Multi-modal liver ultrasound conformal discrimination method based on evidence conflict perception

Abstract

The application discloses a multi-modal liver ultrasound conformal discrimination method based on evidence conflict perception, which comprises eight steps: multi-modal acquisition and pretreatment, modal evidence coding, evidence conflict perception fusion, uncertainty decomposition and active review, self-calibration non-consistency scoring, class conditional risk controllable conformal calibration, set operation and double-layer feedback, and clinical decision output; the application decomposes uncertainty into two parts of accident and cognition, and drives active review by cognitive uncertainty; a correction term composed of conflict quality and cognitive uncertainty and having a coefficient self-calibrated according to calibration set statistics is introduced on a probability type non-consistency term, and a class conditional risk controllable conformal prediction is adopted, a more strict significance level is applied to a malignant class, and a prediction set with coverage guarantee is generated; and the system outputs fusion probability, a prediction set and its size, accidental / cognitive uncertainty and modal conflict prompt, and is suitable for small sample, multi-modal and high class imbalance liver ultrasound auxiliary analysis.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging artificial intelligence technology, and relates to multimodal liver ultrasound image processing, uncertainty quantification of evidence deep learning, modal conflict perception fusion, and risk-controllable conformal prediction calibration technology, especially a multimodal liver ultrasound conformal discrimination method based on evidence conflict perception. Background Technology

[0002] Differentiating between benign and malignant liver tumors is crucial for subsequent treatment planning and patient prognosis assessment. Utilizing deep learning models for ultrasound image-assisted analysis is gradually becoming an important research direction in the field of medical image-assisted analysis.

[0003] Existing deep learning-assisted analysis methods have the following shortcomings in liver ultrasound image-assisted analysis: First, the output format is limited, providing only a single classification result or probability value, failing to reflect diagnostic confidence and lacking statistical coverage. For complex cases with poor image quality or atypical lesion features, the output results can easily mislead physicians and are insufficient to assist in case review and judgment.

[0004] Second, modal conflicts are easily masked. Traditional fusion methods only pursue classification accuracy and mask multimodal diagnostic discrepancies through feature splicing and probability averaging, which cannot reflect the sufficiency of diagnostic evidence.

[0005] Third, the uncertainty cannot be subdivided, and only a single quantitative result is output. It cannot distinguish between the inherent random uncertainty of the case and the cognitive uncertainty of the model, and cannot guide supplementary clinical examinations.

[0006] Fourth, the parameter design is unreasonable. The fusion weights and discrimination thresholds rely on manual settings without mechanistic support, and the symmetrical discrimination criteria for benign and malignant diseases do not meet the clinical scenario of high risk of missed diagnosis of malignant diseases.

[0007] Therefore, how to quantify prediction uncertainty with theoretical basis, explicitly perceive intermodal conflicts, and provide risk-oriented coverage guarantees for auxiliary judgment results based on making full use of multimodal information has become a technical problem that needs further research and solutions. Summary of the Invention

[0008] This invention proposes a multimodal liver ultrasound conformal discrimination method based on evidence conflict perception. It combines evidence-based modeling of multimodal ultrasound images, modal conflict quality naturally emerging from evidence combination rules, self-calibrated inconsistency scoring, and risk-controllable conformal prediction based on category conditions. This allows the system to output not only auxiliary lesion benign / malignant discrimination results, but also a prediction set with statistical coverage guarantees, uncertainty decomposable into accidental / cognitive components, and modal conflict hints. Specifically: 1. Evidence-based deep learning replaces single probability output, enabling each modality to provide a "belief-uncertainty" opinion, with modal conflict quality naturally generated by Dempster combination rules, replacing manually set conflict metrics. 2. The fusion weights, conflict penalty coefficients, and discrimination thresholds are driven by calibration set statistics and adaptively adapted sample-by-sample, eliminating the problem of "relying solely on parameter tuning." Third, adopting Mondrian risk-controllable conformal prediction, we impose stricter coverage constraints on malignant cases, providing statistically significant coverage assurance oriented towards clinical risk; Fourth, we decompose uncertainty into two parts: chance and cognition, and use cognitive uncertainty to drive active review, transforming modal conflict, a "harmful signal," into a "useful signal" guiding supplementary examinations.

[0009] This invention proposes a multimodal liver ultrasound conformal discrimination method based on evidence conflict perception, including: S-1, Multimodal Acquisition and Preprocessing: Multimodal liver ultrasound images were acquired for the cases to be processed, including two-dimensional grayscale ultrasound images of the lesions (B-mode ultrasound), color Doppler flow imaging (CDFI) images, and liver background images. Clinical features such as AFP and serological indicators were obtained when necessary. After acquisition, the images were standardized, and data enhancement such as random cropping, rotation, and brightness adjustment were applied to the B-mode ultrasound and CDFI images; adaptive histogram equalization (CLAHE) was also applied to the liver background images.

[0010] S-2, Modal Evidence Encoding: The output of each modal encoder is not directly used as a probability, but is transformed into evidence through a non-negative transformation. Dirichlet distribution parameters are constructed to provide a "belief quality - uncertainty quality" opinion for each modality: belief reflects the modality's support for good / bad, and uncertainty quality reflects whether the modality's own evidence is sufficient. Thus, each modality not only provides a predicted probability, but also explicitly gives its reliability.

[0011] S-3, Evidence Conflict Perception Fusion: Based on evidence opinions from various modalities, evidence combination rules are used for fusion. The quality of conflict between modalities emerges naturally from the combination rules, without the need for manually setting conflict metrics. Simultaneously, reliability weights are calculated for each sample according to the uncertainty quality of each modality, resulting in an adaptive fusion probability of reliability, which automatically reduces the weight of weak or noisy modalities when their evidence is insufficient. This invention can also use JS divergence to measure modal divergence at the representation layer as a supplement to conflict quality.

[0012] S-4, Uncertainty Decomposition and Active Review: The total uncertainty of the fusion output is decomposed into two parts: accidental uncertainty (inherent to the case itself and irreducible) and cognitive uncertainty (insufficient model evidence, which can be reduced by supplementing information); the score of active review is obtained by constructing cognitive uncertainty and modal conflict quality. When it exceeds the threshold, the system actively prompts supplementary examination or expert review, transforming the harmful signal of modal conflict into a useful signal to guide clinical decision-making.

[0013] S-5, Self-calibrated Inconsistency Score: Based on the ordinary conformal prediction probability-based inconsistency term, a correction term composed of modal conflict quality and cognitive uncertainty is introduced. The correction coefficient is driven by the running statistics of conflict and uncertainty on the calibration set and is adaptive on a sample-by-sample basis, eliminating the need for manual setting for each case. When modal consistency is high and evidence is sufficient, the correction term is small, tending to output a single-class set; when modal divergence is significant or evidence is insufficient, the score increases, and the set is more likely to include both benign and malignant categories, thus indicating uncertainty.

[0014] S-6, Category-Conditioned Risk-Controllable Conformal Calibration: Category-conditional (Mondrian) segmentation conformal prediction is employed. Conformal thresholds are determined for each category's respective calibration subset, and a stricter significance level is applied to the malignant class than the benign class, thus providing a higher, statistically significant coverage guarantee for the malignant class (addressing the clinical risk structure where "missed malignant diagnoses are more costly"). For samples to be discriminated, their self-calibrated inconsistency scores for each candidate category are evaluated, and categories that meet the corresponding category thresholds are included in the prediction set.

[0015] S-7, Ensemble Operations and Two-Layer Feedback: In addition to conformal calibration at the fusion level, conformal calibration can be performed on each single modality separately, and the prediction sets can be calibrated by intersection (high deterministic cross-validation scenario) or union (high coverage initial screening scenario). When the intersection is empty, the union is used as a fallback. Furthermore, using the size of the prediction set on the validation set and various coverage biases as signals, the scale prior of the fusion weights and scoring coefficients is calibrated by two-layer optimization to form an adaptive feedback loop of "downstream ensemble performance → upstream fusion and scoring".

[0016] S-8, Clinical Decision Output: The system's final output includes the reliability-adaptive fusion probability, the prediction set with coverage guarantee and its size, the decomposed random / cognitive uncertainty, and modal conflict warnings and proactive review suggestions. When the prediction set is a single case, the system indicates that the results are relatively clear at the preset risk level; when it includes multiple categories, the system indicates uncertainty and provides a suggestion for further investigation based on cognitive uncertainty. In optional implementations, the system supports the joint fusion of image fusion results with clinical laboratory indicators (such as AFP).

[0017] Preferably, the S-2 modal evidence encoding involves inputting each modal image into its corresponding encoder, and then performing a non-negative transformation on the encoder output to obtain the evidence. in This indicates the current case samples that need to be processed; Indicates modal index, Represent the category index, construct the Dirichlet parameters, and give the belief and uncertainty mass of the mode: ; in, It is the first A modal encoder for the case Belongs to the The class output is the raw class score without probability normalization. Softplus represents the activation transformation that maps real-valued class scores to non-negative values. In the Dirichlet distribution, the first... The parameters for each category; adding 1 to the evidence value keeps the constructed Dirichlet parameters positive; The formulas for calculating total modal strength, belief mass, and uncertainty mass are as follows: ; in Indicates the first One modality for cases The total Dirichlet intensity is obtained by summing the Dirichlet parameters of all classes of this mode, where the summation formula contains... Iterate through each category in turn; Indicates the first The mode is assigned to the first The quality of a class's beliefs is used to reflect the proportion of support it has for that class. This indicates the uncertain mass of the mode that is not assigned to any specific category; K represents the total number of categories; ; The larger the value, the less sufficient the evidence for that modality itself; Each modality pairs candidate categories The predicted probability is taken ; Indicates the candidate category currently being evaluated. In the category index The time represents the Dirichlet parameter corresponding to the candidate category.

[0018] Preferably, the S-3 evidence conflict perception fusion uses the modal belief quality and uncertainty quality obtained in S-2 as input, for any two different modalities The pairwise conflict quality is calculated based on the belief quality assigned to different categories by both individuals, and the calculation results for all non-repetitive modal pairs are summarized to obtain the case-level conflict quality. and They represent the first and second parts of the comparison. The first mode and the first One mode; ; ; in, Indicates the first The modality and the first A modality in the case The quality of pairwise conflicts; and For category indexing, the condition "k ≠ l" means that the product of belief quality is accumulated only when the two modalities support different categories respectively; The corresponding category of belief quality obtained in S-2; Indicates the quality of case-level conflict; This represents the set of modes participating in the fusion, and the count term represents the number of modes contained in the set. The condition i < j is used to avoid the same mode pair being counted repeatedly. When the value is relatively large, it indicates that there is a significant discrepancy in the support of different modalities for case categories; After obtaining the uncertain mass of each mode, with As the first Each modality provides a reliability measure for the current case, and the reliability measures of each modality are normalized on a case-by-case basis to form modal weights for probability fusion: ; ; in, Indicates the first Modality for cases The normalized reliability weights are such that the sum of the modal weights is 1; This indicates the modal index used when summing the denominators; The modal uncertainty mass is obtained from S-2. Therefore, modes with relatively high uncertainty mass usually receive relatively low fusion weights. Indicates the first Each modality pairs candidate categories The predicted probability; This represents the fusion category probability obtained by weighting the reliability of each modality. use This measures the difference between the complete class probability distributions for each modality, where JS represents the Jensen-Shannon divergence, and the dots represent the probability vector composed of all class probabilities; the divergence results are compared with... They can be used interchangeably, or, after being pre-set or selected based on the validation set, combined weights can be used together to characterize the divergence between modes.

[0019] Preferably, the S-4 uncertainty decomposition and active verification decompose the total uncertainty into a random component and a cognitive component; the random component is the reliability-weighted expectation of the prediction entropy of each modality, reflecting inherent ambiguity; the cognitive component is the fused distribution entropy minus the random component, reflecting insufficient evidence. Random uncertainty: ; Cognitive uncertainty: ; The entropy function represents the probability distribution. It is calculated by summing the probabilities of all categories in the distribution and using a consistent logarithmic base in the same implementation. Indicates the first The predicted probability distribution of each modality with respect to all categories, where the dots in the formula represent taking all categories in turn; This represents the probability distribution of all categories after fusion; The modal reliability weights are determined for S-3; It represents the random uncertainty of cases and is used to characterize the degree of ambiguity of categories within each modality; The cognitive uncertainty of a case is used to characterize the degree of additional uncertainty in the fusion probability distribution relative to the uncertainty within each modality; both random uncertainty and cognitive uncertainty are specific to the current case. Calculation; Active review is scored by combining cognitive uncertainty and case-level conflict quality; when this score exceeds the active review threshold... In such cases, the system may provide prompts for supplementary checks or expert review: ; in, Indicates case Active review to obtain points; To understand uncertainty; For case-level conflict quality obtained from S-3; It is a non-negative tradeoff coefficient used to adjust the relative role of conflict quality in obtaining scores.

[0020] Preferably, the S-5 self-calibration non-consistency score is based on... This is a basic probability-type inconsistent term; for cases... and candidate categories to be evaluated The inconsistency score and corresponding self-calibration coefficient are calculated by combining case-level conflict quality and cognitive uncertainty: ; ; ; in Indicates case In candidate categories The non-consistency score is a factor used to indicate the degree of consistency between the case and the candidate category. A relatively large score indicates that the case is relatively less consistent with the candidate category. The fusion probability obtained from S-3; and These represent case-level conflict quality and cognitive uncertainty, respectively. and These are the conflict correction coefficient and the cognitive uncertainty correction coefficient corresponding to the current case, respectively. and To control the prior parameters of the overall scale of the two types of corrections; These represent the running mean of case-level conflict quality and cognitive uncertainty on the calibration set, respectively. The corresponding values ​​are averaged over all calibration samples after the calibration set is determined and kept constant during the discrimination phase. Selection is based on metrics such as coverage and prediction set size on independent calibration or validation sets. During calculation, the running mean should be a positive number that can be divided; when it is zero or close to zero, a non-zero lower bound can be used to reduce the number of divisions to zero while maintaining the original scoring structure. When modal consistency is relatively high and evidence is relatively sufficient, the correction term is usually small; when modal divergence is high or evidence is relatively insufficient, the correction term can be increased accordingly.

[0021] Preferably, the S-6 category-conditional risk-controllable conformal calibration employs category-conditional Mondrian segmentation conformal prediction; for each category In the corresponding calibration subset The conformal thresholds are determined above; the calibration subset is composed of true labels as categories. The calibration samples consist of [samples], and the number of samples is denoted as [sample number]. ; Indicate category The corresponding significance level ranges between 0 and 1; considering that malignant missed diagnoses usually have a high clinical risk, a significance level of 0 is set. To make the malignant category adopt a relatively stricter significance level, the conformal threshold for each category is determined according to the following formula: ; in, Indicate category The corresponding conformal threshold; Quantile represents the empirical quantile operation; This represents the inconsistency score calculated for the i-th calibration sample under its true label; This represents the case input for the corresponding calibration sample; the round-up sign is used to determine the empirical quantile position under finite sample conditions; after obtaining the threshold for each category, the inconsistency score of the case to be discriminated under each candidate category is calculated, and the candidate categories that meet the corresponding thresholds are included in the prediction set: ; in, Indicates case The category-conditional conformal prediction set; This represents a parameter composed of the significance levels of each category; Take the benign category and the malignant category in sequence; when Not greater than the corresponding category threshold When, classify Incorporate into the prediction set; when the calibration samples and the samples to be judged meet the conformal prediction application conditions such as commutativity, the following can be obtained: The category coverage is represented by Pr, which represents the conditional probability, and y, which represents the true category of the case. When the prediction set contains only one category, it can indicate that the result is relatively clear under the preset risk level. When the prediction set contains both benign and malignant categories, it can indicate that there is still some uncertainty in the current evidence, and it is recommended to combine it with other test results or clinical experience for comprehensive evaluation.

[0022] Preferably, the S-7 set operation and two-layer feedback include single-modality prediction set operation and parameter feedback selection; each single mode forms a prediction set based on its calibration result; in high-determinism cross-validation scenarios, the intersection of each single-modality prediction set can be taken; in high-coverage initial screening scenarios, the union can be taken; when the intersection is empty, the union can be used as a fallback to reduce the number of candidate categories not output; on the calibration set and validation set, the scale parameter is selected with the premise of satisfying the coverage effectiveness constraint and the goal of reducing the average prediction set size. ;in, Depend on , and The components correspond to the scale of the conflict correction term, the scale of the cognitive uncertainty correction term, and the conflict trade-off coefficient in the score obtained through active review. The parameter selection process is as follows: ; in, This represents the parameter combination selected according to the optimization criteria; argmin represents the parameter combination in the candidate parameter space. Find parameters within the objective function that result in a relatively small value; This represents the total number of validation samples participating in the evaluation, and the subscript i in the formula represents the validation sample number; and These represent the verified case input and its actual label, respectively. Indicates in the parameter The following is the prediction set generated for the i-th case; This represents the input from the i-th case; The outer vertical line represents the cardinality of the prediction set, i.e. the number of candidate categories it contains. The average of the cardinality of the set for all validated cases is used to measure the efficiency of the prediction set. This indicates that the true labels in the validation set are categories. The number of samples; The parentheses indicate that the function takes the value 1 when the condition within the parentheses is true, and 0 otherwise; "for all c" means that the coverage constraint applies to each category. Furthermore, the average predicted set size on the validation set and the deviation between the empirical coverage and the target coverage for each category can be used as feedback signals. In the inner layer, the fusion probability, inconsistency score, and conformal threshold are recalculated based on given parameters. In the outer layer, the set efficiency and coverage performance under different parameter settings are compared, and the fusion weights in S-3 are adjusted accordingly. The scoring coefficient scale in S-5 is iteratively compared and selected; the selected parameters and weights can be kept fixed in the discrimination stage, thus forming an adaptive loop from the performance of the downstream prediction set to the upstream fusion and scoring process.

[0023] The system architecture of this invention includes an image preprocessing module, a multimodal evidence encoding module, an evidence conflict perception and fusion module, an uncertainty decomposition and active review module, a self-calibrated inconsistency scoring module, a category condition risk controllable conformal calibration module, and an ensemble operation and clinical decision output module. Each module is sequentially connected and contains a feedback loop to form a complete auxiliary discrimination system. It is suitable for liver ultrasound-assisted discrimination under conditions of small sample size, multimodality, and high class imbalance, and can simultaneously provide discrimination results, decomposable uncertainty, modal conflict indication, and risk-oriented coverage guarantee.

[0024] The beneficial effects of this invention are: First, it replaces heuristics with theoretical basis. This invention uses evidence-based deep learning and Dempster combined rules to replace single probability and manual conflict measurement. The quality of modal conflict emerges naturally from the combined rules, avoiding the problem of "relying solely on manually setting penalty terms and tuning parameters", and significantly improving the interpretability and arguability of the mechanism.

[0025] Second, it provides risk-oriented coverage assurance. Employing category-conditional risk-controlled conformal prediction, a more stringent significance level is applied to malignant cases. This provides a higher, statistically significant coverage assurance for malignant cases while meeting the corresponding assumptions, aligning with the clinical risk structure where "missed diagnoses of malignant cases are more costly." In the experiment, the overall coverage rate of the fusion-level conformal prediction under 5-fold cross-validation was approximately 0.9367, the malignant case coverage rate was 1.000, and the average prediction set size was approximately 1.55, indicating that this method can achieve a reasonable balance between coverage and set size.

[0026] Third, the sources of uncertainty are divisible and can guide action. Uncertainty is decomposed into two parts: chance and cognition. Cognitive uncertainty drives proactive review, enabling the system not only to "know that it is uncertain," but also to distinguish whether the uncertainty can be reduced through supplementary checks. This transforms the harmful signal of modal conflict into a useful signal to guide supplementary checks.

[0027] Fourth, self-calibration of weights and thresholds. The fusion weights, conflict penalties, and discrimination thresholds are driven by calibration set statistics and are adaptive on a sample-by-sample basis, reducing reliance on human experience and improving adaptability across data conditions. For weak modalities (such as liver background), existing strategies such as automatic weight reduction based on evidence uncertainty combined with cost-sensitive loss are used to reduce their adverse impact on the fusion results.

[0028] Fifth, the efficiency is usable. The complete inference process in the experiment took about 225ms per sample, with the generation of conformal sets taking relatively little time. The optional image-clinical joint fusion achieved an AUC of 0.8789 in the supplementary experiment, indicating that clinical indicators can further supplement the image model information in specific scenarios. Attached Figure Description

[0029] Figure 1 This is a block diagram of the system architecture modules of the present invention; Figure 2 This is a flowchart illustrating the mechanism of evidence conflict perception and risk-controllable conformal prediction in this invention. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to specific embodiments.

[0031] This invention proposes a multimodal liver ultrasound conformal discrimination method based on evidence conflict perception. Includes: (i) Data acquisition and preprocessing

[0032] Multimodal liver ultrasound images of the case to be processed can be acquired, including two-dimensional grayscale ultrasound images of the lesion, CDFI images, and liver background images, or some modalities can be selected according to the actual examination conditions. Each modal image is subjected to size standardization, grayscale / color channel normalization, region of interest cropping, and necessary data enhancement; textured images such as the liver background can also be enhanced using histogram equalization. All of the above are implementation methods using existing technologies.

[0033] (ii) Modal evidence coding

[0034] Each modality image is input into its corresponding encoder (a lightweight network such as MobileNetV3-Large can be used, with transfer learning initialization, which is an existing method). Unlike existing methods that directly output probabilities, this implementation causes the encoder output to undergo a non-negative transformation to obtain evidence. Construct the Dirichlet parameters and give the belief and uncertainty mass of this mode: ; ; in For the number of categories, ; A larger value indicates less sufficient evidence for the mode itself. The predicted probability of the mode is taken as... .

[0035] (III) Fusion of Evidence Conflict Perception

[0036] For any two modes The conflict quality is derived from the belief quality, and the sample-level conflict quality is obtained by averaging over all modal pairs. ;

[0037] Conflict quality emerges naturally from the evidence combination rules, eliminating the need for manually setting conflict metrics. The fusion approach employs reliability based on each modality. Sample-by-sample weighting, with weak modalities automatically reduced in weight when evidence is insufficient: ; ;

[0038] As a supplement, modal divergence can also be measured using JS divergence at the presentation layer. ,and They can be substitutes or weighted combinations.

[0039] (iv) Uncertainty decomposition and proactive review

[0040] The total uncertainty is decomposed into a random component (the reliability-weighted expectation of the prediction entropy of each modality, reflecting inherent ambiguity) and a cognitive component (the fused distribution entropy minus the random component, reflecting insufficient evidence): ; ; Active review is conducted based on cognitive uncertainty and conflict quality to obtain a score, exceeding a threshold. Proactively prompt for supplementary examinations or expert review:

[0041] This transforms the harmful signal of modal conflict into a useful signal to guide supplementary investigations.

[0042] (v) Self-calibration inconsistency score

[0043] In the probabilistic non-uniformity term of ordinary conformal prediction Based on this, a correction term consisting of conflict quality and cognitive uncertainty is introduced, and the coefficient is self-calibrated for each sample:

[0044] ;

[0045] ; in The running mean on the calibration set is used to make the correction strength adaptive to the data distribution; scale-only prior. It should be selected from the calibration set to avoid manual setting for each case. When modal consistency is high and evidence is sufficient, the correction term is small, tending towards a single-class set; when modal divergence is high or evidence is insufficient, the score increases, and the set is more likely to include both benign and anti-benign categories.

[0046] (vi) Category-based condition risk-controllable conformal calibration and set generation

[0047] Mondrian segmentation conformal prediction is used for each category: In its calibration subset (sample size) Conformal thresholds were determined for each class, and a more stringent significance level was applied to the malignant class. : ; ; Under the commutativity assumption, there is a class conditional coverage guarantee for each class. This provides a higher coverage guarantee for malignant cases. When the set is a singleton, it indicates that the result is relatively clear under the preset risk level; when it contains both benign and malignant cases, it indicates uncertainty and suggests combining it with other examinations or experience for evaluation.

[0048] (vii) Parameter selection and two-layer feedback

[0049] To avoid relying on human experience for parameters, scale parameters are selected on the calibration / validation set based on the criterion that "ensemble efficiency follows coverage effectiveness". (Using the existing conformal efficiency-efficiency criterion, it does not claim novelty on its own):

[0050] Furthermore, using the size of the validation set and various coverage biases as signals, a two-layer optimization is used to reverse-calibrate the fusion weights. It forms an adaptive feedback loop with the scoring coefficient scale.

[0051] (viii) Modal-level set operations and output

[0052] In addition to the fusion level, conformal calibration can be performed on each individual modality to obtain an independent prediction set. The intersection (for more precise targeting) or union (for higher coverage) can be taken according to the scenario; if the intersection is empty, the union is used as a fallback. The final output includes the fusion probability, the prediction set with coverage guarantee and its size, random / cognitive uncertainty, modal conflict warnings, and proactive review suggestions. When ultrasound and CDFI show different tendencies, the modal conflict warning can remind the user to pay attention to the image quality and consistency of the relevant modalities.

[0053] (ix) Integration of imaging and clinical data When clinical indicators such as AFP and serological data are available, the results of image evidence fusion can be combined with the output of the clinical classifier in a reliability-weighted manner, and category condition risk-controlled conformal calibration can be performed to make the method suitable for both pure imaging and imaging-clinical combined scenarios.

[0054] In summary, this implementation method, through modal evidence coding, evidence conflict perception fusion, uncertainty decomposition and active review, self-calibrated inconsistency scoring, and category-conditional risk-controllable conformal calibration, ensures that the auxiliary discrimination results not only include benign or malignant tendencies, but also decomposable uncertainties, modal conflict alerts, and risk-oriented coverage guarantees. Figure 1 , Figure 2 The architecture and process shown correspond to each other.

[0055] The above description is only a part of the specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal liver ultrasound conformal discrimination method based on evidence conflict perception, characterized in that: It includes 8 steps: S-1, Multimodal Acquisition and Preprocessing: Acquire pathological images and perform standardized processing; S-2, Modal Evidence Encoding: Each modal encoder obtains evidence through non-negative transformation, constructs Dirichlet distribution parameters, and thus provides an opinion on "belief quality - uncertainty quality" for each modality: belief reflects the modality's support for good / bad, and uncertainty quality reflects whether the modality's own evidence is sufficient; S-3, Evidence Conflict Perception Fusion: Based on opinions from various modalities of evidence, evidence combination rules are used for fusion; S-4, Uncertainty Decomposition and Active Review: The total uncertainty of the fusion output is decomposed into two parts: accidental uncertainty and cognitive uncertainty; the score is obtained by constructing an active review based on cognitive uncertainty and modal conflict quality. When it exceeds the threshold, the system actively prompts for supplementary examination or expert review, transforming the harmful signal of modal conflict into a useful signal to guide clinical decision-making. S-5, Self-calibrated Inconsistency Score: Based on the ordinary conformal prediction probability inconsistency term, a correction term composed of modal conflict quality and cognitive uncertainty is introduced. When modal consistency is high and evidence is sufficient, the correction term is small and tends to output a single-class set. When modal divergence is significant or evidence is insufficient, the score increases and the set is more likely to include both benign and malignant categories, thus indicating uncertainty. S-6, Category-conditional risk-controlled conformal calibration: Category-conditional Mondrian segmentation conformal prediction is adopted. Conformal thresholds are determined on the calibration subsets of each category, and a stricter significance level is applied to the malignant class than to the benign class, thereby providing a higher, statistically significant coverage guarantee for the malignant class. For the samples to be discriminated, the self-calibrated inconsistency score when they belong to each candidate class is evaluated, and the classes that meet the corresponding category thresholds are included in the prediction set. S-7, Set Operations and Two-Layer Feedback: In addition to fusion-level conformal calibration, conformal calibration is performed on each single mode and the prediction set is combined with the intersection: high deterministic cross-validation scenario; or union: high coverage initial screening scenario, operation, when the intersection is empty, the union is used to back off; using the size of the prediction set on the validation set and various coverage deviations as signals, the scale prior of the fusion weights and scoring coefficients is calibrated in reverse through two-layer optimization, forming an adaptive feedback loop of "downstream set performance → upstream fusion and scoring"; S-8, Clinical Decision Output: The system output includes the reliability adaptive fusion probability, the prediction set with coverage guarantee and its size, the decomposed random / cognitive uncertainty, and modal conflict prompts and proactive review suggestions; when the prediction set is a single case, it indicates that the result is relatively clear under the preset risk level; when it includes multiple categories, it indicates that there is uncertainty, and provides a prompt on whether to recommend supplementary examinations based on cognitive uncertainty; In optional implementations, the ability to combine image fusion results with clinical laboratory indicators is supported.

2. The multimodal liver ultrasound conformal discrimination method based on evidence conflict perception according to claim 1, characterized in that, The S-2 modal evidence encoding method involves inputting each modal image into its corresponding encoder, and then performing a non-negative transformation on the encoder output to obtain the evidence. in This indicates the current case samples that need to be processed; Indicates modal index, Represent the category index, construct the Dirichlet parameters, and give the belief and uncertainty mass of the mode: ； in, It is the first A modal encoder for the case Belongs to the The class output is the raw class score without probability normalization. Softplus represents the activation transformation that maps real-valued class scores to non-negative values. In the Dirichlet distribution, the first... The parameters for each category; adding 1 to the evidence value keeps the constructed Dirichlet parameters positive; The formulas for calculating total modal strength, belief mass, and uncertainty mass are as follows: ； in Indicates the first One modality for cases The total Dirichlet intensity is obtained by summing the Dirichlet parameters of all classes of this mode, where the summation formula contains... Iterate through each category in turn; Indicates the first The mode is assigned to the first The quality of a class's beliefs is used to reflect the proportion of support it has for that class. This indicates the uncertain mass of the mode that is not assigned to any specific category; K represents the total number of categories; ; The larger the value, the less sufficient the evidence for that modality itself; Each modality pairs candidate categories The predicted probability is taken ; Indicates the candidate category currently being evaluated. In the category index The time represents the Dirichlet parameter corresponding to the candidate category.

3. The multimodal liver ultrasound conformal discrimination method based on evidence conflict perception according to claim 1, characterized in that, The S-3 evidence conflict perception fusion uses the modal belief quality and uncertainty quality obtained from S-2 as inputs for any two different modalities. The pairwise conflict quality is calculated based on the belief quality assigned to different categories by both individuals, and the calculation results for all non-repetitive modal pairs are summarized to obtain the case-level conflict quality. and They represent the first and second parts of the comparison. The first mode and the first One mode; ； ； in, Indicates the first The modality and the first A modality in the case The quality of pairwise conflicts; and For category indexing, the condition "k ≠ l" means that the product of belief quality is only accumulated when the two modalities support different categories respectively; The corresponding category of belief quality obtained in S-2; Indicates the quality of case-level conflict; This represents the set of modes participating in the fusion, and the count term represents the number of modes contained in the set. The condition i < j is used to avoid the same mode pair being counted repeatedly. When the value is relatively large, it indicates that there is a significant discrepancy in the support of different modalities for case categories; After obtaining the uncertain mass of each mode, with As the first Each modality provides a reliability measure for the current case, and the reliability measures of each modality are normalized on a case-by-case basis to form modal weights for probability fusion: ； ； in, Indicates the first Modality for cases The normalized reliability weights are such that the sum of the modal weights is 1; This indicates the modal index used when summing the denominators; The modal uncertainty mass is obtained from S-2. Therefore, modes with relatively high uncertainty mass usually receive relatively low fusion weights. Indicates the first Each modality pairs candidate categories The predicted probability; This represents the fusion category probability obtained by weighting the reliability of each modality. use This measures the difference between the complete class probability distributions for each modality, where JS represents the Jensen-Shannon divergence, and the dots represent the probability vector composed of all class probabilities; the divergence results are compared with... They can be used interchangeably, or, after being pre-set or selected based on the validation set, combined weights can be used together to characterize the divergence between modes.

4. The multimodal liver ultrasound conformal discrimination method based on evidence conflict perception according to claim 1, characterized in that, The S-4 uncertainty decomposition and active verification decompose the total uncertainty into a random part and a cognitive part; the random part is the reliability-weighted expectation of the prediction entropy of each modality, reflecting the inherent ambiguity; The cognitive portion, calculated by subtracting the random component from the fused distribution entropy, reflects insufficient evidence. Random uncertainty: ; Cognitive uncertainty: ; The entropy function represents the probability distribution. It is calculated by summing the probabilities of all categories in the distribution and using a consistent logarithmic base in the same implementation. Indicates the first The predicted probability distribution of each modality with respect to all categories, where the dots in the formula represent taking all categories in turn; This represents the probability distribution of all categories after fusion; The modal reliability weights are determined for S-3; It represents the random uncertainty of cases and is used to characterize the degree of ambiguity of categories within each modality; The cognitive uncertainty of a case is used to characterize the degree of additional uncertainty in the fusion probability distribution relative to the uncertainty within each modality; both random uncertainty and cognitive uncertainty are specific to the current case. Calculation; Active review is scored by combining cognitive uncertainty and case-level conflict quality; when this score exceeds the active review threshold... In such cases, the system may provide prompts for supplementary checks or expert review: ； in, Indicates case Active review to obtain points; To understand uncertainty; For case-level conflict quality obtained from S-3; It is a non-negative tradeoff coefficient used to adjust the relative role of conflict quality in obtaining scores.

5. The multimodal liver ultrasound conformal discrimination method based on evidence conflict perception according to claim 1, characterized in that, The S-5 self-calibration non-consistency score is based on This is a basic probability-type inconsistent term; for cases... and candidate categories to be evaluated The inconsistency score and corresponding self-calibration coefficient are calculated by combining case-level conflict quality and cognitive uncertainty: ； ； ； in Indicates case In candidate categories The non-consistency score is a factor used to indicate the degree of consistency between the case and the candidate category. A relatively large score indicates that the case is relatively less consistent with the candidate category. The fusion probability obtained from S-3; and These represent case-level conflict quality and cognitive uncertainty, respectively. and These are the conflict correction coefficient and the cognitive uncertainty correction coefficient corresponding to the current case, respectively. and To control the prior parameters of the overall scale of the two types of corrections; These represent the running mean of case-level conflict quality and cognitive uncertainty on the calibration set, respectively. The corresponding values ​​are averaged over all calibration samples after the calibration set is determined and kept constant during the discrimination phase. The selection is based on metrics such as coverage and prediction set size on independent calibration or validation sets.

6. The multimodal liver ultrasound conformal discrimination method based on evidence conflict perception according to claim 1, characterized in that, The S-6 category-conditional risk-controlled conformal calibration employs category-conditional Mondrian segmentation conformal prediction; for each category... In the corresponding calibration subset The conformal thresholds are determined above; the calibration subset is composed of true labels as categories. The calibration samples consist of [samples], and the number of samples is denoted as [sample number]. ; Indicate category The corresponding significance level ranges between 0 and 1; considering that malignant missed diagnoses usually have a high clinical risk, a significance level of 0 is set. To make the malignant category adopt a relatively stricter significance level, the conformal threshold for each category is determined according to the following formula: ； in, Indicate category The corresponding conformal threshold; Quantile represents the empirical quantile operation; This represents the inconsistency score calculated for the i-th calibration sample under its true label; This represents the case input for the corresponding calibration sample; the round-up sign is used to determine the empirical quantile position under finite sample conditions; after obtaining the threshold for each category, the inconsistency score of the case to be discriminated under each candidate category is calculated, and the candidate categories that meet the corresponding thresholds are included in the prediction set: ； in, Indicates case The category-conditional conformal prediction set; This represents a parameter composed of the significance levels of each category; Take the benign category and the malignant category in sequence; when Not greater than the corresponding category threshold When, classify Incorporate into the prediction set; when the calibration samples and the samples to be judged meet the conformal prediction application conditions such as commutativity, the following can be obtained: The category coverage is represented by Pr, which represents the conditional probability, and y, which represents the true category of the case. When the prediction set contains only one category, it can indicate that the result is relatively clear under the preset risk level. When the prediction set contains both benign and malignant categories, it can indicate that there is still some uncertainty in the current evidence, and it is recommended to combine it with other test results or clinical experience for comprehensive evaluation.

7. The multimodal liver ultrasound conformal discrimination method based on evidence conflict perception according to claim 1, characterized in that, The S-7 set operation and two-layer feedback include single-modality prediction set operation and parameter feedback selection; each single mode forms a prediction set based on its calibration results; in high-deterministic cross-validation scenarios, the intersection of each single-modality prediction set can be taken; in high-coverage initial screening scenarios, the union can be taken; when the intersection is empty, the union can be used as a fallback to reduce the number of candidate categories not output; on the calibration set and validation set, the scale parameter is selected with the premise of satisfying the coverage effectiveness constraint and the goal of reducing the average prediction set size. ;in, Depend on , and The components correspond to the scale of the conflict correction term, the scale of the cognitive uncertainty correction term, and the conflict trade-off coefficient in the score obtained through active review. The parameter selection process is as follows: ； in, This represents the parameter combination selected according to the optimization criteria; argmin represents the parameter combination in the candidate parameter space. Find parameters within the objective function that result in a relatively small value; This represents the total number of validation samples participating in the evaluation, and the subscript i in the formula represents the validation sample number; and These represent the verified case input and its actual label, respectively. Indicates in the parameter The following is the prediction set generated for the i-th case; This represents the input from the i-th case; The outer vertical line represents the cardinality of the prediction set, i.e. the number of candidate categories it contains. The average of the cardinality of the set for all validated cases is used to measure the efficiency of the prediction set. This indicates that the true labels in the validation set are categories. The number of samples; The parentheses indicate that the function takes the value 1 when the condition within the parentheses is true, and 0 otherwise; "for all c" indicates that the coverage constraint applies to each category. Furthermore, the average predicted set size on the validation set and the deviation between the empirical coverage and the target coverage for each category can be used as feedback signals. In the inner layer, the fusion probability, inconsistency score, and conformal threshold are recalculated based on given parameters. In the outer layer, the set efficiency and coverage performance under different parameter settings are compared, and the fusion weights in S-3 are adjusted accordingly. The scoring coefficient scale in S-5 is iteratively compared and selected; the selected parameters and weights can be kept fixed in the discrimination stage, thus forming an adaptive loop from the performance of the downstream prediction set to the upstream fusion and scoring process.

Images