A multi-model coordinated pathological image segmentation method, device, and medium based on evidence-based deep learning and pathological prior guidance.

By combining Dirichlet distribution and Dempster-Shafer evidence theory with prior pathological knowledge, the problems of inference delay and decision conflict in multi-model collaborative segmentation of pathological images are solved, achieving efficient and accurate pathological image segmentation.

CN122888253BActive Publication Date: 2026-11-10SHENZHEN SHENGQIANG TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611399061.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-10
Estimated Expiration
2046-09-10

AI Technical Summary

Technical Problem

Existing deep learning-based multi-model collaborative segmentation methods for pathological images suffer from problems such as high inference latency, improper handling of decision conflicts between models, and lack of pathological prior knowledge correction, resulting in a lack of objective arbitration in key boundary regions.

Method used

By parameterizing the multi-model branch prediction features into a Dirichlet distribution, using Dempster-Shafer evidence theory for nonlinear evidence fusion, and combining it with a pathological prior knowledge base for conflict correction, we achieve unified modeling of cognitive and data uncertainty and pathological prior guidance.

Benefits of technology

It improves the accuracy and reliability of pathological image segmentation, meets the real-time requirements of clinical practice, provides objective basis for model discrepancy resolution, and reduces the risk of misdiagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122888253B_ABST
    Figure CN122888253B_ABST
Patent Text Reader

Abstract

This invention proposes a multi-model coordinated pathological image segmentation method, device, and medium based on evidence-based deep learning and pathological prior guidance. To address the problem of uncertainty estimation relying on multiple sampling and lacking conflict resolution models based on pathological priors in existing pathological image segmentation, this invention inputs the image to be segmented into a multi-branch network. Each branch's evidence prediction head outputs Dirichlet distribution parameters through a single forward propagation to calculate cognitive and data uncertainty. The evidence quantities of each branch are converted into mass functions and fused using Dempster rules to calculate the conflict coefficient. When the conflict exceeds a threshold, the evidence is corrected based on statistical divergence decay using a pathological prior knowledge base and then re-fused. The segmentation result is output based on the fused evidence, and a three-level reliability assessment map is generated by combining the conflict coefficient. This invention is mainly used for efficient and robust segmentation of tumors and tissues in digital pathological slides and for clinical decision support.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis and computer vision processing technology for digital pathological images, and in particular to a multi-model coordinated pathological image segmentation method, device, and medium based on evidence deep learning and pathological prior guidance. Background Technology

[0002] Semantic segmentation of pathological images is one of the core tasks of digital pathology artificial intelligence analysis, and it is crucial for clinical diagnosis such as tumor infiltration region identification, histological classification, and Ki-67 proliferation index assessment. In recent years, segmentation methods based on pathological baseline models have achieved good performance on pathological segmentation benchmarks due to their strong semantic representation capabilities obtained through large-scale self-supervised pre-training. However, single deep learning models are prone to performance fluctuations and generalization bottlenecks when dealing with different tissue regions, staining batch differences, and multi-source scanning equipment. Clinical applications place extremely high demands on the robustness and reliability of pathological segmentation results, making multi-model collaborative segmentation an important path to improve prediction stability.

[0003] In existing multi-model collaborative segmentation methods, Monte Carlo Dropout is typically used to quantify uncertainty. This involves calculating the information entropy and variance of the prediction results through multiple forward propagations during the model inference phase to characterize the confidence level. Regarding multi-model decision fusion, conventional solutions often employ a linear averaging strategy based on information entropy weighting within the probability space of the Softmax output.

[0004] However, such methods have revealed significant shortcomings in practical applications: First, Monte Carlo sampling causes the computational overhead of inference to increase linearly with the number of samplings, which cannot meet the timeliness requirements of rapid analysis of large field of view in clinical slides, and it fails to decouple the cognitive uncertainty caused by model ignorance and the data uncertainty caused by data noise within the theoretical framework. Secondly, the features extracted by each model branch belong to a heterogeneous feature subspace. The linear weighted average in the probability space ignores the non-independence between evidences and cannot reasonably handle the intense decision-making conflicts and information redundancy between models. Third, pathological images themselves contain rich anatomical prior rules such as tissue spatial co-occurrence rules, boundary morphology smoothness, and cell density gradient. Existing fusion algorithms do not incorporate such prior knowledge into the decision-making mechanism, resulting in a lack of objective arbitration and correction basis in key boundary areas where model disagreements are serious. Summary of the Invention

[0005] This invention provides a multi-model coordinated pathological image segmentation method, device, and medium based on evidence-based deep learning and pathological prior guidance. It addresses the problems of existing deep learning-based multi-model collaborative pathological image segmentation methods, which rely on multiple forward samplings, resulting in high inference latency, and the inability to reasonably resolve decision conflicts and redundancies between models by using linear weighted averaging in the probability space. Furthermore, it lacks pathological prior knowledge to objectively correct discrepancies in model regions.

[0006] The core technology of this invention is to directly decouple the calculation of cognitive and data uncertainty by parameterizing the predicted features of multiple model branches into a Dirichlet distribution through a single forward propagation, mapping the evidence of each branch into a Dempster-Shafer confidence function and performing nonlinear orthogonal fusion and conflict coefficient modeling in the evidence space, and using a pathological prior knowledge base to correct the divergent evidence by statistical divergence negative exponential decay in response to the conflict coefficient exceeding the threshold, and combining the fusion of cognitive uncertainty and conflict coefficient to adaptively output a multi-level reliability assessment.

[0007] In a first aspect, the present invention provides a multi-model coordinated pathological image segmentation method based on evidence deep learning and pathological prior guidance, the method comprising the following steps: Acquire the pathological image to be segmented and input it into a preset multi-branch pathological segmentation network; By using the evidence prediction heads of each model branch in the multi-branch pathological segmentation network, pixel-level prediction features are parameterized into Dirichlet distribution parameters through a single forward propagation, and the amount of evidence and uncertainty index of each model branch at the pixel level are determined based on the Dirichlet distribution parameters. The evidence quantity of each model branch is converted into the confidence quality function in Dempster-Shafer evidence theory. The confidence quality function of multiple model branches is nonlinearly fused using the Dempster combination rule, and the prediction conflict coefficient between model branches is calculated. Determine whether the predicted conflict coefficient is greater than the preset conflict threshold; if it is greater than the preset conflict threshold, retrieve the target pathological prior probability distribution from the pathological prior knowledge base based on the neighboring organization prediction consensus, attenuate and correct the evidence amount of the corresponding model branch according to the statistical divergence deviation between the predicted probability distribution of each model branch and the target pathological prior probability distribution, and re-perform evidence fusion based on the corrected evidence amount to obtain fused evidence. The expected category probability distribution is determined based on fused evidence, and the pathological segmentation results of the pathological image to be segmented are output. The reliability assessment results are generated by combining the uncertainty index of fused evidence and the prediction conflict coefficient.

[0008] Furthermore, through the evidence prediction heads of each model branch in the multi-branch pathological segmentation network, pixel-level predicted features are parameterized into Dirichlet distribution parameters via a single forward propagation. Based on the Dirichlet distribution parameters, the amount of evidence and uncertainty indices of each model branch at the pixel level are determined, including: The evidence prediction head performs nonlinear activation mapping on the decoded features to output a non-negative evidence vector. Each component of the non-negative evidence vector is incremented by 1 to obtain the Dirichlet distribution parameters of each preset pathological category. The Dirichlet distribution parameters of all categories are summed to obtain the Dirichlet distribution intensity. The cognitive uncertainty representing the degree of ignorance of the model is calculated based on the ratio of the total number of categories to the intensity of the Dirichlet distribution. The expected category probability distribution is determined based on the ratio of the Dirichlet distribution parameter to the Dirichlet distribution intensity for each preset pathological category, and the normalized information entropy of the expected category probability distribution is calculated to characterize the data uncertainty caused by inherent noise in the data.

[0009] Furthermore, the evidence quantity of each model branch is converted into the reliability quality function in Dempster-Shafer evidence theory. Nonlinear evidence fusion is then performed on the reliability quality functions of multiple model branches using the Dempster combination rule, and the prediction conflict coefficient between model branches is calculated, including: For each model branch, the ratio of the evidence quantity of each pathological category to the intensity of the Dirichlet distribution is used as the single-point reliability quality of the corresponding category, and the ratio of the total number of categories to the intensity of the Dirichlet distribution is used as the empty set reliability quality representing global uncertainty. The mass function of each model branch is constructed. The sum of the reliability quality products between branch pairs of the model in different categories is calculated using orthogonal summation to determine the prediction conflict coefficient, which characterizes the degree of model divergence. The first-order normalization factor is used to nonlinearly combine the intersection of the class consensus term and the uncertainty term between the model branch pairs, and the fusion class reliability quality and fusion uncertainty quality after the combination of multiple model branches are obtained iteratively. The pixel-level total fusion uncertainty is obtained by summing the combined Dirichlet cognitive uncertainty with the weighted prediction conflict coefficient.

[0010] Furthermore, the pathological prior knowledge base includes statistical priors on the spatial co-occurrence of tissue types, priors on the smoothness of tissue boundary morphology, and priors on the cell density gradient at the interface. If the conflict exceeds a preset threshold, the target pathological prior probability distribution is retrieved from the pathological prior knowledge base based on the consensus prediction of neighboring organizations. The evidence amount of the corresponding model branch is attenuated and corrected according to the statistical divergence deviation between the predicted probability distribution of each model branch and the target pathological prior probability distribution, including: By using a sliding window to statistically analyze the mode of category prediction for each model branch within the local neighborhood of the pixel to be corrected, the consensus on neighborhood tissue type prediction is determined, and the expected anatomical prior probability distribution is obtained by matching and retrieving from the pathological prior knowledge base. Calculate the KL divergence between the predicted probability distribution of each model branch output and the anatomical prior probability distribution; The original evidence quantity of each model branch in each category is multiplied by a negative exponential decay term with KL divergence and a preset prior strength coefficient as exponential factors to decay the evidence quantity of model branches that deviate from the prior distribution.

[0011] Furthermore, reliability assessment results are generated by combining the uncertainty index of fused evidence with the prediction conflict coefficient, including: On the validation set, the pixel-level fusion uncertainty distribution is statistically analyzed, and a preset first quantile threshold is extracted as the high / medium reliability boundary threshold, and a preset second quantile threshold is extracted as the medium / low reliability boundary threshold. The total fusion uncertainty of each pixel in the pathological image to be segmented is compared with the high / medium reliability boundary threshold and the medium / low reliability boundary threshold, and an adaptive three-level reliability evaluation map of high reliability, medium reliability or low reliability is output.

[0012] Furthermore, the multi-branch pathological segmentation network is pre-trained, and its training process includes applying multi-scale boundary-aware spatial consistency smoothing constraints to the evidence space: The first-order difference gradient magnitudes in the horizontal and vertical directions of the input pathological image are calculated, and negative exponential boundary-aware weights are constructed based on the gradient magnitudes. The weights approach zero in the tissue boundary region to suppress smoothing, and the weights approach one in the smooth region inside the tissue to enhance smoothing. The local neighborhood mean of the fused evidence is calculated by sliding across multiple preset scale windows, and the squared difference between the pixel fused evidence and the neighborhood mean at the corresponding scale is measured. The squared differences at each scale are multiplied by the boundary-aware weights and then summed at multiple scales to construct a multi-scale smoothing loss for the evidence space, which directly constrains the Dirichlet evidence quantity distribution.

[0013] Furthermore, the multi-branch pathological segmentation network is pre-trained, and its training process includes: In the Dirichlet parameter space, the analytical closed-form solution of the KL divergence of the Dirichlet distribution between models is calculated based on the Dirichlet distribution parameters of the model branch pairs. Based on the cognitive uncertainty after multi-model fusion, a negatively correlated gating factor is constructed. The gating factor is multiplied with the KL divergence of the Dirichlet distribution and averaged over all model branch pairs to construct the pairwise model mutual promotion loss. During the training phase, the random dropout layer of the backbone network is periodically enabled at a preset iteration interval to perform multiple forward propagations. The statistical prediction entropy and variance are combined and normalized to form the teacher uncertainty signal. The cognitive uncertainty output by the evidence prediction head is aligned with the teacher uncertainty signal by constraining the loss through distillation.

[0014] Secondly, the present invention provides a multi-model coordinated pathological image segmentation device based on evidence deep learning and pathological prior guidance, comprising: The evidence prediction module is used to acquire the pathological image to be segmented and input the pathological image to be segmented into a preset multi-branch pathological segmentation network. Through the evidence prediction heads of each model branch in the multi-branch pathological segmentation network, the pixel-level prediction features are parameterized into Dirichlet distribution parameters through a single forward propagation. Based on the Dirichlet distribution parameters, the amount of evidence and uncertainty index of each model branch at the pixel level are determined. The evidence fusion module is used to convert the evidence quantity of each model branch into the confidence quality function in the Dempster-Shafer evidence theory, perform nonlinear evidence fusion on the confidence quality function of multiple model branches through the Dempster combination rule, and calculate the prediction conflict coefficient between model branches. The pathology prior correction module is used to retrieve the target pathology prior probability distribution from the pathology prior knowledge base based on the neighboring tissue prediction consensus when the prediction conflict coefficient is greater than the preset conflict threshold. It then attenuates and corrects the evidence amount of the corresponding model branch according to the statistical divergence deviation between the predicted probability distribution of each model branch and the target pathology prior probability distribution, and triggers the evidence fusion module to re-execute evidence fusion based on the corrected evidence amount to obtain fused evidence. The segmentation and evaluation module is used to determine the probability distribution of the expected category based on fused evidence and output the pathological segmentation results of the pathological image to be segmented. It also generates a reliability evaluation result by combining the uncertainty index of the fused evidence with the prediction conflict coefficient.

[0015] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to execute the above-described multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance.

[0016] Fourthly, the present invention provides a readable storage medium storing a computer program, which, when executed by a processor, implements the multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described above.

[0017] The main contributions and innovations of this invention are as follows: 1. This invention parameterizes the model output as a Dirichlet distribution, allowing cognitive uncertainty and data uncertainty to be directly analyzed in a single forward propagation, eliminating the need for multiple forward propagations in Monte Carlo Dropout. This significantly improves inference efficiency compared to the original number of samplings, meeting the clinical requirements for real-time segmentation. Furthermore, cognitive uncertainty and data uncertainty are modeled uniformly within the same Dirichlet parameter framework, overcoming the problems of isolated prediction entropy and variance and lack of theoretical consistency.

[0018] 2. This invention treats the predictions of each model as evidence and constructs them as a mass function. It achieves fusion in the evidence space rather than the probability space through the Dempster combination rule. The conflict coefficient naturally quantifies the degree of divergence between models, avoiding the fusion distortion caused by linear weighting imposing the independent and identically distributed assumption on the predictions of heterogeneous models, and improving the accuracy of the fusion segmentation results.

[0019] 3. In areas of high conflict in the model, this invention utilizes pathological prior knowledge such as spatial co-occurrence of tissue types, boundary morphology, and density gradient. Using the KL divergence between the predicted distribution and the prior distribution as a deviation measure, a stronger negative exponential decay is applied to model evidence with greater deviation from the prior. After correction, the evidence combination is re-executed, providing an objective and interpretable basis for adjudicating model discrepancies and suppressing the dominance of erroneous evidence from individual models in the fusion results.

[0020] 4. Monte Carlo Dropout is only periodically enabled during the training phase as an uncertainty teacher signal. Statistical uncertainty knowledge is injected into the evidence prediction head through distillation loss. Dropout is completely disabled during the inference phase, requiring no sampling. This achieves Monte Carlo-level uncertainty estimation quality while maintaining the inference efficiency of a single forward propagation.

[0021] 5. This invention performs reliability grading based on a weighted combination of cognitive uncertainty and conflict coefficient. The grading threshold is automatically learned from the upper and lower tertiary positions of the uncertainty distribution in the validation set, ensuring that the proportion of high, medium, and low reliability regions is stable and does not drift with the dataset, thus overcoming the problem of fixed thresholds failing on different datasets. The prediction of high reliability regions can be directly used to save physicians' image reading time, medium reliability regions prompt boundary review, and low reliability regions prompt manual review or reassessment, effectively avoiding the risk of misdiagnosis and supporting efficient, reliable, and interpretable clinical auxiliary decision-making.

[0022] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance according to an embodiment of the present invention; Figure 2 This is a structural diagram of the model according to an embodiment of the present invention. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0025] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0026] Figure 1 A flowchart illustrating a multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance according to an embodiment of the present invention is shown. Figure 1 As shown, the method includes, in sequence: data preparation and multi-branch basic model construction, evidence deep learning parameterization and direct uncertainty prediction, Dempster-Shafer evidence combination, pathological prior-guided evidence correction, evidence space multi-scale boundary-aware spatial smoothing, uncertainty-gated Dirichlet space model mutual promotion, model training and inference, three-level reliability assessment and adaptive quantile calibration.

[0027] Figure 2 A schematic diagram of the overall structure of a model according to an embodiment of the present invention is shown. Figure 2As shown, immunohistochemical pathological images are input into the shared backbone, and decoder 1 and EDL evidence head 1, decoder 2 and EDL evidence head 2 respectively output pixel-level evidence for each model branch; the evidence from each branch is combined with DS evidence and pathological prior correction to form fused evidence, which is used to output segmentation results, reliability maps and uncertainty heatmaps; during the training phase, the MC distillation training module provides distillation supervision to each evidence head through the training distillation guidance path, and the evidence space multi-scale smoothing and uncertainty gating model mutually promote each other as the training loss term, and the loss is backpropagated to the shared backbone and each branch along the loss propagation path.

[0028] Example 1 This embodiment provides a multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance, using a breast Ki-67 immunohistochemical pathological image and dual-model coordinated segmentation as an example. The method includes the following steps: Step 1: Data Preparation and Construction of Multi-Branch Basic Model A dataset of Ki-67 immunohistochemical pathological images of breast tissue was selected and divided into training and validation sets. During training, the segments were scaled to... Pixel-level data augmentation includes horizontal flipping, vertical flipping, and random rotation.

[0029] A shared backbone strategy is adopted: all branches share the same pre-trained pathological baseline model (e.g., UNI2, using the ViT-Giant / 14 architecture, 14-pixel blocks, 1536 embedding dimensions, 24 depths, and 24 attention heads). After loading the pre-trained weights, the encoder parameters are frozen, and only the decoders and evidence prediction heads for each branch are trained independently. Each branch decoder adopts a 4-layer transposed convolutional upsampling structure with 256, 128, 64, and 32 channels respectively. Each transposed convolutional layer is followed by batch normalization and ReLU activation, and finally, the evidence prediction head outputs pixel-level non-negative evidence.

[0030] By sharing a backbone strategy, multiple branches can reuse the same strong semantic representation and reduce the computational and storage overhead of multi-model coordination, while each retains its own independent decoder feature subspace, so that the pixel-level predictions of each branch become evidence based on different feature subspaces.

[0031] Step 2: Evidence-based Deep Learning Parameterization and Direct Uncertainty Prediction The model output is parameterized from traditional softmax probabilistic parameters to Dirichlet distribution parameters. Specifically, the decoder output features of each branch are processed by the evidence prediction head (a ReLU-activated convolutional layer) to output non-negative evidence. ,in Indicates the first The amount of evidence for a class The number of predefined pathological categories is used, and the evidence quantity is a non-negative real number that represents the strength of evidence held by the model for the corresponding category. ; in and These represent the weights and biases of the evidence prediction head convolutional layer. These are decoding features.

[0032] Add 1 to each component of the nonnegative evidence quantity to obtain the Dirichlet distribution parameters: ; in For the first The Dirichlet distribution parameter of the class is increased by 1 based on the evidence quantity to ensure that the parameter is always positive.

[0033] Dirichlet strength is: ; in The sum of the Dirichlet distribution parameters for all categories represents the total strength of evidence held by the model.

[0034] Two types of uncertainty can be directly calculated analytically from the Dirichlet parameters. Cognitive uncertainty, reflecting model ignorance, is as follows: ; in This represents cognitive uncertainty. When the amount of evidence approaches zero (the model "knows nothing" about that pixel), Approaching , Approaching 1; when the amount of evidence is large (the model is highly confident). Much larger , Approaching 0.

[0035] The data uncertainty reflecting data noise is: ; ; in For the first The expected class probability of a class. To measure the irreducible uncertainty caused by inherent noise in the normalized data.

[0036] This step requires only a single forward propagation during inference, eliminating the need for multiple Monte Carlo samplings, and simultaneously obtains cognitive uncertainty and data uncertainty, significantly improving computational efficiency. Furthermore, the two types of uncertainty are modeled uniformly within the same Dirichlet parameter framework, overcoming the shortcomings of prediction entropy only reflecting data uncertainty, variance only reflecting model uncertainty, and the two not being modeled uniformly within a theoretical framework.

[0037] Step 3: Dempster-Shafer evidence combination Multi-model nonlinear fusion is implemented based on Dempster-Shafer evidence theory. The Dirichlet evidence for each model is converted into a mass function. Categories reliability quality and uncertainty quality They are respectively: ; ; in Indicates assignment to a single class The reliability quality, Indicates assignment to the recognition framework The uncertainty quality (i.e., "unable to determine a specific category") satisfies both. .

[0038] mass function for any two models and The conflict coefficient of Dempster's combinatorial rule is: ; in The conflict coefficient, obtained by summing the products of the reliability qualities of the two models in different categories, quantifies the degree of discrepancy in predictions between the models. A value close to 0 indicates high consensus in the model, while a value close to 1 indicates significant disagreement.

[0039] Combined reliability Combination uncertainty They are respectively: ; ; The combined reliability is composed of the product of the reliability of the two models of the same type and the product of the quality of each model's confidence and the uncertainty quality of the other, divided by the normalization factor. After deducting conflicting terms, the combined uncertainty consists only of the product of the uncertainties of the two models.

[0040] right Each model iteration applies the above combination rule, where The number of model branches. Total uncertainty after fusion. for: ; in The combined Dirichlet strength, The conflict weighting coefficients. The total uncertainty after fusion also includes Dirichlet cognitive uncertainty (…). ) and DS conflict uncertainty ( The model consists of two parts, with the conflict coefficient serving as an additional uncertainty signal, which naturally quantifies the degree of divergence between models and achieves nonlinear fusion in the evidence space rather than the probability space.

[0041] Step 4: Pathological Prior Guidance-Driven Evidence Correction In areas of high conflict in the model, pathological priors are introduced to correct for evidence. When the conflict coefficient... At that time, among them To trigger the preset conflict threshold, the following correction process is initiated: (1) Construction of the prior knowledge base: It includes three types of pathological priors. The first is the spatial co-occurrence statistics of tissue types, such as the conditional probability of the occurrence of stroma around tumor nests, which is used to characterize the co-occurrence pattern of different tissue types in spatial location; the second is the boundary morphology prior, such as the expected smoothness and concavity range of the tumor boundary, which is used to characterize the morphological constraints of the tissue boundary; the third is the cell density gradient prior, such as the expected density transition pattern at the tissue interface, which is used to characterize the cell density change pattern at the tissue interface. The above priors can be statistically obtained from the training set data and pre-constructed into a prior knowledge base.

[0042] (2) Prior context retrieval: By using a sliding window to statistically analyze the mode of class predictions of each model branch in the local neighborhood of the pixel to be corrected, the consensus of neighborhood tissue type prediction is determined, and the expected anatomical prior probability distribution is retrieved from the prior knowledge base. .

[0043] (3) Evidence Correction: For each model, the evidence is corrected according to the degree of deviation between its prediction and the pathological prior. ; in For the model In pixels First The amount of original evidence for a class The revised amount of evidence. Predict probability distribution for the model Prior distribution of pathology KL divergence between them This represents the prior strength coefficient. The greater the deviation of the prediction from the prior, the more its evidence is attenuated, thus suppressing erroneous evidence from individual models from dominating the fusion result. After correction, the Dempster combination rule is re-executed to obtain the fusion evidence after prior correction.

[0044] Through the above mechanism, the model divergence region obtains an objective and interpretable basis for adjudication, overcoming the shortcomings of existing methods that do not incorporate pathological priors into the uncertainty estimation and fusion process.

[0045] Step 5: Multi-scale boundary perception spatial smoothing in the evidence space A multi-scale boundary-aware spatial consistency constraint is applied to the evidence space to suppress isolated noise in the evidence map and maintain the evidence sharpness of tissue boundaries. This step is implemented as a loss term during the training phase and specifically includes: (1) Image gradient calculation: After converting the input pathological image into a grayscale image, the first-order difference gradient is calculated along the horizontal and vertical directions respectively. and The gradient magnitude map is obtained as follows: ; in This represents the gradient magnitude of the image, with larger values ​​at tissue boundaries and smaller values ​​in uniform regions within the tissue.

[0046] (2) Boundary-aware weight map construction: Boundary weights are calculated based on image gradient magnitude. ; in For boundary-aware weights, is the boundary sensitivity coefficient. Where the image gradient is large (i.e., at the tissue boundary), the boundary weight approaches 0, relaxing the smoothing constraint to preserve the boundary sharpness; where the image gradient is small (i.e., in a uniform region within the tissue), the boundary weight approaches 1, applying a stronger smoothing constraint to suppress evidence noise.

[0047] (3) Multi-scale neighborhood consistency measure: Let the fused evidence be... For multiple neighborhood scales (The number of preset scales is recorded as) ), respectively Sliding window calculation of neighborhood mean And calculate the difference between the current pixel evidence and the neighborhood mean: ; The multi-scale consistency loss is a weighted sum of the consistency at each scale: ; in For scale Scale weights, This indicates that the average value of the pixels is calculated.

[0048] This loss imposes constraints on the evidence space rather than the probability space, allowing the smoothing process to directly affect the Dirichlet evidence quantity, thus preserving the semantic integrity of the evidence theory. The multi-scale mechanism simultaneously captures local details (small-scale neighborhood) and regional trends (large-scale neighborhood), while boundary-aware weights ensure that abrupt changes in evidence at organizational boundaries are not erroneously smoothed.

[0049] Step 6: Mutual Promotion of Uncertainty-Gated Dirichlet Space Models Based on uncertainty-gated model mutual promotion loss, adaptive knowledge sharing among multiple models is achieved in the Dirichlet parameter space. The core idea is: in low uncertainty regions (high model confidence), the Dirichlet distributions of different models are forced to align to achieve knowledge sharing and consistency; in high uncertainty regions (low model confidence), the diversity of distributions among models is preserved to avoid error propagation. This step is also implemented as a loss term during the training phase, specifically including: (1) Dirichlet KL divergence calculation: For any two model branches and Dirichlet parameters and Calculate the KL divergence between Dirichlet distributions. This measures the difference in the predicted distributions of the two models at that pixel. The Dirichlet KL divergence has a closed-loop solution and does not require sampling estimation.

[0050] (2) Construction of uncertainty gating factors: based on fusion of cognitive uncertainty Constructing gating factors : ; in This indicates that the value will be truncated to... Interval. Regions approaching 0 (high model confidence), gating factor Approaching 1, the mutual reinforcement loss is fully applied; in Regions approaching 1 (low model confidence), gating factor As the value approaches zero, the mutual promotion loss is suppressed.

[0051] (3) Pairwise mutual promotion loss aggregation: For all model pairs Calculate the gated KL divergence and take the average: ; in For the number of model pairs.

[0052] This mechanism implements an adaptive and mutually reinforcing strategy of "knowledge sharing in certain regions and maintaining diversity in uncertain regions," avoiding the error propagation problem caused by forced alignment in high-uncertainty regions due to traditional consistency loss, while utilizing the closed-form solution of Dirichlet KL divergence to ensure computational efficiency.

[0053] Step 7: Model Training and Inference A strategy of "guided training and no sampling during inference" is adopted: (1) Training phase: every In the next iteration, the Dropout layer of the backbone network is enabled. Monte Carlo forward propagation, among which To preset the iteration interval, Number of Monte Carlo samplings; Calculation of statistical uncertainty As a "signal of uncertainty for teachers" This is a normalized combination of prediction entropy and variance.

[0054] (2) Distillation loss: The evidence prediction head is trained by distillation loss to reduce the cognitive uncertainty of its predictions. With MC statistical uncertainty Alignment: ; in To provide evidence for predicting cognitive uncertainty in the head output, , where is the KL loss weighting coefficient.

[0055] The evidence-based deep learning loss employs Type II maximum likelihood loss and KL regularization: ; in The Dirichlet parameter after removing evidence of the correct category. For pixel-based true labels, This is the annealing coefficient. The distribution is uniform Dirichlet. Type II maximum likelihood loss constrains the expected class probability to approximate the true label, KL regularization penalizes evidence accumulation on incorrect classes, and the annealing coefficient... The regularization strength is adjusted as the training progresses.

[0056] The total training loss is: ; in For evidence-based deep learning loss, The loss is reweighted for uncertainty, with the weights being... This allows for stronger supervision of training in the model's confidence region; For Dice's loss; For step 5, the evidence space multi-scale boundary perception space smoothing loss; The mutual promotion loss of the Dirichlet space model with uncertainty gated in step 6; Distillation loss from MC to EDL (training only); , , These are the weighting coefficients for the corresponding loss terms.

[0057] The above hyperparameters , , , , , , , , , , , and scale set The specific values ​​can be determined by those skilled in the art through conventional hyperparameter tuning methods such as grid search or Bayesian optimization on the validation set, so as to obtain the optimal parameter combination suitable for specific pathological image segmentation tasks without excessive experiments.

[0058] (3) Inference stage: Dropout is completely disabled. Dirichlet parameters and uncertainties are directly output through a single forward propagation of the evidence prediction head. Based on the expected class probability distribution determined by the fused evidence, the class with the highest probability is taken as the segmentation class of the corresponding pixel, and the pathological image segmentation result is output. This achieves the efficient strategy of "statistical guidance during training and direct prediction during inference", and the inference efficiency is improved by several times.

[0059] Step 8: Level 3 Reliability Assessment and Adaptive Quantile Calibration Based on fusion uncertainty, a three-level reliability classification is adopted, and the reliability classification rules are as follows: ; in For pixels Reliability level, For pixels The uncertainty of integration This is the high / medium reliability dividing threshold. This is the threshold for distinguishing between medium and low reliability.

[0060] Adaptive quantile calibration mechanism: The threshold no longer relies on empirical values, but is automatically learned from the uncertainty distribution of the validation set. Specifically, the fusion uncertainty of all pixels is collected on the validation set, and the third percentile (33%) is taken as the high / medium reliability cutoff threshold. The upper third (67th percentile) is taken as the threshold for medium / low reliability. : ; ; in and These are the 33rd and 67th percentiles of the fusion uncertainty distribution of the validation set, respectively.

[0061] This mechanism ensures that the prior proportions of the reliability distribution remain stable (approximately one-third, one-third, one-third), without drifting with the dataset. Predictions for high-reliability regions can be used directly, saving pathologists time; boundary review is recommended for medium-reliability regions, with a focus on morphological transition areas; manual review or reassessment is recommended for low-reliability regions to avoid the risk of misdiagnosis. The threshold adaptively adjusts with the validation set distribution, avoiding the problem of fixed thresholds failing on different datasets.

[0062] Example 2 Based on the same inventive concept, such as Figure 2 As shown, this embodiment provides a multi-model coordinated pathological image segmentation device based on evidence deep learning and pathological prior guidance, including: The evidence prediction module is used to acquire the pathological image to be segmented and input the pathological image to be segmented into a preset multi-branch pathological segmentation network. Through the evidence prediction heads of each model branch in the multi-branch pathological segmentation network, the pixel-level prediction features are parameterized into Dirichlet distribution parameters through a single forward propagation. Based on the Dirichlet distribution parameters, the amount of evidence and uncertainty index of each model branch at the pixel level are determined. Its specific implementation method is the same as steps 1 and 2 of Embodiment 1.

[0063] The evidence fusion module is used to convert the evidence quantity of each model branch into the confidence quality function in the Dempster-Shafer evidence theory. It performs nonlinear evidence fusion on the confidence quality function of multiple model branches through the Dempster combination rule and calculates the prediction conflict coefficient between model branches. Its specific implementation method is the same as step 3 of Example 1.

[0064] The pathology prior correction module is used to retrieve the target pathology prior probability distribution from the pathology prior knowledge base based on the neighboring tissue prediction consensus when the prediction conflict coefficient is greater than the preset conflict threshold. It then attenuates and corrects the evidence amount of the corresponding model branch according to the statistical divergence deviation between the predicted probability distribution of each model branch and the target pathology prior probability distribution, and triggers the evidence fusion module to re-execute evidence fusion based on the corrected evidence amount to obtain fused evidence. Its specific implementation method is the same as step 4 of Example 1.

[0065] The segmentation and evaluation module is used to determine the expected category probability distribution based on fused evidence and output the pathological segmentation results of the pathological image to be segmented. It combines the uncertainty index of fused evidence and the prediction conflict coefficient to generate a reliability evaluation result. Its specific implementation method is the same as step 8 of Example 1.

[0066] The data transfer relationship between the modules is as follows: the output of the evidence prediction module is connected to the input of the evidence fusion module; the output of the evidence fusion module is connected to the inputs of the pathology prior correction module and the segmentation and evaluation module, respectively; the output of the pathology prior correction module flows back to the input of the evidence fusion module to trigger re-fusion. The specific working process is as follows: IHC pathological images first enter the evidence prediction module, where feature extraction and Dirichlet parameterization are performed through a shared backbone network, branch decoders, and evidence prediction heads, outputting the Dirichlet distribution parameters and uncertainty indices of each model branch; subsequently, the evidence fusion module constructs a mass function based on the Dirichlet evidence of each branch and fuses it using the Dempster combination rule, calculating the conflict coefficient; if the conflict coefficient is greater than a preset conflict threshold, the pathology prior correction module is triggered. This module retrieves the expected probability distribution from the pathology prior knowledge base based on the consensus of neighboring tissue predictions and attenuates and corrects the evidence of each branch according to the degree of deviation between the prediction and the prior distribution. The corrected evidence is then re-inputted into the evidence fusion module for re-fusion; finally, the fused evidence is input into the segmentation and evaluation module, which determines the expected category probability distribution and outputs the segmentation result and reliability evaluation result.

[0067] Furthermore, the device may also include an evidence space smoothing module, a model mutual promotion module, and an MC distillation training module. The evidence space smoothing module applies multi-scale boundary-aware spatial consistency smoothing constraints to the fused evidence during training. The model mutual promotion module enables knowledge sharing between models in the Dirichlet parameter space using uncertainty gating during training. The MC distillation training module periodically enables Dropout layers at preset iteration intervals during training to perform multiple forward propagations to statistically analyze uncertain teacher signals and train the evidence prediction head through distillation loss. The specific implementation of these three modules is consistent with steps 5, 6, and 7 of Embodiment 1, respectively. As functional modules in the training phase, they participate in the training of the multi-branch network through loss backpropagation and do not need to be executed during the inference phase.

[0068] Example 3 This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0069] Specifically, the processor may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0070] The memory may include a large-capacity storage device for data or instructions. For example, and not limitingly, the memory may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory may include removable or non-removable (or fixed) media. Where appropriate, the memory may be internal or external to the data processing device. In a particular embodiment, the memory is non-volatile memory. In a particular embodiment, the memory includes read-only memory (ROM) and random access memory (RAM).

[0071] The memory can be used to store or cache various data files that need to be processed and / or used for communication, as well as possible computer program instructions executed by the processor. The processor reads and executes the computer program instructions stored in the memory to implement any of the multi-model coordinated pathological image segmentation methods based on evidence-based deep learning and pathological prior guidance in the above embodiments.

[0072] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor. The transmission device can be used to receive or send data via a network. The input / output device is used to input or output information.

[0073] Example 4 This embodiment also provides a computer-readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the multi-model coordinated pathological image segmentation method based on evidence deep learning and pathological prior guidance according to Embodiment 1.

[0074] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0075] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others are implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0076] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof.

[0077] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance, characterized in that, Includes the following steps: Acquire the pathological image to be segmented, and input the pathological image to be segmented into a preset multi-branch pathological segmentation network; The evidence prediction heads of each model branch in the multi-branch pathological segmentation network are used to parameterize the pixel-level prediction features into Dirichlet distribution parameters through a single forward propagation, and the amount of evidence and uncertainty index of each model branch at the pixel level are determined based on the Dirichlet distribution parameters. The evidence quantity of each model branch is converted into the confidence quality function in Dempster-Shafer evidence theory. The confidence quality function of multiple model branches is nonlinearly fused using the Dempster combination rule, and the prediction conflict coefficient between model branches is calculated. Determine whether the predicted conflict coefficient is greater than a preset conflict threshold; if it is greater than the preset conflict threshold, retrieve the target pathological prior probability distribution from the pathological prior knowledge base based on the neighborhood tissue prediction consensus, and attenuate the evidence amount of the corresponding model branch according to the statistical divergence deviation between the predicted probability distribution of each model branch and the target pathological prior probability distribution, including: statistically analyzing the category prediction mode of each model branch in the local neighborhood of the pixel to be corrected through a sliding window to determine the neighborhood tissue type prediction consensus, and matching and retrieving the expected anatomical prior probability distribution from the pathological prior knowledge base; calculating the KL divergence between the predicted probability distribution output by each model branch and the anatomical prior probability distribution; multiplying the original evidence amount of each model branch in each category by a negative exponential attenuation term with the KL divergence and the preset prior strength coefficient as exponential factors to attenuate the model branch evidence amount that deviates from the prior distribution, and re-performing evidence fusion based on the attenuated model branch evidence amount to obtain fused evidence; wherein, the pathological prior knowledge base includes tissue type spatial co-occurrence statistical prior, tissue boundary morphological smoothness prior, and interface cell density gradient prior. Based on the fused evidence, the expected category probability distribution is determined, and the pathological segmentation result of the pathological image to be segmented is output. The reliability assessment result is generated by combining the uncertainty index of the fused evidence with the prediction conflict coefficient.

2. The multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in claim 1, characterized in that, Using the evidence prediction heads of each model branch in the multi-branch pathological segmentation network, pixel-level predicted features are parameterized into Dirichlet distribution parameters through a single forward propagation. Based on the Dirichlet distribution parameters, the amount of evidence and uncertainty indices of each model branch at the pixel level are determined, including: The evidence prediction head performs nonlinear activation mapping on the decoded features to output a non-negative evidence vector. Each component of the non-negative evidence vector is incremented by 1 to obtain the Dirichlet distribution parameters of each preset pathological category. The Dirichlet distribution parameters of all categories are summed to obtain the Dirichlet distribution intensity. The cognitive uncertainty representing the degree of ignorance of the model is calculated based on the ratio of the total number of categories to the intensity of the Dirichlet distribution. The expected category probability distribution is determined based on the ratio of the Dirichlet distribution parameter of each preset pathological category to the Dirichlet distribution intensity, and the normalized information entropy of the expected category probability distribution is calculated to characterize the data uncertainty of inherent noise in the data.

3. The multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in claim 1, characterized in that, The evidence quantities of each model branch are converted into the reliability quality function in Dempster-Shafer evidence theory. Nonlinear evidence fusion is then performed on the reliability quality functions of multiple model branches using the Dempster combination rule, and the prediction conflict coefficients between model branches are calculated, including: For each model branch, the ratio of the evidence quantity of each pathological category to the intensity of the Dirichlet distribution is used as the single-point reliability quality of the corresponding category, and the ratio of the total number of categories to the intensity of the Dirichlet distribution is used as the empty set reliability quality representing global uncertainty. The mass function of each model branch is constructed. The sum of the reliability quality products between branch pairs of the model in different categories is calculated using orthogonal summation to determine the prediction conflict coefficient, which characterizes the degree of model divergence. The first-order normalization factor is used to nonlinearly combine the intersection of the class consensus term and the uncertainty term between the model branch pairs, and the fusion class reliability quality and fusion uncertainty quality after the combination of multiple model branches are obtained iteratively. The pixel-level total fusion uncertainty is obtained by summing the combined Dirichlet cognitive uncertainty with the weighted prediction conflict coefficient.

4. The multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in claim 1, characterized in that, A reliability assessment result is generated by combining the uncertainty index of the fused evidence with the predicted conflict coefficient, including: On the validation set, the pixel-level fusion uncertainty distribution is statistically analyzed, and a preset first quantile threshold is extracted as the high / medium reliability boundary threshold, and a preset second quantile threshold is extracted as the medium / low reliability boundary threshold. The total fusion uncertainty of each pixel in the pathological image to be segmented is compared with the high / medium reliability boundary threshold and the medium / low reliability boundary threshold, and an adaptive three-level reliability evaluation map of high reliability, medium reliability or low reliability is output.

5. The multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in claim 1, characterized in that, The multi-branch pathological segmentation network was pre-trained, and its training process included applying multi-scale boundary-aware spatial consistency smoothing constraints to the evidence space. Calculate the first-order difference gradient magnitudes in the horizontal and vertical directions of the input pathological image, and construct negative exponential boundary-aware weights based on the gradient magnitudes; In this process, the weights in the boundary regions of the organization approach zero to suppress smoothing, while the weights in the smooth regions within the organization approach one to enhance smoothing. The local neighborhood mean of the fused evidence is calculated by sliding across multiple preset scale windows, and the squared difference between the pixel fused evidence and the neighborhood mean at the corresponding scale is measured. The squared differences at each scale are multiplied by the boundary-aware weights and then summed at multiple scales to construct a multi-scale smoothing loss for the evidence space, which directly constrains the Dirichlet evidence quantity distribution.

6. The multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in claim 1, characterized in that, The multi-branch pathological segmentation network is pre-trained, and its training process includes: In the Dirichlet parameter space, the analytical closed-form solution of the KL divergence of the Dirichlet distribution between models is calculated based on the Dirichlet distribution parameters of the model branch pairs. Based on the cognitive uncertainty after multi-model fusion, a negatively correlated gating factor is constructed. The gating factor is multiplied by the KL divergence of the Dirichlet distribution and averaged over all model branch pairs to construct the pairwise model mutual promotion loss. During the training phase, the random dropout layer of the backbone network is periodically enabled at a preset iteration interval to perform multiple forward propagations. The statistical prediction entropy and variance are combined and normalized to form a teacher uncertainty signal. The cognitive uncertainty output by the evidence prediction head is aligned with the teacher uncertainty signal by constraining the loss through distillation.

7. An apparatus for implementing the multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in any one of claims 1 to 6, characterized in that, include: The evidence prediction module is used to acquire the pathological image to be segmented and input the pathological image to be segmented into a preset multi-branch pathological segmentation network. Through the evidence prediction head of each model branch in the multi-branch pathological segmentation network, the pixel-level prediction features are parameterized into Dirichlet distribution parameters through a single forward propagation. Based on the Dirichlet distribution parameters, the amount of evidence and uncertainty index of each model branch at the pixel level are determined. The evidence fusion module is used to convert the evidence quantity of each model branch into the confidence quality function in the Dempster-Shafer evidence theory, perform nonlinear evidence fusion on the confidence quality function of multiple model branches through the Dempster combination rule, and calculate the prediction conflict coefficient between model branches. The pathology prior correction module is used to retrieve the target pathology prior probability distribution from the pathology prior knowledge base based on the neighboring tissue prediction consensus when the prediction conflict coefficient is greater than the preset conflict threshold. It then performs attenuation correction on the evidence amount of the corresponding model branch according to the statistical divergence deviation between the prediction probability distribution of each model branch and the target pathology prior probability distribution, and triggers the evidence fusion module to re-execute evidence fusion based on the corrected evidence amount to obtain fused evidence. The segmentation and evaluation module is used to determine the expected category probability distribution based on the fused evidence, output the pathological segmentation result of the pathological image to be segmented, and generate a reliability evaluation result by combining the uncertainty index of the fused evidence with the prediction conflict coefficient.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in any one of claims 1 to 6.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the multi-model coordinated pathological image segmentation method based on evidence-based deep learning and pathological prior guidance as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Cross-modal medical image segmentation method based on pathological anchoring

    CN122289680A

  • System and method for image registration using nonparametric priors and statistical learning techniques

    US20070165943A1