A fusion image quality evaluation method based on bipolar hesitant fuzzy preference ordering
Patent Information
- Application Number
- CN202611415864.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-14
- Publication Date
- 2026-10-23
AI Technical Summary
[0005]本发明的目的在于提供一种基于双极犹豫模糊偏好排序的融合图像质量评估方法，解决了现有技术中多模态医学图像融合质量评估难以同时处理双向不确定性、评价犹豫性及决策者风险偏好的问题
[0037]本发明的一种基于双极犹豫模糊偏好排序的融合图像质量评估方法，本发明针对CT-MRI、PET-CT、PET-MRI等多模态医学图像融合场景，将模糊集理论、行为决策科学与多属性决策分析方法有机结合，首先通过模糊化处理与均值比较机制将各融合算法对应的客观评价指标值转化为正负隶属度信息，构建犹豫双极模糊集决策矩阵，有效保留了医学图像融合评价过程中存在的双向不确定性和评价犹豫性；随后利用双极犹豫模糊熵自动确定各评价指标的客观权重，减少了人为经验赋权带来的主观偏差，提高了评价结果的客观性与一致性；进而综合运用累计前景理论价值函数刻画决策者面对收益和损失时的不对称风险偏好，通过概率权重函数将客观属性权重转化为反映心理风险的决策权重，并基于对偶偏差构建高斯型偏好函数计算算法间的偏好指数，最终通过正超越流、负超越流和净流实现融合算法的排序优选，从而能够有效处理多模态医学图像融合质量评价中的模糊性、不确定性和评价犹豫性，同时充分考虑医学专家在收益与风险感知方面的心理偏好，为多模态医学图像融合算法的客观评价、性能比较及优选提供了一种科学、可靠且具有良好可解释性的决策框架，有助于提升融合结果在临床辅助诊断、病灶检测、精准治疗规划等医学应用中的实用价值，并可进一步推广至其他医学影像智能决策与多属性评价场景。
Smart Images
Figure CN122966643A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a method for evaluating the quality of fused images based on bipolar hesitant fuzzy preference ranking. Background Technology
[0002] Currently, with the rapid development of artificial intelligence and medical imaging technology, multimodal medical image fusion has become an important technical means to improve the accuracy of disease diagnosis, lesion localization, and treatment planning. Different medical image modalities (such as CT, MRI, PET, and SPECT) can provide complementary information from multiple dimensions such as anatomical structure, tissue features, and functional metabolism, while different fusion algorithms have their own advantages in terms of structural detail preservation, edge information enhancement, functional information expression, and artifact suppression. Therefore, how to objectively and comprehensively evaluate the quality of fusion results and select the most suitable fusion scheme for clinical application from among many algorithms has become a key issue in the research and application of multimodal medical image fusion.
[0003] Existing traditional medical image fusion evaluation methods typically rely on objective indicators such as mutual information (MI), structural similarity (SSIM), peak signal-to-noise ratio (PSNR), and edge preservation index for independent evaluation or weighted synthesis. However, these methods still have the following shortcomings: First, the dimensions, distribution characteristics, and clinical significance of different evaluation indicators vary significantly, and simple weighting cannot accurately reflect the complex nonlinear relationships between indicators; second, under different patient datasets, different lesion types, and multiple experimental conditions, the evaluation performance of fusion results often exhibits strong fuzziness and uncertainty, and a single deterministic numerical value is insufficient to fully characterize the fluctuation characteristics of the evaluation results; third, when evaluating fusion quality, clinical experts and researchers have different cognitive preferences regarding the benefits and risks of "better than average" and "worse than average," and traditional evaluation models usually ignore this psychological factor in decision-making.
[0004] Existing multi-attribute decision-making methods often use deterministic numerical values or simple fuzzy sets to represent the quality assessment of medical image fusion, which makes it difficult to simultaneously characterize positive and negative preference information, evaluation hesitation, and decision-maker risk attitude during the fusion evaluation process. Summary of the Invention
[0005] The purpose of this invention is to provide a fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking, which solves the problem in the existing technology that multimodal medical image fusion quality assessment is difficult to simultaneously handle bidirectional uncertainty, evaluation hesitancy and decision-maker risk preference.
[0006] To achieve the above objectives, this invention provides a method for evaluating the quality of fused images based on bipolar hesitant fuzzy preference ranking, comprising the following steps:
[0007] A hesitant bipolar fuzzy set decision matrix is constructed through fuzzification, averaging, and comparison operations.
[0008] The uncertainty and fuzziness of each algorithm under each attribute are calculated by bipolar hesitant fuzzy entropy, and then the objective weights of each attribute are obtained by normalization.
[0009] The reference point is determined by equalization, the gains and losses are calculated, the prospect value is obtained by the cumulative prospect theoretical value function, and finally the duality deviation between algorithms is calculated.
[0010] A preference function based on a Gaussian distribution is constructed by calculating the mean and standard deviation of the dual bias.
[0011] The objective attribute weights are transformed into decision weights that reflect psychological risk through a probability weight function, and then the preference index between algorithms is calculated.
[0012] The net flow is obtained by calculating the positive and negative transcendental flows, and the final priority order of the fusion algorithm is obtained by sorting the net flows from largest to smallest.
[0013] The process of constructing a hesitant bipolar fuzzy set decision matrix through fuzzification, averaging, and comparison operations includes the following steps:
[0014] The blurring result is obtained by normalizing the original index values of each algorithm under each attribute and each image;
[0015] The average blur value of each algorithm for each attribute is obtained by averaging the blurring results of all images under each attribute for each algorithm.
[0016] Each fuzzification result is compared with the corresponding average fuzzy value. Those with values greater than or equal to the mean are assigned to the positive membership set, and those with values less than the mean are assigned to the negative membership set, thus obtaining bipolar hesitant fuzzy elements. Then, a bipolar fuzzy set decision matrix is constructed.
[0017] The step of calculating the uncertainty and fuzziness of each algorithm under each attribute by using bipolar hesitant fuzzy entropy, and then obtaining the objective weights of each attribute through normalization, includes the following steps:
[0018] The hesitant fuzzy entropy is calculated for the positive membership set and the negative membership set using the sinusoidal entropy formula, respectively, to measure the uncertainty and fuzziness of all test images under each attribute for each algorithm.
[0019] The average entropy is obtained by averaging the entropy values of all algorithms under each attribute. Then, based on the principle that the smaller the entropy value, the greater the information content, the objective weight of each attribute is obtained through normalization calculation.
[0020] The process of determining a reference point through equalization, calculating gains and losses, obtaining the prospect value through the cumulative prospect theoretical value function, and finally calculating the duality deviation between algorithms includes the following steps:
[0021] By using the longest sequence length as a benchmark and repeating the maximum value at the end of the shorter sequence, the original data sequences of each algorithm are lengthened, and then the average comprehensive score of all algorithms under each attribute is calculated as a reference point.
[0022] The difference between the combined score of each algorithm and the reference point is used to obtain the gain or loss value. A positive value indicates that the algorithm is better than the average level, and a negative value indicates that the algorithm is worse than the average level.
[0023] The prospect value is obtained by performing nonlinear transformations on the gains and losses using the cumulative prospect theory value function. The gains part is transformed using a power function (concave function), and the losses part is transformed using a power function (convex function) and multiplied by the loss aversion coefficient.
[0024] The difference between the prospect values of the two algorithms is calculated to determine the duality of the two algorithms in terms of attributes.
[0025] The step of constructing a preference function based on a Gaussian distribution by calculating the mean and standard deviation of the dual biases includes the following steps:
[0026] The average value of the dual bias for each attribute is obtained by taking the arithmetic mean of the dual biases over all different algorithm pairs.
[0027] The standard deviation of the dual deviation for each attribute is obtained by taking the square root of the square mean of the difference between the dual deviation and the mean for all different algorithm pairs.
[0028] By constructing a piecewise function, the preference value is zero when the dual bias is negative, and a Gaussian exponential function with standard deviation as the parameter is used when the dual bias is non-negative, thus obtaining the degree to which the algorithm is superior to the algorithm in terms of attributes.
[0029] The process of converting objective attribute weights into decision weights reflecting psychological risk through a probability weight function, and then calculating the preference index between algorithms, includes the following steps:
[0030] By performing convex and concave transformations on the objective weights using the Tversky-Kahneman probability weight function, the psychological weights of decision-makers when making comparative advantage decisions and when making comparative disadvantage decisions can be obtained.
[0031] The preference function value is weighted and summed with the advantage psychological weight and the disadvantage psychological weight respectively to obtain the preference index and preference index of the algorithm relative to the algorithm.
[0032] The step of obtaining the net flow by calculating the positive and negative transcendental flows, and sorting the net flows from largest to smallest to obtain the final priority order of the fusion algorithm, includes the following steps:
[0033] The positive transcendental flow is obtained by averaging the preference index of the algorithm relative to all other algorithms.
[0034] The negative transcendental flow is obtained by averaging the preference indices of all other algorithms relative to the algorithm.
[0035] The net flow is obtained by subtracting the negative excess flow from the positive excess flow;
[0036] All candidate algorithms are sorted in descending order of net flow to obtain the final priority order. When net flows are equal, the order is determined by the positive overshoot flow level or the algorithms are kept in parallel.
[0037] This invention discloses a fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking. Targeting multimodal medical image fusion scenarios such as CT-MRI, PET-CT, and PET-MRI, this invention organically combines fuzzy set theory, behavioral decision science, and multi-attribute decision analysis methods. First, through fuzzification processing and mean comparison mechanisms, the objective evaluation index values corresponding to each fusion algorithm are transformed into positive and negative membership information, constructing a hesitant bipolar fuzzy set decision matrix, effectively preserving the bidirectional uncertainty and evaluation hesitancy present in the medical image fusion evaluation process. Then, the objective weights of each evaluation index are automatically determined using bipolar hesitant fuzzy entropy, reducing subjective bias caused by human experience-based weighting and improving the objectivity and consistency of the evaluation results. Finally, the cumulative prospect theory value function is comprehensively applied to characterize the decision-maker's reaction to gains and losses. This paper addresses the asymmetric risk preference problem by transforming objective attribute weights into decision weights reflecting psychological risk through a probability weight function. A Gaussian preference function is constructed based on dual bias to calculate the preference index between algorithms. Finally, positive transcendental flow, negative transcendental flow, and net flow are used to achieve the optimal ranking of fusion algorithms. This effectively handles the fuzziness, uncertainty, and evaluation hesitation in the quality evaluation of multimodal medical image fusion. It also fully considers the psychological preferences of medical experts in terms of benefit and risk perception. This provides a scientific, reliable, and well-interpretable decision framework for the objective evaluation, performance comparison, and optimization of multimodal medical image fusion algorithms. It helps improve the practical value of fusion results in clinical auxiliary diagnosis, lesion detection, and precision treatment planning, and can be further extended to other intelligent decision-making and multi-attribute evaluation scenarios in medical imaging. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0039] Figure 1This is a flowchart of a fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking provided by the present invention.
[0040] Figure 2 This is a flowchart of constructing a hesitant bipolar fuzzy set decision matrix through fuzzification, averaging, and comparison operations.
[0041] Figure 3 This is a flowchart that calculates the uncertainty and fuzziness of each algorithm under each attribute by using bipolar hesitant fuzzy entropy, and then obtains the objective weights of each attribute through normalization.
[0042] Figure 4 It is a flowchart that determines the reference point through equalization, calculates the gains and losses, obtains the prospect value through the cumulative prospect theoretical value function, and finally calculates the dual deviation between algorithms.
[0043] Figure 5 It is a flowchart that constructs a preference function based on a Gaussian distribution by calculating the mean and standard deviation of the dual biases.
[0044] Figure 6 It is a flowchart that transforms objective attribute weights into decision weights that reflect psychological risk through a probability weight function, and then calculates the preference index between algorithms.
[0045] Figure 7 This is a flowchart showing how to obtain the net flow by calculating the positive and negative transcendental flows, and then sorting the net flows from largest to smallest to obtain the final priority order of the fusion algorithm. Detailed Implementation
[0046] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.
[0047] Please see Figures 1 to 7 This invention provides a method for evaluating the quality of fused images based on bipolar hesitant fuzzy preference ranking, comprising the following steps:
[0048] S1 constructs a hesitant bipolar fuzzy set decision matrix through fuzzification, averaging, and comparison operations;
[0049] S11 normalizes the original index values of each algorithm under each attribute and each image to obtain the blurring result;
[0050] S12 calculates the average blur value of each algorithm for each attribute by averaging the blurring results of all images under each attribute for each algorithm.
[0051] S13 compares each fuzzification result with the corresponding average fuzzy value. Those with values greater than or equal to the average are assigned to the positive membership set, and those with values less than the average are assigned to the negative membership set, thus obtaining bipolar hesitant fuzzy elements, and then constructing a bipolar fuzzy set decision matrix.
[0052] Specifically, the set of dummy equipment options is as follows: (each) (representing the i-th fusion algorithm), there are m fusion algorithms in total. Each method is evaluated using multiple evaluation metrics, i.e., the attribute set is... (each) Let represent the j-th evaluation attribute. There are n evaluation attributes in total, and each fusion method is verified on K sets of source images (each set of source images generates one fused image under each fusion algorithm), numbered k=1,2,…,K. For the i-th fusion algorithm, the j-th evaluation attribute, and the k-th fusion result image, the original value measured using objective indicators is denoted as . ,and, .
[0053] For each evaluation attribute j, the evaluation results of each algorithm are fuzzed.
[0054]
[0055] in, This represents the original value of the i-th algorithm for the j-th attribute in the k-th image group. This represents the minimum value obtained by traversing all algorithms for the k-th image and the j-th attribute. This represents the maximum value obtained by traversing all algorithms for the j-th attribute in the k-th image group.
[0056] Then, for each evaluation attribute j, the mean of the evaluation results for each algorithm is calculated.
[0057]
[0058] Where K represents a total of K sets of images, The blurring result of the i-th fusion algorithm for the k-th image and the j-th attribute.
[0059] To obtain the positive and negative membership degrees for each scheme, the fuzzification result is subtracted from the mean.
[0060] if ,So ;
[0061] if ,So ;
[0062] in, The set representing positive membership degree. The set representing negative membership degree. This represents the blurring result of the i-th algorithm for the k-th image, the j-th evaluation attribute. This represents the mean of all blurring results for the j-th attribute and the i-th fusion algorithm.
[0063] Then we can obtain the bipolar hesitant fuzzy element at position (i, j) of the matrix:
[0064]
[0065] Therefore, the decision matrix of a bipolar fuzzy set can be represented as follows:
[0066]
[0067] S2 calculates the uncertainty and fuzziness of each algorithm under each attribute by using bipolar hesitant fuzzy entropy, and then obtains the objective weight of each attribute by normalization.
[0068] S21 calculates the hesitant fuzzy entropy for the positive membership set and the negative membership set respectively using the sinusoidal entropy formula, and measures the uncertainty and fuzziness of all test images under each attribute for each algorithm.
[0069] S22 obtains the average entropy by averaging the entropy values of all algorithms under each attribute, and then, based on the principle that the smaller the entropy value, the greater the information content, it obtains the objective weight of each attribute through normalization calculation.
[0070] Specifically, in order to measure the first The algorithm in the th... For each attribute, calculate the uncertainty and fuzziness degree of all test images (i.e., multiple values in the hesitant fuzziness set), and then calculate the hesitant fuzziness entropy.
[0071]
[0072] in, It is the length of the set of positive membership degrees (that is, in the attribute). Algorithm Among all the corresponding images, it falls into (Number of hesitant and ambiguous elements in the text) It is the length of the set of negative membership degrees (similarly, it is the length of the set of negative membership degrees). The number of hesitant and ambiguous elements in the text. For the algorithm index (the first) (a fusion algorithm) For attribute index, Pi This represents the sorted index. This represents the set of positive membership degrees, sorted in ascending order. The specific value of each element This represents the set of negative membership degrees, sorted in ascending order. The specific value of each element Used to symmetrically select the first element in the set. Xiaohe Di Larger elements are used to capture the overall symmetry within the interval.
[0073]
[0074] in, The algorithm calculated for 2.1 In attributes The entropy value under the condition, This represents the total number of alternative solutions (fusion algorithms). Then,
[0075]
[0076] in, For the attributes calculated above The average entropy, This indicates all attributes arrive of Summation, used as the normalized denominator, Indicates the first The final objective weights of each attribute result in the set of attribute weight values.
[0077] S3 determines the reference point through equalization, calculates the gains and losses, obtains the prospect value through the cumulative prospect theoretical value function, and finally calculates the duality deviation between algorithms.
[0078] S31 uses the longest sequence length as a benchmark and repeats the maximum value at the end of the shorter sequence to equalize the length of the original data sequence of each algorithm, and then calculates the average comprehensive score of all algorithms under each attribute as a reference point.
[0079] S32 calculates the difference between the combined score of each algorithm and the reference point to obtain the gain or loss value. A positive value indicates that the algorithm is better than the average level, and a negative value indicates that the algorithm is worse than the average level.
[0080] S33 performs nonlinear transformations on gains and losses using the cumulative prospect theory value function. The gain part uses a power function (concave) and the loss part uses a power function (convex) and is multiplied by a loss aversion coefficient to obtain the prospect value.
[0081] S34 calculates the difference between the foreground values of the two algorithms and the duality of the two algorithms in terms of attributes.
[0082] Specifically, considering the different properties of different algorithms Below Since their lengths differ, direct averaging is not possible. Therefore, a unified reference point needs to be obtained. First, the original data sequences of each algorithm are length-matched. Specifically, the longest sequence among all fusion algorithms is used as the standard, and the maximum value of the shorter sequences is repeated at the end to make all sequences equal in length. The average comprehensive score of all algorithms under each attribute is calculated as a reference point.
[0083]
[0084] in, Representation matrix The bipolar hesitant fuzzy element of position, This represents the total number of alternative solutions (fusion algorithms).
[0085] To obtain the gains (advantages) and disadvantages, the difference between the original value and the reference point is calculated.
[0086]
[0087] in, Representation matrix The bipolar hesitant fuzzy element of position, The reference point calculated in 3.1, if Representation Algorithm In attributes The performance was better than average, if Representation Algorithm In attributes Their performance was below average.
[0088] To quantify decision-makers' asymmetric risk preferences when facing gains and losses, prospect value is calculated:
[0089]
[0090] in, , These represent the degree of risk preference under gain and loss scenarios, respectively, and are typically 1 / 2 to 2 / 3. The higher the value, the more risk-averse the decision-maker is; they are more risk-averse when faced with gains and more risk-averse when faced with losses. θ is the loss aversion coefficient. A larger θ indicates a greater sensitivity to loss for the decision-maker, suggesting a more conservative approach. , The objective gains (advantages) or losses (disadvantages) calculated for 3.2.
[0091] To represent the algorithm on attribute j Compared to algorithms The dual bias is calculated by subtracting the foreground value (since it calculates the difference between different algorithm pairs, it does not include comparisons between itself, i.e.) ),
[0092]
[0093] in, The results, calculated from 3.3, are the algorithms respectively. In attributes Value and Algorithm In attributes The value of the above The first The, the Algorithms.
[0094] S4 constructs a preference function based on a Gaussian distribution by calculating the mean and standard deviation of the dual biases;
[0095] S41 obtains the average value of the dual deviation for each attribute by taking the arithmetic mean of the dual deviations over all different algorithm pairs;
[0096] S42 obtains the standard deviation of the dual deviation for each attribute by taking the square root of the square mean of the difference between the dual deviation and the mean for all different algorithm pairs.
[0097] S43 constructs a piecewise function. When the dual bias is negative, the preference value is zero. When the dual bias is non-negative, a Gaussian exponential function with standard deviation as the parameter is used to obtain the degree of preference that the algorithm is superior to the algorithm in terms of attributes.
[0098] Specifically, in all algorithm pairs ( and On ), calculate a certain attribute The difference The average value,
[0099]
[0100] in, The dual bias is calculated in section 3.4, and m is the total number of alternative solutions (fusion algorithm). and The indices representing the algorithms are all taken from 1 to 1. ,and .
[0101] To calculate the standard deviation of all dual deviations under the j-th attribute
[0102]
[0103] in, The result calculated for 4.1 across all algorithm pairs ( and The difference of a certain attribute j on the () The average value, This represents the total number of alternative solutions (fusion algorithms). These are the indices for the first algorithm and the second algorithm, respectively, both taken from... ,in It represents the total number of algorithms. This indicates that comparisons between the same algorithm and itself are not performed.
[0104] To calculate in the attribute Below, algorithm Superior to Algorithm To what extent,
[0105]
[0106] in, The dual deviation calculated in 3.4 The standard deviation of all dual deviations under the j-th attribute is calculated in section 4.2. It is an exponential function.
[0107] S5 transforms objective attribute weights into decision weights that reflect psychological risk through a probability weight function, and then calculates the preference index between algorithms.
[0108] S51 uses the Tversky-Kahneman probability weight function to perform convex and concave transformations on the objective weights respectively, to obtain the psychological weights of decision-makers when making comparative advantage decisions and when making comparative disadvantage decisions.
[0109] S52 calculates the sum of the preference function value with the advantage psychological weight and the disadvantage psychological weight, respectively, to obtain the preference index and preference index of the algorithm relative to the algorithm.
[0110] Specifically, the weight of objective attributes Transformed into decision weights that reflect psychological risk ,
[0111]
[0112]
[0113] in, The weight value of attribute j calculated in section 2.2. The bending coefficient of the probability weighting function is usually taken as... . Weighted by objective attributes After passing through the probability weight function (parameters) This is derived from the transformation, and it represents the decision-maker's preference for attributes when making "comparative advantage" decisions. The psychological weight, Weighted by objective attributes After passing through the probability weight function (parameters) The transformation yields the attribute, representing the decision-maker's perception of the attribute when making a "comparative disadvantage" decision. The psychological weight.
[0114]
[0115]
[0116] in, The preference values obtained in 4.3 , Calculated from 5.1, it represents the decision-maker's attitude towards attributes when making decisions about "comparative advantage" and "comparative disadvantage," respectively. The psychological weight.
[0117] S6 calculates the net flow by measuring the positive and negative transcendental flows, and sorts the net flows from largest to smallest to obtain the final priority order of the fusion algorithm.
[0118] S61 averages the preference index of the algorithm relative to all other algorithms to obtain the positive transcendental flow;
[0119] S62 averages the preference indices of all other algorithms relative to the algorithm to obtain the negative transcendental flow;
[0120] S63 obtains net flow by subtracting negative overcurrent from positive overcurrent;
[0121] S64 arranges all candidate algorithms in descending order of net flow to obtain the final priority order. When net flows are equal, the order is determined by the positive overshoot flow level or the algorithms remain in the same order.
[0122] Specifically, calculate the positive transcendental flow:
[0123]
[0124]
[0125] in, The first The, the A fusion algorithm for individuals, Represents the set of all algorithms (the number of which is...) ), This represents the total number of alternative solutions (fusion algorithms). The algorithm is calculated in section 5.2. Compared to The higher the preference index (the greater the preference index, the better). The better ),on the contrary, express Compared to The preference index.
[0126]
[0127] in, The forward current calculated in 6.1 The negative current is calculated as 6.2.
[0128] All alternative solutions (fusion algorithms) are ranked by net flow. Sort them in descending order of size to obtain the final priority order.
[0129]
[0130] in, The algorithm that maximizes net flow (optimal solution) This represents the algorithm that minimizes net flow (worst-case scenario). This indicates "superior". If the net flows of the two algorithms are equal... If they are, then they are considered equivalent in the ordering (which can be further determined based on positive transcendental flow). The order of importance is determined by the ranking, or they remain equal.
[0131] Beneficial effects:
[0132] This invention addresses multimodal medical image fusion scenarios such as CT-MRI, PET-CT, and PET-MRI, and comprehensively analyzes the fusion results across multiple evaluation dimensions, including structural information preservation, edge detail retention, functional information expression, information entropy enhancement, and visual quality.
[0133] First, by using fuzzification and mean comparison mechanisms, the objective evaluation index values corresponding to each fusion algorithm are transformed into positive and negative membership information, and a hesitant bipolar fuzzy set decision matrix is constructed, thereby effectively preserving the bidirectional uncertainty and evaluation hesitancy that exist in the medical image fusion evaluation process.
[0134] Subsequently, the objective weights of each evaluation index are automatically determined using bipolar hesitant fuzzy entropy, reducing subjective bias caused by human experience in weighting and improving the objectivity and consistency of the evaluation results.
[0135] This invention comprehensively utilizes bipolar hesitant fuzzy set theory, behavioral decision science, and multi-attribute decision analysis methods to effectively handle fuzziness, uncertainty, and evaluation hesitation in the quality evaluation of multimodal medical image fusion, while fully considering the psychological preferences of medical experts in terms of benefit and risk perception.
[0136] This invention provides a scientific, reliable, and well-interpretable decision-making framework for the objective evaluation, performance comparison, and optimization of multimodal medical image fusion algorithms, which helps to enhance the practical value of fusion results in medical applications such as clinical auxiliary diagnosis, lesion detection, and precision treatment planning.
[0137] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.
Claims
1. A method for evaluating the quality of fused images based on bipolar hesitant fuzzy preference ranking, characterized in that, Includes the following steps: A hesitant bipolar fuzzy set decision matrix is constructed through fuzzification, averaging, and comparison operations. The uncertainty and fuzziness of each algorithm under each attribute are calculated by bipolar hesitant fuzzy entropy, and then the objective weights of each attribute are obtained by normalization. The reference point is determined by equalization, the gains and losses are calculated, the prospect value is obtained by the cumulative prospect theoretical value function, and finally the duality deviation between algorithms is calculated. A preference function based on a Gaussian distribution is constructed by calculating the mean and standard deviation of the dual bias. The objective attribute weights are transformed into decision weights that reflect psychological risk through a probability weight function, and then the preference index between algorithms is calculated. The net flow is obtained by calculating the positive and negative transcendental flows, and the final priority order of the fusion algorithm is obtained by sorting the net flows from largest to smallest.
2. The fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking as described in claim 1, characterized in that, The process of constructing a hesitant bipolar fuzzy set decision matrix through fuzzification, averaging, and comparison operations includes the following steps: The blurring result is obtained by normalizing the original index values of each algorithm under each attribute and each image; The average blur value of each algorithm for each attribute is obtained by averaging the blurring results of all images under each attribute for each algorithm. Each fuzzification result is compared with the corresponding average fuzzy value. Those with values greater than or equal to the mean are assigned to the positive membership set, and those with values less than the mean are assigned to the negative membership set, thus obtaining bipolar hesitant fuzzy elements. Then, a bipolar fuzzy set decision matrix is constructed.
3. The fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking as described in claim 1, characterized in that, The process of calculating the uncertainty and fuzziness of each algorithm under each attribute using bipolar hesitant fuzzy entropy, and then obtaining the objective weights of each attribute through normalization, includes the following steps: The hesitant fuzzy entropy is calculated for the positive membership set and the negative membership set using the sinusoidal entropy formula, respectively, to measure the uncertainty and fuzziness of all test images under each attribute for each algorithm. The average entropy is obtained by averaging the entropy values of all algorithms under each attribute. Then, based on the principle that the smaller the entropy value, the greater the information content, the objective weight of each attribute is obtained through normalization calculation.
4. The fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking as described in claim 1, characterized in that, The process of determining a reference point through equalization, calculating gains and losses, obtaining the prospect value through the cumulative prospect theoretical value function, and finally calculating the duality deviation between algorithms includes the following steps: By using the longest sequence length as a benchmark and repeating the maximum value at the end of the shorter sequence, the original data sequences of each algorithm are lengthened, and then the average comprehensive score of all algorithms under each attribute is calculated as a reference point. The difference between the combined score of each algorithm and the reference point is used to obtain the gain or loss value. A positive value indicates that the algorithm is better than the average level, and a negative value indicates that the algorithm is worse than the average level. The prospect value is obtained by performing nonlinear transformations on the gains and losses using the cumulative prospect theory value function. The gains part is transformed using a power function (concave function), and the losses part is transformed using a power function (convex function) and multiplied by the loss aversion coefficient. The difference between the prospect values of the two algorithms is calculated to determine the duality of the two algorithms in terms of attributes.
5. The fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking as described in claim 1, characterized in that, The process of constructing a preference function based on a Gaussian distribution by calculating the mean and standard deviation of the dual bias includes the following steps: The average value of the dual bias for each attribute is obtained by taking the arithmetic mean of the dual biases over all different algorithm pairs. The standard deviation of the dual deviation for each attribute is obtained by taking the square root of the square mean of the difference between the dual deviation and the mean for all different algorithm pairs. By constructing a piecewise function, the preference value is zero when the dual bias is negative, and a Gaussian exponential function with standard deviation as the parameter is used when the dual bias is non-negative, thus obtaining the degree to which the algorithm is superior to the algorithm in terms of attributes.
6. The fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking as described in claim 1, characterized in that, The process of converting objective attribute weights into decision weights reflecting psychological risk through a probability weight function, and then calculating the preference index between algorithms, includes the following steps: By performing convex and concave transformations on the objective weights using the Tversky-Kahneman probability weight function, the psychological weights of decision-makers when making comparative advantage decisions and when making comparative disadvantage decisions can be obtained. The preference function value is weighted and summed with the advantage psychological weight and the disadvantage psychological weight respectively to obtain the preference index and preference index of the algorithm relative to the algorithm.
7. The fusion image quality assessment method based on bipolar hesitant fuzzy preference ranking as described in claim 1, characterized in that, The process of calculating the net flow by positive and negative transcendental flows, and then sorting the net flows from largest to smallest to obtain the final priority order of the fusion algorithm, includes the following steps: The positive transcendental flow is obtained by averaging the preference index of the algorithm relative to all other algorithms. The negative transcendental flow is obtained by averaging the preference indices of all other algorithms relative to the algorithm. The net flow is obtained by subtracting the negative excess flow from the positive excess flow; All candidate algorithms are sorted in descending order of net flow to obtain the final priority order. When net flows are equal, the order is determined by the positive overshoot flow level or the algorithms are kept in parallel.