Medical image denoising system based on dynamic Harris eagle optimization

Abstract

The invention discloses a medical image denoising system based on dynamic Harris eagle optimization, relates to the technical field of image denoising, and aims to solve the problem of low medical diagnosis accuracy caused by a large number of irrelevant features in existing medical images. And a population is generated by calculating symmetric uncertainty, so that the overall quality of the population is improved by using a Harlisia eagle algorithm, and a high-quality feature subset is obtained. And finally, through wavelet inverse transformation, obtaining a denoised medical image. According to the technical scheme of the invention, the denoised medical image greatly reduces the influence of irrelevant features in the medical image, and improves the accuracy of medical diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising, and in particular to a medical image denoising system based on dynamic Harris Hawk optimization. Background Art

[0002] With the continuous advancement of medical technology, medical image analysis plays an increasingly important role in medical diagnosis and treatment. Medical image analysis refers to the process of processing and analyzing medical imaging data (such as X-rays, CT scans, MRI scans, and ultrasound) to help doctors diagnose diseases, assess disease progression, and evaluate treatment effectiveness. However, medical images contain a large number of irrelevant features that can seriously interfere with diagnostic accuracy. Summary of the Invention

[0003] The purpose of the present invention is to provide a medical image denoising system based on dynamic Harris Hawk optimization to address the problem that a large number of irrelevant features exist in existing medical images, which leads to low medical diagnosis accuracy.

[0004] The technical solution adopted by the present invention to solve the above technical problems is:

[0005] A medical image denoising system based on dynamic Harris Eagle optimization includes a feature extraction module, a ranking table generation module and a denoising module;

[0006] The feature extraction module specifically performs the following steps:

[0007] Acquiring a medical image containing a label and preprocessing the medical image, wherein the preprocessing includes normalization and noise reduction, and then extracting frequency domain features of the preprocessed medical image using wavelet transform to obtain multiple medical feature vectors;

[0008] The sorting table generation module specifically performs the following steps:

[0009] Obtain the symmetric uncertainty corresponding to each two medical feature vectors respectively, retain the corresponding feature index, and generate the SU ranking table according to the frequency of occurrence of the feature index;

[0010] The denoising module specifically performs the following steps:

[0011] According to the SU sorting table, the feature indexes are grouped, and the position vectors of the eagle group are initialized according to the grouping results to perform Harris eagle optimization. Then, the fitness value of each eagle is obtained, and the medical feature subset is determined according to the fitness value of each eagle. The medical feature subset is then subjected to inverse wavelet transform to obtain the denoised medical image.

[0012] Furthermore, the Harris Hawk optimization process is specifically as follows:

[0013] Step 1: Calculate the prey energy E, when When the eagles enter the exploration phase, they search for prey and potential medical features. When the eagles enter the exploitation phase, the eagles quickly approach the prey and dig out key medical features;

[0014] Entering the exploration phase, if an individual has not been updated for m consecutive iterations, quantum perturbation is triggered, and each position vector in the eagle group is updated to search for high-quality medical image features. The updated position vector is expressed as:

[0015]

[0016] in, 、 and is the disturbance coefficient, , , the value range is (0.1, 0.2), () is the Levy flight function, which provides a position perturbation with a step size of (1,2). is the optimal position vector of the previous iteration;

[0017] If the quantum perturbation is not triggered, the eagle group uses a random exploration method to search for prey and updates the position vector, which is expressed as:

[0018]

[0019] in, and are the current and next iteration numbers of Harris Hawk, For the Eagle The random position of the generation, For prey The position of the generation, For Harris Hawk The average position of the generation, and are the lower and upper bounds of the search space, 、 、 、 、 A random number between 0 and 1;

[0020] Entering the development phase, the eagles use four hunting methods to chase prey and search for medical characteristics, specifically:

[0021] First generate a random number r, ,when , When the prey has enough energy to escape, the hawks approach the prey in a soft encirclement manner. First, calculate the jumping distance J when the prey escapes. is a random number between (0,1):

[0022]

[0023] Then calculate the difference between the prey's position and the current position :

[0024]

[0025] in, is the difference between the eagle's current position and the prey's position, For prey location;

[0026] Utilize the difference between the prey's location and the current location Calculate the target position of the eagle after iteration, expressed as:

[0027]

[0028] Where t represents the current generation;

[0029] when , When the prey does not have enough energy to escape, the hawks use a hard encirclement method to approach the prey, which can be expressed as:

[0030]

[0031] when , When the prey is exhausted, the eagles use a rapid dive to softly surround it. First, the gradual dive distance Y is calculated. The gradual dive distance Y is expressed as:

[0032]

[0033] Calculate the wandering distance of the current position vector based on the progressive dive distance Y , the wandering distance of the current position vector Expressed as:

[0034]

[0035] Where LF is the Levy flight constant, S is the random vector;

[0036] After that, the eagle’s position vector is updated and relevant medical features are searched near the feature subset, which can be expressed as:

[0037]

[0038] in, represents the fitness function, when When the Harris Hawk's position is updated to Y, When , the position of Harris Hawk is updated to Z;

[0039] when , When the energy of the prey drops to a lower level, Harris's hawk adopts a hard encirclement strategy combined with progressive rapid dives, which can be expressed as:

[0040]

[0041] in, represents the average position of the eagle group;

[0042] Step 2: Calculate the fitness value of each eagle;

[0043] Step 3: After sorting the fitness values ​​of each eagle, generate a fitness value table. Then, determine whether there are individuals with the same fitness value in the fitness value table. If not, execute step 7. If so, dynamically update the position vector of the eagle group and execute step 4.

[0044] Step 4: Get the feature subset corresponding to each individual with the same fitness value and calculate the Hamming distance D between the two feature subsets H If there are more than two individuals with the same fitness value, the Hamming distance D is calculated for each pair. H ;

[0045] Step 5: Convert the Hamming distance into a similarity measure. When the similarity measure S H =1, perform dynamic update and go to step 6, otherwise go to step 7;

[0046] Step 6: Randomly remove one of the two eagle position vectors, and add a new position vector according to the initialization method in the denoising module, and execute step 7;

[0047] Step 7: According to the fitness value table, get the current optimal fitness value and compare the current optimal fitness value with the historical optimal fitness value. If the current optimal fitness value is greater than the historical optimal fitness value, use the current optimal fitness value to replace the historical optimal fitness value. Otherwise, do not replace it. If replacement is made, the feature subset x corresponding to the position vector of the current optimal individual is best That is, the selected feature subset;

[0048] Step 8: Determine whether the current number of iterations has reached the maximum number of iterations. If so, stop the iteration and the selected feature subset obtained by the current iteration will be the final selected feature subset. Otherwise, repeat steps 1 to 7 to iterate.

[0049] At each iteration, there is a probability Skip calculating the prey's energy E and go directly to the exploration phase;

[0050] When the current iteration number t is in the range of (0, 0.25T), the eagle swarm adopts the soft rime iteration strategy in the Rime algorithm to diffuse, so that the eagle swarm quickly spreads throughout the feature search space, thereby performing a large-scale search for medical features, which can be expressed as:

[0051]

[0052] Where T represents the maximum number of iterations, and for the index and dimension index, is the new location after SR search, For the The best individual position of dimension, It is a random number ranging from -1 to 1, used to control the direction of particle movement. is an environmental variable, and its change is related to the number of iterations. is a random number in [0,1], and are the current number of iterations and the maximum number of iterations, respectively. is the weight parameter used to control The range of change, is the angle parameter, is a random number in the range of [0,1], the cohesion coefficient of the particle and Synergistic effect, which determines the diffusion behavior of particles;

[0053] When the current iteration number t is not in the range of (0, 0.25T), the diffusion strategy is not used;

[0054] When the number of iterations t reaches half of the maximum number of iterations T, after the exploration phase, a diversion exploration strategy is introduced, which performs the following steps:

[0055] Step a: Divide the eagle group with fitness values ​​in the bottom 30%;

[0056] Step b: From the last 30% of hawks, select the last 20% of hawks again, without considering the prey energy, and randomly cross each coordinate in the position vector. For the last 10% of hawks, add a new position vector according to the initialization method in the denoising module.

[0057] Step c: For the remaining 70% of the eagle group, use the steps of the exploration phase to explore.

[0058] Furthermore, the prey energy E is expressed as:

[0059]

[0060] Among them, E0 is the escape energy of the initial state, which changes randomly within the numerical range of (-1,1).

[0061] Furthermore, the probability Expressed as:

[0062]

[0063] in, 、 is the regulating factor.

[0064] Furthermore, the fitness value is expressed as:

[0065]

[0066] in, and is the weight coefficient, is the classification error rate, is the number of features selected in the current feature subset, is the total number of features in the original dataset.

[0067] Furthermore, the Hamming distance DH is expressed as:

[0068]

[0069] Where d is the dimension of the feature vector, is the index of the vector dimension, is the element difference function.

[0070] Furthermore, the similarity measure is expressed as:

[0071] .

[0072] Furthermore, the wavelet transform is a Symlet wavelet basis.

[0073] Furthermore, the symmetric uncertainty is expressed as:

[0074]

[0075] in, is the mutual information of two medical feature vectors, and is the information entropy of two medical feature vectors.

[0076] Furthermore, the feature indexes are grouped as follows:

[0077] According to the SU ranking table, the top 30%, 30% to 70% and the bottom 30% are grouped.

[0078] The beneficial effects of the present invention are:

[0079] This application encodes labeled medical images into feature vectors, generates a population by calculating symmetric uncertainty, and then uses the Harris Eagle algorithm to improve the overall quality of the population, resulting in a high-quality feature subset. Finally, an inverse wavelet transform is performed to obtain a denoised medical image. The denoised medical image achieved by this technical solution significantly reduces the influence of irrelevant features in the medical image, improving the accuracy of medical diagnoses. DETAILED DESCRIPTION

[0080] It should be noted that, unless there is any conflict, the various embodiments disclosed in this application can be combined with each other.

[0081] Specific embodiment 1: This embodiment describes a medical image denoising system based on dynamic Harris Hawk optimization. The system first preprocesses the input image. The preprocessing steps are as follows:

[0082] Obtain medical images containing labels, perform normalization and denoising processing, and reduce basic noise interference.

[0083] The image is encoded and its frequency domain features are extracted through wavelet transform, using the Symlet wavelet basis. Multi-level wavelet decomposition is performed, generating a high-dimensional feature vector for each image data point. After multiple processing steps, the dataset generates multiple sets of feature vectors for subsequent processing.

[0084] After pre-processing, the system specifically performs the following steps:

[0085] Step (1): Through the above steps, we get a feature set with a size of K, a feature dimension of D, and instance labels. Where X represents a feature instance without a label, f(X) is the optimization target, D is the dimension of the dataset, and k is the index of the corresponding feature. In the feature subset, x k A value of 1 indicates that the feature with index k is selected, and a value of 0 indicates that it is not selected.

[0086] (1)

[0087] Step (2): First, discretize the continuous features and check for missing values ​​to facilitate the subsequent calculation of information measures. Let the upper bound of the eigenvalue be U and the lower bound be L. The threshold criterion for the eigenvalue is TH = (U - L) / 2.

[0088] Step (3): Calculate the information entropy H(X) of a single feature vector. The more critical medical image features have higher information entropy.

[0089] (2)

[0090] Step (4): For a dataset with feature dimension D, calculate the joint entropy H(X,Y) of every two features (each feature vector contains features with D dimensions).

[0091] (3)

[0092] Step (5): Calculate the mutual information I(X;Y) of the two features. Feature pairs with larger mutual information contain high-value potential medical information.

[0093] (4)

[0094] Step (6): Calculate the symmetric uncertainty SU(X;Y) corresponding to these two features.

[0095] (5)

[0096] Step (7): Generate an information table containing the symmetric uncertainty corresponding to each pair of features, with a numerical range of [0, 1], and each value on the diagonal is 1.00. Keep the index of the feature pair and generate a SU ranking table. According to the SU value, find the index with the highest frequency and sort the feature index accordingly.

[0097] Step (8): Divide the feature index into three groups, with the division criteria being the first 30%, 30% to 70%, and the last 30% of SU frequencies.

[0098] Step (9): Initialize the position vector of the eagle group according to the grouping criteria. Among them, the feature index with a higher SU ranking has a greater probability of generating a larger initial position. S1, S2, and S3 are the medical image features ranked in the top 30%, 30% to 70%, and bottom 30%, respectively, and s represents a specific feature.

[0099] (6)

[0100] Step (10): Calculate the prey energy E and determine whether the algorithm enters the exploration phase or the development phase. When the eagles enter the exploration phase, they search for prey and search for potential features over a large area; when When the HHO enters the development phase, the eagles quickly approach the prey and explore key medical features in a small area.

[0101] (7)

[0102] Where E0 is the initial escape energy, which varies randomly within the range of (-1, 1). T represents the maximum number of iterations, and t represents the current number of iterations.

[0103] Step (11): The algorithm enters the exploration phase and adopts a quantum perturbation strategy. If the individual has not been updated for m consecutive iterations, it means that the algorithm has stagnated and quantum perturbation is triggered. M is an integer ranging from 3 to 5. A quantum leap is generated to generate a leapfrog solution and adjust the position value. .

[0104] Update each position vector in the eagle group and try to search for high-quality medical image features.

[0105] (8)

[0106] in, is the perturbation coefficient, the value range is (0.1, 0.2), () is the Levy flight function, which provides a position perturbation with a step size of (1,2). , , is the optimal position vector of the previous iteration.

[0107] Step (12): If the algorithm does not use the quantum perturbation strategy during exploration, the hawk group uses the original exploration method to search for prey. At this time, it is necessary to update the other position vectors in the hawk group.

[0108] (8)

[0109] in and are the current and next iteration numbers of Harris Hawk, It's the eagle The random position of the generation, It's the prey The position of the generation, It's Harris Hawk in The average position of the generation, and are the lower and upper bounds of the search space, 、 、 、 、 is a random number.

[0110] Step (13): In the early stage of iteration, the current iteration number t is in the range of (0, 0.25T). The eagle swarm adopts the soft rime iteration strategy (SR) in the Rime algorithm to spread, so that the eagle swarm quickly spreads across the feature search space, thereby performing a large-scale search for medical features.

[0111] (9)

[0112] Among them, i and j are index and dimension index, The new location after the SR search. represents the best individual position in the jth dimension. It is a random number ranging from -1 to 1, used to control the direction of particle movement. It is an environment variable whose change is related to the number of iterations. Randomly set in the range [0,1]. and Represent the current iteration number and the maximum iteration number respectively. Is a random number in the range [0,1]. The cohesion coefficient of the particle and The synergistic effect determines the diffusion behavior of particles.

[0113] Step (14): If the algorithm enters the development phase, the hawks use four hunting methods to chase prey and search for medical features. First, a random number r is generated to decide whether to adopt a surprise attack strategy.

[0114] Step (15): When , When the prey has enough energy to escape, the hawks approach the prey using a soft encirclement method. The algorithm first calculates the jump distance J when the prey escapes. is a random number between (0,1).

[0115] (10)

[0116] Step (16): Next, calculate the difference between the prey's position and the current position.

[0117] (11)

[0118] in, is the difference between the eagle's current position and the prey's position, For prey location.

[0119] Step (17): Calculate the target position of the eagle after iteration. t represents the current generation.

[0120] (12)

[0121] Step (18): When , When the prey does not have enough energy to escape, the eagles will use a hard encirclement method to approach the prey.

[0122] (13)

[0123] Step (19): When , When the prey is exhausted, it still has a chance to escape. The hawks use a rapid dive to softly surround the prey. First, they calculate the progressive dive distance Y. Then, they calculate Z based on Y.

[0124] (14)

[0125] Step (20): Calculate the wandering distance of the current position vector.

[0126] (15)

[0127] Where LF is the Levy flight constant and S is a random vector.

[0128] Step (21): Afterwards, the eagle's position vector is updated in two ways, attempting to search for relevant medical features near the feature subset.

[0129] (16)

[0130] Represents the fitness function, which is used to calculate the fitness value corresponding to the current position vector. When , the Harris Hawk moves to position Y; otherwise, if , update the position of Harris Hawk to Z.

[0131] Step (22): When , When the prey's energy level drops to a low level and it is very vulnerable, Harris's hawks use a hard encirclement strategy combined with gradual, rapid dives.

[0132] (17)

[0133] in, represents the average position of the hawk group, and the rest of the process is similar to the soft encirclement strategy of gradual rapid dive.

[0134] Step (23): In order to balance exploration and exploitation, an adaptive balancing strategy based on space reconstruction is adopted. The algorithm divides the search space into high fitness areas and low fitness areas according to the distribution of fitness values.

[0135] Step (24): In order to ensure the stage balance of the algorithm, the next generation has a certain probability Skip the prey energy calculation phase and go directly to the exploration phase. This prevents the algorithm from over-exploiting in the late iterations. To explore the probability, 、 Regulatory factors 、 The value ranges are (3,5) and (0.2,0.5), Adjust the initial exploration probability, Controls the rate at which the probability decreases.

[0136] (18)

[0137] Step (25): When the iteration number t reaches half of the maximum iteration number T, the diversion exploration strategy is introduced in the exploration phase. This strategy is executed after the original exploration strategy.

[0138] Step (26): First, divide the hawk population with fitness values ​​in the bottom 30%. Among the hawk population ranked in the bottom 20%, regardless of prey energy, directly perform random crossover on each coordinate in the position vector to increase the diversity of the population.

[0139] Step (27): For the last 10% of the eagle population, use the method in step (9), that is, the initialization method based on symmetric uncertainty, to directly generate new individuals and strengthen exploration.

[0140] Step (28): The first 70% of the eagles use the normal exploration method of step (11).

[0141] Step (29): The algorithm begins to evaluate the quality of the currently selected feature subset. First, the upper and lower bounds of the position vector and the threshold are determined. Let the upper bound be UB and the lower bound be LB. The threshold is (UB-LB) / 2. The threshold is used to determine whether to select the feature subset corresponding to the position vector.

[0142] Step (30): Prepare to train the KNN classifier by first dividing the training set into a feature instance matrix X and a label instance vector Y.

[0143] Step (31): Based on the selected feature subset, the training set and the validation set retain the selected feature part.

[0144] Step (32): Set the KNN parameters, set the K value to 5, input the processed training set, and train the KNN classifier.

[0145] Step (33): Use the trained KNN classifier to predict on the processed validation set and calculate the classification accuracy acc.

[0146] Step (34): Calculate the classification error rate The higher the classification accuracy, the better the classification effect of the feature subset is, and the key feature information of the medical image is obtained.

[0147] Step (35): Let s be the size of the feature subset and n be the total size of the medical image features, and calculate the ratio of the selected features .

[0148] Step (36): Calculate the fitness value of each eagle using the parameters obtained in steps (34) and (35).

[0149] (19)

[0150] Step (37): Generate a fitness table based on the fitness value of each eagle and sort it.

[0151] Step (38): If there is no individual with the same fitness value in the fitness value table, then jump to step (42); otherwise, dynamically update the position vector of the eagle group.

[0152] Step (39): Calculate the Hamming distance D between the two feature subsets H .

[0153] (20)

[0154] Convert the Hamming distance to a similarity measure, where L is the size of the current feature subset.

[0155] (twenty one)

[0156] Step (40): Determine whether individuals with the same fitness value have selected exactly the same feature subset. H =1, indicating that the feature subsets are exactly the same, and dynamic update is performed; otherwise, jump to step (42).

[0157] Step (41): The algorithm randomly removes one of the two position vectors and uses the initialization method based on the symmetric uncertainty table in step (9) to add a new position vector. Continue to step (42).

[0158] Step (42): Compare the currently calculated optimal fitness value with the historical optimal value. If the current optimal value is greater than the historical optimal value, replace the historical optimal value with the current optimal fitness value.

[0159] Step (43): If a replacement is made, record the information of the current optimal individual and the feature subset x represented by its position vector best That is, the selected feature subset.

[0160] Step (44): Set the maximum number of iterations T = 200. If the current iteration number t = T, then stop the iteration. Otherwise, return to step (10) and repeat the iteration.

[0161] Step (45): After the iteration, x best The feature subset selected by the algorithm has strong classification capabilities and low dimensionality, making it easier to mine and analyze the feature data. Furthermore, the reduced dimensionality of the features reduces the efficiency and cost of training machine learning models.

[0162] Step (46): Set the experimental parameters for medical feature selection, including the information of the comparison algorithm, as shown in Table 1. This algorithm is a feature selection method based on dynamic Harris Hawks optimization algorithm (DHHO)

[0163] Table 1 Parameter information of this algorithm and the comparison algorithm

[0164]

[0165] All comparison algorithms in this application use the KNN classifier with K=5 to evaluate the classification ability of feature subsets. The data set is divided into 80%~20%, with 80% divided into the training set and 20% divided into the validation set.

[0166] Step (47): To further explore the algorithm’s feature selection effect, eight datasets were selected from the UCI medical feature dataset and the ASU high-dimensional dataset to conduct medical feature selection experiments. The dataset information is shown in Table 2.

[0167] Table 2 Information of feature selection dataset

[0168]

[0169] Step (48): The experimental results of this application are shown in Table 3. This experiment includes the ablation version of the algorithm. DHHO-A is the version of the algorithm that only introduces quantum perturbations, DHHO-B is the version of the algorithm that only introduces the adaptive balancing strategy, and DHHO-C is the version of the algorithm that only introduces the initialization strategy based on symmetric uncertainty. The experimental results are averaged after 10 runs, and the best results for each dataset are highlighted in bold.

[0170] Table 3 Average fitness results of the algorithms

[0171]

[0172] Step (49): In order to verify the feasibility and experimental effect of the algorithm, the experimental results in Table 3 were analyzed. DHHO achieved the lowest fitness value on all eight data sets, indicating that the quality of its selected feature subset was the highest. Compared with BHHO, the fitness value of DHHO decreased by 8.61% to 46.30%, and this result was obtained on the DARWIN and TOX_171 data sets, respectively. Compared with DHHO-B and DHHO-C, DHHO-A had the highest improvement on BHHO, with a fitness value reduction of 13.59% on the Lymphoma data set and a fitness value reduction of 28.08% on the Arrhythmia data set. This result shows that the algorithm has good optimization and feature selection capabilities after improvement, and also proves the effectiveness of the algorithm improvement.

[0173] Step (50): To verify the algorithm’s optimization and feature selection capabilities, the classification accuracy of the algorithm during the iteration process was recorded. The experiment was run 10 times, and the results were averaged, as shown in Table 4.

[0174] Table 4 Average classification accuracy of DHHO and comparison algorithms (%)

[0175]

[0176] Step (51): In order to verify the algorithm's optimization and feature selection capabilities, the experimental results in Table 4 are analyzed. Compared with BHHO, the average classification accuracy of DHHO increased by 2.76% to 14.97% on the seven datasets. Among them, the improvement on the TOX_171 and Prostate_GE high-dimensional datasets was 11.27% and 8.31% respectively, proving that DHHO still has good feature selection capabilities in high-dimensional scenarios. Although DHHO's performance on the DARWIN dataset is not as good as OGGWO, the results are not much different. DHHO's performance on other datasets is the best. Compared with other algorithms, DHHO still has an advantage. On the Colon data, compared with OGGWO and DSCA, the average classification accuracy of DHHO increased by 5.36% and 3.72%. On the Prostate high-dimensional dataset, compared with DMBA, the average classification accuracy of DHHO increased by 3.68% and 3.32%. The above analysis proves that the classification ability of the feature subset selected by DHHO is strong and is generally better than the cutting-edge comparison algorithm.

[0177] Step (52): In order to verify the feature dimensionality reduction capability of the algorithm, the dimensionality reduction rate of the algorithm during the iteration process is recorded. ,The experiment was run 10 times and the results were averaged, as shown in Table 5.

[0178] Table 5 Average dimension reduction rate (%) of DHHO and comparison algorithms

[0179]

[0180] Step (53): In order to further compare the effectiveness of the algorithm improvement, the experimental results in Table 5 are analyzed. Compared with BHHO, the average dimensionality reduction rate of DHHO on the seven datasets increased by 10.02%~34.98%. Among them, the improvement on the TOX_171 and Prostate_GE high-dimensional datasets was 6.89% and 10.87% respectively, indicating that DHHO has a strong dimensionality reduction ability in the high-dimensional feature selection problem. Although DHHO's performance on the TOX_171 dataset is not as good as DSCA, the results are not much different. DHHO's dimensionality reduction performance on other datasets is the best. Compared with other algorithms, DHHO's dimensionality reduction ability is still leading. On the Arrhythmia dataset, compared with OGGWO, DHHO's average dimensionality reduction rate increased by 5.09% and 3.76%. On the Prostate high-dimensional dataset, compared with DMBA and DSCA, DHHO's average dimensionality reduction rate increased by 9.86% and 2.35%. The above analysis proves that DHHO has a high dimensionality reduction rate for feature datasets and its dimensionality reduction capability is better than that of advanced comparative algorithms.

[0181] Step (54): In order to reflect the comprehensive advantages of this algorithm, a comprehensive analysis of the experiments of this method is conducted. Each experimental result is obtained by taking the average value of 10 experiments. From Table 3, compared with BHHO, the fitness value of DHHO is reduced by 8.61%~46.30% overall. From Table 4, compared with BHHO, the average classification accuracy of DHHO is improved by 2.76%~14.97% on 7 data sets. From Table 5, compared with BHHO, the average dimension reduction rate of DHHO is improved by 10.02%~34.98% on 7 data sets. The above data proves the effectiveness of the improvement of DHHO, and the improvement strategy can improve the feature selection ability of the algorithm from multiple angles. Compared with other advanced algorithms, the feature subset selected by DHHO is also in a leading position, indicating that the quality of its selected feature subset is high and can effectively solve the feature selection problem.

[0182] Step (55): Relevant feature data is retained, and the remaining features are subjected to soft threshold processing to reduce the influence of noise or irrelevant features by shrinking the wavelet coefficients to zero. The medical image is reconstructed through inverse wavelet transform.

[0183] Step (56): By combining the wavelet transform with the improved optimization algorithm through the above process, the medical image is denoised, and a medical image with lower noise can be quickly generated. The medical image generated by this application eliminates a large number of invalid features while slightly losing clarity and retaining features with original information, which is beneficial for subsequent diagnosis and analysis of medical conditions.

[0184] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solutions of the present invention and cannot be used to limit the scope of protection. Any minor changes made based on the claims and description of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. A medical image denoising system based on dynamic Harris Hawk optimization, characterized by It includes feature extraction module, sorting table generation module and denoising module; The feature extraction module specifically performs the following steps: Acquiring a medical image containing a label and preprocessing the medical image, wherein the preprocessing includes normalization and noise reduction, and then extracting frequency domain features of the preprocessed medical image using wavelet transform to obtain multiple medical feature vectors; The sorting table generation module specifically performs the following steps: Obtain the symmetric uncertainty corresponding to each two medical feature vectors respectively, retain the corresponding feature index, and generate the SU ranking table according to the frequency of occurrence of the feature index; The denoising module specifically performs the following steps: According to the SU sorting table, the feature indexes are grouped, and the position vectors of the eagle group are initialized according to the grouping results to perform Harris eagle optimization. Then, the fitness value of each eagle is obtained, and the medical feature subset is determined according to the fitness value of each eagle. The medical feature subset is then subjected to inverse wavelet transform to obtain the denoised medical image.

2. A medical image denoising system based on dynamic Harris Hawk optimization according to claim 1, characterized in that The Harris Hawk optimization process is specifically as follows: Step 1: Calculate the prey energy E, when When the eagles enter the exploration phase, they search for prey and potential medical features. When the eagles enter the exploitation phase, the eagles quickly approach the prey and dig out key medical features; Entering the exploration phase, if an individual has not been updated for m consecutive iterations, quantum perturbation is triggered, and each position vector in the eagle group is updated to search for high-quality medical image features. The updated position vector is expressed as: in, 、 and is the disturbance coefficient, , , the value range is (0.1, 0.2), () is the Levy flight function, which provides a position perturbation with a step size of (1,2). is the optimal position vector of the previous iteration; If the quantum perturbation is not triggered, the eagle group uses a random exploration method to search for prey and updates the position vector, which is expressed as: in, and are the current and next iteration numbers of Harris Hawk, For the Eagle The random position of the generation, For prey The position of the generation, For Harris Hawk The average position of the generation, and are the lower and upper bounds of the search space, 、 、 、 、 A random number between 0 and 1; Entering the development phase, the eagles use four hunting methods to chase prey and search for medical characteristics, specifically: First generate a random number r, ,when , When the prey has enough energy to escape, the hawks approach the prey in a soft encirclement manner. First, calculate the jumping distance J when the prey escapes. is a random number between (0,1): Then calculate the difference between the prey's position and the current position : in, is the difference between the eagle's current position and the prey's position, For prey location; Utilize the difference between the prey's location and the current location Calculate the target position of the eagle after iteration, expressed as: Where t represents the current generation; when , When the prey does not have enough energy to escape, the hawks use a hard encirclement method to approach the prey, which can be expressed as: when , When the prey is exhausted, the eagles use a rapid dive to softly surround it. First, the gradual dive distance Y is calculated. The gradual dive distance Y is expressed as: Calculate the wandering distance of the current position vector based on the progressive dive distance Y , the wandering distance of the current position vector Expressed as: Where LF is the Levy flight constant, S is the random vector; After that, the eagle’s position vector is updated and relevant medical features are searched near the feature subset, which can be expressed as: in, represents the fitness function, when When the Harris Hawk's position is updated to Y, When , the position of Harris Hawk is updated to Z; when , When the energy of the prey drops to a lower level, Harris's hawk adopts a hard encirclement strategy combined with progressive rapid dives, which can be expressed as: in, represents the average position of the eagle group; Step 2: Calculate the fitness value of each eagle; Step 3: After sorting the fitness values ​​of each eagle, generate a fitness value table. Then, determine whether there are individuals with the same fitness value in the fitness value table. If not, execute step 7. If so, dynamically update the position vector of the eagle group and execute step 4. Step 4: Get the feature subset corresponding to each individual with the same fitness value and calculate the Hamming distance D between the two feature subsets H If there are more than two individuals with the same fitness value, the Hamming distance D is calculated for each pair. H ; Step 5: Convert the Hamming distance into a similarity measure. When the similarity measure S H =1, perform dynamic update and go to step 6, otherwise go to step 7; Step 6: Randomly remove one of the two eagle position vectors, and add a new position vector according to the initialization method in the denoising module, and execute step 7; Step 7: According to the fitness value table, get the current optimal fitness value and compare the current optimal fitness value with the historical optimal fitness value. If the current optimal fitness value is greater than the historical optimal fitness value, use the current optimal fitness value to replace the historical optimal fitness value. Otherwise, do not replace it. If replacement is made, the feature subset x corresponding to the position vector of the current optimal individual is best That is, the selected feature subset; Step 8: Determine whether the current number of iterations has reached the maximum number of iterations. If so, stop the iteration and the selected feature subset obtained by the current iteration is the final selected feature subset. Otherwise, repeat steps 1 to 7 for iteration. At each iteration, there is a probability Skip calculating the prey's energy E and go directly to the exploration phase; When the current iteration number t is in the range of (0, 0.25T), the eagle swarm adopts the soft rime iteration strategy in the Rime algorithm to diffuse, so that the eagle swarm quickly spreads throughout the feature search space, thereby performing a large-scale search for medical features, which can be expressed as: Where T represents the maximum number of iterations, and for the index and dimension index, is the new location after SR search, For the The best individual position of dimension, It is a random number ranging from -1 to 1, used to control the direction of particle movement. is an environmental variable, and its change is related to the number of iterations. is a random number in [0,1], and are the current number of iterations and the maximum number of iterations, respectively. is the weight parameter used to control The range of change, is the angle parameter, is a random number in the range of [0,1], the cohesion coefficient of the particle and Synergistic effect, which determines the diffusion behavior of particles; When the current iteration number t is not in the range of (0, 0.25T), the diffusion strategy is not used; When the number of iterations t reaches half of the maximum number of iterations T, after the exploration phase, a diversion exploration strategy is introduced, which performs the following steps: Step a: Divide the eagle group with fitness values ​​in the bottom 30%; Step b: From the last 30% of hawks, select the last 20% of hawks again, without considering the prey energy, and randomly cross each coordinate in the position vector. For the last 10% of hawks, add a new position vector according to the initialization method in the denoising module. Step c: For the remaining 70% of the eagle group, use the steps of the exploration phase to explore.

3. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 2, characterized in that: The prey energy E is expressed as: Among them, E0 is the escape energy of the initial state, which changes randomly within the numerical range of (-1,1).

4. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 3, characterized in that: The probability Expressed as: in, 、 is the regulating factor.

5. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 4, characterized in that: The fitness value is expressed as: in, and is the weight coefficient, is the classification error rate, is the number of features selected in the current feature subset, is the total number of features in the original dataset.

6. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 5, characterized in that: The Hamming distance DH is expressed as: Where d is the dimension of the feature vector, is the index of the vector dimension, is the element difference function.

7. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 6, characterized in that: The similarity measure is expressed as: 。 8. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 1, characterized in that: The wavelet transform is a Symlet wavelet basis.

9. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 1, characterized in that: The symmetric uncertainty is expressed as: in, is the mutual information of two medical feature vectors, and is the information entropy of two medical feature vectors.

10. The medical image denoising system based on dynamic Harris Hawk optimization according to claim 1, characterized in that: The grouping of feature indexes is specifically as follows: According to the SU ranking table, the top 30%, 30% to 70% and the bottom 30% are grouped.

Images