A method, apparatus, medium and device for segmenting an intracranial hemorrhage CT image

Abstract

The application discloses an intracranial hemorrhage CT image segmentation method and device, medium and equipment, and relates to the technical field of image segmentation. In the threshold segmentation optimization process, the Barebones mechanism and the orthogonal learning mechanism are integrated into the African vulture optimization algorithm. The search space is expanded in the search population corresponding to the segmentation threshold through Gaussian mutation, and the ability of the threshold optimization algorithm to get rid of local optimization is enhanced. The orthogonal learning mechanism promotes the resource sharing between the search population individuals corresponding to the segmentation threshold, and improves the search diversity of the segmentation threshold. Through multiple rounds of iteration optimization, the optimal individual that makes the Rényi entropy of the two-dimensional histogram based on the threshold segmentation maximum is obtained as the optimal segmentation threshold of the intracranial hemorrhage CT image. The application improves the population diversity and convergence speed in the threshold segmentation optimization of the intracranial hemorrhage CT image, and improves the image segmentation effect of the intracranial hemorrhage CT image.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to a method, apparatus, medium and device for segmenting CT images of intracranial hemorrhage. Background Technology

[0002] Intracranial hemorrhage (ICH) is a serious cerebrovascular disease characterized by the rupture of a blood vessel in the brain, leading to blood entering the cranial cavity and causing brain tissue damage and neurological dysfunction. The disease has a sudden onset and rapid progression, and is closely associated with extremely high mortality and disability rates. Furthermore, it is a significant factor contributing to stroke-related deaths and dependency. Studies show that the overall incidence of ICH over the past 40 years is 29.9 cases per 100,000 people, and has not shown a decreasing trend. Given the current lack of a specific treatment for ICH and its severe clinical consequences, early and accurate diagnosis is crucial for improving patient prognosis.

[0003] Currently, the diagnosis of intracranial hemorrhage mainly relies on imaging techniques designed to detect hemorrhage, identify potential causes, and guide treatment. CT scans are fast, highly sensitive, and widely available; magnetic resonance imaging (MRI) has extremely high sensitivity for detecting small hemorrhages; and digital subtraction angiography (DSA) plays an irreplaceable role in determining the cause of intracranial hemorrhage and is considered the "gold standard." However, these imaging techniques have some limitations: CT is not sensitive to early ischemic changes, MRI is time-consuming and expensive, and DSA is an invasive procedure. Researchers have also studied the diagnostic effectiveness of CT, MRI / MRA, and DSA in patients with intracranial hemorrhage. The results showed that CT has moderate accuracy as a preliminary examination, while MRI / MRA and DSA need to be used as further evaluation methods.

[0004] Despite its limitations, CT scans remain the preferred tool for diagnosing intracranial hemorrhage due to their affordability and safety. In recent years, with the development of artificial intelligence (AI) technologies—including deep learning (DL), machine learning (ML), and metaheuristic optimization algorithms (MOA)—AI has been widely applied to the CT diagnosis of intracranial hemorrhage, significantly improving diagnostic accuracy and efficiency.

[0005] In existing technologies, AI-assisted CT diagnosis of intracranial hemorrhage can be divided into three main categories: detection, classification, and segmentation. Segmentation techniques are further subdivided into semantic segmentation and threshold segmentation. Semantic segmentation primarily relies on deep learning techniques, but its accuracy and reliability are limited due to the scarcity of labeled data for intracranial hemorrhage CT segmentation and the fact that its results typically exhibit binary features (foreground and background). These limitations make it difficult to address challenges such as variations in CT equipment, the complexity of intracranial hemorrhage etiologies, and individual differences. Detection, classification, and semantic segmentation are outcome-oriented technologies that can provide doctors with specific lesion localization or disease prediction. Threshold segmentation, on the other hand, divides the image into multiple regions using preset thresholds (Multi-Threshold Image Segmentation, MTIS). While this technology cannot precisely locate lesions or predict disease outcomes, it effectively processes complex images, extracts more valuable information, improves diagnostic efficiency, and is easy for doctors to operate. MTIS has attracted widespread attention in the academic community due to its unique advantages.

[0006] MTIS is essentially an optimization problem, the most crucial aspect of which is determining a set of thresholds. Exhaustive search methods are time-consuming and often ineffective. Mathematical methods frequently struggle to escape local optima. Metaheuristic optimization algorithms, as powerful tools for solving such complex optimization problems, have been extensively studied. MOA is an optimization technique inspired by various natural phenomena. For example, Particle Swarm Optimization (PSO) simulates the foraging behavior of birds; Mixed Frog Leaping Algorithm (SFLA) draws on the theory of natural memes; Rime Optimization Algorithm (RIME) is based on the physical phenomenon of rime formation; Escape Algorithm (ESC) simulates crowd evacuation behavior; Salicylic Aquatic Optimization Algorithm (SSA) is inspired by the navigation and foraging behavior of salicylic aquatic colonies in the ocean; Harris Eagle Optimization Algorithm (HHO) imitates the hunting behavior of eagles; Slime Mold Algorithm (SMA) simulates the foraging behavior of slime molds; Whale Optimization Algorithm (WOA) simulates the social behavior of humpback whales; Simulated Annealing Algorithm (SA) originates from the annealing principle in solid-state physics; Multiverse Optimization Algorithm (MVO) is based on the multiverse hypothesis; and Differential Evolution Algorithm (DE) is based on a specific evolutionary mechanism.

[0007] These algorithms typically do not depend on the specific structure of the problem, but instead employ heuristic methods to perform global exploration and local optimization within the search space to find an approximate optimal solution. They are widely used in engineering optimization, machine learning, and other fields, and are also extensively applied in MTIS (Meta-Inductively Coupled Image Processing). For example, one research scheme proposed an improved algorithm, CBQMVO, which successfully achieved efficient segmentation of breast cancer pathology images; another scheme improved the initialization strategy of the ant colony optimization algorithm and introduced the Levy flight mechanism to enhance its optimization capabilities; this improved algorithm was applied to Otsu's method for image segmentation, successfully segmenting brain MRI images; yet another research scheme integrated the PSO and SSA algorithms and combined them with Otsu's method to achieve segmentation of diabetic retinal vascular images. This method enhances the visibility of major blood vessels through Gaussian filters and improves the clarity of peripheral blood vessels using top-hat transformation techniques. These studies demonstrate that using meta-heuristic optimization algorithms to solve for threshold sets is a feasible approach.

[0008] However, existing metaheuristic optimization algorithms still face challenges in dealing with the complex scenarios of intracranial hemorrhage CT images. They suffer from problems such as insufficient population diversity, difficulty in escaping local optima, and poor convergence speed, making it difficult to obtain effective segmentation thresholds, resulting in unsatisfactory image segmentation results for intracranial hemorrhage CT images. Summary of the Invention

[0009] Therefore, it is necessary to provide a method, apparatus, medium, and device for segmenting CT images of intracranial hemorrhage to address the aforementioned technical problems.

[0010] The present invention adopts the following technical solution: This invention provides a method for segmenting CT images of intracranial hemorrhage. First, the CT image of the patient's intracranial hemorrhage is acquired, converted to grayscale, and denoised to obtain a grayscale image and a denoised image. A two-dimensional histogram is constructed based on the pixel pairs of the denoised image and the grayscale image. Then, multiple thresholds are randomly generated for threshold segmentation of the two-dimensional histogram as individuals in the search population to initialize the search population. The fitness value of each individual in the search population is determined based on the Rényi entropy of the two-dimensional histogram after segmentation based on each individual in the search population. Finally, the optimal and second-best individuals in the search population are determined based on their fitness values, and the African Vulture optimization algorithm is used to further optimize each individual. The algorithm optimizes each individual and updates the best individual after optimization. Then, based on the current best individual, it updates each individual using the Barebones mechanism, retaining those with better fitness values ​​before and after the update, and updating the best individual again. Based on the current best individual, it optimizes each individual using an orthogonal learning mechanism, retaining those with better fitness values ​​before and after optimization. Finally, it re-determines the best and second-best individuals in the search population and iteratively optimizes them using the African vulture optimization algorithm, the Barebones mechanism, and the orthogonal learning mechanism. The two-dimensional histogram is then segmented based on multiple thresholds corresponding to the best individuals after multiple rounds of iterative optimization.

[0011] This invention provides a segmentation device for CT images of intracranial hemorrhage, comprising: The acquisition module is used to acquire CT images of the patient's intracranial hemorrhage, convert them to grayscale, and denoise them to obtain grayscale images and denoised images; a two-dimensional histogram is constructed based on the pixel pairs of the grayscale images and denoised images; The initialization module is used to randomly generate multiple thresholds for thresholding the two-dimensional histogram as individuals in the search population to initialize the search population. Based on the Rényi entropy of the two-dimensional histogram after partitioning each individual in the search population, the fitness value of each individual in the search population is determined. The first optimization module is used to determine the best and second-best individuals in the search population based on their fitness values, so as to optimize each individual using the African vulture optimization algorithm, and update the best individual after individual optimization. The second optimization module is used to update each individual based on the current best individual using the Barebones mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the update and updating the best individual; and to optimize each individual based on the current best individual using the orthogonal learning mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the optimization. The threshold segmentation module is used to redetermine the best and second-best individuals in the search population for iterative optimization using the African vulture optimization algorithm, the Barebones mechanism, and the orthogonal learning mechanism. It performs threshold segmentation on the two-dimensional histogram based on multiple thresholds corresponding to the best individuals after multiple rounds of iterative optimization.

[0012] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for segmenting intracranial hemorrhage CT images.

[0013] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned method for segmenting intracranial hemorrhage CT images.

[0014] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: This invention proposes a threshold segmentation method for intracranial hemorrhage CT images. It integrates the Barebones mechanism and orthogonal learning mechanism into the African Vulture optimization algorithm. Gaussian mutation expands the search space within the search population corresponding to the segmentation threshold, enhancing the algorithm's ability to escape local optima. The orthogonal learning mechanism promotes resource sharing among individuals in the search population corresponding to the segmentation threshold, increasing the search diversity of the segmentation threshold. Through multiple rounds of iterative optimization, the optimal individual that maximizes the Rényi entropy after dividing the 2D histogram based on the threshold is obtained and used as the optimal segmentation threshold for the intracranial hemorrhage CT image. This invention improves population diversity and convergence speed in the segmentation threshold optimization of intracranial hemorrhage CT images, thereby improving the image segmentation effect. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic flowchart of a method for segmenting CT images of intracranial hemorrhage provided by the present invention; Figure 2 A schematic diagram of a BR-MTIS process provided by the present invention; Figure 3 This is a schematic diagram of a segmentation device for intracranial hemorrhage CT images provided by the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0017] Current research shows that AI assistance significantly improves the accuracy of interpreting brain CT scan results for both non-radiologists and radiologists. Table 1 illustrates some results of CT images used in this invention for assisting in the diagnosis of intracranial hemorrhage.

[0018] Table 1. Several results of CT imaging in the auxiliary diagnosis of intracranial hemorrhage

[0019] Inspired by various natural phenomena, MOA optimization techniques are widely applied in MTIS. For example, one research scheme proposed a method based on SSA (Salivatapetus Optimization Algorithm), which achieves independent segmentation of the skull, trunk, and lower limbs in skeletal imaging images by maximizing the Tsallis entropy value. Another research scheme significantly improved the optimization performance of the original PSO algorithm by integrating techniques such as dynamic parameters, gain sharing strategies, random position updates, and vertical crossover mutation, thus proposing the GSRPSO algorithm. Subsequently, researchers combined this algorithm with two-dimensional Kapur entropy to achieve efficient segmentation of cervical cancer pathology images. Yet another research scheme proposed an improved multiverse optimization algorithm (MVO), named MOLMVO. This method successfully achieved segmentation of prostate cancer pathology images by employing multi-population topology and orthogonal learning techniques, combined with Rény entropy. Still another research scheme improved the RIME algorithm by adopting a cooperative adaptive hunting and dispersed foraging strategy, and combined it with two-dimensional Kapur entropy to propose the CDRIME-MTIS method for COVID-19 X-ray image segmentation.

[0020] Other research proposals have developed an improved algorithm, CBQMVO, which successfully achieved efficient segmentation of breast cancer pathology images. Further research has improved the initialization strategy of the Ant Colony Optimization (ACO) algorithm and introduced a Levy flight mechanism to enhance its optimization capabilities. This improved algorithm was applied to Otsu's method for image segmentation, successfully segmenting brain MRI images. Still other research has integrated PSO and SSA techniques, combining them with Otsu's method to achieve segmentation of diabetic retinal vascular images. This method enhances the visibility of major vessels through Gaussian filters and employs top-cap transformation techniques to improve the clarity of peripheral vessels. Table 2 illustrates some MTIS studies in this invention.

[0021] Table 2 Some MTIS studies based on mechanism of action

[0022] AVOA, proposed by Abdollahzadeh et al. in 2021, is a model of multi-modal optimization (MOA) that has gained widespread attention and application in multiple fields. Other research projects include the Binary African Vulture Optimization Algorithm (BAVOA) and its application to solving engineering and combinatorial optimization problems; the introduction of three strategies into AVOA—chaotic elite anti-learning, simplex method, and differential evolution—resulting in CESDAVO and its application to feature selection engineering; the IAVOA algorithm and its application to the shop floor scheduling problem; improvements to the AVOA algorithm through quantum double-chain encoding, quantum rotation gates, and elite mutation, and its application to the production scheduling problem; the HWEAVOA algorithm and its advantages in complex multimodal problems verified using benchmark function sets; an enhanced version of AVOA to improve the classification ability of multilayer perceptrons on cancer datasets; and the successful segmentation of brain MRI and dermoscopy images using an algorithm that minimizes generalized cross-entropy (MGCE). While the AVOA algorithm performs well in various optimization problems, it still faces challenges in complex scenarios, including insufficient population diversity, difficulty escaping local optima, and slow convergence speed. Furthermore, the "no free lunch" theorem states that no single algorithm can solve all problems. Therefore, continuous improvement of the AVOA algorithm is crucial for segmentation scenarios involving intracranial hemorrhage CT images.

[0023] To assist physicians in diagnosing intracranial hemorrhage (ICH), this invention proposes a MTIS method called BR-MTIS, specifically designed for processing ICH CT images. This method is based on BOAVOA and Rény entropy. The proposed BOAVOA is an improved algorithm based on AVOA, integrating the Barebones mechanism (BM) and the orthogonal learning mechanism (OLM) into AVOA. Gaussian mutations in the BM mechanism expand the search space within the population, enhancing the algorithm's ability to escape local optima; while the OLM mechanism promotes resource sharing among population particles, increasing population diversity. Subsequently, the BOAVOA algorithm iteratively calculates the maximum Rény entropy, thereby determining the optimal threshold combination for multi-threshold segmentation of ICH CT images.

[0024] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram of a segmentation method for CT images of intracranial hemorrhage according to the present invention, which specifically includes the following steps: S101: Acquire CT images of the patient's intracranial hemorrhage, convert them to grayscale, and denoise them to obtain grayscale and denoised images; construct a two-dimensional histogram based on the pixel pairs of the grayscale and denoised images.

[0026] S102: Generate multiple thresholds for thresholding the two-dimensional histogram as individuals in the search population to initialize the search population. Determine the fitness value of each individual in the search population based on the Rényi entropy of the two-dimensional histogram after partitioning each individual in the search population.

[0027] S103: Determine the best and second-best individuals in the search population based on fitness values ​​to optimize each individual using the African vulture optimization algorithm, and update the best individual after individual optimization.

[0028] S104: Update each individual based on the current best individual using the Barebones mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the update, and update the best individual; optimize each individual based on the current best individual using the orthogonal learning mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the optimization.

[0029] S105: Re-determine the best and second-best individuals in the search population for iterative optimization using the African vulture optimization algorithm, the Barebones mechanism, and the orthogonal learning mechanism. Then, perform threshold segmentation on the two-dimensional histogram based on multiple thresholds corresponding to the best individuals after multiple rounds of iterative optimization.

[0030] For ease of explanation, the following description focuses solely on the server as the executing entity. The server mentioned in this invention can be a server set up on a business platform, or a device such as a desktop computer or laptop computer capable of executing the solution of this invention.

[0031] In the process of multi-threshold image segmentation, it is usually necessary to optimize the segmentation threshold, which involves determining the fitness. In one or more embodiments of the present invention, the fitness value in the search can be determined based on the Rényi entropy of the segmentation threshold.

[0032] Rényi entropy is essentially a type of information entropy, a commonly used information metric. Therefore, the MTIS algorithm based on Rényi entropy in this invention judges the quality of the threshold set by evaluating the total information content of different sub-images after segmentation: the larger the Rényi entropy value, the more total information content, and the better the quality of the threshold set.

[0033] For the patient's original CT images of intracranial hemorrhage, the threshold set in MTIS can consist of simple one-dimensional data, utilizing only the image's standard gray-level histogram information. Taking Rényi entropy as an example, specifically assuming the image gray-level value range is 0 to 255, a threshold containing... A threshold set of thresholds: According to the theory of MTIS, this Each threshold divides the grayscale information of the image into three parts. Areas: Calculate the total Rényi entropy. The overall formula is shown in equation (1), where Indicates grayscale value The number of pixels, Indicates the first region The total number of pixels (calculated by formula (2)), item Indicates the area The Rényi entropy. Similarly, Indicates the first Sub-regions The total number of pixels, and This represents the Rényi entropy of that region. Finally, Indicates the first each region The total number of pixels, This is the Rényi entropy of that region.

[0034] (1).

[0035] (2).

[0036] This method is simple and fast to implement, but it only considers the grayscale information of each pixel in the image and lacks spatial context (i.e., it only focuses on the pixels themselves and ignores the relationships between them). Therefore, this method has low segmentation accuracy on some images and is relatively sensitive to noise. In 1989, Abutealb et al. introduced the concept of a two-dimensional histogram: by associating the grayscale value of a specific pixel with the average value of its surrounding neighborhood, a more informative image representation is created. Experiments have shown that MTIS based on this two-dimensional histogram has better segmentation performance than methods relying on standard histograms, especially exhibiting stronger noise resistance in images with high signal-to-noise ratios.

[0037] The noise resistance of the two-dimensional histogram proposed by Abutaleb mainly relies on a simple averaging of pixel neighborhoods. This averaging technique assumes that all neighboring pixels are equally important, but because it cannot effectively distinguish the importance of different pixels, it often leads to poor preservation of edge features. Conversely, medical images have complex features and are significantly affected by non-uniform noise, making direct use of local means for denoising limited. A more effective denoising method is nonlocal means (NLM). NLM can not only effectively reduce noise but also preserve edge details better, thus having a significant advantage in medical imaging. Therefore, this invention will use the original grayscale image and the image processed by NLM to jointly construct a two-dimensional histogram and use it as input to the Rényi entropy.

[0038] Assume the grayscale image of the original image is denoted as... Pixel values ​​are used The image obtained after nonlocal means (NLM) denoising is shown below. Pixel values ​​are used This indicates that the pixel pair Represents a set of joint pixels. Based on pixel pairs. The Rényi entropy of the constructed two-dimensional histogram is shown in equation (3), where This indicates that the pixel value in the original image is And the corresponding denoised image pixel value The number of pixels. In one or more embodiments of the present invention, the parameter It will be set to 0.5.

[0039] (3).

[0040] Furthermore, in one or more embodiments of the present invention, the search population is initially optimized based on the African Vulture Optimization Algorithm (AVOA): AVOA (African Vulture Optimization Algorithm) is a swarm intelligence optimization algorithm based on the foraging behavior of African vultures, proposed by Abdollahzadeh et al. in 2021. This algorithm simulates the foraging strategies of vultures in nature, utilizing group cooperation and information sharing mechanisms to search for the optimal solution.

[0041] African vultures are divided into three categories, each exhibiting different characteristics during foraging: the strongest vultures possess excellent hunting abilities, easily capturing food and often dominating resource competition; the second strongest vultures, while not as powerful as the former, still possess a certain hunting ability and can maintain a place in the competition; the weakest vultures are relatively lacking in strength and ability, and usually rely on the scraps of other vultures to survive.

[0042] In their natural environment, when food is spotted, vultures will fly towards their companions who have successfully caught prey. During this process, fierce conflicts and competition may erupt between different types of vultures over food. Hungry, weaker vultures, in order to survive, become more aggressive. They will gather around healthy vultures, exhausting and trapping them through continuous pressure, ultimately stealing the food that should belong to the healthy vultures. This strategy demonstrates their wisdom in the struggle for survival. Furthermore, vultures exhibit unique flight behavior, primarily employing a spiral flight strategy, which is effective in searching for food in vast airspaces. This flight pattern allows vultures to maintain stability at high altitudes while quickly adjusting their direction to find food or avoid potential threats.

[0043] AVOA Implementation Steps: The core of the AVOA algorithm comprises five steps, which will be described in detail in this invention. First, this invention defines the population size as... , dimension ,remember Representative population (i.e.) It is The matrix), with a maximum number of iterations. The current iteration number is .

[0044] Step 1: Initialize the population: The first population was generated using a random initialization strategy. Specifically, multiple thresholds for thresholding the two-dimensional histogram can be generated as individuals in the search population to initialize the search population, as shown in equation (4). In the matrix In the diagram, each row represents a candidate solution (i.e., the position of a vulture in a flock of vultures).

[0045] (4).

[0046] Step 2: Population grouping: In AVOA, the population is divided into two groups, referred to as Group 1 and Group 2. Each group has its own evolutionary direction. Vultures in Group 1 evolve based on the best vulture (the best individual) in the entire population, while vultures in Group 2 evolve based on the second best vulture (the second-best individual) in the population. The specific evolutionary direction of the vultures is determined by the following formula: (5).

[0047] In the formula, This represents the best vulture in the current population, while It refers to the second-best vulture. and These are two constants between 0 and 1, and their sum equals 1. (Based on a probability set) A roulette wheel selection strategy is used to determine Take still Then establish the current evolutionary direction of the vulture. .if and This means the algorithm will focus entirely on accelerating the convergence of the entire population; all vultures will quickly move towards the current optimal solution to rapidly find the global optimum. However, this can also lead the algorithm into local optima, limiting its exploration capabilities. Conversely, if and The algorithm emphasizes maintaining population diversity, allowing vultures to explore the solution space more randomly, rather than focusing excessively on a single optimal solution. This helps avoid premature convergence and getting trapped in local optima, enabling the algorithm to search for potential better solutions in a broader search space, thereby improving overall performance. In this invention, the invention will... and The values ​​were set to 0.8 and 0.2 respectively. This setting ensures that the algorithm maintains a certain speed during the optimization process, while fully protecting the diversity of the population and achieving a good balance between convergence and exploration.

[0048] Step 3: Calculate the hunger rate: In nature, when a vulture is extremely hungry, it faces the problem of lacking the strength to fly long distances in search of food. This situation forces the vulture to adopt a more aggressive strategy to obtain food. At this time, these hungry vultures become very aggressive, often approaching other vultures that are eating, attempting to circle around them to steal their food. On the other hand, if a vulture is not too hungry, it can conserve enough strength and stamina to fly longer distances in search of food. In this case, the vulture's behavior may be more cautious, choosing to forage in more favorable environments rather than directly competing fiercely with other individuals. Therefore, a vulture's foraging behavior is primarily influenced by its state of hunger. The hunger rate is calculated based on its state of hunger using a formula.

[0049] (6) In the formula, It represents a constant. , and These are random numbers in the ranges of 0 to 1, -1 to 1, and -2 to 2, respectively. This represents the vulture's state of hunger, when When vultures are hungry, they may exhibit more aggressive behavior in order to obtain food; conversely, when... At this point, it means the vulture is no longer hungry and may adopt a more cautious foraging strategy. Hunger rate This directly determines the vulture's subsequent foraging behavior. When This means that the vulture has enough energy to explore more food resources, so the AVOA will enter the exploration phase; when If this happens, it indicates that the vulture is low on energy and will forage for food near its current location instead of searching extensively for new food sources.

[0050] Step 4: Exploration Phase when At this time, the vulture is in the exploration phase. During this phase, the vulture is relatively less hungry and can fly freely, circling in the air to better observe its surroundings. This flight not only expands its field of vision but also allows it to effectively search for potential food sources. By observing from high altitudes, the vulture can spot animal carcasses or other possible food resources on the ground and react quickly. Equation (7) describes this process.

[0051] (7).

[0052] (8).

[0053] In the formula, It is obtained from equation (5) and represents the reference vulture of the current vulture; while It can then be calculated using equation (6). It is a preset constant. The distance between the current vulture and its reference vulture is measured and calculated using equation (8). and These are the upper and lower bounds of the solution space, respectively, which have been pre-defined. The remaining variables... , , and It is a random number between 0 and 1.

[0054] Step 5: Development Phase when When the value is between 0.5 and 1, the vulture enters the first mining stage. The evolutionary pattern at this time is defined by equation (9).

[0055] (9).

[0056] (10).

[0057] (11).

[0058] (12).

[0059] In the formula, and The meaning and calculation formula are the same as before. This is another way to measure the distance between the current vulture and the reference vulture, obtained by equation (10). Calculate using equation (11), and Then it is calculated using equation (12). Similar to equation (7), the control parameters... It's pre-defined. Variable , , and It is a random number between 0 and 1. When Between 0 and 0.5, the vulture enters the second developmental stage, at which point the evolutionary pattern is defined by equation (13).

[0060] (13).

[0061] (14).

[0062] (15).

[0063] (16).

[0064] In the above formula, It is a stochastic process known as Levy flight, and its calculation method is shown in Equation (16). Levy flight is a special type of random walk model, often used to simulate animal foraging behavior, optimization algorithms, and other natural phenomena. In Equation (16), the variables are... and Defined as containing A vector of n elements, the values ​​of which are randomly generated and range from 0 to 1. It is a preset constant with a value of 1.5. It is used as a scaling factor in the calculation of Levy flight to adjust the size of the random step size, ensuring that the flight process has a certain degree of randomness while avoiding excessive dispersion. The gamma function, found in the MATLAB toolbox, is a mathematical function widely used in probability theory and statistics, providing essential mathematical support for complex calculations.

[0065] After optimizing each individual using the African vulture optimization algorithm, to avoid the optimization process getting stuck in local optima, in one or more embodiments of this invention, the server can combine the Barebones mechanism to further optimize the optimized individuals. Specifically, the server can first determine the current best individual and its mean and standard deviation for each individual in the search population. Then, it can generate multiple thresholds corresponding to the individual based on a Gaussian distribution of the mean and standard deviation to update the individual.

[0066] The Barebones Mechanism (BM) was first introduced in the BBPSO algorithm developed by Kennedy in 2003. The standard Particle Swarm Optimization (PSO) algorithm has been theoretically proven to guide the trajectory of individuals by updating the average of their positions toward their individual optimal positions and the global optimal positions. That is, the global optimal solution and the individual optimal solution in the population can guide the trajectory of individuals, as shown in equation (17): (17).

[0067] in, and These are two learning factors (hyperparameters) in PSO. Indicates the first Individuals first The optimal position obtained in the generation, and This represents the best position in the entire population so far. In BBPSO, the velocity term in the standard PSO is removed, and the individual update formula is Equation (18): (18).

[0068] in, Indicates the first The updated results for each individual This represents a specific Gaussian distribution with a mean of 1 / 2. The standard deviation is This update mechanism is called the "Barebones mechanism" (BM). Experimental results show that BM can significantly improve the convergence speed and accuracy of the algorithm. This invention applies it to the expansion of the search space optimized by the African Vulture optimization algorithm in the segmentation threshold search, so as to enhance the ability of the threshold optimization algorithm to get rid of local optima.

[0069] Furthermore, to improve population diversity and enhance the convergence speed of the segmentation threshold optimization, in one or more embodiments of the present invention, the server can further optimize the search population based on an orthogonal learning mechanism. Specifically, the server can first obtain a pre-determined orthogonal array; then, for each individual in the search population, the search space corresponding to that individual is determined based on the current optimal solution and the maximum and minimum values ​​of the thresholds in each dimension; thereby, based on the correspondence between factors in the orthogonal array and the thresholds in each dimension of the individual, the factor level combination corresponding to each row of trials in the orthogonal array is mapped to the search space to obtain a preset number of trial solutions, and the fitness value of each trial solution is determined; then, based on the fitness value of each trial solution, the average response of the thresholds in each dimension at different levels is determined, the optimal level of each dimension threshold is determined and combined to obtain the optimal trial solution corresponding to that individual; finally, the individual and its corresponding trial solutions and the optimal trial solution with the best fitness value are taken as the optimized result of that individual.

[0070] Orthogonal Learning Mechanism (OLM): When the experiment is subjected to Factors have an influence, and each factor has When there are multiple levels (candidate values), obtaining the optimal combination through experimental testing becomes extremely expensive, and the number of combinations grows exponentially with the number of factors and levels. Orthogonal design is a solution that significantly reduces the number of experiments by selecting a representative set of experimental combinations, while effectively analyzing the impact of each factor and its interactions on the experimental results. One commonly used orthogonal array is... For example, as shown in equation (19): (19).

[0071] This table supports designs with up to 4 factors (F=4) and 3 levels per factor (Q=3). When there are only 3 levels, the 4th column can be ignored, requiring only 9 experiments. Some researchers have introduced orthogonal learning methods (OLM) to improve the overall performance of metaheuristic optimization algorithms. In the first stage of this algorithm, OLM provides important guiding information, making the search process more efficient; in the second stage, OLM is used to construct a feature vector pool, promoting information exchange within the population and greatly enhancing the collaborative evolutionary ability among individuals.

[0072] The following explains the specific process of the orthogonal learning mechanism based on the orthogonal arrays mentioned above: 1. Parameter boundary construction: In an OLM call, the search space for an individual is defined by the maximum and minimum values ​​of the thresholds for each dimension between the current individual and the globally optimal individual. For example, the total maximum value of the current individual and the globally optimal individual across all dimensions can be used as the upper bound of the search thresholds for each dimension, and the total minimum value across all dimensions can be used as the lower bound of the search thresholds for each dimension.

[0073] 2. Horizontal numbering → actual parameter value mapping: 2.1 If (Dimensional mapping): .

[0074] In the formula, For the mapped first The first experimental solution Dimension threshold, For the first The lower bound of the dimensional threshold search. For the first The upper bound of the dimensional threshold search. The orthogonal array is the Line number The value of the column, This represents the total number of threshold values ​​in an individual.

[0075] 2.2, if (Segmented mapping): Divide the dimensional interval into part: .

[0076] In the formula, For the first part For example, when there are 10 thresholds, the thresholds can be divided into segments of three, until all 10 thresholds are mapped to 4 segments.

[0077] Then the mapped first A test solution No. The threshold for the segment is: .

[0078] 3. Objective function evaluation: The fitness of each orthogonal experiment is: .

[0079] In the formula, This is the aforementioned function for calculating fitness values.

[0080] 4. Main Effects Analysis (Data Statistical Process): Construct the indicator matrix: , .

[0081] For each factor Each level Calculate the water-mean response: .

[0082] Therefore, the optimal level of each factor .

[0083] 5. Predicting the generation and re-evaluation of optimal solutions: The optimal level of each factor can be combined to obtain a vector: .

[0084] Mapping to the true parameters yields the optimal experimental solution for that individual: .

[0085] Unified comparison The output is obtained from 9 candidate solutions (9 rows of the orthogonal array corresponding to 9 experimental solutions and the optimal experimental solution corresponding to the predicted optimal combination): .

[0086] In the formula, express The optimal individual among the candidate solutions. This represents the fitness value of the optimal individual among the candidate solutions. Finally, the individual can be compared before and after the orthogonal learning mechanism to determine the optimal result as the optimized outcome for that individual.

[0087] In the original African vulture optimization algorithm, each individual generates only one candidate solution in a single update. This invention introduces an orthogonal learning mechanism (OLM), enabling each individual to perform structured sampling and main effects analysis within its local search interval. Using orthogonal arrays, the optimal combination of parameters for each dimension can be identified with only a finite number of additional objective function evaluations, forming a predicted optimal solution. This significantly enhances mining capabilities while maintaining exploratory nature. Therefore, the update process no longer relies on random trial and error, but becomes "statistically driven targeted mining," significantly improving the optimization algorithm's search depth, convergence speed, and ability to escape local traps.

[0088] The following details the improvements to AVOA by introducing OLM. AVOA improves by adjusting parameters. , , , , and To achieve a balance between the exploration and development phases, this invention effectively avoids the algorithm getting trapped in local optima. On the one hand, AVOA exhibits high efficiency under specific conditions or for certain problems; on the other hand, AVOA is highly sensitive to parameter settings, carries the risk of getting trapped in local minima, and its convergence speed still has significant room for improvement in some complex scenarios. To address these issues, this invention proposes an improved algorithm, BOAVOA, which combines BM and OLM to overcome the shortcomings of the original AVOA.

[0089] BM uses Gaussian sampling to update individuals, weights the current optimal solution, and introduces randomness, giving the algorithm a stronger ability to escape local minima and thus improving its robustness. After standard AVOA completes one population evolution, BOAVOA uses BM to update the current position. By expanding, new candidate solutions are obtained. Then evaluate. and The fitness values ​​are then used to retain the better ones. Next, OLM is used for further optimization. An orthogonal array as shown in equation (19) can be used here. Among the factors It is 4, level The value is 3. For ease of description, this orthogonal array is called . The number of experiments was The optimal position in the experimental solution is generated through mapping. Then a greedy strategy is adopted to evaluate... and The solution with better fitness is retained. The pseudocode for BOAVOA (Algorithm 1) is as follows: Input parameter: Population size Problem Dimensions Initial population Maximum number of iterations Current iteration number Other parameters.

[0090] Output: Optimal solution.

[0091] Initialize population ; While ( ; Population assessment The fitness of; Find the optimal vulture and suboptimal vulture ; For ; Calculation reference vulture (Formula (5)); Calculate the hunger rate (Formula (6)); If ; Update according to formula (7) ; End; If ; Update according to formula (9) ; End; If ; Update according to formula (13) ; End; / / Use BM to expand the current position; generate via BM; Assess fitness ; If ; ; = ; End; / / Use OLM optimization; generate via OLM; Assess fitness ; If ; ; = ; End; Update the optimal vulture and suboptimal vulture ; End for; = +1; End while.

[0092] The BR-MTIS proposed in this invention serves as the basic framework of this invention, using Rényi entropy as the fitness function to evaluate the segmentation threshold quality in MTIS. The proposed BOAVOA aims to solve the key problem of searching for the optimal segmentation threshold. Specifically, this invention uses Rényi entropy (Formula (3)) as the objective function of the optimization algorithm. The system pre-sets the threshold level corresponding to the number of thresholds (NT), thereby determining the dimension of the optimization problem as dim=2NT (the setting of 2 is due to the input data using a two-dimensional histogram). Through iterative calculation, the threshold set is gradually optimized to maximize the objective function, thereby obtaining a higher Rényi entropy. Figure 2 This is a schematic diagram of a BR-MTIS process according to the present invention.

[0093] based on Figure 1 This invention proposes a threshold segmentation method for intracranial hemorrhage CT images. This method integrates the Barebones mechanism and orthogonal learning mechanism into the African Vulture optimization algorithm. Gaussian mutation expands the search space in the search population, enhancing the threshold optimization algorithm's ability to escape local optima. The orthogonal learning mechanism promotes resource sharing among particles in the search population, increasing its diversity. Through multiple rounds of iterative optimization, the optimal individual that maximizes the Rényi entropy after threshold-based partitioning of the 2D histogram is obtained, serving as the optimal segmentation threshold for the intracranial hemorrhage CT image. This invention improves population diversity and convergence speed in threshold optimization for intracranial hemorrhage CT image segmentation, thereby enhancing the image segmentation performance.

[0094] This invention proposes an improved version of the AVOA algorithm—BOAVOA. This algorithm expands the search space through BM and utilizes OLM to promote information sharing among the population, thus effectively solving the common problems of slow convergence and getting trapped in local optima in complex optimization problems. On the IEEE CEC2020 benchmark function, the proposed BOAVOA is compared with eight other optimization methods, verifying the enhanced optimization capabilities of BOAVOA. This invention also proposes a MTIS method specifically for CT images of cerebral hemorrhage, BR-MTIS. This method is based on Rény entropy and utilizes BOAVOA to optimize threshold quality. Segmentation experiments were conducted on CT images of cerebral hemorrhage. The results of PSNR, SSIM, and FSIM indicators clearly show that BOAVOA performs significantly better in segmentation ability compared to other segmentation methods. Non-parametric statistical tests such as WSRT and FT were used to statistically evaluate the experimental results at a significance level of 0.05.

[0095] When applying the segmentation method for intracranial hemorrhage CT images provided by this invention, it is not necessary to consider... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0096] To verify the effectiveness of the proposed BOAVOA algorithm, this invention uses the IEEE CEC2020 benchmark function for comparative experiments and employs the Wilcoxon Sign-Rank Test (WSRT) and Friedman Test (FT) for statistical analysis. Furthermore, this invention conducts segmentation experiments on intracranial hemorrhage CT images with five threshold levels (10, 15, 20, 25, and 30), and evaluates the results using three full-reference image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Feature Similarity (FSIM).

[0097] On the same benchmark function, the BOAVOA algorithm demonstrates superior performance compared to similar algorithms, further confirming its advantages in the field of global optimization. Segmentation of CT images of cerebral hemorrhage was successfully achieved based on the BR-MTIS algorithm. Experimental results show that the proposed BR-MTIS algorithm has significant advantages in multiple evaluation metrics, including mean absolute error (PSNR), root mean square error (SSIM), and free mean square error (FSIM). This method provides an effective tool for the clinical diagnosis of cerebral hemorrhage and has broad application prospects.

[0098] The above describes a method for segmenting intracranial hemorrhage CT images according to one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for segmenting intracranial hemorrhage CT images, such as... Figure 2 As shown.

[0099] Figure 2 A schematic diagram of a segmentation device for intracranial hemorrhage CT images provided by the present invention includes: The acquisition module 201 is used to acquire the patient's intracranial hemorrhage CT image, convert it to grayscale and denoise it to obtain a grayscale image and a denoised image; and to construct a two-dimensional histogram based on the pixel pairs of the grayscale image and the denoised image. The initialization module 202 is used to generate multiple thresholds for thresholding the two-dimensional histogram as individuals in the search population to initialize the search population, and to determine the fitness value of each individual in the search population based on the Rényi entropy of the two-dimensional histogram after partitioning based on each individual in the search population. The first optimization module 203 is used to determine the best and second-best individuals in the search population based on the fitness value so as to optimize each individual using the African vulture optimization algorithm, and to update the best individual after individual optimization; The second optimization module 204 is used to update each individual based on the current best individual through the Barebones mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the update and updating the best individual; and to optimize each individual based on the current best individual through the orthogonal learning mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the optimization. The threshold segmentation module 205 is used to redetermine the best and second-best individuals in the search population for iterative optimization using the African vulture optimization algorithm, the Barebones mechanism, and the orthogonal learning mechanism. The two-dimensional histogram is segmented based on multiple thresholds corresponding to the best individuals after multiple rounds of iterative optimization.

[0100] Specific limitations regarding the segmentation device for intracranial hemorrhage CT images can be found in the above description of the segmentation method for intracranial hemorrhage CT images, and will not be repeated here. Each module in the aforementioned segmentation device for intracranial hemorrhage CT images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0101] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided segmentation method for CT images of intracranial hemorrhage.

[0102] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1The provided segmentation method for CT images of intracranial hemorrhage.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for segmenting CT images of intracranial hemorrhage, characterized in that, include: The patient's intracranial hemorrhage CT images were acquired, converted to grayscale, and denoised to obtain grayscale and denoised images; a two-dimensional histogram was constructed based on the pixel pairs of the grayscale and denoised images. Multiple thresholds for thresholding the two-dimensional histogram are randomly generated as individuals in the search population to initialize the search population. The fitness value of each individual in the search population is determined based on the Rényi entropy of the two-dimensional histogram after partitioning each individual in the search population. The best and second-best individuals in the search population are determined based on their fitness values, and each individual is optimized using the African vulture optimization algorithm. The best individual is then updated after optimization. Based on the current best individual, each individual is updated using the Barebones mechanism, retaining the cases where the fitness value of each individual is better before and after the update, and updating the best individual; based on the current best individual, each individual is optimized using the orthogonal learning mechanism, retaining the cases where the fitness value of each individual is better before and after the optimization; The optimal and second-best individuals in the search population are re-determined for iterative optimization using the African vulture optimization algorithm, the Barebones mechanism, and the orthogonal learning mechanism. The two-dimensional histogram is then segmented based on multiple thresholds corresponding to the optimal individuals after multiple rounds of iterative optimization.

2. The segmentation method for intracranial hemorrhage CT images as described in claim 1, characterized in that, The step of determining the fitness value of each individual in the search population based on the Rényi entropy obtained by partitioning the individual individuals in the search population using a two-dimensional histogram specifically includes: The Rényi entropy of the two-dimensional histogram, after partitioning the individual population in the search population, is determined by the following formula and used as the fitness value of each individual in the search population: ； in, For each individual, a threshold pair. This is the threshold set for thresholding grayscale images. This is the threshold set for threshold segmentation of the denoised image. Rényi entropy is the result of dividing a two-dimensional histogram based on individuals. To adjust the parameters, The pixel value in the grayscale image of intracranial hemorrhage CT scan And the corresponding pixel value in the denoised image The number of pixels, The second 2D histogram is the result of dividing individuals. Sub-regions The total number of pixels.

3. The segmentation method for intracranial hemorrhage CT images as described in claim 1, characterized in that, The optimization of each individual using the African vulture optimization algorithm specifically includes: The search population is divided into two groups according to the roulette wheel selection strategy using the following formula: the first group takes the best individual in the search population as the evolutionary direction, and the second group takes the second best individual in the search population as the evolutionary direction: ， ， ； Based on the evolutionary direction of each individual in the grouped search population, the African vulture optimization algorithm is used to optimize each individual; in, For the first The evolutionary direction of each individual To search for the optimal individual in the population. To search for the second-best individual in the population, For the first Individuals based on probability groups The value to be selected using the roulette wheel selection strategy.

4. The segmentation method for intracranial hemorrhage CT images as described in claim 1, characterized in that, The process of updating each individual based on the current optimal individual using the Barebones mechanism specifically includes: For each individual in the search population, determine the current best individual and the mean and standard deviation of that individual; The update of an individual is completed by generating multiple thresholds corresponding to that individual based on a Gaussian distribution of the mean and the standard deviation.

5. The segmentation method for intracranial hemorrhage CT images as described in claim 1, characterized in that, The process of optimizing each individual using an orthogonal learning mechanism based on the current best individual, while retaining the cases where the fitness value of each individual is better before and after optimization, specifically includes: Obtain a pre-determined orthogonal array; For each individual in the search population, the search space corresponding to that individual is determined based on the current optimal solution and the maximum and minimum values ​​of the thresholds in each dimension. Based on the correspondence between factors in the orthogonal array and the threshold values ​​of each dimension in the individual, the factor level combination corresponding to each row of the experiment in the orthogonal array is mapped to the search space to obtain a preset number of experimental solutions, and the fitness value of each experimental solution is determined. Based on the fitness values ​​of each experimental solution, determine the average response of each dimension threshold at different levels, determine the optimal level of each dimension threshold and combine them to obtain the optimal experimental solution corresponding to the individual. The optimal fitness value among the individual, its corresponding experimental solutions, and the best experimental solution is taken as the result of the individual optimization.

6. The segmentation method for intracranial hemorrhage CT images as described in claim 1, characterized in that, The process of acquiring and denoising the patient's intracranial hemorrhage CT images to obtain grayscale and denoised images specifically includes: The patient's intracranial hemorrhage CT images were acquired, and the intracranial hemorrhage CT images were converted to grayscale to obtain grayscale images. The grayscale images were then denoised using the nonlocal mean method to obtain denoised images.

7. A segmentation device for CT images of intracranial hemorrhage, characterized in that, include: The acquisition module is used to acquire the patient's intracranial hemorrhage CT images, convert them to grayscale and denoise them to obtain grayscale images and denoised images; and to construct a two-dimensional histogram based on the pixel pairs of the grayscale images and denoised images. The initialization module is used to randomly generate multiple thresholds for thresholding the two-dimensional histogram as individuals in the search population to initialize the search population. Based on the Rényi entropy of the two-dimensional histogram after partitioning each individual in the search population, the fitness value of each individual in the search population is determined. The first optimization module is used to determine the best and second-best individuals in the search population based on their fitness values, so as to optimize each individual using the African vulture optimization algorithm, and update the best individual after individual optimization. The second optimization module is used to update each individual based on the current best individual using the Barebones mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the update and updating the best individual; and to optimize each individual based on the current best individual using the orthogonal learning mechanism, retaining the cases where the fitness values ​​of each individual are better before and after the optimization. The threshold segmentation module is used to redetermine the best and second-best individuals in the search population for iterative optimization using the African vulture optimization algorithm, the Barebones mechanism, and the orthogonal learning mechanism. It performs threshold segmentation on the two-dimensional histogram based on multiple thresholds corresponding to the best individuals after multiple rounds of iterative optimization.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method as described in any one of claims 1 to 6.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 6.

Images