A credible medical image segmentation method decoupling non-independent evidence fusion from uncertainty

Abstract

This invention discloses a reliable medical image segmentation method that integrates non-independent evidence fusion and decouples uncertainty, addressing the issues of overconfidence and evidence explosion caused by non-independent evidence in high-risk medical scenarios. The method utilizes an improved UNet++ to extract spatial and semantic perspective features and maps them to Dirichlet distribution parameters. It generates discount factors through recursive discount fusion, performing attenuation processing on incremental beliefs to achieve proactive evidence fusion. Based on Bayesian variance decomposition theory, it explicitly decouples data and model uncertainties within a second-order probability framework, using data uncertainty to optimize the segmentation loss and model uncertainty to guide redundant discounting. A discount regularization term is introduced to reduce redundant evidence while preventing model forgetting. A joint optimization objective function is constructed, the model is trained in stages, and the segmentation results and pixel-level uncertainty assessment results are output. This invention provides uncertainty assessment while maintaining high-precision segmentation, significantly improving the safety of clinical applications.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical image processing technology, specifically to a reliable medical image segmentation method that integrates non-independent evidence fusion and decoupling of uncertainty. It is applicable to segmentation tasks of medical images such as dermatoscopes, CT, and MRI, and can be applied to clinical scenarios such as computer-aided diagnosis, surgical planning, and radiotherapy navigation. Background Technology

[0002] Medical image segmentation is a core task in clinical medical decision support. Convolutional neural networks, represented by UNet++, have been widely used in this field due to their nested, dense skip connection structure, which alleviates the problem of semantic inconsistency between deep and shallow features. However, such pure convolutional architectures still have inherent defects: the locality of convolution operations leads to a limited receptive field and insufficient global modeling ability; the "stitching + convolution" of skip connections is a passive feature fusion and cannot identify and suppress noise features in the image.

[0003] To compensate for locality defects, methods such as TransUnet and SwinUnet introduce the Transformer attention mechanism. Although this improves segmentation accuracy, the computational complexity increases quadratically with image resolution, making it difficult to handle ultra-high resolution medical images. Moreover, such networks generally exhibit "overconfidence" and lack a sound uncertainty assessment mechanism, making it unable to identify and predict potential risks and failing to meet the safety requirements of high-risk medical scenarios.

[0004] To construct reliable segmentation models, researchers have introduced uncertainty quantification mechanisms: early Bayesian methods such as MCDropout require multiple forward propagation samplings, resulting in huge computational overhead; Evidence-based deep learning (EDL) quantifies overall uncertainty through a single forward propagation, significantly improving efficiency, and methods such as I-EDL and Devis have been introduced into the field of medical image segmentation. However, existing EDL-related methods still face two key technical bottlenecks in complex clinical scenarios:

[0005] The lack of explicit decoupling and targeted utilization of evidence uncertainty: Although existing methods model under the second-order probability framework of multi-class distribution-Dirichlet prior, they do not explicitly decouple the total variance and cannot distinguish between data uncertainty and model uncertainty. This leads to the confidence calibration penalty term incorrectly suppressing the model's behavior of outputting reasonable high uncertainty in high-noise regions, which reduces the robustness of the model and the reliability of calibration.

[0006] Limitations of feature fusion paradigm and evidence explosion caused by non-independent evidence: Traditional passive feature fusion cannot identify feature quality, while multi-scale features in medical images are strongly correlated, which does not meet the independence assumption of traditional DS fusion. Direct fusion is prone to induce "evidence explosion", amplifying the prediction confidence level and exacerbating the problem of overconfidence.

[0007] Therefore, given the inconsistent quality of medical images and the extremely high requirements for clinical safety, there is an urgent need for a reliable medical image segmentation method that can achieve adaptive discount fusion of non-independent evidence and explicitly utilize the uncertainty after decoupling. Summary of the Invention

[0008] To address the aforementioned issues, this invention aims to propose a reliable medical image segmentation method that integrates non-independent evidence fusion and uncertainty decoupling. Targeting high-risk clinical decision-making scenarios in medical image segmentation, it seeks to transform feature integration from "passive stitching" to "active reliable fusion," and utilizes explicitly decoupled data uncertainty to guide optimization, thereby enhancing noise robustness, mitigating overconfidence, and providing reliable evidence for high-precision segmentation results and pixel-level uncertainty assessment for clinical auxiliary diagnosis.

[0009] To achieve the above objectives, this invention proposes a reliable medical image segmentation method based on non-independent evidence fusion and uncertainty decoupling. This method is implemented through a non-independent evidence recursive fusion module and an uncertainty decoupling optimization module, and includes the following steps:

[0010] Step S1: Input the medical image to be segmented into the trusted segmentation model, extract multi-scale feature maps using the nested connection structure inside the model, reconstruct the shallow features of the encoder into a spatial perspective with fine contours, and reconstruct the deep upsampled features of the decoder into a semantic perspective with global semantic consensus.

[0011] Step S2: Using the evidence head module, the spatial perspective feature map and the speech perspective feature map obtained in step S1 are mapped and processed respectively, and the corresponding Dirichlet distribution parameters are output. Based on subjective logic theory, the Dirichlet distribution parameters are converted into subjective opinions under the second-order probability framework. The subjective opinions include belief quality and uncertainty quality.

[0012] Step S3: A recursive fusion mechanism is adopted. First, the encoder spatial view feature map obtained in step S1 is fused with the semantic view feature map of the first decoder. Then, the previous fusion branch is recursively fused with the current decoder semantic view. During the fusion process, the convolutional network is used to dynamically capture the semantic overlap between views represented by the Dirichlet distribution parameters output in step S2 to generate a pixel-level discount factor. For non-independent evidence, the operation of first direct fusion and then discount fusion calibration is performed.

[0013] Step S4: Based on Bayesian variance decomposition theory, under the second-order probability framework of the predicted distribution being a multi-class distribution and the prior distribution being the Dirichlet distribution corresponding to the fused Dirichlet distribution parameters output in step S3, the closed-form solution of the total variance of the predicted distribution is derived, and the total variance is explicitly decoupled into the data uncertainty of quantified data noise and the model uncertainty of quantified knowledge deficiency.

[0014] Step S5: Construct a dynamic pixel-level weight map using the data uncertainty decoupled in Step S4, and perform attenuation optimization on the segmentation loss; fine-tune the model, after enabling the discount fusion in Step S3, introduce discount constraint terms and regularization terms to drive the model to adaptively reduce the discount factor, reduce the redundant belief caused by non-independent perspectives, and force the discounted distribution to maintain the same predictive semantics as the original distribution after direct fusion in Step S3.

[0015] Step S6: Construct a joint optimization objective function that includes segmentation loss, data uncertainty attenuation term from step S5, non-truth class distribution constraint term, and discount regularization term from step S5. Optimize model parameters using a phased end-to-end training method. After model training is completed, output the predicted segmentation result of the medical image to be segmented and the pixel-level uncertainty evaluation result obtained based on the decoupling uncertainty obtained in step S4.

[0016] Furthermore, in step S1, the reconstruction of the spatial perspective and the semantic perspective specifically involves:

[0017] S11. Spatial perspective construction: The shallow detail feature map extracted by the first layer node of the model encoder branch through double convolution is defined as the spatial perspective, which includes pixel-level texture and fine contour features.

[0018] S12. Semantic Perspective Construction: The deep feature map obtained by nested dense skip connections and upsampling operations from the model decoder branch is reconstructed into a semantic perspective, which contains global high-level semantic information.

[0019] Furthermore, in step S2, the specific calculation process for constructing subjective opinions is as follows:

[0020] S21. Dirichlet parameter calculation: using the evidence header... Convolutional layer and Softplus activation function output Branch feature map Corresponding Dirichlet distribution parameters The calculation formula is:

[0021]

[0022] in, H and W are the height and width of the feature map, respectively, and K is the number of categories;

[0023] S22. Subjective Opinion Generation: Calculating the total intensity of the Dirichlet distribution using subjective logic theory. And the first The Dirichlet distribution parameters of each branch are mapped to subjective opinions. Among them, b i For the belief quality of the i-th branch, u iLet k be the uncertain mass of the i-th branch, and k be the value of K. The calculation formula is:

[0024]

[0025] (3)

[0026] in, For the belief quality of the i-th branch category k, .

[0027] Furthermore, in step S3, the recursive discount fusion mechanism specifically includes:

[0028] S31. Discount Factor Generation: This method guides evidence calibration by explicitly modeling the distributional exclusivity between perspectives. It combines the Dirichlet distribution parameters of the fused previous branch, the Dirichlet distribution parameters of the current branch, and the initial conflict coefficient representing the mutual exclusivity of belief distributions between the two branches. Cascade, in which, , For the belief quality of the i-th branch category k1, For the belief quality of the previous fusion branch category k2, By learning semantic overlap and belief exclusivity features at the evidence level between perspectives through convolutional networks, pixel-level discount intensity coefficients are generated after Sigmoid activation. The calculation formula is shown in (4):

[0029]

[0030] in, The Dirichlet distribution parameters are those obtained after the fusion of the previous branch. Here are the Dirichlet distribution parameters for the current branch, where k1 and k2 are the class values. ,, K is the number of categories;

[0031] S32. Belief Discount Processing: Utilizing Discount Strength Coefficient The incremental belief quality from the current perspective is discounted and decayed, and the reduced belief amount is compensated for by the uncertainty quality. In this approach, by suppressing the explosion of evidence caused by an independent perspective, a discounted subjective opinion is ultimately obtained. The calculation formula is:

[0032]

[0033]

[0034] in, For the belief quality of the i-th branch after discount, Let the uncertainty quality of the i-th branch after the discount be... , ;

[0035] S33, DS Recursive Fusion: The conflict coefficient C is calculated using formula (7), and then the DS combination rules of formulas (8) and (9) are used to discount the quality of beliefs from the current perspective. Quality of belief after integration with the higher level Perform recursive fusion to obtain the fusion belief quality from the current perspective. The calculation formula is:

[0036]

[0037]

[0038] in, To determine the belief quality of the i-th branch after fusion, To represent the uncertainty quality of the i-th branch after fusion, , , k1 and k2 are the category values. , K is the number of categories;

[0039] S34. Parameter Restoration: Restore the fusion belief quality to the parameters of the fused Dirichlet distribution. The calculation formula is:

[0040]

[0041] in, .

[0042] Furthermore, in step S4, the uncertainty decoupling under second-order probability modeling specifically includes:

[0043] S41. Problem Definition and Probabilistic Modeling: Assume that the pixel class prediction value Y of the medical image to be segmented follows a multi-class distribution with probability P. The prior distribution of the predicted distribution is the Dirichlet distribution. Given a probability vector Under the given conditions, predict the conditional expectation of the distribution. Conditional variance The results are obtained by formula (11) and formula (12) respectively:

[0044]

[0045] The diagonal component of the conditional variance can be simplified as follows: ,in For element-wise multiplication, and the probability vector ;

[0046] S42. Total variance decomposition based on the total variance theorem: Using formula (13) of the variance decomposition theorem, the total variance of the random variable is decomposed. Total variance Explicitly decompose data uncertainty into quantified data noise Model uncertainty due to lack of quantitative knowledge :

[0047]

[0048] Data uncertainty Defined as conditional variance matrix Diagonal components under prior distribution The mathematical expectation is defined by formula (14):

[0049]

[0050] Using the properties of the Dirichlet distribution, we obtain the first-order origin moment as shown in formula (15). With formula (16) second-order origin moment Substituting into formula (14), we obtain the closed-form solution for the data uncertainty in formula (17). , :

[0051]

[0052]

[0053]

[0054] in, Here, S represents the Dirichlet distribution parameters, and S represents the total intensity.

[0055] Formula (18) addresses the model uncertainty. Defined as the variance of the conditional expectation under the prior distribution. By calculating the expected prediction probability covariance matrix The trace is realized. The calculation formula is formula (19):

[0056]

[0057]

[0058] Wherein, the second-order origin moment matrix The elements are calculated according to the properties of Dirichlet integrals, i.e., when the index is... Its value at that time ;when Its value at that time ,in and To assign a value to the category, , K is the number of categories, thus obtaining the second-order original moment matrix of formula (20). :

[0059]

[0060] Subsequently, with the first moment Substituting the squares into formula (19) yields formula (21) for the covariance matrix. The calculation results are then substituted into formula (19) to obtain formula (22) model uncertainty. , :

[0061]

[0062]

[0063] in, For the Dirichlet distribution parameters, This represents the total intensity.

[0064] Furthermore, in step S5, the uncertainty-guided loss optimization specifically involves:

[0065] S51. Segmentation Loss Decay Optimization: Constructing a Weight Decay Optimization Term Using Data Uncertainty ,in Let i be the basic segmentation loss for the i-th branch. For the weighting coefficients of data uncertainty, This is the dynamic pixel-level weight map for the i-th branch. This optimization term automatically reduces the weight of blurred boundaries or high-noise regions in gradient updates.

[0066] S52. Discount Regularization Constraints: Constructing Discount Regularization Terms + ,in For the discount constraint term of the i-th branch, construct spatial guided weights using model uncertainty. ,pass By increasing the penalty for locations where the total strength after discounting and fusion exceeds the maximum single-source strength before fusion, the model is driven to adaptively reduce the discount factor, thereby reducing redundant beliefs caused by non-independent perspectives and confining the evidence space within reasonable physical boundaries. For the discount consistency term of the i-th branch, through the discount consistency term The distribution before and after the forced discount maintains consistency with the predicted semantic direction, ensuring that the operation only reduces the magnitude without changing the prediction direction. The calculation formula is as follows:

[0067]

[0068]

[0069] in, The expected distribution of the predicted distribution of category k obtained by direct DS fusion. Let $\frac{k}{k}$ be the expected distribution of the predicted distribution of category $k$ obtained by discount fusion.

[0070] Furthermore, in step S6, the construction and phased training of the joint optimization objective function are specifically as follows:

[0071] S61. Joint Optimization Objective Function: The overall optimization objective is the sum of the losses of all fusion branches, calculated using the following formula:

[0072]

[0073] in, This is the pixel-level data uncertainty map for the i-th branch. Let i be the segmentation loss for the i-th branch. For the data uncertainty regularization term hyperparameter, For the non-truth class distribution constraint term of the i-th fusion branch, For hyperparameters of non-truth-valued constraint optimization terms, + Let the discount constraint term and discount regularization term be the i-th branch. The hyperparameter of the discount regularization term, t is the training time, which is related to the number of training epochs;

[0074] S62. Phased Training:

[0075] Phase 1: Setting , , A stable feature representation benchmark is established through direct recursive fusion;

[0076] Phase Two: Separating Data Uncertainty Using Bayesian Variance Decomposition It automatically reduces the weight of blurred boundaries or noisy regions in gradient updates, allowing the model to learn to "reject" forced fitting of unreliable noise while maintaining the integrity of multi-scale features;

[0077] Phase 3: After the model has acquired noise resistance, fine-tune the existing fusion model, enable discount fusion, and activate the discount constraint term. With discount regularization Optimize and mitigate the problems of evidence overload and overconfidence caused by evidence not being independent.

[0078] Furthermore, the basic segmentation loss It consists of the Dice loss under Dirichlet expectation and the cross-entropy loss, and the calculation formula is as follows:

[0079]

[0080] in, For Dice's loss, The cross-entropy loss is expected to be calculated using the Digamma function of the Dirichlet distribution.

[0081] Beneficial effects: (1) This invention realizes active calibration and redundancy suppression of evidence under the same non-independent perspective: According to the prior of belief conflict between perspectives and the Dirichlet distribution, the convolutional network is used to adaptively generate pixel-level discount factors; this mechanism can identify and reduce the amount of redundant beliefs from the same multi-scale features, and constrain the total evidence strength after recursive fusion within a reasonable range, thus alleviating the problem of "evidence explosion" and overconfidence caused by direct DS fusion of non-independent evidence.

[0082] (2) This invention establishes an uncertainty decoupling and collaborative optimization mechanism under second-order probability modeling of evidence: This invention uses the data uncertainty explicitly decoupled under the second-order probability framework to construct a dynamic weight graph, and performs attenuation optimization on the segmentation loss to reduce the contribution of high-noise regions to the gradient and improve the noise robustness of the model; during fine-tuning, the model uncertainty is used to guide the model to discount in the region of evidence redundancy to achieve model calibration.

[0083] (3) The present invention improves the reliability of evaluation while maintaining high-precision segmentation: The present invention maintains segmentation accuracy comparable to existing advanced methods in normal image scenarios. In damaged image scenarios such as Gaussian noise, Gaussian blur, and random occlusion, the segmentation index (Dice, ASSD) and confidence calibration index (ECE) are significantly better than existing methods. Moreover, the uncertainty discrimination index (AUROC, AUPRC) is greatly improved, which can accurately identify and predict erroneous regions and output early warnings, reducing the risk of missed diagnosis and misdiagnosis in clinical practice.

[0084] (4) This invention achieves a universal improvement to the underlying structure of mainstream segmentation paradigms: This invention aims to solve the common underlying challenge of "failure of the evidence independence assumption" caused by the endogenous correlation between hierarchical features during the transformation of encoder-decoder structures from "passive splicing" to "active evidence-level fusion" of multi-scale features. The design of this scheme does not depend on a specific segmentation benchmark model or a specific training dataset, and can be widely adapted to various medical image segmentation tasks with hierarchical feature extraction requirements, possessing structural adaptability and modular universality. In addition, when dealing with covariate offset data commonly encountered in clinical practice, this invention demonstrates high generalization ability and robustness in uncertainty quantification, ensuring reliable evaluation and decision security of the model under complex and abnormal inputs. Attached Figure Description

[0085] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0086] Figure 1 This is a flowchart of the credible medical image segmentation method for non-independent evidence fusion and uncertainty decoupling described in the embodiments of the present invention;

[0087] Figure 2 This is a flowchart of the main algorithm of the credible medical image segmentation method with non-independent evidence fusion and uncertainty decoupling described in the embodiments of the present invention;

[0088] The upper part of the figure represents Stage 1 (credible fusion): The image to be segmented is input into the evidence UNet++ model. A multi-scale feature perspective is constructed using a nested connection structure and mapped to subjective opinions. Evidence from each perspective is directly and recursively fused to decouple uncertainty, optimize segmentation loss, and establish stable feature extraction capabilities. The lower part represents Stage 2 (discounted fusion calibration): After warm-up, the discounted evidence fusion mode between perspectives is enabled. The discount factor dynamically generated by the convolutional network is used to reduce the amount of redundant evidence from non-independent perspectives. At the same time, discount constraint terms and discount regularization terms are introduced to reduce redundant evidence and force the model to maintain consistency in the predicted semantic direction.

[0089] Figure 3 This is a schematic diagram of the overall architecture of the credible segmentation model of the credible medical image segmentation method that decouples non-independent evidence fusion and uncertainty according to an embodiment of the present invention.

[0090] This figure illustrates the complete topology from the input image to multi-scale feature extraction, then to evidence head mapping and the recursive integration of the fusion module, and finally to the output segmentation result through the uncertainty decoupling module;

[0091] Figure 4This is a diagram showing the internal structure and phased logic of the discounted evidence fusion module in the credible medical image segmentation method for decoupling non-independent evidence fusion and uncertainty as described in an embodiment of the present invention.

[0092] The figure details the process of discount fusion, in which discount coefficients are generated by a convolutional network using the distribution parameters of the previous fusion branch and the current decoding upsampling branch; it also distinguishes the output differences between the direct fusion in stage one and the discount calibration in stage two.

[0093] Figure 5 The visualization of the segmentation results and uncertainties of the reliable medical image segmentation method with non-independent evidence fusion and uncertainty decoupling described in the embodiments of the present invention in Gaussian noise and Gaussian blur scenes;

[0094] Figure 6 This is a visualization of the segmentation results and uncertainties of the reliable medical image segmentation method with non-independent evidence fusion and uncertainty decoupling described in this embodiment of the invention under random occlusion scenarios. Detailed Implementation

[0095] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0096] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0097] Example 1

[0098] See Figures 1-6 A reliable medical image segmentation method based on non-independent evidence fusion and uncertainty decoupling is implemented through a reliable segmentation model that includes a non-independent evidence recursive fusion module and an uncertainty decoupling optimization module, comprising the following steps:

[0099] Step S1: Input the medical image to be segmented into the trusted segmentation model, extract multi-scale feature maps using the nested connection structure inside the model, reconstruct the shallow features of the encoder into a spatial perspective with fine contours, and reconstruct the deep upsampled features of the decoder into a semantic perspective with global semantic consensus.

[0100] Step S2: Using the evidence head module, the spatial perspective feature map and the speech perspective feature map obtained in step S1 are mapped and processed respectively, and the corresponding Dirichlet distribution parameters are output. Based on subjective logic theory, the Dirichlet distribution parameters are converted into subjective opinions under the second-order probability framework. The subjective opinions include belief quality and uncertainty quality.

[0101] Step S3: A recursive fusion mechanism is adopted. First, the encoder spatial view feature map obtained in step S1 is fused with the semantic view feature map of the first decoder. Then, the previous fusion branch is recursively fused with the current decoder semantic view. During the fusion process, the convolutional network is used to dynamically capture the semantic overlap between views represented by the Dirichlet distribution parameters output in step S2 to generate a pixel-level discount factor. For non-independent evidence, the operation of first direct fusion and then discount fusion calibration is performed.

[0102] Step S4: Based on Bayesian variance decomposition theory, under the second-order probability framework of the predicted distribution being a multi-class distribution and the prior distribution being the Dirichlet distribution corresponding to the fused Dirichlet distribution parameters output in step S3, the closed-form solution of the total variance of the predicted distribution is derived, and the total variance is explicitly decoupled into the data uncertainty of quantified data noise and the model uncertainty of quantified knowledge deficiency.

[0103] Step S5: Construct a dynamic pixel-level weight map using the data uncertainty decoupled in Step S4, and perform attenuation optimization on the segmentation loss; fine-tune the model, after enabling the discount fusion in Step S3, introduce discount constraint terms and regularization terms to drive the model to adaptively reduce the discount factor, reduce the redundant belief caused by non-independent perspectives, and force the discounted distribution to maintain the same predictive semantics as the original distribution after direct fusion in Step S3.

[0104] Step S6: Construct a joint optimization objective function that includes segmentation loss, data uncertainty attenuation term from step S5, non-truth class distribution constraint term, and discount regularization term from step S5. Optimize model parameters using a phased end-to-end training method. After model training is completed, output the predicted segmentation result of the medical image to be segmented and the pixel-level uncertainty evaluation result obtained based on the decoupling uncertainty obtained in step S4.

[0105] It should be noted that the passive fusion paradigm of "stitching + convolution" used in the feature aggregation stage of traditional medical image segmentation architectures, represented by UNet++, has the disadvantage of failing to effectively identify and suppress noise features. At the same time, due to the lack of explicit decoupling of total variance in existing evidence-based deep learning methods, they cannot effectively distinguish between data uncertainty and model uncertainty, which in turn induces the problem of incorrect suppression of the "reasonably high uncertainty" output of the confidence calibration penalty term in high-noise regions. The overall solution proposed in this embodiment realizes the paradigm innovation of moving from passive accumulation at the feature level to active and credible fusion at the evidence level by constructing a "conflict prior-guided recursive discount fusion mechanism" and an "uncertainty explicit decoupling paradigm based on Bayesian variance decomposition". By using the decoupled data uncertainty to drive loss decay to suppress noise fitting, and combining model uncertainty to guide evidence discount to alleviate the overconfidence problem caused by non-independent perspectives, it provides uncertainty assessment results with high robustness and clinical safety warning feedback while ensuring high-precision segmentation performance.

[0106] In a specific example, in step S1, the reconstruction of the spatial perspective and the semantic perspective is as follows:

[0107] S11. Spatial perspective construction: The shallow detail feature map extracted by the first layer node of the model encoder branch through double convolution is defined as the spatial perspective, which includes pixel-level texture and fine contour features.

[0108] S12. Semantic Perspective Construction: The deep feature map obtained by nested dense skip connections and upsampling operations from the model decoder branch is reconstructed into a semantic perspective, which contains global high-level semantic information.

[0109] In the specific implementation, it is based on the improved UNet++ multi-scale feature extraction framework:

[0110] The architecture first extracts multi-scale features through encoder branches: let the encoder node be... ,in Represents network hierarchy, For the corresponding number of convolution channels, the th The dimension of the feature map output by the layer encoding node is represented as: The corresponding number of convolution channels can be set to Before viewpoint reconstruction, the decoder utilizes nested dense skip connections for feature fusion to fill semantic gaps in the deconvolution process. Let... hierarchical ,order The fusion process of the fusion nodes follows the following formula:

[0111]

[0112] in, This is a two-layer convolution operation. This indicates the concatenation of channel dimensions. This is a double upsampling operation.

[0113] Spatial Perspective Construction: Unlike traditional UNet++ which only performs passive splicing at the output layer, this invention constructs the encoder branch first-layer node. Defined as a spatial perspective. This perspective originates directly from shallow convolution extraction and contains the most original pixel-level texture and fine contour features, providing crucial boundary details for subsequent fusion.

[0114] Semantic perspective reconstruction: The semantic perspective is reconstructed using intermediate nodes with global consensus in the first level of the decoder branch. Specifically, this involves reconstructing the semantic perspective through nested connections between nodes. (in Perform double upsampling to obtain the reconstructed feature map. This perspective incorporates high-level semantic information learned by deep networks, forming a complementary "detail-semantic" internal perspective together with the spatial perspective. This example achieves a paradigm shift from "passive feature stacking" to "active evidence-level modeling" by explicitly reconstructing shallow, fine contours and deep, global semantics. This overcomes the limitations of traditional "stitching + convolution" fusion strategies, which struggle to suppress noise interference due to a lack of quality discrimination, at the underlying architectural level. This perspective-based representation lays the foundation for uncertainty quantification in subsequent Dirichlet recursive fusion, enabling the model to adaptively fuse evidence from different internal perspectives when dealing with extreme medical noise or regions with blurred boundaries. Furthermore, within a second-order probabilistic framework, it can transform "overconfident" predictions into physically meaningful pixel-level uncertainty assessments.

[0115] In a specific example, the specific calculation process for constructing subjective opinions in step S2 is as follows:

[0116] S21. Dirichlet parameter calculation: using the evidence header... Convolutional layer and Softplus activation function output Branch feature map Corresponding Dirichlet distribution parameters The calculation formula is:

[0117]

[0118] in, H and W are the height and width of the feature map, respectively, and K is the number of categories;

[0119] S22. Subjective Opinion Generation: Calculating the total intensity of the Dirichlet distribution parameters using subjective logic theory. and will the The Dirichlet distribution parameters of each branch are mapped to subjective opinions. Among them, b i For the belief quality of the i-th branch, u i Let k be the uncertain mass of the i-th branch, and k be the value of K. The calculation formula is:

[0120]

[0121] (3)

[0122] in, , For the belief quality of the i-th branch category k, .

[0123] It should be noted that this embodiment draws on the second-order probabilistic modeling concept in existing evidence deep learning (EDL) to establish a unified evidence representation foundation for the spatial-semantic dual perspective reconstructed in step S1. By introducing Dirichlet distribution parameters and mapping them to standardized subjective opinions that include belief quality and uncertainty quality, this invention explicitly transforms the originally implicit multi-scale features into second-order probability distributions with clear physical meaning, providing a mathematical basis for the calculation of conflict coefficients and the generation of discount factors in subsequent step S3.

[0124] In a specific example, in step S3, the recursive discount fusion mechanism specifically includes:

[0125] S31. Discount Factor Generation: This method guides evidence calibration by explicitly modeling the distributional exclusivity between perspectives. It combines the Dirichlet distribution parameters of the fused previous branch, the Dirichlet distribution parameters of the current branch, and the initial conflict coefficient representing the mutual exclusivity of belief distributions between the two branches. Cascade, in which , For the belief quality of the i-th branch category k1, For the belief quality of the previous fusion branch category k2, By learning semantic overlap and belief exclusivity features at the evidence level between perspectives through convolutional networks, pixel-level discount intensity coefficients are generated after Sigmoid activation. The calculation formula is shown in (4):

[0126]

[0127] in, The Dirichlet distribution parameters are those obtained after the fusion of the previous branch. Let be the Dirichlet distribution parameters for the current branch. and To assign a value to the category, , K is the number of categories;

[0128] S32. Belief Discount Processing: Utilizing Discount Strength Coefficient The incremental belief quality from the current perspective is discounted and decayed, and the reduced belief amount is compensated for by the uncertainty quality. In the middle, the subjective opinion after the discount. The calculation formula is:

[0129]

[0130]

[0131] in, For the belief quality of the i-th branch after discount, Let the uncertainty quality of the i-th branch after the discount be... , ;

[0132] S33, DS Recursive Fusion: The conflict coefficient C is calculated using formula (7), and then the DS combination rules of formulas (8) and (9) are used to discount the quality of beliefs from the current perspective. Belief quality after merging with the higher level Perform recursive fusion to obtain the fusion belief quality from the current perspective. The calculation formula is:

[0133]

[0134]

[0135] in, To determine the belief quality of the i-th branch after fusion, To represent the uncertainty quality of the i-th branch after fusion, , , and To assign a value to the category, , K is the number of categories;

[0136] S34. Parameter Restoration: Restore the fusion belief quality to the parameters of the fused Dirichlet distribution. The calculation formula is:

[0137]

[0138] in, .

[0139] It should be noted that, addressing the problems of traditional architectures like UNet++, which passively employ a "concatenation + convolution" paradigm when aggregating multi-scale features, lacking a feature quality identification mechanism, and the inherent tendency of traditional DS fusion to induce "evidence explosion" and overconfidence when handling homologous non-independent features due to the forced assumption of independent perspectives, this embodiment utilizes the Dirichlet distributions of different perspectives and inter-perspective conflicts to adaptively generate pixel-level discount factors through convolutional networks, reducing redundant belief and compensating for it in uncertainty quality. This mechanism aims to alleviate the evidence explosion problem of non-independent features within the DS theoretical framework, effectively suppressing illegal confidence inflation while preserving the semantic complementarity of hierarchical features. This invention improves the adaptability of traditional fusion paradigms in homologous multi-scale feature scenarios, enhancing the calibration accuracy of medical image segmentation results and the model's risk identification capability when facing unknown or damaged data by reasonably transforming redundant evidence into quantifiable uncertainty.

[0140] In a specific example, in step S4, the uncertainty decoupling under second-order probabilistic modeling specifically includes:

[0141] S41. Problem Definition and Probabilistic Modeling: Assume that the pixel class prediction value Y of the medical image to be segmented follows a multi-class distribution with probability P. The prior distribution of the predicted distribution is the Dirichlet distribution. Given a probability vector Under the given conditions, predict the conditional expectation of the distribution. Conditional variance The results are obtained by formula (11) and formula (12) respectively:

[0142] The diagonal component of the conditional variance can be simplified as follows: ,in For element-wise multiplication, and the probability vector ;

[0143] S42. Total variance decomposition based on the full variance theorem: Using formula (13) of the full variance theorem, the total variance of the random variable is decomposed. Total variance Explicitly decompose data uncertainty into quantified data noise Model uncertainty due to lack of quantitative knowledge :

[0144]

[0145] Data uncertainty Defined as conditional variance matrix Diagonal components under prior distribution The mathematical expectation is defined by formula (14):

[0146]

[0147] Using the properties of the Dirichlet distribution, we obtain the first-order moment of the origin, formula (15). With formula (16) second-order origin moment Substituting into formula (14), we obtain the closed-form solution for the data uncertainty in formula (17). , :

[0148]

[0149]

[0150]

[0151] in, For the Dirichlet distribution parameters, Total strength;

[0152] Formula (18) addresses the model uncertainty. Defined as the variance of the conditional expectation under the prior distribution. By calculating the expected prediction probability covariance matrix The trace is realized. The calculation formula is formula (19):

[0153]

[0154]

[0155] Wherein, the second-order origin moment matrix The elements are calculated according to the properties of Dirichlet integrals, i.e., when the index is... Its value at that time ;when Its value at that time ,in and To assign a value to the category, , K is the number of categories, thus obtaining the second-order original moment matrix of formula (20). :

[0156]

[0157] Subsequently, with the first moment Substituting the squares into formula (19) yields formula (21) for the covariance matrix. The calculation results are then substituted into formula (19) to obtain formula (22) model uncertainty. , :

[0158]

[0159]

[0160] in, For the Dirichlet distribution parameters, This represents the total intensity.

[0161] Compared to existing evidence-based deep learning methods, which model within a second-order probability framework of multi-class distributions and Dirichlet priors, these methods do not explicitly decouple the total variance, failing to distinguish between data uncertainty and model uncertainty and thus hindering their targeted and appropriate utilization. This embodiment, based on Bayesian variance decomposition theory, further derives a closed-form solution for predicting the total variance within the existing second-order probability framework, achieving explicit decoupling between data uncertainty and model uncertainty. This lays the theoretical foundation for subsequent collaborative optimization mechanisms.

[0162] In a specific example, in step S5, the uncertainty-guided loss optimization is as follows:

[0163] S51. Segmentation Loss Decay Optimization: Constructing a Weight Decay Optimization Term Using Data Uncertainty ,in Let i be the basic segmentation loss for the i-th branch. For the weighting coefficients of data uncertainty, This is the dynamic pixel-level weight map for the i-th branch. This optimization term automatically reduces the weight of blurred boundaries or high-noise regions in gradient updates.

[0164] S52. Discount Regularization Constraints: Constructing Discount Regularization Terms + ,in Let be the discount constraint term for the i-th branch. This method assumes that the model exhibits extremely low uncertainty after a large amount of repeated evidence accumulation, leading to evidence redundancy and "overconfidence." Therefore, the model utilizes model uncertainty to construct spatial guiding weights. ,pass Increase the penalty for areas where the total strength after discount fusion is greater than the maximum single-source strength before fusion, drive the model to adaptively reduce the discount factor, reduce the amount of redundant beliefs caused by non-independent perspectives, and thus constrain the evidence space within a reasonable physical boundary. For the discount consistency term of the i-th branch, through the discount consistency term The distribution before and after the forced discount maintains consistency with the predicted semantic direction, ensuring that the operation only reduces the magnitude without changing the prediction direction. The calculation formula is as follows:

[0165]

[0166]

[0167] in, The expected distribution of the predicted distribution of category k obtained by direct DS fusion. Let $\frac{k}{k}$ be the expected distribution of the predicted distribution of category $k$ obtained by discount fusion.

[0168] To address the issues of existing loss optimization methods failing to decouple uncertainty, leading to errors in confidence calibration penalty terms and the model's tendency to output reasonably high uncertainty in noisy regions, this embodiment utilizes decoupled data uncertainty to construct a dynamic weight graph. This adaptively reduces gradient contributions in noisy regions, fundamentally suppressing the model's forced fitting to noise. Simultaneously, a discount constraint term is introduced during the fine-tuning phase, using model uncertainty to guide evidence discounting, ensuring the rationality of the confidence distribution and mitigating evidence explosion. Furthermore, a discount regularization term ensures semantic consistency before and after discounting, preventing model forgetting. This achieves the dual effect of "noise-resistant optimization without sacrificing segmentation accuracy, and discount fusion without altering predictive semantics."

[0169] In a specific example, in step S6, the construction and phased training of the joint optimization objective function are as follows:

[0170] S61. Joint Optimization Objective Function: The overall optimization objective is the sum of the losses of all fusion branches, calculated using the following formula:

[0171]

[0172] in, This is the pixel-level data uncertainty map for the i-th branch. Let i be the segmentation loss for the i-th branch. For the data uncertainty regularization term hyperparameter, For the non-truth class distribution constraint term of the i-th fusion branch, For hyperparameters of non-truth-valued constraint optimization terms, + Let the discount constraint term and discount regularization term be the i-th branch. The hyperparameter of the discount regularization term, t, is the training time, which is related to the number of training epochs.

[0173] S62. Phased Training:

[0174] Phase 1 (Warm-up Period): Setting , , A stable feature representation benchmark is established through direct recursive fusion;

[0175] Phase Two (Noise Resilience Enhancement Period): Utilizing Bayesian variance decomposition to separate data uncertainties It automatically reduces the weight of blurred boundaries or noisy regions in gradient updates, allowing the model to learn to "reject" forced fitting of unreliable noise while maintaining the integrity of multi-scale features;

[0176] Phase 3: After the model has acquired noise resistance, fine-tune the existing fusion model, enable discount fusion, and activate the discount constraint term. With discount regularization Optimize and mitigate the problems of evidence overload and overconfidence caused by evidence not being independent.

[0177] Compared to existing models that use a single loss and single-stage training, which easily leads to an imbalance between segmentation accuracy and confidence calibration, this embodiment constructs a multi-loss term joint optimization objective function, integrating segmentation loss, uncertainty decay term, non-truth class constraint term, and discount regularization term. It also employs a staged training strategy, gradually activating each optimization module according to the model's learning pattern: first establishing a stable feature benchmark, then improving noise robustness, and finally calibrating confidence. This solves the problems of "multi-objective optimization conflict and unstable training process" in traditional training methods, achieving global optimization of segmentation performance, noise robustness, and confidence calibration, and alleviating model overconfidence. In a specific example, the basic segmentation loss... It consists of the Dice loss under Dirichlet expectation and the cross-entropy loss, and the calculation formula is as follows:

[0178]

[0179] in, For Dice's loss, The cross-entropy loss is expected to be calculated using the Digamma function of the Dirichlet distribution.

[0180] It should be noted that in this embodiment... as well as The calculation is performed using the following formula:

[0181]

[0182]

[0183] Where n is the number of pixels in a single sample. Let p be the Digamma function, and p be the predicted probability vector. Let y be the Dirichlet distribution parameter and y be the true label.

[0184] This embodiment also requires the construction of non-truth value class constraint optimization terms; the KL divergence between the predicted distribution and the uniform distribution is used to constrain the non-truth value classes. To minimize the amount of evidence for error categories and prevent overconfidence, The calculation method for non-truth value constraint optimization terms in a single branch is shown in the following formula:

[0185]

[0186] It should be noted that traditional segmentation loss functions (such as Dice and cross-entropy loss) are only applicable to point estimation of classification probabilities and are difficult to directly apply to the technical bottlenecks under the second-order probabilistic framework. This embodiment draws on and adopts the expectation optimization paradigm in evidence-based deep learning. By constructing a composite loss function under the expectation of the Dirichlet distribution and using the Digamma function to realize the mathematical mapping of the objective function to the evidence parameter space, this invention achieves paradigm alignment between traditional segmentation evaluation indicators and the second-order probabilistic framework. By synergistically optimizing pixel-level classification accuracy and region-level boundary overlap, it provides a complete mathematical foundation for building a high-precision medical image segmentation system with uncertainty perception capabilities.

[0187] To verify the effectiveness of this invention in complex medical imaging and high-risk clinical decision-making scenarios, this embodiment constructs a detailed experimental scenario based on the ISIC2018 publicly available skin lesion segmentation dataset. This dataset contains 16,072 dermoscopic images, covering both foreground lesion regions and background. In the experiment, the training, validation, and test sets were strictly divided in a 7:1.5:1.5 ratio, and the input images were uniformly adjusted to (3,224,256) for processing. To ensure the model's robustness to covariate shifts, this experiment constructs a dataset containing Gaussian noise (…). Gaussian blur ( ) and random occlusion ( The impaired test set is used to evaluate whether the model can proactively exhibit caution and output accurate uncertainty feedback rather than blindly fitting noise when image quality deteriorates. The performance evaluation system consists of basic segmentation metrics, confidence calibration metrics, and uncertainty discrimination metrics. Among them, AUROC is used to measure the probability that the uncertainty of mispredicted pixels is higher than that of correctly predicted pixels, while AUPRC represents the average value of actual prediction errors in high uncertainty regions.

[0188] Table 1

[0189]

[0190] Based on the experimental results in Table 1, under normal test sets, this method maintains high-precision segmentation performance comparable to the original UNet++. Furthermore, the robustness improvement of this method is particularly pronounced under damaged test set scenarios. For ease of analysis, the model terminology used below is defined as follows: Devis represents the comparative model using existing confidence-calibrated loss; AUDC represents the data uncertainty-guided weighted loss used in this invention; UPPM refers to the model incorporating the evidence multi-view fusion (TMC) mechanism; UPPM(Disc) represents the complete model with the proposed "discount fusion" mechanism enabled during the fine-tuning phase.

[0191] 1. Effectiveness of Data Uncertainty Weighting (AUDCE): Comparing Upp and Upp+AUDCE, it can be found that the model's performance is improved at all noise levels after introducing data uncertainty guided loss. For example, under NoiseL1, Dice increases from 76.9377% to 79.8439%, indicating that AUDCE effectively enhances the model's noise resistance by explicitly modeling image noise and reducing the gradient contribution in noisy regions.

[0192] 2. Performance Gains of Multi-View Evidence Fusion (UPPM): A comparison of the results of Upp+AUDCE and UppM+AUDCE shows that the introduction of the multi-view evidence fusion mechanism significantly improves segmentation accuracy and stability. Under NoiseL1, UppM+AUDCE achieves a Dice metric of 84.8555%, far exceeding the 79.8439% of the single-view fusion. This demonstrates that reconstructing multi-scale features within UNet++ into mutually verifying evidence perspectives is more effective than traditional feature layer concatenation (Concat) in extracting complementary information, thus overcoming the limitation of passive feature fusion in identifying feature quality.

[0193] 3. The effect of the discount fusion mechanism (Disc) on suppressing overconfidence: Comparing UppM+AUDCE with the full model UPPM(Disc)+AUDCE, it can be found that enabling the discount fusion mechanism optimizes the model's reliability in extremely damaged scenarios. Especially in the extreme case of MaskL3 (random occlusion), the Dice metric of UPPM(Disc) (46.8155%) significantly surpasses that of the model without discount (30.7374%).

[0194] Meanwhile, regarding calibration metrics, taking NoiseL1 as an example, the ECE value of UPPM(Disc)+AUDCE dropped significantly from 0.1451 in the original model to 0.0374. This indicates that the discount mechanism effectively suppresses the overconfidence problem caused by "evidence explosion" by dynamically reducing the amount of redundant beliefs generated by non-independent perspectives, making the uncertainty of the model output highly aligned with the actual predicted risk.

[0195] In summary, this invention addresses the robust training problem under "noise interference" by using decoupled data uncertainty to guide segmentation loss, achieves "multi-view information synergy" through DS fusion, and alleviates overconfidence through non-independent evidence fusion and model uncertainty-guided redundancy discount. Ultimately, while maintaining high accuracy, it significantly improves the evaluation reliability and early warning capability of medical image segmentation in complex clinical environments.

[0196] Table 2

[0197]

[0198] As can be seen from the experimental results in Table 2, Gaussian noise ( In this scenario, the AUROC of our method reached 0.9301, higher than the 0.8255 of the DEVIS method. This indicates that our invention, through evidence recursion fusion and uncertainty decoupling mechanism, enables the model to honestly output high uncertainty when faced with prediction errors caused by high-intensity noise. The high AUPRC (0.3914) further proves that our scheme can effectively filter high-risk areas and use them as automatic early warning signals, significantly reducing the risk of missed diagnoses in clinical practice.

[0199] Figures 5-6 The uncertainty graph of the model segmentation results is shown. Experimental results fully demonstrate that the method proposed in this invention can achieve more stable and reliable medical image segmentation under extreme noise and complex clinical interference conditions. In summary, through the synergistic optimization of evidence recursive fusion and uncertainty decoupling, this method significantly enhances the model's calibration reliability and OOD detection capability while improving segmentation accuracy, proving its practical value in identifying risk areas and reducing missed diagnoses and misdiagnoses.

[0200] Those skilled in the art should understand that although this embodiment uses skin lesion images as the experimental benchmark, the non-independent evidence recursive fusion and uncertainty decoupling optimization mechanism proposed in this invention targets the underlying common challenges that are prevalent in medical imaging. It is not limited to dermoscopy images, but can also be widely applied to medical image segmentation tasks of various modalities such as CT, MRI, ultrasound, and PET. Its technical effect of improving segmentation accuracy and evaluation reliability has universality across datasets.

[0201] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of trustworthy medical image segmentation with non-independent evidence fusion and uncertainty decoupling, characterized in that, This is achieved through a reliable segmentation model that includes a recursive fusion module for non-independent evidence and an uncertainty decoupling optimization module, comprising the following steps: Step S1: Input the medical image to be segmented into the trusted segmentation model, extract multi-scale feature maps using the nested connection structure inside the model, reconstruct the shallow features of the encoder into a spatial perspective with fine contours, and reconstruct the deep upsampled features of the decoder into a semantic perspective with global semantic consensus. Step S2: Using the evidence head module, the spatial perspective feature map and the speech perspective feature map obtained in step S1 are mapped and processed respectively, and the corresponding Dirichlet distribution parameters are output. Based on subjective logic theory, the Dirichlet distribution parameters are converted into subjective opinions under the second-order probability framework. The subjective opinions include belief quality and uncertainty quality. In step S2, the specific calculation process for constructing subjective opinions is as follows: S21, Dirichlet parameter calculation: use the evidence head Convolutional layer and Softplus activation function output the i-th branch feature map X i The corresponding Dirichlet distribution parameter The calculation formula is: in, H and W are the height and width of the feature map, respectively, and K is the number of categories; S22. Subjective Opinion Generation: Calculating the total intensity of the Dirichlet distribution using subjective logic theory. Map the Dirichlet distribution parameters of the i-th branch to subjective opinions. , where b i For the belief quality of the i-th branch, u i Let k be the uncertain mass of the i-th branch, and k be the value of K. The calculation formula is: （3） in, , For the belief quality of the i-th branch category k, ; Step S3: A recursive fusion mechanism is adopted. First, the encoder spatial view feature map obtained in step S1 is fused with the semantic view feature map of the first decoder. Then, the previous fusion branch is recursively fused with the current decoder semantic view. During the fusion process, the convolutional network is used to dynamically capture the semantic overlap between views represented by the Dirichlet distribution parameters output in step S2 to generate a pixel-level discount factor. For non-independent evidence, the operation of first direct fusion and then discount fusion calibration is performed. Step S4: Based on Bayesian variance decomposition theory, under the second-order probability framework of the predicted distribution being a multi-class distribution and the prior distribution being the Dirichlet distribution corresponding to the fused Dirichlet distribution parameters output in step S3, the closed-form solution of the total variance of the predicted distribution is derived, and the total variance is explicitly decoupled into the data uncertainty of quantified data noise and the model uncertainty of quantified knowledge deficiency. Step S5: Construct a dynamic pixel-level weight map using the data uncertainty decoupled in Step S4, and perform attenuation optimization on the segmentation loss; fine-tune the model, after enabling the discount fusion in Step S3, introduce discount constraint terms and regularization terms to drive the model to adaptively reduce the discount factor, reduce the redundant belief caused by non-independent perspectives, and force the discounted distribution to maintain the same predictive semantics as the original distribution after direct fusion in Step S3. Step S6: Construct a joint optimization objective function that includes segmentation loss, data uncertainty attenuation term from step S5, non-truth class distribution constraint term, and discount regularization term from step S5. Optimize model parameters using a phased end-to-end training method. After model training is completed, output the predicted segmentation result of the medical image to be segmented and the pixel-level uncertainty evaluation result obtained based on the decoupling uncertainty obtained in step S4.

2. The reliable medical image segmentation method based on non-independent evidence fusion and uncertainty decoupling according to claim 1, characterized in that, In step S1, the reconstruction of the spatial and semantic perspectives specifically involves: S11. Spatial perspective construction: The shallow detail feature map extracted by the first layer node of the model encoder branch through double convolution is defined as the spatial perspective, which includes pixel-level texture and fine contour features. S12. Semantic Perspective Construction: The deep feature map obtained by nested dense skip connections and upsampling operations from the model decoder branch is reconstructed into a semantic perspective, which contains global high-level semantic information.

3. The reliable medical image segmentation method based on the fusion of non-independent evidence and decoupling of uncertainty as described in claim 1, characterized in that, In step S3, the recursive discount fusion mechanism specifically includes: S31. Discount Factor Generation: This method guides evidence calibration by explicitly modeling the distributional exclusivity between perspectives. It combines the Dirichlet distribution parameters of the fused previous branch, the Dirichlet distribution parameters of the current branch, and the initial conflict coefficient representing the mutual exclusivity of belief distributions between the two branches. Cascade, in which, , For the belief quality of the i-th branch category k1, For the belief quality of the previous fusion branch category k2, By learning semantic overlap and belief exclusivity features at the evidence level between perspectives through convolutional networks, pixel-level discount intensity coefficients are generated after Sigmoid activation. The calculation formula is shown in (4): in, The Dirichlet distribution parameters are those obtained after the fusion of the previous branch. Here are the Dirichlet distribution parameters for the current branch, where k1 and k2 are the class values. , K is the number of categories; S32. Belief Discount Processing: Utilizing Discount Strength Coefficient The incremental belief quality from the current perspective is discounted and decayed, and the reduced belief amount is compensated for by the uncertainty quality. In this approach, by suppressing the explosion of evidence caused by an independent perspective, a discounted subjective opinion is ultimately obtained. The calculation formula is: in, For the belief quality of the i-th branch after discount, Let the uncertainty quality of the i-th branch after the discount be... , ; S33, DS Recursive Fusion: The conflict coefficient C is calculated using formula (7), and then the DS combination rules of formulas (8) and (9) are used to discount the quality of beliefs from the current perspective. Belief quality after merging with the higher level Perform recursive fusion to obtain the fusion belief quality from the current perspective. The calculation formula is: in, To determine the belief quality of the i-th branch after fusion, To represent the uncertainty quality of the i-th branch after fusion, , k1 and k2 are the category values. , K is the number of categories; S34. Parameter Restoration: Restore the fusion belief quality to the parameters of the fused Dirichlet distribution. The calculation formula is: in, .

4. The reliable medical image segmentation method based on the fusion of non-independent evidence and decoupling of uncertainty as described in claim 1, characterized in that, In step S4, the uncertainty decoupling under second-order probability modeling specifically includes: S41. Problem Definition and Probabilistic Modeling: Assume that the pixel class prediction value Y of the medical image to be segmented follows a multi-class distribution with probability P. The prior distribution of the predicted distribution is the Dirichlet distribution. Given a probability vector Under the given conditions, predict the conditional expectation of the distribution. Conditional variance The results are obtained by formula (11) and formula (12) respectively: The diagonal component of the conditional variance can be simplified as follows: ,in For element-wise multiplication, and the probability vector ; S42. Total variance decomposition based on the total variance theorem: Using formula (13) of the variance decomposition theorem, the total variance of the random variable is decomposed. Total variance Explicitly decompose data uncertainty into quantified data noise Model uncertainty due to lack of quantitative knowledge : Data uncertainty Defined as conditional variance matrix Diagonal components under prior distribution The mathematical expectation is defined by formula (14): Using the properties of the Dirichlet distribution, we obtain the first-order origin moment as shown in formula (15). With formula (16) second-order origin moment Substituting into formula (14), we obtain the closed-form solution for the data uncertainty in formula (17). , : in, Here, S represents the Dirichlet distribution parameters, and S represents the total intensity. Formula (18) addresses the model uncertainty. Defined as the variance of the conditional expectation under the prior distribution. By calculating the expected prediction probability covariance matrix The trace is realized. The calculation formula is formula (19): Wherein, the second-order origin moment matrix The elements are calculated according to the properties of Dirichlet integrals, i.e., when the index is... Its value at that time ;when Its value at that time ,in and To assign a value to the category, , K is the number of categories, thus obtaining the second-order original moment matrix of formula (20). : Subsequently, with the first moment Substituting the squares into formula (19) yields formula (21) for the covariance matrix. The calculation results are then substituted into formula (19) to obtain formula (22) model uncertainty. , : in, For the Dirichlet distribution parameters, This represents the total intensity.

5. The reliable medical image segmentation method based on non-independent evidence fusion and uncertainty decoupling according to claim 1, characterized in that, In step S5, the uncertainty-guided loss optimization is specifically as follows: S51. Segmentation Loss Decay Optimization: Constructing a Weight Decay Optimization Term Using Data Uncertainty ,in Let i be the basic segmentation loss for the i-th branch. For the weighting coefficients of data uncertainty, This is the dynamic pixel-level weight map for the i-th branch. This optimization term automatically reduces the weight of blurred boundaries or high-noise regions in gradient updates. S52. Discount Regularization Constraints: Constructing Discount Regularization Terms + ,in For the discount constraint term of the i-th branch, construct spatial guided weights using model uncertainty. ,pass By increasing the penalty for locations where the total strength after discounting and fusion exceeds the maximum single-source strength before fusion, the model is driven to adaptively reduce the discount factor, thereby reducing redundant beliefs caused by non-independent perspectives and confining the evidence space within reasonable physical boundaries. For the discount consistency term of the i-th branch, through the discount consistency term The distribution before and after the forced discount maintains consistency with the predicted semantic direction, ensuring that the operation only reduces the magnitude without changing the prediction direction. The calculation formula is as follows: in, The expected distribution of the predicted distribution of category k obtained by direct DS fusion. Let $\frac{k}{k}$ be the expected distribution of the predicted distribution of category $k$ obtained by discount fusion.

6. The reliable medical image segmentation method based on non-independent evidence fusion and uncertainty decoupling according to claim 1, characterized in that, In step S6, the construction and phased training of the joint optimization objective function are as follows: S61. Joint Optimization Objective Function: The overall optimization objective is the sum of the losses of all fusion branches, calculated using the following formula: in, This is the pixel-level data uncertainty map for the i-th branch. Let i be the segmentation loss for the i-th branch. For the data uncertainty regularization term hyperparameter, For the non-truth class distribution constraint term of the i-th fusion branch, For hyperparameters of non-truth-valued constraint optimization terms, + Let the discount constraint term and discount regularization term be the i-th branch. The hyperparameter of the discount regularization term, t is the training time, which is related to the number of training epochs; S62. Phased Training: Phase 1: Setting , , A stable feature representation benchmark is established through direct recursive fusion; Phase Two: Separating Data Uncertainty Using Bayesian Variance Decomposition It automatically reduces the weight of blurred boundaries or noisy regions in gradient updates, allowing the model to learn to "reject" forced fitting of unreliable noise while maintaining the integrity of multi-scale features; Phase 3: After the model has acquired noise resistance, fine-tune the existing fusion model, enable discount fusion, and activate the discount constraint term. With discount regularization Optimize and mitigate the problems of evidence overload and overconfidence caused by evidence not being independent.

7. The reliable medical image segmentation method based on non-independent evidence fusion and uncertainty decoupling according to claim 5, characterized in that, The basic segmentation loss It consists of the Dice loss under Dirichlet expectation and the cross-entropy loss, and the calculation formula is as follows: in, For Dice's loss, The cross-entropy loss is expected to be calculated using the Digamma function of the Dirichlet distribution.

Images