprehensive eigenvector;

[0062] The SVN classifier is trained using the comprehensive feature vector to obtain the second model.

[0063] Based on the above embodiment, the sixth processing module is specifically configured to obtain an original set of abdominal medical sample images, and filter out images with a slice thickness of 5 mm from the original set of abdominal medical sample images;

[0064] The screened images were subjected to Gaussian filtering to reduce noise interference, and the screened images were adjusted by setting the window width to 3200 and the window level to -1400 to obtain the abdominal medical sample image set.

[0065] In a third aspect, the present invention further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements a deep learning-based vascular embolism identification method described in any one of the first aspects.

[0066] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a deep learning-based vascular embolism identification method as described in any one of the first aspects.

[0067] The present invention incorporates abdominal enhanced CT imaging data into the analysis, combines multiple laboratory test indicators, vital signs indicators and demographic information, and constructs a multi-dimensional machine learning model. This comprehensive data method can more comprehensively capture the influencing factors of acute superior mesenteric artery thrombosis and screen out predictive factors with strong correlation, thereby assisting clinicians in making more accurate diagnoses and providing a scientific basis for clinical prevention and treatment. A rapid positioning and identification method for acute superior mesenteric artery thrombosis based on a supervised machine learning classification algorithm was constructed, which can better locate and identify acute superior mesenteric artery thrombosis, helping clinicians to identify potential acute superior mesenteric artery thrombosis as early as possible, and ultimately achieve the purpose of early detection, early diagnosis, and early treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings, which constitute part of the present invention, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flowchart of a method for identifying vascular embolism based on deep learning provided by an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of a deep learning-based vascular embolism identification device provided by another embodiment of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other.

[0072] The following detailed description is an exemplary description and is intended to provide further detailed description of the present invention. Unless otherwise indicated, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present invention belongs. The terms used in the present invention are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0073] It should be understood that superior mesenteric artery thrombosis (SMART) is a condition caused by an embolus entering the superior mesenteric artery, resulting in obstruction. This condition can lead to vascular intestinal obstruction and, in turn, acute intestinal ischemia and necrosis. In recent years, with the continuous improvement of diagnostic technology and the increasing incidence of vascular diseases, acute abdominal pain due to vascular causes has become increasingly common. Mesenteric artery ischemia is the most common, with SMART thrombosis accounting for approximately 60%-70% of cases. Emboli often originate from valvular heart disease, atrial fibrillation, myocardial infarction, bacterial endocarditis, and other conditions, with SMART thrombosis being the most common. The mesenteric vessels consist of three branches arising from the abdominal aorta and their collateral circulation: the celiac trunk (CA), the superior mesenteric artery (SMA), the inferior mesenteric artery (IMA), and the mesenteric vein (MV). The SMA supplies the duodenum, jejunum, ileum, ascending colon, and the proximal 1 / 3 of the transverse colon. The IMA supplies the remaining transverse colon, descending colon, sigmoid colon, and proximal rectum. In addition to being the main blood vessel supplying the intestine, the SMA also establishes three important collateral circulations: (1) The inferior pancreaticoduodenal artery, the superior pancreaticoduodenal artery, the gastroduodenal artery, and the celiac artery form a collateral circulation. (2) The middle colic artery, the colic artery, the ileocolic artery, and the inferior mesenteric artery form a collateral circulation. (3) The distal mesenteric artery sends branches to form a collateral circulation with the retroperitoneum and the greater omentum. When the mesenteric vessels are blocked by more than 70%, the collateral circulation opens in large quantities. In addition, the mesenteric venous system is different from other veins and has a corresponding nutrient supply function. It collects venous blood from the intestine and returns it to the portal vein system. 25% to 30% of the blood supply to the liver comes from the hepatic artery, and 70% to 75% comes from the portal vein. The mesenteric blood supply accounts for 15% to 20% of cardiac output during fasting and 35% after eating. Therefore, due to the dual nature and unique characteristics of the intestinal blood supply, significant ischemia only occurs when the intestinal blood supply is reduced by 50% and the mesenteric arterial pressure drops below 45 mmHg (1 mmHg = 0.133 kPa). However, a blood supply of less than 75% should be maintained for 12 hours. Therefore, under normal circumstances, intestinal ischemia is not easily precipitated in the human body. However, if acute ischemia does occur and is not promptly corrected, the consequences can be severe.

[0074] The occurrence of mesenteric artery thrombosis is also related to the anatomical structure of the superior mesenteric artery: the superior mesenteric artery branches off from the abdominal aorta at an acute angle, running parallel to the aorta. Its lumen is relatively large, aligning with the direction of blood flow in the abdominal aorta. This makes it easy for dislodged emboli to enter, leading to thrombosis at vascular stenosis or bifurcations. This thrombosis is most common at or below the origin of the middle colic artery. The intestinal mucosa is not well tolerated by ischemia. If ischemia lasts longer than 15 minutes, the villi of the small intestinal mucosa will be destroyed and detached. This leads to blood stasis in the intestinal wall, resulting in congestion and edema. Loss of intestinal tone and cyanosis and edema can occur, and large amounts of plasma can infiltrate the intestinal wall, causing hemorrhagic necrosis. This massive extravasation of plasma into the abdominal and intestinal cavities leads to a sharp decrease in circulating blood volume, rapid bacterial proliferation within the intestinal lumen, and the continuous absorption of toxic metabolites resulting from intestinal ischemia and hypoxia, leading to hypovolemia and toxic shock. Intestinal necrosis causes intestinal dilation and loss of peristalsis, manifesting as vascular intestinal obstruction.

[0075] CT angiography (CTA) is currently the preferred diagnostic method for mesenteric artery thrombosis (MAE), with a sensitivity of 93%, a specificity of 100%, a positive rate of 94%, and a rejection rate of 100%. Ultrasound examination is also helpful in diagnosing this condition. Although once considered the gold standard, mesenteric angiography is no longer a routine option due to various factors and is only used during surgical intervention.

[0076] Generally, patients with unexplained acute abdominal pain, such as severe pain refractory to standard medications and symptoms disproportionate to physical signs, should be highly suspected of mesenteric vascular disease. A detailed medical history and a thorough physical examination are essential. A thorough history, combined with specific symptoms and signs, often aids in the initial diagnosis of mesenteric vascular disease. A history of heart disease, severe gastrointestinal emptying symptoms, and a disproportionate symptom-sign relationship constitute the triad of features that support the early diagnosis of acute superior mesenteric artery thrombosis. Acute superior mesenteric artery thrombosis often develops secondary to preexisting chronic arterial stenosis and therefore often presents with evidence of ischemia in other arteries or a history of arterial surgery. Once acute mesenteric vascular disease develops, it can present with severe abdominal pain that is refractory to standard medications, but with no obvious abdominal signs or only mild tenderness and hyperactive bowel sounds. This represents the early stages of intestinal ischemia, making surgical treatment optimal. However, it can be difficult to differentiate from other acute abdominal conditions. Laparotomy often reveals only pale, constricted, or dilated bowel, making it prone to misdiagnosis and treatment errors. As the disease progresses rapidly, physical signs gradually worsen, with tenderness, rebound tenderness, and muscle tension appearing. Intestinal smooth muscle paralysis develops due to increased ischemia and hypoxia, manifesting as abdominal distension and significantly weakened or absent bowel sounds. This is accompanied by frequent vomiting, often bloody, which often indicates intestinal necrosis. Once necrosis occurs, large amounts of fluid quickly leak into the intestinal and peritoneal cavities, leading to severe water and electrolyte imbalances and acid-base imbalances. Patients may then experience symptoms of toxic shock, such as a rapid heart rate, hypotension, and acidosis.

[0077] Laboratory tests for acute mesenteric vascular disease lack specificity and sensitivity. Common manifestations include elevated white blood cell and neutrophil counts, and during hemoconcentration, elevated red blood cell count, hematocrit, and serum amylase. As the disease progresses, metabolic acidosis, electrolyte imbalances, and enzymatic changes may also develop, but these changes can be easily misdiagnosed as intra-abdominal infection. Elevated serum lactate levels and metabolic acidosis often indicate intestinal obstruction, while elevated plasma glutathione S-transferase isoenzymes often indicate intestinal ischemia. Plasma D-dimer is nonspecific for acute intravascular thrombosis or thrombosis, but it is highly sensitive. A 5-fold or greater increase in D-dimer results should raise suspicion of mesenteric vascular thrombosis or thrombotic disease.

[0078] Early selection of appropriate imaging studies often facilitates rapid diagnosis. In the early stages of the disease, abdominal X-rays may be unremarkable, occasionally revealing dilated small intestine and colon, rigid folds, and thickened, unchanged bowel loops. As the disease progresses, intraluminal gas, fluid-air levels, and free air in the peritoneal cavity may be revealed. Although abnormalities are present on abdominal X-rays in most patients, they lack specificity. Doppler ultrasound is a simple, effective, and inexpensive screening method. It not only measures intestinal wall thickness and intravascular emboli, but also comprehensively assesses the degree of vascular stenosis, the location and extent of obstruction, and hemodynamic changes, thereby assessing the severity of vascular obstruction and the degree of intestinal ischemia. However, in some patients, intestinal distension can affect ultrasound findings. Therefore, if ultrasound findings are unclear, CT should be performed. CT helps differentiate from other acute abdominal conditions and can demonstrate focal or segmental thickening of the bowel wall, submucosal edema and hemorrhage, free intraperitoneal air, gas within the mesenteric veins, and ascites. Enhanced CT and CT angiography (CTA) can achieve a diagnostic rate of over 90% for mesenteric vascular disease. Digital subtraction angiography (DSA), a diagnostic procedure specialized in vascular surgery, is currently the gold standard for diagnosing mesenteric vascular obstruction. However, due to its specialized and invasive nature, and the fact that the entire mesenteric DSA process, from preparation to surgery, can take several hours, often delaying the patient's condition, many hospitals find it difficult to perform in emergency situations. Magnetic resonance angiography (MRA) is also an option for those allergic to contrast agents. While MRA has similar sensitivity and specificity to CTA for diagnosing mesenteric vascular disease, it lacks the ability to identify non-obstructive mesenteric ischemia and distal arterial thrombosis.

[0079] For thrombosis CT images with a single component, grayscale characteristics are often used as a starting point. By traversing the field around the pixel, the pixel value distribution law is found to find the edge of the lesion, which belongs to a two-category task. At present, there are mainly two major methods in the field of image classification. The first is a classification method based on statistical learning, and the second is a classification method based on deep learning. The classification method based on statistical learning often requires three steps: pre-processing, feature extraction and classification. The classification method based on deep learning is one of the research hotspots in the field of image classification in recent years, and is mainly based on the architecture of convolutional neural network (CNN).

[0080] The present invention is designed to solve the above-mentioned problems and aims to provide a method for rapid location and identification of acute superior mesenteric artery thrombosis based on deep learning, specifically comprising:

[0081] Data Collection and Organization: Based on inclusion and exclusion criteria, clinical records of patients with acute superior mesenteric artery thrombosis between January 1, 2015, and December 31, 2022, were identified. The relevant medical records were downloaded and preprocessed. Inclusion criteria included: ① patients with pathologically confirmed acute superior mesenteric artery thrombosis; ② patients with pretreatment abdominal contrast-enhanced CT data. Exclusion criteria included: ① patients with poor image quality or lack of preoperative imaging data, making postoperative imaging evaluation impossible. The processed data were divided into training and test sets.

[0082] Whole-abdomen contrast-enhanced CT was performed using a Siemens Somatom Sensation 64 (Siemens, Forchheim); a GE Discovery CT 750m (GE Healthcare, Princeton, NJ); or a GE LightSpeed ​​VCT (GE Healthcare, Princeton, NJ). Spiral CT parameters were: tube voltage 120 kV, automatic milliampere-second technology (effective milliampere-seconds 210-250 mAs), pitch 1, rotation time 0.5 s, field of view 350 mm × 350 mm, matrix 512 × 512, slice thickness 5 mm, and interval 5 mm. Patients were required to fast for at least 3 hours and to drink 500 ml of water orally 5-10 minutes before the scan. Contrast-enhanced scanning was performed with intravenous iodine-containing nonionic contrast agent (300-370 mgI / ml) at a flow rate of 3 ml / s. The arterial phase scan used bolus tracking technology, the monitoring plane was the level of the celiac trunk opening of the abdominal aorta, and the scan was performed 12 seconds after the trigger threshold (100Hu). The venous phase and delayed phase were delayed by 28 seconds and 50 seconds, respectively.

[0083] Image data preprocessing automatically selects images with a slice thickness of exactly 5mm. Image processing techniques, such as Gaussian filtering, are used to reduce noise interference. Images are adjusted by setting a window width of 3200 (ranging from -3000 to 200) and a window level of -1400 to make the thrombus more prominent on the images. Deep learning models, such as convolutional neural networks (CNNs), are deployed to automatically detect and segment thrombi. These models accurately identify thrombus-containing regions, known as regions of interest (ROIs), and automatically create corresponding mask files, providing a solid data foundation for further analysis. The system automatically labels thrombus locations and assigns probability scores, assisting physicians in more accurate image interpretation. To ensure the accuracy and stability of the deep learning model, it is trained using a large amount of annotated medical image data, and cross-validation and other methods are used to evaluate its performance and ensure its effectiveness in real-world applications. The system automatically captures the identified mesenteric artery thrombus site and saves it in 512x512 pixel PNG format for further analysis and documentation.

[0084] Artificial intelligence model building includes:

[0085] 3D image reconstruction of the abdominal artery tree: Seed points are found, and the abdominal aorta is selected as the seed point. A dataset is constructed and a UNet network is used for learning to obtain a model (existing methods require manual seed selection). The model accuracy reaches 95% (acc = value of the pixel with 1 in the test set × training result at the corresponding coordinate / value of the pixel with 1 in the test set). This means that the training result can extract almost all pixels corresponding to the abdominal aorta. After obtaining the mask value of the abdominal aorta, the coordinates of the middle pixel are selected as the seed point.

[0086] 3D reconstruction is performed using the GrowCut algorithm. The GrowCut algorithm uses iterative growing algorithms to label each pixel as foreground or background. The steps include: 1) Initialization: a seed point is labeled as foreground, and all other pixels are labeled as background. 2) Growing iterations: the algorithm considers the pixel's neighborhood (26 neighborhoods) and expands the region based on pixel similarity. Similarity is typically calculated based on pixel value (color, grayscale, etc.) and spatial position. 3) Iterative convergence: the algorithm stops when no pixel label changes during an iteration. 4) Output: the image is segmented into foreground and background based on the final labeling.

[0087] The GrowCut algorithm is a graph-based segmentation method for 3D reconstruction, particularly well-suited for images with complex boundaries and textures. It demonstrates significant advantages in thrombus localization in the superior mesenteric artery and 3D reconstruction of the abdominal arterial tree. Its unsupervised nature effectively reduces manual intervention and significantly improves automation, making it particularly suitable for rapid diagnosis in emergency situations. Through a growing and iterative approach, the algorithm effectively handles complex vascular structures, incrementally improving segmentation results to accurately delineate vascular networks and locate thrombi. Furthermore, the GrowCut algorithm accounts for image noise and inhomogeneity, demonstrating robustness and enhancing segmentation accuracy, particularly in conditions with low contrast between vessels and surrounding tissue. Compared to other computationally intensive algorithms, GrowCut maintains high segmentation accuracy while maintaining high computational efficiency. This makes it suitable for clinical applications requiring rapid reconstruction and real-time feedback, such as real-time 3D reconstruction during surgery, enabling rapid decision-making. Notably, the GrowCut algorithm offers excellent refinement and scalability, allowing for tailoring and optimization to specific application scenarios to improve the identification of thrombus regions. These technical advantages jointly promote the accuracy and efficiency of diagnosis, enabling the growcut algorithm to play an important role in clinical practice.

[0088] Calculate the vessel centerline and locate the endpoint of the superior mesenteric artery. Use the centerlinecomputation library of VTK to obtain the RAS coordinates of the endpoint from the table.

[0089] A model was trained to learn features such as LBP and Haralick for normal and thrombosed vessels. The accuracy of the SVM classifier was 0.766, and the accuracy of the random forest classifier was 0.83. Further image augmentation and network replacement will be required to increase training accuracy.

[0090] Starting from the current coordinate (x, y), the above model is used to determine whether the blood vessels in a series of nearby slices are thrombi. The judgment range is [[x-10, x+10], [y-10, y+10]].

[0091] The present invention provides a method for rapid positioning and identification of acute superior mesenteric artery thrombosis based on deep learning, which has the following characteristics: according to medical record data, medical record data is trained based on a deep convolutional neural network algorithm to obtain rapid positioning of acute superior mesenteric artery thrombosis, further training to obtain a fusion model, and saving the qualitative model and fusion model of acute superior mesenteric artery thrombosis as a callable automatic discrimination artificial intelligence model; calling the automatic discrimination artificial intelligence model, inputting the medical record data to be predicted for prediction, and obtaining the identification of acute superior mesenteric artery thrombosis.

[0092] The present invention incorporates abdominal enhanced CT imaging data into the analysis, combines multiple laboratory examination indicators, vital signs indicators and demographic information, and constructs a multi-dimensional machine learning model. This comprehensive data method can more comprehensively capture the influencing factors of acute superior mesenteric artery thrombosis and screen out predictive factors with strong correlation, thereby assisting clinicians in making more accurate diagnoses and providing a scientific basis for clinical prevention and treatment. A rapid positioning and identification method for acute superior mesenteric artery thrombosis based on a supervised machine learning classification algorithm was constructed, which can better locate and identify acute superior mesenteric artery thrombosis, helping clinicians to identify potential acute superior mesenteric artery thrombosis as early as possible, and ultimately achieve the purpose of early detection, early diagnosis and early treatment. The indicator mIoU, which is not commonly used in diagnosis, was introduced as training data, and its technical effect was proved, which not only enhanced the predictive ability of the model, but also provided a new perspective for clinical diagnosis.

[0093] The following will be combined with the Figure 1 , a method for identifying vascular embolism based on deep learning provided by an embodiment of the present invention is described, comprising the following steps:

[0094] 110. Extract a regional image from the abdominal medical image to be identified, where the regional image is an abdominal aorta blood vessel image.

[0095] 120. Based on the abdominal aorta vascular image and the trained first model, the seed point of the abdominal medical image to be identified is obtained. The first model is constructed based on the Unet network and is trained using the abdominal medical sample image set. The seed point is the pixel point in the abdominal aorta pixel in the medical image to be identified.

[0096] 130. The mesenteric artery region in the regional image is semi-automatically segmented using the GrowCut algorithm and the image of the mesenteric artery region is extracted.

[0097] 140. The image of the mesenteric artery region is imaged using the VTK platform to delineate the arterial centerline, wherein the VTK platform is an application for three-dimensional computer graphics.

[0098] 150. An image of the mesenteric artery region with the arterial centerline delineated is input into the second model to determine whether there is a thrombus and the location of the thrombus in the abdominal medical image to be identified. The second model is established based on the SVN classifier and is used to determine whether there is a thrombus based on the normal vascular image characteristics and the thrombus vascular image characteristics.

[0099] Based on the above embodiment, step 120 specifically includes:

[0100] 121. Input the abdominal aorta blood vessel image into the trained first model to obtain the corresponding pixel points of the abdominal aorta in the abdominal aorta blood vessel image.

[0101] 122. Determine the seed point of the abdominal medical image to be identified based on the corresponding pixel points of the abdominal aorta in the abdominal aorta vascular image.

[0102] Based on the above embodiment, step 122 specifically includes:

[0103] The sliding window range is determined according to the position of each pixel point in the abdominal aorta blood vessel image.

[0104] Traverse the pixel points within the sliding window range and get the average value of the pixel values ​​within the sliding window range.

[0105] The original seed point in the abdominal aorta image is determined according to the average value of pixel values ​​within the range of all sliding windows.

[0106] Calculate the weighted distance between each pixel in the abdominal aorta image and the original seed point.

[0107] The seed point of the abdominal medical image to be identified is determined according to the pixel value of each pixel point in the abdominal aorta blood vessel image, and the weighted distance, maximum weighted distance and minimum weighted distance between each pixel point and the original seed point.

[0108] Specifically, for example, when the pixel value of each pixel point in the abdominal aorta image is greater than a preset pixel value and the weighted distance between the pixel point and the original seed point is between the maximum weighted distance and the minimum weighted distance, the pixel point is determined to be a seed point.

[0109] Based on the above embodiment, step 130 specifically includes:

[0110] 131. Mark the seed point of the abdominal aorta image as the foreground point, and mark the other pixel points as the background point.

[0111] 132. Obtain neighboring points of each foreground point, and calculate similarity values ​​between the foreground point and each neighboring point, wherein the similarity values ​​include calculating color similarity, gray value similarity, and spatial position similarity between the foreground point and each neighboring point.

[0112] 133. Expand the neighborhood points whose similarity values ​​are greater than a preset threshold into foreground points, and return to step 132.

[0113] 134. In this round of iteration, no neighboring points are expanded to foreground points. All foreground points in the abdominal aorta image are extracted to obtain the foreground image. The remaining pixels are the background image, where the foreground image is the image of the mesenteric artery area.

[0114] Based on the above embodiment, step 140 specifically includes:

[0115] Extract the CSV file of the image of the mesenteric artery region and convert the RAS coordinates of the end point of the superior mesenteric artery in the CSV file. The RSA coordinates are right superior anterior coordinates used in the medical field.

[0116] Convert the RAS coordinates of the end point of the superior mesenteric artery into a data object in the VTK platform.

[0117] The center line of the superior mesenteric artery is calculated using the polygonal data object of the superior mesenteric artery in the VTK platform and the data object of the end point of the superior mesenteric artery in the VTK platform.

[0118] Based on the above embodiment, the present invention further includes:

[0119] A sample image of the mesenteric artery region is obtained from an abdominal medical sample image set, where the arterial vessel centerline is delineated.

[0120] Extract the local binary pattern feature vector of the sample image of the mesenteric artery region to delineate the arterial vessel centerline.

[0121] Extract the gray-level co-occurrence matrix eigenvectors of the sample images of the mesenteric artery region to delineate the arterial centerline.

[0122] The local binary pattern eigenvector and the gray-level co-occurrence matrix eigenvector are combined to obtain the comprehensive eigenvector.

[0123] The SVN classifier is trained using the comprehensive feature vector to obtain the second model.

[0124] Based on the above embodiment, the present invention further includes:

[0125] An original set of abdominal medical sample images is obtained, and images with a slice thickness of 5 mm are screened out from the original set of abdominal medical sample images.

[0126] The screened images were subjected to Gaussian filtering to reduce noise interference, and the screened images were adjusted by setting the window width to 3200 and the window level to -1400 to obtain an abdominal medical sample image set.

[0127] The present invention incorporates abdominal enhanced CT imaging data into the analysis, combines multiple laboratory examination indicators, vital signs indicators and demographic information, and constructs a multi-dimensional machine learning model. This comprehensive data method can more comprehensively capture the influencing factors of acute superior mesenteric artery thrombosis and screen out predictive factors with strong correlation, thereby assisting clinicians in making more accurate diagnoses and providing a scientific basis for clinical prevention and treatment. A rapid positioning and identification method for acute superior mesenteric artery thrombosis based on a supervised machine learning classification algorithm was constructed, which can better locate and identify acute superior mesenteric artery thrombosis, helping clinicians to identify potential acute superior mesenteric artery thrombosis as early as possible, and ultimately achieve the purpose of early detection, early diagnosis and early treatment. The indicator mIoU, which is not commonly used in diagnosis, was introduced as training data, and its technical effect was proved, which not only enhanced the predictive ability of the model, but also provided a new perspective for clinical diagnosis.

[0128] like Figure 2 As shown, the present invention also provides a device for identifying vascular embolism based on deep learning, the device comprising:

[0129] The first processing module is used to extract a regional image from the abdominal medical image to be identified, where the regional image is an abdominal aorta blood vessel image.

[0130] The second processing module is used to obtain the seed points of the abdominal medical image to be identified based on the abdominal aorta vascular image and the trained first model. The first model is constructed based on the Unet network and is trained using the abdominal medical sample image set. The seed points are the pixels in the abdominal aorta pixels in the medical image to be identified.

[0131] The third processing module is used to use the GrowCut algorithm to semi-automatically segment the mesenteric artery region in the regional image using the seed points, and extract the image of the mesenteric artery region.

[0132] a fourth processing module for delineating an arterial centerline of an image of the mesenteric artery region using a VTK platform, wherein the VTK platform is an application for three-dimensional computer graphics;

[0133] The fifth processing module is used to input the image of the mesenteric artery area that delineates the centerline of the arterial blood vessel into the second model to determine whether there is a thrombus and the location of the thrombus in the abdominal medical image to be identified. The second model is established based on the SVN classifier and is used to determine whether there is a thrombus based on the normal blood vessel image characteristics and the thrombus blood vessel image characteristics.

[0134] Based on the above embodiment, the second processing module is specifically configured to input the abdominal aorta blood vessel image into the trained first model to obtain the corresponding pixel points of the abdominal aorta in the abdominal aorta blood vessel image.

[0135] The seed point of the abdominal medical image to be identified is determined based on the corresponding pixel points of the abdominal aorta in the abdominal aorta blood vessel image.

[0136] Based on the above embodiment, the second processing module is specifically configured to determine the sliding window range according to the position of each pixel point in the abdominal aorta blood vessel image.

[0137] Traverse the pixel points within the sliding window range and get the average value of the pixel values ​​within the sliding window range.

[0138] The original seed point in the abdominal aorta image is determined according to the average value of pixel values ​​within the range of all sliding windows.

[0139] Calculate the weighted distance between each pixel in the abdominal aorta image and the original seed point.

[0140] The seed point of the abdominal medical image to be identified is determined according to the pixel value of each pixel point in the abdominal aorta blood vessel image, and the weighted distance, maximum weighted distance and minimum weighted distance between each pixel point and the original seed point.

[0141] Based on the above embodiment, the third processing module includes a first subunit, a second subunit, a third subunit, and a fourth subunit;

[0142] The first subunit is used to mark the seed point of the abdominal aorta blood vessel image as a foreground point and mark other pixel points as background points.

[0143] The second subunit is used to obtain the neighborhood points of each foreground point and calculate the similarity value between the foreground point and each neighborhood point, wherein the similarity value includes calculating the color similarity, gray value similarity and spatial position similarity between the foreground point and each neighborhood point.

[0144] The third subunit is configured to expand the neighborhood points whose similarity values ​​are greater than a preset threshold into foreground points and return the points to the second subunit.

[0145] The fourth subunit is used to extract all foreground points from the abdominal aorta image when no neighboring points are expanded to foreground points in this iteration, and obtain a foreground image. The remaining pixels are the background image, where the foreground image is the image of the mesenteric artery area.

[0146] Based on the above embodiment, the fourth processing module is specifically used to extract the CSV file of the image of the mesenteric artery region, and obtain the RAS coordinates of the end point of the superior mesenteric artery in the CSV file. The RSA coordinates are the right upper anterior coordinates used in the medical field.

[0147] Convert the RAS coordinates of the end point of the superior mesenteric artery into a data object in the VTK platform.

[0148] The center line of the superior mesenteric artery is calculated using the polygonal data object of the superior mesenteric artery in the VTK platform and the data object of the end point of the superior mesenteric artery in the VTK platform.

[0149] Based on the above embodiment, a sixth processing module is further included, which is used to obtain a sample image of the mesenteric artery region in which the center line of the arterial vessel is delineated in the abdominal medical sample image set.

[0150] Extract the local binary pattern feature vector of the sample image of the mesenteric artery region to delineate the arterial vessel centerline.

[0151] Extract the gray-level co-occurrence matrix eigenvectors of the sample images of the mesenteric artery region to delineate the arterial centerline.

[0152] The local binary pattern eigenvector and the gray-level co-occurrence matrix eigenvector are combined to obtain the comprehensive eigenvector.

[0153] The SVN classifier is trained using the comprehensive feature vector to obtain the second model.

[0154] Based on the above embodiment, the sixth processing module is specifically configured to obtain an original set of abdominal medical sample images, and filter out images with a slice thickness of 5 mm from the original set of abdominal medical sample images.

[0155] The screened images were subjected to Gaussian filtering to reduce noise interference, and the screened images were adjusted by setting the window width to 3200 and the window level to -1400 to obtain an abdominal medical sample image set.

[0156] The application incorporates abdominal enhanced CT image data into analysis, combines multiple laboratory examination indexes, vital sign indexes and demographic information, and constructs a multi-dimensional machine learning model. This comprehensive data method can more comprehensively capture the influencing factors of acute superior mesenteric artery thrombosis, screen out the strong correlation prediction factors, thereby assisting clinicians to make more accurate diagnosis, and providing a scientific basis for clinical prevention and treatment. A rapid positioning and identification method of acute superior mesenteric artery thrombosis based on a supervised machine learning classification algorithm is constructed, which can better locate and identify acute superior mesenteric artery thrombosis, and is helpful for clinicians to identify potential acute superior mesenteric artery thrombosis as soon as possible, and ultimately achieve the purpose of early detection, early diagnosis and early treatment. The index mIoU not commonly used in diagnosis is introduced as training data, and its technical effect is proved, which not only enhances the prediction ability of the model, but also provides a new perspective for clinical diagnosis.

[0157] In addition, the embodiment of the application comprises a computer device comprising a memory and a processor, the memory stores a computer program, and the processor implements the blood vessel embolism identification method based on deep learning according to any one of the technical solutions when executing the computer program.

[0158] The embodiment of the application further comprises a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the blood vessel embolism identification method based on deep learning according to any one of the technical solutions.

[0159] The above embodiments are only used to illustrate the technical solutions of the application but not to limit it, and although the application has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the specific embodiments of the application can be modified or replaced equivalently without departing from the spirit and scope of the application, and any modification or equivalent replacement without departing from the spirit and scope of the application should be covered in the protection scope of the claims of the application.

Claims

1. A method for identifying vascular embolism based on deep learning, characterized in that: The method comprises: Extracting a regional image from an abdominal medical image to be identified, wherein the regional image is an abdominal aorta blood vessel image; Obtaining seed points of the abdominal medical image to be identified based on the abdominal aorta vascular image and a trained first model, wherein the first model is constructed based on a Unet network and trained using a set of abdominal medical sample images, and the seed points are pixels of the abdominal aorta in the abdominal medical image to be identified; Semi-automatically segmenting the mesenteric artery region in the regional image using the GrowCut algorithm using the seed points to extract an image of the mesenteric artery region; Delineating an arterial centerline of the image of the mesenteric artery region using a VTK platform, wherein the VTK platform is an application for three-dimensional computer graphics; Inputting an image of the mesenteric artery region with the arterial centerline delineated into a second model to determine whether a thrombus is present and the location of the thrombus in the abdominal medical image to be identified, wherein the second model is established based on a SVM classifier and is used to determine whether a thrombus is present based on normal vascular image features and thrombus vascular image features; Obtaining the seed point of the to-be-identified abdominal medical image based on the abdominal aorta image and the trained first model specifically includes: Inputting the abdominal aorta blood vessel image into the trained first model to obtain the corresponding pixel points of the abdominal aorta in the abdominal aorta blood vessel image; Determining a seed point of the abdominal medical image to be identified based on a pixel point corresponding to the abdominal aorta in the abdominal aorta blood vessel image; The step of determining the seed point of the abdominal medical image to be identified based on the corresponding pixel point of the abdominal aorta in the abdominal aorta blood vessel image specifically includes: determining a sliding window range according to a position of each pixel point in the abdominal aorta blood vessel image; Traversing the pixel points within the sliding window to obtain the average value of the pixel values ​​within the sliding window; determining an original seed point in the abdominal aorta image according to an average value of all pixel values ​​within the sliding window; Calculating the weighted distance between each pixel point in the abdominal aorta blood vessel image and the original seed point; The seed point of the abdominal medical image to be identified is determined according to the pixel value of each pixel point in the abdominal aorta blood vessel image, and the weighted distance, maximum weighted distance and minimum weighted distance between each pixel point and the original seed point.

2. The method according to claim 1, characterized in that The semi-automatic segmentation of the mesenteric artery region in the regional image using the GrowCut algorithm on the seed point to extract the image of the mesenteric artery region specifically includes: S1. Marking the seed point of the abdominal aorta image as a foreground point, and marking the other pixel points as background points; S2. Obtaining neighborhood points of each foreground point, and calculating similarity values ​​between the foreground point and each neighborhood point, wherein the similarity values ​​include calculating color similarity, grayscale value similarity, and spatial position similarity between the foreground point and each of the neighborhood points; S3, expanding the neighborhood points whose similarity values ​​are greater than a preset threshold into the foreground points, and returning to step S2; S4. In this round of iteration, no neighboring points are expanded to foreground points. All the foreground points in the abdominal aorta image are extracted to obtain a foreground image, and the remaining pixels are the background image, wherein the foreground image is an image of the mesenteric artery region.

3. The method according to claim 1, characterized in that The image of the mesenteric artery region is delineated using the VTK platform, wherein the VTK platform is an application for three-dimensional computer graphics, specifically including: Extracting a CSV file of an image of the mesenteric artery region, and obtaining the RAS coordinates of the end point of the superior mesenteric artery in the CSV file, where the RSA coordinates are right superior anterior coordinates used in the medical field; Converting the RAS coordinates of the end point of the superior mesenteric artery into a data object in the VTK platform; The centerline of the artery is calculated using the polygonal data object of the superior mesenteric artery in the VTK platform and the data object of the end point of the superior mesenteric artery in the VTK platform.

4. The method according to claim 1, wherein The method further comprises: Acquire a sample image of the mesenteric artery region where the arterial blood vessel centerline is delineated in the abdominal medical sample image set; Extracting a local binary pattern feature vector of a sample image of the mesenteric artery region delineating the arterial vessel centerline; Extracting a gray level co-occurrence matrix eigenvector of a sample image of the mesenteric artery region delineating the arterial vessel centerline; Combining the local binary pattern eigenvector and the gray level co-occurrence matrix eigenvector to obtain a comprehensive eigenvector; The comprehensive feature vector is used to train an SVM classifier to obtain the second model.

5. The method according to claim 1, wherein The method further comprises: Acquire an original set of abdominal medical sample images, and screen out images with a slice thickness of 5 mm from the original set of abdominal medical sample images; The screened images were subjected to Gaussian filtering to reduce noise interference, and the screened images were adjusted by setting the window width to 3200 and the window level to -1400 to obtain the abdominal medical sample image set.

6. A deep learning-based vascular embolism identification device, characterized in that: The device comprises: A first processing module is used to extract a regional image from the abdominal medical image to be identified, wherein the regional image is an abdominal aorta blood vessel image; a second processing module, configured to obtain seed points of the abdominal medical image to be identified based on the abdominal aorta vascular image and a trained first model, wherein the first model is constructed based on a Unet network and trained using a set of abdominal medical sample images, and the seed points are pixels of the abdominal aorta in the abdominal medical image to be identified; a third processing module, configured to semi-automatically segment the mesenteric artery region in the regional image using the seed points using a GrowCut algorithm, and extract an image of the mesenteric artery region; a fourth processing module, configured to delineate an arterial centerline of the image of the mesenteric artery region using a VTK platform, wherein the VTK platform is an application for three-dimensional computer graphics; a fifth processing module, configured to input an image of the mesenteric artery region, which delineates the arterial centerline, into a second model to determine whether a thrombus is present and the location of the thrombus in the to-be-identified abdominal medical image, wherein the second model is established based on a SVM classifier and is configured to determine whether a thrombus is present based on normal vascular image features and thrombus vascular image features; Obtaining the seed point of the to-be-identified abdominal medical image based on the abdominal aorta image and the trained first model specifically includes: Inputting the abdominal aorta blood vessel image into the trained first model to obtain the corresponding pixel points of the abdominal aorta in the abdominal aorta blood vessel image; Determining a seed point of the abdominal medical image to be identified based on a pixel point corresponding to the abdominal aorta in the abdominal aorta blood vessel image; The step of determining the seed point of the abdominal medical image to be identified based on the corresponding pixel point of the abdominal aorta in the abdominal aorta blood vessel image specifically includes: determining a sliding window range according to a position of each pixel point in the abdominal aorta blood vessel image; Traversing the pixel points within the sliding window to obtain the average value of the pixel values ​​within the sliding window; determining an original seed point in the abdominal aorta image according to an average value of all pixel values ​​within the sliding window; Calculating the weighted distance between each pixel point in the abdominal aorta blood vessel image and the original seed point; The seed point of the abdominal medical image to be identified is determined according to the pixel value of each pixel point in the abdominal aorta blood vessel image, and the weighted distance, maximum weighted distance and minimum weighted distance between each pixel point and the original seed point.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the deep learning-based vascular embolism identification method according to any one of claims 1 to 5 is 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 deep learning-based vascular embolism identification method according to any one of claims 1 to 5 is implemented.

Images