Multimodal medical image imaging information processing method and system

Abstract

The present application belongs to the technical field of medical image processing, and specifically relates to a multi-modal medical image imaging information processing method and system, which solves the problems of unquantified quality influence, insufficient correlation mining, uncontrolled uncertainty, contradiction between cross-institutional collaboration and privacy protection, and lack of continuous evolution capability in the prior art. The present application first acquires images and imaging context information, generates a dynamic quality evaluation flow through multi-scale quality perception evaluation; then, based on the prior causal knowledge base and the evaluation flow, the strictness of the causal discovery algorithm is regulated, and a personalized causal network is constructed; subsequently, the network is relied on to perform counterfactual reasoning to predict the processing effect, and multi-level risk quantification including quality-driven components is completed; further, an imaging information processing interface is generated, and cross-institutional knowledge fusion and model updating, dynamic updating of the prior causal knowledge base are realized. The present application realizes precise processing of multi-modal images, controllable risk and continuous optimization, improves processing reliability and efficiency, and reduces the threshold of engineering application.

Description

Technical Field

[0001] This invention belongs to the field of information processing technology, and specifically relates to a method and system for processing multimodal medical image information. Background Technology

[0002] Medical imaging serves as a crucial reference for clinical diagnosis and treatment. With the rapid development of digital technology, the volume of medical imaging data has exploded. Achieving efficient processing, orderly management, and precise application of this data has become a key requirement in the medical field. To address this need, existing technologies have developed processing solutions for medical imaging information. For example, the invention patent with publication number CN121011315A proposes a multi-module medical imaging information processing method, system, device, and storage medium. This technology, through a processing flow of "interface reception - extraction output - definition storage - construction comparison - traversal correction - traversal transmission," realizes the reception of DICOM format image files, the extraction and storage of imaging information (patient, examination, sequence, image-level information), and file modification and transmission based on key information matching rules. This solves the problem in traditional technologies of being unable to extract and store information from massive amounts of image data in a searchable and statistically feasible manner, providing a fundamental solution for the standardized processing of image data.

[0003] While the aforementioned existing technologies have made some progress in the structured storage and basic processing of image data, they still have many key shortcomings in practical applications for precise imaging information processing. These shortcomings make it difficult to meet the high-level requirements of imaging information processing for data reliability, depth of correlation analysis, risk controllability, and dynamic adaptability. Specific shortcomings are as follows:

[0004] 1. Ignoring the impact of image quality on processing results, the reliability of imaging information processing is not guaranteed: Existing technologies only achieve standardized reception and information extraction of image data, without evaluating and quantifying the quality of input images (such as motion artifacts, signal-to-noise ratio, anatomical coverage integrity, etc.), and without establishing a correlation between image quality and the reliability of subsequent information extraction, storage, and matching comparison results. When low-quality images are input, the extracted imaging information is prone to deviation, which in turn leads to errors in subsequent key information matching and file modification. The reliability of the entire imaging information processing process cannot be guaranteed, and there is a lack of corresponding quality risk warning mechanisms.

[0005] 2. Limited to information extraction and matching, lacking the ability to mine imaging information correlations: The core logic of existing technologies is based on predefined rules for information extraction and key information comparison, a simple "rule matching" process. It fails to delve into the intrinsic relationships between different dimensions of imaging information, or between imaging information and imaging parameters, and cannot construct the causal logic in the imaging information generation process. The processing only completes basic information extraction and matching, failing to achieve in-depth analysis and value mining of imaging information, thus limiting the deep application of this technology in high-precision imaging information processing scenarios.

[0006] 3. Lack of consideration for uncertainties in imaging information processing and lack of risk management capabilities: Existing technologies only output deterministic information extraction results and file modification / sending commands, without quantifying and decomposing the uncertainties that may exist in the imaging information processing process (such as information extraction deviations caused by image quality defects, ambiguity in rule matching, and information deviations caused by fluctuations in imaging parameters). Staff cannot know the reliability of the imaging information processing results and the sources of potential risks, making it difficult to formulate targeted risk avoidance strategies. This limits its application in high-precision imaging information processing scenarios.

[0007] 4. Insufficient collaboration in cross-institutional imaging information sharing and privacy protection, and lack of continuous evolution capability: Existing technologies do not consider imaging information sharing scenarios across medical institutions. If directly applied to multi-institutional collaboration, the privacy of raw imaging data is easily leaked. At the same time, its predefined extraction and matching rules are statically set and cannot be dynamically optimized based on the differentiated imaging data quality of each institution, the differences in imaging equipment parameters, and the new imaging data types. As imaging technology is updated and imaging information processing needs change, this technology is prone to performance degradation and cannot adapt to dynamic imaging information processing application scenarios.

[0008] To address the core shortcomings of existing technologies, there is an urgent need to construct an imaging information processing method and system that uses image quality as the core control element, causal reasoning as the basis for correlation mining, privacy-preserving federated learning as a collaborative means, and closed-loop feedback as the driving force for evolution. This system would specifically solve problems such as neglect of quality, lack of deep correlation mining, lack of risk control, and inability to evolve dynamically. It would achieve precise imaging information processing throughout the entire process of "quality perception - deep analysis - collaborative optimization - continuous evolution," thereby improving the reliability, depth, and applicability of medical imaging technology in the field of imaging information processing. Summary of the Invention

[0009] To address the aforementioned problems in existing technologies, namely, the lack of quantification of quality impact, insufficient correlation mining of imaging information, uncontrolled processing uncertainty, conflict between cross-institutional data sharing and privacy protection, and lack of continuous evolution capabilities, which make it difficult to meet the high-level requirements of imaging information processing for data reliability, correlation mining depth, risk controllability, and dynamic adaptability, the present invention, in its first aspect, proposes a multi-module medical image processing method, comprising the following steps:

[0010] Obtain raw medical image data and associated imaging context information;

[0011] The original medical image data is subjected to multi-scale quality perception assessment to generate a dynamic quality assessment stream;

[0012] A prior causal knowledge base is acquired, and based on the dynamic quality assessment flow, the strictness of the causal discovery algorithm is adaptively adjusted to construct a personalized causal network model for characterizing the relationship between image features and imaging parameters.

[0013] Using the personalized causal network model as an interpretable structure, combined with the dynamic quality assessment flow, counterfactual reasoning is performed to predict the expected processing effect of adjusting image features or imaging parameters, and the expected processing effect is subjected to multi-level uncertainty quantization decomposition to obtain uncertainty decomposition results; the uncertainty quantization decomposition includes a quality-driven operational uncertainty component derived from image quality variation propagated through the personalized causal network model.

[0014] Based on the original medical image data, the personalized causal network model, the expected processing effect, and the uncertainty decomposition result, an interactive and visual imaging information processing interface is generated.

[0015] In a distributed learning scenario, the aggregation process of the local personalized causal network model is differentially weighted based on the historical quality assessment profiles of each participating node, thereby achieving cross-node knowledge fusion and updating of the local personalized causal network model under privacy protection; wherein, the historical quality assessment profiles are constructed based on the dynamic quality assessment streams generated by each node in the past.

[0016] The interactive feedback data in the interactive visualization imaging information processing interface is acquired, and the feedback data is calibrated according to the dynamic quality assessment stream corresponding to the source of the interactive feedback data, thereby driving the dynamic update of the prior causal knowledge base.

[0017] In some preferred embodiments, a personalized causal network model for characterizing the relationship between image features and imaging parameters is further constructed, and the method is as follows:

[0018] Obtain deterministic causal edges, prohibited edges, and temporal constraints from the prior causal knowledge base;

[0019] A fully connected graph is constructed using image features and imaging parameters as nodes, and the fully connected graph is forcibly modified according to the prior causal knowledge base, including forcibly retaining the deterministic causal relationship edges and forcibly deleting the prohibited edges, to obtain a prior constraint graph.

[0020] Based on the prior constraint graph, an improved PC algorithm framework is used for causal discovery, and an undirected skeleton graph is constructed through quality-aware conditional independence tests. In each conditional independence test, the significance level of the statistical test is dynamically adjusted according to the dynamic quality assessment flow, so that the test strictness is negatively correlated with data quality.

[0021] Based on V-structure recognition, directional propagation rules, and temporal relationships in the prior causal knowledge base, causal direction arbitration is performed on the undirected skeleton graph to form a directed acyclic personalized causal network model, and causal strength and reliability weights are assigned to the causal edges in the personalized causal network model.

[0022] In some preferred embodiments, the method for determining the quality-driven operational uncertainty component derived from image quality variations propagated through the personalized causal network model is as follows:

[0023] A measurement error model for key image features is established, wherein the standard deviation of the feature measurement error is negatively correlated with the dynamic quality assessment flow.

[0024] Based on the measurement error model, the current image feature vector is sampled multiple times using Monte Carlo perturbation to generate a feature vector simulating quality defects.

[0025] The feature vectors of the simulated quality defects are respectively input into the counterfactual prediction model constructed based on the personalized causal network model to obtain the corresponding set of expected processing effect values.

[0026] Calculate the variance of the expected processing effect value set as the uncertainty component of the quality-driven operation.

[0027] In some preferred embodiments, the uncertainty quantification decomposition further includes:

[0028] By resampling the training data using the bootstrap method and reconstructing multiple bootstrap reconstructed causal network models, the variance of the expected processing effect under each bootstrap reconstructed causal network model is calculated, and the model structure uncertainty component is obtained.

[0029] The modal alignment uncertainty component is obtained by calculating the difference between the expected processing effects of sub-models from different image modalities;

[0030] By finding similar historical cases in the feature space, the residual variance between the actual imaging processing results and the expected processing effect predicted based on the localized personalized causal network model is calculated to obtain the individual heterogeneity uncertainty component.

[0031] In some preferred embodiments, an interactive decision-making sandbox is generated as the interactive visual imaging information processing interface, the sandbox comprising:

[0032] An interactive causal network diagram is used to display the personalized causal network model.

[0033] Multimodal image maps are used to display the original medical image data and its associated anatomical structure annotations;

[0034] The processing effect analysis chart is used to show the expected processing effect and the uncertainty decomposition results;

[0035] In this context, a bidirectional interactive mapping is established between the nodes in the interactive causal network graph and the corresponding anatomical region segmentation mask in the multimodal image graph.

[0036] The processing effect analysis diagram includes: a control for interactively adjusting the hypothesis value of the dynamic quality assessment flow, which is configured to update the uncertainty range of the expected processing effect in real time based on the adjusted value; and a hypothesis simulation engine, which is configured to respond to user operations, trigger real-time counterfactual reasoning for different parameter adjustments, and dynamically update the visualization of the expected processing effect and its multi-level uncertainty decomposition results.

[0037] In some preferred embodiments, the method for achieving privacy-preserving cross-node knowledge fusion and local personalized causal network model update is as follows:

[0038] Each participating node performs privacy processing on the parameters of the locally trained personalized causal network model.

[0039] The federated coordination server calculates the differential weight of each node in the local personalized causal network model aggregation based on the local data volume of each node and the average value of its historical dynamic quality assessment stream.

[0040] Based on the aforementioned differential weights, the privacy-enhanced parameters of the local personalized causal network model are securely aggregated to generate a global causal model.

[0041] Based on the historical quality assessment profiles of each node, the mixing coefficients are determined, and the global causal model is fused with the local personalized causal network model to obtain the updated local personalized causal network model.

[0042] In some preferred embodiments, the method for driving the dynamic update of the prior causal knowledge base is as follows:

[0043] The user's operational feedback behavior in the interactive visual imaging information processing interface is transformed into feedback evidence for a specific causal path.

[0044] The strength of the feedback evidence is multiplied by the corresponding dynamic quality assessment stream that generated the feedback to perform evidence quality calibration.

[0045] The confidence levels of relevant causal relationships in the prior causal knowledge base are updated using a Bayesian update method with calibrated evidence.

[0046] When the change in the confidence level exceeds the threshold, an update to the prior causal knowledge base is confirmed.

[0047] In some preferred embodiments, a bidirectional interactive mapping is established between nodes in the interactive causal network graph and corresponding anatomical region segmentation masks in the multimodal image image, the method being as follows:

[0048] When an image feature node is selected in the causal network diagram, the corresponding anatomical region segmentation mask is highlighted in the multimodal image; and

[0049] When an image region is selected in the multimodal image map, the corresponding image feature node is highlighted in the causal network map.

[0050] In some preferred embodiments, causal strength and reliability weights are assigned to the causal edges in the personalized causal network model, specifically including:

[0051] For each directed edge X→Y in the directed acyclic graph, perform the following calculation:

[0052] Estimate the average causal effect value of the side, and combine the absolute value of the effect value with the minimum p-value obtained when the side is tested for conditional independence to calculate the causal strength of the side;

[0053] Based on the dynamic quality assessment stream corresponding to the image data that generates the current personalized causal network model, and the minimum p-value, the reliability weight of the edge is calculated; wherein, the reliability weight is positively correlated with the value of the dynamic quality assessment stream.

[0054] The second invention is a multi-module medical image processing system, characterized in that the system comprises:

[0055] The data acquisition module is configured to acquire raw medical image data and associated imaging context information.

[0056] The quality perception assessment module is configured to perform multi-scale quality perception assessment on the original medical image data and generate a dynamic quality assessment stream.

[0057] The causal discovery and network construction module is configured to: acquire a prior causal knowledge base, and adaptively adjust the strictness of the causal discovery algorithm based on the dynamic quality assessment flow, thereby constructing a personalized causal network model to characterize the relationship between image features and imaging parameters.

[0058] The counterfactual reasoning and uncertainty quantification module is configured to: use the personalized causal network model as an interpretable structure, combine it with the dynamic quality assessment flow, perform counterfactual reasoning to predict the expected processing effect of adjusting image features or imaging parameters, and perform multi-level uncertainty quantification decomposition on the expected processing effect to obtain uncertainty decomposition results; wherein, the quantification decomposition includes a quality-driven operational uncertainty component derived from image quality variation propagated through the personalized causal network model.

[0059] The interactive processing interface generation module is configured to generate an interactive visual imaging information processing interface based on the original medical image data, the personalized causal network model, the expected processing effect, and the uncertainty decomposition result.

[0060] The federated knowledge fusion and update module is configured as follows: In a distributed learning scenario, based on the historical quality assessment profiles of each participating node, the aggregation process of the local personalized causal network model is differentially weighted to achieve cross-node knowledge fusion and update of the local personalized causal network model under privacy protection; wherein, the historical quality assessment profiles are constructed based on the dynamic quality assessment streams generated in the history of each node;

[0061] The feedback-driven knowledge base evolution module is configured to: acquire interactive feedback data in the interactive visual imaging information processing interface, calibrate the feedback data according to the dynamic quality assessment stream corresponding to the source of the interactive feedback data, and drive the dynamic update of the prior causal knowledge base.

[0062] The beneficial effects of this invention are:

[0063] 1) A dynamic quality assessment flow is generated through multi-scale quality perception assessment, which runs through the entire process of causal discovery and uncertainty quantification, so as to achieve dynamic matching between algorithm rigor and data quality; a personalized causal network is used as an interpretable structure to clearly represent the causal relationship between image features and imaging parameters, and the effect of parameter adjustment is intuitively analyzed by combining counterfactual reasoning, which greatly improves the reliability, interpretability and traceability of the processing results.

[0064] 2) Multi-level uncertainty quantification + cross-institutional privacy collaboration to enhance risk management and handling capabilities: Construct a four-level quantification framework with quality-driven operational uncertainty components to accurately identify and handle risk sources; adopt a quality-data-volume weighted federated aggregation strategy, combined with privacy protection technology, to achieve high-quality knowledge fusion of multiple institutions under the premise of ensuring data security, avoid low-quality data pollution, and improve the overall imaging information processing level.

[0065] 3) Collect user feedback through interactive sandbox, combine it with case quality calibration feedback evidence, and use Bayesian methods to dynamically update the prior causal knowledge base to achieve continuous self-improvement of the system through "data-knowledge-processing-feedback-optimization" and adapt to the ever-updating imaging technology and processing needs in the long term. Attached Figure Description

[0066] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0067] Figure 1 This is a flowchart of the steps of the multi-module medical image processing method of the present invention.

[0068] Figure 2 This is a flowchart of the personalized causal network model construction process for the multi-module medical image information processing method of the present invention.

[0069] Figure 3 This is a flowchart illustrating the dynamic update of the prior causal knowledge base of the multi-module medical image processing method of this invention.

[0070] Figure 4 This is a module diagram of the multi-module medical imaging information processing system of the present invention. Detailed Implementation

[0071] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0072] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0073] To more clearly explain the multi-module medical image processing method of the present invention, the following will be combined with... Figures 1 to 4 The steps in the embodiments of the present invention will be described in detail below.

[0074] The first embodiment of the present invention, a multi-module medical image processing method, see [link to relevant documentation]. Figure 1 Taking lung cancer CT+PET-CT multimodal image processing as an application scenario, the process is described in detail step by step, as follows:

[0075] S1, acquire raw medical image data (e.g., lung cancer CT and PET-CT multimodal DICOM image data streams from imaging equipment) and associated imaging context information (acquired from the hospital HIS / RIS system, including the purpose of the examination being lung cancer-related image quality assessment, the requesting department being the radiology department, and imaging equipment parameter information, such as CT tube voltage, tube current, PET-CT tracer dosage, etc.).

[0076] S2, perform multi-scale quality perception assessment on the original medical image data (quantify the image data quality from both technical quality and applicability dimensions), and generate a dynamic quality assessment stream;

[0077] In this embodiment, the dynamic quality assessment flow is generated as follows:

[0078] S21, extract the technical quality parameters of the original medical image data. These parameters include at least the motion artifact index M (obtained by quantifying the motion blur of CT images through frequency domain analysis and edge detection algorithms; typically M=3, 0-10 points, 0 for no artifacts), signal-to-noise ratio (SNR) (calculated as the ratio of the mean signal intensity to the standard deviation of background noise in a uniform lung muscle region, yielding SNR=28, reference value SNR_ref=30), and contrast-to-noise ratio (CNR) (calculated as the ratio of the difference in mean signal intensity between the lung cancer lesion and adjacent normal lung tissue regions to the standard deviation of background noise, yielding CNR=15, reference value CNR_ref=18). Normalize and weightedly fuse these parameters to generate a technical quality score Q_tech. The expression for Q_tech is:

[0079] Q_tech = 0.4×(1−M / 10) + 0.3×(SNR / SNR_ref) + 0.3×(CNR / CNR_ref);

[0080] Substituting the data, we get: Q_tech = 0.4×0.7 + 0.3×0.933 + 0.3×0.833 ≈ 0.81;

[0081] S22. To assess the suitability of images for clinical treatment needs and ensure that images contain key anatomical information to meet the basic requirements for subsequent causal relationship mining, a pre-trained anatomical structure segmentation model is adopted to evaluate the coverage integrity and boundary clarity of the original medical image data for key anatomical structures and generate an applicability score.

[0082] The anatomical coverage integrity score (S_cov) was calculated based on the segmentation of key lung anatomical structures (left lung, right lung, and tumor lesion) using a pre-trained nnU-Net model. The volume coverage was 92%, and the average gradient magnitude for boundary clarity was 250, resulting in an overall S_cov score of 8.5 (0-10). The Sigmoid function was used to map the result to Q_clin.

[0083] Q_clin = 1 / [1+exp(−0.5×(S_cov−6))];

[0084] Substituting the data, we calculate: Q_clin≈0.778;

[0085] S23. The harmonic mean method (which highlights the impact of low-scoring indicators and avoids high scores in a single dimension from masking overall quality defects) is used to fuse the technical quality score and the applicability score to generate the dynamic quality assessment flow Q_fused.

[0086] Q_fused = (Q_tech×Q_clin) / (Q_tech+Q_clin−Q_tech×Q_clin);

[0087] Substituting the data, we calculate: Q_fused≈0.658;

[0088] S24. Based on the comparison result between the dynamic quality assessment stream and the preset threshold, trigger structured quality alarms of different levels;

[0089] The preset quality threshold is 0.7. A level 2 alarm is triggered when Q_fused < 0.7 (0.5 ≤ Q_fused < 0.6), which notifies the user to verify the image quality. A level 1 alarm is triggered when Q_fused < 0.5, indicating that the image cannot be used for subsequent processing. This is to prevent low-quality data from entering the subsequent processing stage and causing erroneous results.

[0090] S3. Obtain a prior causal knowledge base and, based on the dynamic quality assessment flow, adaptively adjust the strictness of the causal discovery algorithm to construct a personalized causal network model for characterizing the relationship between image features and imaging parameters; provide an interpretable structural basis for subsequent counterfactual reasoning and solve the problem of insufficient correlation mining in traditional methods.

[0091] In this embodiment, see Figure 2 Furthermore, a personalized causal network model is constructed to characterize the relationship between image features and imaging parameters. The method is as follows:

[0092] S31, obtain deterministic causal relationship edges, prohibited edges, and temporal relationship constraints from the prior causal knowledge base; the prior causal knowledge base is structured as a "node-relationship-attribute" triple format, where nodes include image features (signal-to-noise ratio, artifact index, etc.) and imaging parameters (tube voltage, tube current, etc.), and relations include deterministic causal edges (marked "certainly exist"), prohibited edges (marked "certainly do not exist"), and temporal relationships (marked "prior to" and "subsequent to"); the knowledge base is constructed using a "knowledge input + historical data verification" method, where the radiology department first sorts out the core causal relationships, then verifies the validity of the relationships through 100,000 historical image data, and finally forms a structured knowledge base;

[0093] S32, construct a fully connected graph G_initial using image features and imaging parameters as nodes, and forcibly modify the fully connected graph according to the prior causal knowledge base, including forcibly retaining the deterministic causal relationship edges ("tube voltage → signal-to-noise ratio", "tube current → artifact index") and forcibly deleting the prohibited edges ("tracer dose → tumor gray value"), to obtain a prior constraint graph; used to reduce the computational load of subsequent causal discovery, while avoiding obviously erroneous associations;

[0094] S33, based on the prior constraint graph, an improved PC algorithm framework is used to perform causal discovery, and an undirected skeleton graph is constructed through quality-aware conditional independence tests; wherein, when performing each conditional independence test, the significance level of the statistical test is dynamically adjusted according to the dynamic quality assessment flow, so that the test strictness is negatively correlated with data quality.

[0095] In this embodiment, low-quality image data corresponds to lower test strictness, reducing false positives of spurious associations; high-quality data corresponds to higher test strictness, ensuring the reliability of causal associations. The improved PC algorithm is an improvement on the traditional PC algorithm (Peter-Clark algorithm), addressing the issue of large data quality differences in medical image processing scenarios by adding a dynamic adjustment mechanism for the significance level of quality perception. The core process of the traditional PC algorithm is "initializing a completely undirected graph → deleting edges through conditional independence test → identifying V-structures → orienting edges." The improved PC algorithm retains this core process, but the key improvement lies in integrating a dynamic quality assessment flow (Q_fused) into the conditional independence test stage, achieving adaptive matching between test strictness and data quality. The specific adaptation logic is as follows:

[0096] Dynamic adjustment of significance level: The adjusted significance level α_adj = α_base / Q_fused is calculated based on Q_fused, where α_base = 0.05. Substituting this into the equation, we get α_adj ≈ 0.076 (limited to [0.001, 0.1]), thus achieving a negative correlation between test strictness and data quality.

[0097] Conditional independence test implementation: The Pearson partial correlation coefficient is used for the test. The calculation method is as follows: For nodes X, Y and conditional subset Z, the influence of Z on X and Y is first eliminated by linear regression to obtain residuals e_X and e_Y. Then, the Pearson correlation coefficient r of e_X and e_Y is calculated. If the p value corresponding to |r| is greater than α_adj, then X and Y are determined to be independent under condition Z, and the edge is deleted.

[0098] The test loop is executed as follows: The size of the separation set d is gradually increased from 0 to 3. For each pair of adjacent nodes in the prior constraint graph, the corresponding condition subset Z is found. The conditional independence is tested by the Pearson correlation coefficient. For example, if the p value of "pipe current-artifact index" is 0.085 > α_adj = 0.076, the conditional independence is determined and the edge is deleted. Finally, the undirected skeleton graph is obtained.

[0099] S34, based on V-structure recognition, directional propagation rules and temporal relationships in the prior causal knowledge base, causal direction arbitration is performed on the undirected skeleton graph to form the directed acyclic personalized causal network model, and causal strength and reliability weights are assigned to the causal edges in the personalized causal network model.

[0100] In this embodiment, the undirected skeleton graph is arbitrated for causal orientation. Specifically, the V-structure is first identified (e.g., “tube voltage-signal-noise ratio-artifact index” is oriented as “tube voltage → signal-noise ratio ← artifact index”), and then the temporal relationship “imaging parameter setting → image generation” is combined to orient “tube voltage → tumor region gray value” to avoid loop structures and form a directed acyclic personalized causal network model graph G_k.

[0101] Assigning causal strength and reliability weights to the causal edges in the personalized causal network model, specifically including:

[0102] For each directed edge X→Y in the directed acyclic graph, perform the following calculation:

[0103] Estimate the average causal effect value of this side. The absolute value of the effect is then combined with the minimum p-value (p_min) obtained when performing the conditional independence test for that edge to calculate the causal strength of that edge, as shown in the following formula:

[0104] Strength = |τ̂| × (1 - p_min);

[0105] Taking the "tube voltage → signal-to-noise ratio" side as an example, the average causal effect value is estimated using DML. =0.42 (meaning that for every 10kV increase in tube voltage, the signal-to-noise ratio increases by an average of 0.42 times). The minimum p-value obtained by this side in the conditional independence test is p_min=0.02. Substituting it into the formula, we get: Strength=|0.42|×(1-0.02)=0.42×0.98≈0.412;

[0106] Based on the dynamic quality assessment stream corresponding to the image data that generated the current personalized causal network model, and the minimum p-value, the reliability weight of this edge is calculated; wherein, the reliability weight W is positively correlated with the value of the dynamic quality assessment stream and is also affected by the significance test, and the calculation formula is as follows:

[0107] W = Q_fused × (1 - p_min);

[0108] Taking the "tube voltage → signal-to-noise ratio" side as an example, with Q_fused = 0.658 and p_min = 0.02, substituting into the formula, we get: W = 0.658 × (1 - 0.02) = 0.658 × 0.98 ≈ 0.645. This result meets the positive correlation requirement that "the higher the Q_fused, the greater the reliability weight".

[0109] In summary, this step outputs a personalized causal graph G_k (containing 20 nodes and 18 directed edges, each edge containing direction, causal strength, and reliability weight) and a causal discovery report (120 tests, 15 edges deleted, key causal paths: "tube voltage → signal-to-noise ratio → image sharpness" and "tube current → artifact index → ​​image quality score"), achieving accurate representation of causal relationships.

[0110] S4. To address the issue of uncontrolled uncertainty in traditional methods, the following approach is adopted: using the personalized causal network model as an interpretable structure, combined with the dynamic quality assessment flow, counterfactual reasoning is performed to predict the expected processing effect of adjusting image features or imaging parameters, and the expected processing effect is subjected to multi-level uncertainty quantification decomposition to obtain the uncertainty decomposition result; the uncertainty quantification decomposition includes a quality-driven operational uncertainty component derived from image quality variation propagated through the personalized causal network model.

[0111] In this embodiment, the uncertainty decomposition result is obtained as follows:

[0112] S41, obtain input parameters, including the current image feature vector x (20-dimensional, including tumor volume = 8cm³, PET metabolic value = 12SUV, etc.), the intervention option to be evaluated T=t (parameter adjustment corresponding to targeted therapy), the target-related outcome variable Y (1-year survival rate), the personalized causal network model graph G_k, and the dynamic quality assessment flow Q_fused=0.658;

[0113] S42, based on G_k, identify the parent nodes (tumor volume, PET metabolic value, age) of T (parameter adjustment corresponding to targeted therapy) and Y (1-year survival rate), and construct the covariate set C;

[0114] S43 uses a dual machine learning (DML) framework to calculate the expected processing effect, and the method is as follows:

[0115] Step 1: Fit the intervention allocation model g(C)=E[T|C] and the potential outcome model m(C)=E[Y|C] using a gradient boosting tree (parameter configuration: tree depth 5, learning rate 0.1, number of iterations 100, minimum number of sample splits 20);

[0116] Step 2: Calculate the orthogonal score φ = (Ym(C)) × (Tg(C));

[0117] Step 3: Estimate the expected treatment effect using the effect size calculation formula, as follows:

[0118] =E[φ] / E[(Tg(C))²] ;

[0119] Substituting the data, we calculated μ̂=0.42, which corresponds to the expected treatment effect of a 42% increase in the 1-year survival rate.

[0120] In this embodiment, the expected processing effect is decomposed into multiple levels of uncertainty quantification to obtain the uncertainty decomposition result, including a four-level uncertainty quantification decomposition, specifically:

[0121] The model structure uncertainty component (σ²_structure) is obtained by resampling the training data using the bootstrap method and reconstructing multiple personalized causal network models. The variance of the expected processing effect under each personalized causal network model is calculated, and the model structure uncertainty component is obtained. For example, if Bootstrap sampling B=100 times is performed from the original training data to generate 100 bootstrap sample sets, and S3 is re-executed for each sample set to obtain 100 causal graphs, and S4 is executed to obtain the expected processing effect, the variance is then calculated to obtain σ²_structure=0.008 and σ_structure=0.089.

[0122] The modal alignment uncertainty component (σ²_alignment) is obtained by calculating the difference between the expected processing effects of sub-models from different image modalities. For example, by constructing a CT single-modal sub-model and a PET-CT fusion modal sub-model respectively, and calculating the variance of the expected processing effects of the CT sub-model and the PET-CT fusion sub-model, we get σ²_alignment=0.005 and σ_alignment=0.071.

[0123] The individual heterogeneity uncertainty component (σ²_heterogeneity) is obtained by finding similar historical cases in the feature space and calculating the residual variance between their actual imaging processing results and the expected processing effect predicted based on the personalized causal network model. For example, the K-nearest neighbor algorithm (K=20) is used to find 20 most similar historical cases in the feature space, and the residual variance between their actual outcomes and the model predicted effect values ​​is calculated. After quality calibration, σ²_heterogeneity=0.006 and σ_heterogeneity=0.077 are obtained.

[0124] The quality-driven operational uncertainty component (σ²_operation) is derived from image quality variations propagated through the personalized causal network model, and its method is as follows:

[0125] 1. Establish a measurement error model for key image features: σ_i = α_i + β_i × (1 − Q_fused), where the standard deviation of feature measurement error is negatively correlated with the dynamic quality assessment flow; based on historical data calibration, the tumor volume features are α = 0.2 and β = 0.5. Substituting the data, we calculate σ_tumor volume = 0.2 + 0.5 × (1 − 0.658) = 0.2 + 0.5 × 0.342 = 0.371;

[0126] 2. Based on the measurement error model, perform multiple Monte Carlo perturbation samplings on the current image feature vector (the sampling distribution is a normal distribution N(f_i,σ_i), and the perturbation range is constrained to ±10% of the actual physical range of the feature to avoid the sampled values ​​from exceeding a reasonable range), to generate feature vectors simulating quality defects; for example, Monte Carlo perturbation sampling M=1000 times, each time sampling key features such as tumor volume in x from N(f_i,σ_i), generating 1000 feature vectors simulating quality defects;

[0127] 3. Input the feature vectors of the simulated quality defects into the counterfactual prediction model constructed based on the personalized causal network model to obtain the corresponding set of expected processing effect values; accordingly, input 1000 feature vectors into the counterfactual prediction model to obtain 1000 expected processing effect values, and calculate the variance to obtain σ²_operation=0.012, σ_operation=0.109;

[0128] Calculate the variance of the expected processing effect value set as the uncertainty component of the quality-driven operation;

[0129] Furthermore, synthesizing multidimensional uncertainty yields total uncertainty, and confidence intervals quantify the range of fluctuations in expected effects, providing an intuitive risk reference for decision-making;

[0130] Specifically, the total uncertainty is synthesized using the formula for calculating total uncertainty, as follows:

[0131] U_total=√(σ²_structure+σ²_operation+σ²_alignment+σ²_heterogeneity);

[0132] Substituting the data, we get:

[0133] U_total=√(0.008+0.012+0.005+0.006)=√0.031≈0.176;

[0134] The 90% confidence interval is calculated using the confidence interval formula, as follows:

[0135] CI_90=[μ̂−1.645×U_total, μ̂+1.645×U_total];

[0136] Substituting the data, we get: CI_90≈[0.42−0.289, 0.42+0.289]≈[0.131, 0.709];

[0137] The contribution ratio of uncertainty in quality-driven operations is calculated using a proportional formula, as follows:

[0138] p_operation=σ²_operation / (σ²_structure+σ²_operation+σ²_alignment+σ²_heterogeneity);

[0139] Substituting the data, we calculate: p_operation = 0.012 / 0.031 ≈ 38.7%;

[0140] In summary, this step outputs the associated expected treatment effect of a 42% increase in the expected 1-year survival rate, with a 90% confidence interval [13.1%, 70.9%], and a Level 4 uncertainty decomposition report (including the values ​​of each component and their contribution ratios, with quality-driven operational uncertainty accounting for 38.7%).

[0141] S5. Based on the original medical image data, the personalized causal network model, the expected processing effect, and the uncertainty decomposition result, an interactive and visual imaging information processing interface is generated; this enables deep human-computer interaction, improves the intuitiveness and efficiency of processing decisions, and reduces manual workload.

[0142] In this embodiment, an interactive decision-making sandbox is generated as the interactive visual imaging information processing interface. The sandbox includes:

[0143] An interactive causal network graph is used to display the personalized causal network model. Specifically, a force-directed graph algorithm (layout parameters: repulsion coefficient 500, attraction coefficient 20, iteration count 100) can be used to visualize the personalized causal network model G_k. The node size is proportional to the causal strength, and the edge thickness is positively correlated with the reliability weight, highlighting core nodes such as "tube voltage", "signal-to-noise ratio", and "artifact index". Basic interactions such as node clicking, dragging, and zooming are supported. Clicking a node allows you to view detailed information about the corresponding indicator (such as the current value of signal-to-noise ratio, tube voltage parameter range, etc.).

[0144] Multimodal image maps are used to display the original medical image data and its associated anatomical structure annotations; load and display original lung cancer CT and PET-CT multimodal image data, complete the segmentation and annotation of key lung anatomical structures (left lung, right lung, tumor lesion) based on the nnU-Net pre-trained model, generate anatomical region segmentation masks and overlay them; support image scaling, translation, slice thickness adjustment and switching display of different modal images;

[0145] The processing effect analysis chart is used to show the expected processing effect and the uncertainty decomposition results;

[0146] In this context, a bidirectional interactive mapping is established between the nodes in the interactive causal network graph and the corresponding anatomical region segmentation mask in the multimodal image image. The method is as follows:

[0147] When an image feature node is selected in the causal network diagram, the corresponding anatomical region segmentation mask is highlighted in the multimodal image diagram; and when an image region is selected in the multimodal image diagram, the corresponding image feature node is highlighted in the causal network diagram.

[0148] The processing effect analysis diagram includes: a control for interactively adjusting the hypothesis value of the dynamic quality assessment flow, which is configured to update the uncertainty interval of the expected processing effect in real time based on the adjusted value; and a hypothesis simulation engine, which is configured to respond to user operations, trigger real-time counterfactual inference for different parameter adjustments, and dynamically update the visualization of the expected processing effect and its multi-level uncertainty decomposition results; for example, a Q_fused adjustment slider (0.5-1.0) is provided. When it is adjusted to 0.8, the front end calls the fast simulation API to recalculate, and the total uncertainty U_total is updated to 0.14, and the confidence interval is updated to [0.184, 0.656];

[0149] Three imaging parameter optimization schemes were implemented: tube voltage adjustment (120kV→140kV), tube current adjustment (200mA→250mA), and scanning slice thickness adjustment (1mm→0.8mm). After selecting "tube current adjustment" and executing the S3 simplified process, the expected artifact index was reduced by 28% (confidence interval [8.2%, 47.8%]). The expected effects of each scheme are compared and shown in a bar chart with error bars.

[0150] Anonymous imaging information summaries of 20 similar historical cases (same equipment, same site) were displayed. After excluding 3 cases with significant differences in imaging conditions, σ²_heterogeneity was recalculated to be 0.004, and the total uncertainty was updated to 0.16.

[0151] Automatically generate structured summaries containing core causal relationships ("tube voltage → signal-to-noise ratio" causal path strength 0.412), image quality evidence (current signal-to-noise ratio 28, presence of mild motion artifacts, etc.), and explanation of parameter adjustment effects (tube voltage adjustment is expected to improve signal-to-noise ratio by 42%, quality-driven uncertainty accounts for 38.7%, originating from mild motion artifacts).

[0152] S6, under the premise of ensuring data privacy and security, achieves the sharing and fusion of high-quality knowledge from multiple institutions. Simultaneously, based on dynamic weighting of node historical quality, it improves the overall model quality, resolving the contradiction between cross-institutional data sharing and privacy protection. In a distributed learning scenario, it differentiates the weighting of the local personalized causal network model aggregation process based on the historical quality assessment profiles of each participating node, achieving cross-node knowledge fusion and updating of the local personalized causal network model while protecting privacy. The historical quality assessment profiles are constructed based on the dynamic quality assessment streams generated historically by each node.

[0153] In this embodiment, the method for achieving cross-node knowledge fusion under privacy protection and updating the local personalized causal network model is as follows:

[0154] S61, each participating node performs privacy processing on the local personalized causal network model parameters obtained from local training; each node adds Laplace noise that satisfies (ε=1.0, δ=1e-5) differential privacy to the local model parameters (the noise intensity is adaptively adjusted according to the parameter magnitude to ensure that the noise does not mask the effective information of the parameters), and then ensures the security of parameter transmission through encryption processing (such as using the Paillier homomorphic encryption algorithm with a key length of 2048 bits);

[0155] The federated coordination server calculates the differentiated weights of each node in the local personalized causal network model aggregation based on the local data volume of each node and the average of its historical dynamic quality assessment streams. The historical quality assessment profile is composed of the mean, variance, and trend characteristics of each node's historical dynamic quality assessment streams, used to characterize the long-term data quality level of the nodes. The weight calculation formula is as follows:

[0156] ω_i=(n_i×Q̄_fused_i) / Σ(n_j×Q̄_fused_j);

[0157] In the formula, n_i is the local data volume of node i, and Q̄_fused_i is the historical dynamic quality assessment stream mean of node i; ensuring that nodes with larger data volume and higher historical imaging quality contribute more to the global model.

[0158] Based on the aforementioned differential weights, the privacy-enhanced parameters of the local personalized causal network model are securely aggregated to generate a global causal model.

[0159] Based on the historical quality assessment profile of each node, the mixing coefficient λ is determined, and the global causal model is personalized and fused with the local personalized causal network model to obtain the updated local personalized causal network model.

[0160] The federated coordination server uses homomorphic encryption technology to obtain the encrypted global model parameter θ_global by weighting the encryption parameters according to ω_A, ω_B, and ω_C.

[0161] The mixing coefficient λ is determined based on the historical mean of Q_fused for each node (positively correlated with Q̄_fused_i; the higher Q̄_fused_i, the larger λ, indicating stronger global model adaptability), with A=0.6, B=0.4, and C=0.7. A personalized model is generated by θ_final=λ×θ_global+(1−λ)×θ_local and distributed to the corresponding nodes. In this way, the global model and the local model are merged to obtain a local personalized causal network model that adapts to the local data characteristics of each node, thereby improving the local applicability of the model.

[0162] S7. In order to achieve continuous evolution of the system and solve the problem that traditional methods lack continuous evolution capabilities, the interactive feedback data in the interactive visualization imaging information processing interface is obtained, and the feedback data is calibrated according to the dynamic quality assessment flow corresponding to the source of the interactive feedback data, thereby driving the dynamic update of the prior causal knowledge base.

[0163] In this embodiment, see Figure 3 The method for driving the dynamic update of the prior causal knowledge base is as follows:

[0164] The user's operational feedback behavior in the interactive visual imaging information processing interface is transformed into feedback evidence for a specific causal path; for example, the interactive behavior of "acknowledging the causal path and the prediction of the effect of parameter adjustment" is transformed into positive feedback evidence E=1.0, "denying" is transformed into negative feedback evidence E=-1.0, and "doubting" is transformed into neutral feedback evidence E=0.

[0165] The strength of the feedback evidence is multiplied by the corresponding dynamic quality assessment stream that generated the feedback to perform evidence quality calibration (the feedback corresponding to low-quality images has low credibility, and calibration reduces its impact on knowledge base updates); E_calibrated=E×Q_fused=1.0×0.658=0.658;

[0166] The confidence levels of relevant causal relationships in the prior causal knowledge base are updated using a Bayesian update method with calibrated evidence.

[0167] When the change in confidence exceeds a threshold, an update to the prior causal knowledge base is confirmed.

[0168] For the causal relationship of "tube voltage → signal-to-noise ratio", let the prior confidence level C_prior = 0.7, and the likelihood function be:

[0169] L(E_calibrated|e)∝exp(γ·E_calibrated);

[0170] L(E_calibrated|¬e)∝exp(0)=1;

[0171] Where γ=0.8 is the feedback sensitivity coefficient, determined by experts in the imaging field through calibration (the larger γ is, the more significant the impact of feedback on confidence), calculated as follows:

[0172] L(E_calibrated|e)∝exp(0.8×0.658)≈1.692;

[0173] Calculate the posterior confidence level using Bayes' theorem:

[0174] C_posterior = [C_prior×L(E_calibrated|e)] / [C_prior×L(E_calibrated|e) + (1−C_prior)×L(E_calibrated|¬e)] ;

[0175] Substituting the data, we calculate C_posterior≈0.798;

[0176] When the change in confidence exceeds a threshold, an update to the prior causal knowledge base is confirmed.

[0177] If the confidence change ΔC = 0.098 < the preset threshold of 0.2, no expert review will be triggered, and the confidence of that edge in the prior causal knowledge base will be directly updated to 0.798; if ΔC ≥ 0.2, the expert review process will be triggered, and domain experts will review the data before deciding whether to update it.

[0178] Thus, through quality perception assessment, the impact of quality was accurately quantified, significantly reducing the processing error of low-quality image data; personalized causal reasoning enhanced the depth of imaging information correlation mining, significantly improving the feature-parameter correlation analysis capability of high-quality images; multi-level uncertainty quantification enabled controllable management of processing risks, providing quantitative confidence support for decision-making, effectively avoiding the risk of misjudgment based on fuzzy information, and comprehensively ensuring the reliability and accuracy of imaging information processing; the federated knowledge fusion function enabled the sharing of high-quality knowledge among multiple institutions while ensuring privacy and security, resolving the core contradiction between cross-institutional data sharing and privacy protection; the feedback-driven knowledge base evolution function endowed the system with continuous adaptability, and combined with the local rapid update mechanism of the model, avoided the waste of resources in global reconstruction, greatly improving the efficiency of the system in adapting to the characteristics of imaging equipment of different institutions and the needs of dynamic imaging technologies. At the same time, the interactive visualization sandbox realized the deep linkage between causal networks, multimodal images and processing effects, and the two-way interactive mapping mechanism and hypothesis simulation engine greatly improved the intuitiveness and flexibility of human-machine collaborative decision-making, significantly reducing the workload of manual parameter debugging and result verification. Furthermore, the method does not rely on proprietary imaging equipment or special data formats, has strong cross-platform adaptability, and its core parameters can be automatically adjusted based on quality assessment without manual intervention, which greatly reduces the threshold for engineering applications and provides reliable technical support for the intelligent deep processing and precise application of massive multimodal medical images.

[0179] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related explanations of the methods described above can be found in the corresponding processes in the foregoing system embodiments, and will not be repeated here.

[0180] The second embodiment of the present invention provides a multi-module medical image processing system, based on the aforementioned multi-module medical image processing method, see [link to relevant documentation]. Figure 4 The system includes:

[0181] The data acquisition module is configured to acquire raw medical image data and associated imaging context information.

[0182] The quality perception assessment module is configured to perform multi-scale quality perception assessment on the original medical image data and generate a dynamic quality assessment stream.

[0183] The causal discovery and network construction module is configured to: acquire a prior causal knowledge base, and adaptively adjust the strictness of the causal discovery algorithm based on the dynamic quality assessment flow, thereby constructing a personalized causal network model to characterize the relationship between image features and imaging parameters.

[0184] The counterfactual reasoning and uncertainty quantification module is configured to: use the personalized causal network model as an interpretable structure, combine it with the dynamic quality assessment flow, perform counterfactual reasoning to predict the expected processing effect of adjusting image features or imaging parameters, and perform multi-level uncertainty quantification decomposition on the expected processing effect to obtain uncertainty decomposition results; wherein, the quantification decomposition includes a quality-driven operational uncertainty component derived from image quality variation propagated through the personalized causal network model.

[0185] The interactive processing interface generation module is configured to generate an interactive visual imaging information processing interface based on the original medical image data, the personalized causal network model, the expected processing effect, and the uncertainty decomposition result.

[0186] The federated knowledge fusion and update module is configured as follows: In a distributed learning scenario, based on the historical quality assessment profiles of each participating node, the aggregation process of the local personalized causal network model is differentially weighted to achieve cross-node knowledge fusion and update of the local personalized causal network model under privacy protection; wherein, the historical quality assessment profiles are constructed based on the dynamic quality assessment streams generated in the history of each node;

[0187] The feedback-driven knowledge base evolution module is configured to: acquire interactive feedback data in the interactive visual imaging information processing interface, calibrate the feedback data according to the dynamic quality assessment stream corresponding to the source of the interactive feedback data, and drive the dynamic update of the prior causal knowledge base.

[0188] It should be noted that the multi-module medical image processing system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0189] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0190] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0191] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0192] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for processing multimodal medical image information, characterized in that, Includes the following steps: Obtain raw medical image data and associated imaging context information; The original medical image data is subjected to multi-scale quality perception assessment to generate a dynamic quality assessment stream; A prior causal knowledge base is acquired, and based on the dynamic quality assessment flow, the strictness of the causal discovery algorithm is adaptively adjusted to construct a personalized causal network model for characterizing the relationship between image features and imaging parameters. Using the personalized causal network model as an interpretable structure, combined with the dynamic quality assessment flow, counterfactual reasoning is performed to predict the expected processing effect of adjusting image features or imaging parameters, and the expected processing effect is subjected to multi-level uncertainty quantization decomposition to obtain uncertainty decomposition results; the uncertainty quantization decomposition includes a quality-driven operational uncertainty component derived from image quality variation propagated through the personalized causal network model. Based on the original medical image data, the personalized causal network model, the expected processing effect, and the uncertainty decomposition result, an interactive and visual imaging information processing interface is generated. In a distributed learning scenario, the aggregation process of the personalized causal network model is differentially weighted based on the historical quality assessment profiles of each participating node, thereby achieving cross-node knowledge fusion and updating of the personalized causal network model under privacy protection; wherein, the historical quality assessment profiles are constructed based on the dynamic quality assessment streams generated in the history of each node. The interactive feedback data in the interactive visualization imaging information processing interface is acquired, and the feedback data is calibrated according to the dynamic quality assessment stream corresponding to the source of the interactive feedback data, thereby driving the dynamic update of the prior causal knowledge base. Furthermore, a personalized causal network model is constructed to characterize the relationship between image features and imaging parameters. The method is as follows: Obtain deterministic causal edges, prohibited edges, and temporal constraints from the prior causal knowledge base; A fully connected graph is constructed using image features and imaging parameters as nodes, and the fully connected graph is forcibly modified according to the prior causal knowledge base, including forcibly retaining the deterministic causal relationship edges and forcibly deleting the prohibited edges, to obtain a prior constraint graph. Based on the prior constraint graph, an improved PC algorithm framework is used for causal discovery, and an undirected skeleton graph is constructed through quality-aware conditional independence tests. In each conditional independence test, the significance level of the statistical test is dynamically adjusted according to the dynamic quality assessment flow, so that the test strictness is negatively correlated with data quality. Based on V-structure recognition, directional propagation rules, and temporal relationships in the prior causal knowledge base, causal direction arbitration is performed on the undirected skeleton graph to form a directed acyclic personalized causal network model, and causal strength and reliability weights are assigned to the causal edges in the personalized causal network model. The method for defining the quality-driven operational uncertainty component derived from image quality variations propagated through the personalized causal network model is as follows: A measurement error model for key image features is established, wherein the standard deviation of the feature measurement error is negatively correlated with the dynamic quality assessment flow. Based on the measurement error model, the current image feature vector is sampled multiple times using Monte Carlo perturbation to generate a feature vector simulating quality defects. The feature vectors of the simulated quality defects are respectively input into the counterfactual prediction model constructed based on the personalized causal network model to obtain the corresponding set of expected processing effect values. Calculate the variance of the expected processing effect value set as the uncertainty component of the quality-driven operation.

2. The multimodal medical image processing method according to claim 1, characterized in that, The uncertainty quantification decomposition also includes: By resampling the training data using the bootstrap method and reconstructing multiple bootstrap reconstructed causal network models, the variance of the expected processing effect under each bootstrap reconstructed causal network model is calculated to obtain the model structure uncertainty component. The modal alignment uncertainty component is obtained by calculating the difference between the expected processing effects of sub-models from different image modalities; By finding similar historical cases in the feature space, the residual variance between the actual imaging processing results and the expected processing effect predicted based on the personalized causal network model is calculated to obtain the individual heterogeneity uncertainty component.

3. The multimodal medical image processing method according to claim 1, characterized in that, An interactive decision-making sandbox is generated as the interactive visual imaging information processing interface. This sandbox includes: An interactive causal network diagram is used to display the personalized causal network model. Multimodal image maps are used to display the original medical image data and its associated anatomical structure annotations; The processing effect analysis chart is used to show the expected processing effect and the uncertainty decomposition results; In this context, a bidirectional interactive mapping is established between the nodes in the interactive causal network graph and the corresponding anatomical region segmentation mask in the multimodal image graph. The processing effect analysis diagram includes: a control for interactively adjusting the hypothesis value of the dynamic quality assessment flow, which is configured to update the uncertainty range of the expected processing effect in real time based on the adjusted value; and a hypothesis simulation engine, which is configured to respond to user operations, trigger real-time counterfactual reasoning for different parameter adjustments, and dynamically update the visualization of the expected processing effect and its multi-level uncertainty decomposition results.

4. The multimodal medical image processing method according to claim 1, characterized in that, The method for achieving cross-node knowledge fusion and local personalized causal network model update under privacy protection is as follows: Each participating node performs privacy processing on the parameters of the locally trained personalized causal network model. The federated coordination server calculates the differential weight of each node in the local personalized causal network model aggregation based on the local data volume of each node and the average value of its historical dynamic quality assessment stream. Based on the aforementioned differential weights, the privacy-enhanced parameters of the local personalized causal network model are securely aggregated to generate a global causal model. Based on the historical quality assessment profiles of each node, the mixing coefficients are determined, and the global causal model is fused with the local personalized causal network model to obtain the updated local personalized causal network model.

5. The multimodal medical image processing method according to claim 1, characterized in that, The method for driving the dynamic update of the prior causal knowledge base is as follows: The user's operational feedback behavior in the interactive visual imaging information processing interface is transformed into feedback evidence for a specific causal path. The strength of the feedback evidence is multiplied by the corresponding dynamic quality assessment stream that generated the feedback to perform evidence quality calibration. The confidence levels of relevant causal relationships in the prior causal knowledge base are updated using a Bayesian update method with calibrated evidence. When the change in the confidence level exceeds the threshold, an update to the prior causal knowledge base is confirmed.

6. The multimodal medical image processing method according to claim 3, characterized in that, A bidirectional interactive mapping is established between the nodes in the interactive causal network graph and the corresponding anatomical region segmentation mask in the multimodal image image. The method is as follows: When an image feature node is selected in the causal network diagram, the corresponding anatomical region segmentation mask is highlighted in the multimodal image; and When an image region is selected in the multimodal image map, the corresponding image feature node is highlighted in the causal network map.

7. The multimodal medical image processing method according to claim 1, characterized in that, Assigning causal strength and reliability weights to the causal edges in the personalized causal network model, specifically including: For each directed edge X→Y in the directed acyclic graph, perform the following calculation: Estimate the average causal effect value of the side, and combine the absolute value of the effect value with the minimum p-value obtained when the side is tested for conditional independence to calculate the causal strength of the side; Based on the dynamic quality assessment stream corresponding to the image data that generates the current personalized causal network model, and the minimum p-value, the reliability weight of the edge is calculated; wherein, the reliability weight is positively correlated with the value of the dynamic quality assessment stream.

8. A multimodal medical image processing system, comprising the multimodal medical image processing method according to any one of claims 1-7, characterized in that, The system includes: The data acquisition module is configured to acquire raw medical image data and associated imaging context information. The quality perception assessment module is configured to perform multi-scale quality perception assessment on the original medical image data and generate a dynamic quality assessment stream. The causal discovery and network construction module is configured to: acquire a prior causal knowledge base, and adaptively adjust the strictness of the causal discovery algorithm based on the dynamic quality assessment flow, thereby constructing a personalized causal network model to characterize the relationship between image features and imaging parameters. The counterfactual reasoning and uncertainty quantification module is configured to: use the personalized causal network model as an interpretable structure, combine it with the dynamic quality assessment flow, perform counterfactual reasoning to predict the expected processing effect of adjusting image features or imaging parameters, and perform multi-level uncertainty quantification decomposition on the expected processing effect to obtain uncertainty decomposition results; wherein, the quantification decomposition includes a quality-driven operational uncertainty component derived from image quality variation propagated through the personalized causal network model. The interactive processing interface generation module is configured to generate an interactive visual imaging information processing interface based on the original medical image data, the personalized causal network model, the expected processing effect, and the uncertainty decomposition result. The federated knowledge fusion and update module is configured as follows: In a distributed learning scenario, the local personalized causal network model aggregation process is differentially weighted based on the historical quality assessment profiles of each participating node, so as to realize cross-node knowledge fusion and update of the local personalized causal network model under privacy protection; wherein, the historical quality assessment profiles are constructed based on the dynamic quality assessment streams generated in the history of each node. The feedback-driven knowledge base evolution module is configured to: acquire interactive feedback data in the interactive visual imaging information processing interface, calibrate the feedback data according to the dynamic quality assessment stream corresponding to the source of the interactive feedback data, and drive the dynamic update of the prior causal knowledge base.

Images