A multi-modal ultrasound fusion system based on dynamic conflict perception and a method thereof

Abstract

The application discloses a multi-modal ultrasound fusion system and method based on dynamic conflict perception, and in view of the dynamic characteristics of ultrasound imaging and the complex effects thereof in liver cancer, a method capable of evaluating and calibrating modal uncertainty is provided, and a prognosis reaction conflict is solved in combination with the calibration result. The method realizes efficient integration of cross-modal prognosis data, significantly improves the accuracy of survival analysis, supports individualized treatment selection for liver cancer patients, and has important clinical application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical ultrasound multi-modal data processing, and particularly relates to a multi-modal ultrasound fusion system based on dynamic conflict perception and a method thereof. BACKGROUND

[0002] Hepatocellular carcinoma (HCC) is a common malignant tumor of the digestive tract, accounting for the fifth place in the global incidence of malignant tumors. The number of new cases of liver cancer in China accounts for more than 50% of the total number of cases worldwide each year, and is the second largest tumor cause of death in China. Despite the continuous progress of treatment technology, the overall prognosis of liver cancer patients is still not ideal, with a five-year survival rate of only 14%. The survival rate of patients who receive radical surgical resection is as high as 69% to 86%, but only about 20% of patients meet the indications for surgical treatment at the time of initial diagnosis. Early detection and treatment of liver cancer is the key to improving the survival rate of liver cancer patients and is one of the important directions of current clinical research.

[0003] The choice of early liver cancer treatment plan is mainly based on the patient's liver function reserve, the presence or absence of large vessel invasion, and intrahepatic and extrahepatic metastasis, etc. Surgical resection and radiofrequency ablation are the most common treatment methods, but due to the heterogeneity and complexity of liver tumors, the prognosis of these two treatments has a large degree of uncertainty. Therefore, it is necessary to develop precise survival analysis tools and personalized treatment selection methods to optimize patient treatment plans.

[0004] Survival analysis is a key link in the personalized treatment of liver cancer patients, which can help clinicians predict the risk of disease recurrence or death by analyzing the survival probability of patients within a specific time period. Cox proportional hazards model is the standard method for survival analysis at present, which can not only quantify the survival risk, but also evaluate the impact of various clinical variables (such as gender, age, tumor size) on prognosis. However, traditional survival analysis models mainly rely on single modal data, which is difficult to fully capture the complex characteristics of liver cancer.

[0005] Multi-modal ultrasound imaging technology combines B-mode ultrasound, contrast-enhanced ultrasound, and clinical information, providing a comprehensive perspective for the diagnosis and treatment of liver tumors. B-mode ultrasound can visualize tumor size, shape, location, and surrounding tissue morphology, and is the main tool for basic image analysis of liver cancer; contrast-enhanced ultrasound enhances the observation of tumor blood supply by dynamically displaying tumor microcirculation characteristics in real time; pathological examination and other clinical information serve as key indicators for evaluating pathological changes such as fibrosis or cirrhosis. By integrating morphological, functional, and clinical data, multi-modal ultrasound improves the accuracy of survival analysis and supports personalized treatment selection.

[0006] Although multimodal ultrasound imaging has shown great potential in liver cancer survival analysis and personalized treatment, its practical application still faces many challenges. First, the instability of modality quality is one of the key problems. In the actual clinical environment, the quality and consistency of the data may be affected by many factors such as imaging plane selection, device parameter adjustment and imaging artifacts, which is difficult to meet the assumptions of traditional fusion methods on data stability and semantic alignment. Second, the inconsistency of modality information further aggravates the difficulty of fusion. For example, B-ultrasound may show a blurred tumor margin, suggesting a poor prognosis, while contrast-enhanced ultrasound may show a uniform high-enhanced pattern, pointing to a better prognosis, leading to conflicting prognostic information between different modalities. In addition, existing multi-modal fusion methods mostly rely on fixed or heuristic weight allocation, lack of dynamic adjustment ability to data credibility, and are difficult to effectively deal with conflicts or uncertainties between modalities. These problems seriously limit the application of multi-modal ultrasound technology in liver cancer prognosis analysis, and an innovative solution is needed to improve the credibility and adaptability of fusion, effectively model modality-specific uncertainty, and dynamically adjust modality contribution, so as to extract stable and accurate prognostic information from multi-modal ultrasound data, and provide solid technical support for personalized treatment selection for liver cancer patients. SUMMARY

[0007] In view of the shortcomings of the existing multi-modal fusion method relying on fixed or heuristic weight allocation, the present application proposes a multi-modal ultrasound fusion system and method based on dynamic conflict perception, which is used for reliable liver cancer survival analysis and treatment selection.

[0008] To achieve the above object, the technical scheme adopted by the present application is:

[0009] A multi-modal ultrasound fusion system based on dynamic conflict perception, comprising a B-mode ultrasound image acquisition module, a contrast-enhanced ultrasound image acquisition module, and a clinical data acquisition module, further comprising:

[0010] A self-calibration uncertainty estimation module is used to propose a cross-modal uncertainty calibration regularizer to align the modality uncertainty with its prediction effect;

[0011] A dynamic conflict fusion module is used to propose a dynamic conflict perception fusion mechanism to dynamically adjust the weights of different modalities according to the cross-modal consistency and modality uncertainty;

[0012] An individual survival risk prediction module is used for individual survival risk prediction, and risk stratification for liver cancer patients;

[0013] A personalized treatment plan recommendation module is used for treatment plan recommendation for liver cancer patients.

[0014] A method of a multi-modal ultrasound fusion system based on dynamic conflict perception, comprising the following steps:

[0015] S1: B-mode ultrasound image, contrast-enhanced ultrasound image and clinical data collection,

[0016] S2: Self-calibration uncertainty estimation: a cross-modality uncertainty calibration regularizer is proposed to align the modality uncertainty with its prediction effect;

[0017] S3: Dynamic conflict fusion: a dynamic conflict-aware fusion mechanism is proposed to dynamically adjust the weights of different modalities according to cross-modality consistency and modality uncertainty;

[0018] S4: Individual survival risk prediction, risk stratification and treatment recommendation for liver cancer patients.

[0019] Further, the method of step S2 self-calibration uncertainty estimation is as follows:

[0020] According to the characteristics of the Cox proportional hazards model for estimating relative risk, a partial likelihood loss function is used instead of the commonly used mean square error (MSE); the loss function quantifies the prediction accuracy by calculating the relative risk between the individual and the risk set; the formula is defined as follows:

[0021]

[0022] Where: i represents the ith individual, j represents the jth individual, m represents the modality, u i,m is the uncertainty constraint, E i,m is the prediction error, r i,m is the prediction risk function of the ith individual, r j,m is the prediction risk function of the jth individual, and the risk set R(·) refers to the individual still existing at the event occurrence time t i , δ i =1 indicates that the ith individual has an event, the formula ∝ represents direct proportion, -log(·) is the logarithmic function, and ∑ is the summation of all occurring individuals, , exp(r i,m ) and exp(r j,m ) represent the exponential form of the prediction risk score of the ith individual and the jth individual in modality m; this alignment design can ensure that in the subsequent modality fusion process, the modality with larger error is given higher uncertainty, thereby reducing its negative impact on the overall prediction result;

[0023] In batch data processing, the modality uncertainty u i,m is constrained to be proportional to the prediction error E i,m , and they are converted into discrete distributions U and E through normalization processing, defined as follows:

[0024]

[0025] Where: B is the batch size, and M is the number of modes. This represents the uncertainty value of the last element among all samples and modes. The uncertainty vector U and error vector E, which represent the prediction error of the last element in all samples and modes, are normalized to eliminate the interference of the number of modes and batch size on uncertainty estimation, ensuring the consistency and stability of the distribution.

[0026] Furthermore, the method for dynamic conflict fusion in step S3 is as follows:

[0027] To quantify the conflict between different modal risk predictions, a multimodal conflict matrix S is defined, where each element of S represents the prognostic difference s between mode m1 and mode m2. ij s ij =|r i -r j |, where: r i Let r represent the prediction risk function for mode m1. j The predictive risk function for mode m2 is represented by the conflict matrix S, which quantifies cross-modal inconsistency by calculating the difference in prognostic outcomes between modes.

[0028] To adjust the differences between modes based on the confidence level of the modal predictions. ij Furthermore, a conflict adjustment factor d was introduced. ij If the uncertainty of mode m1 is u i Uncertainty u below mode m2 j If mode m1 is more reliable, then the fusion is optimized by reducing the difference contribution from the less reliable mode m2; conversely, if mode m2 has lower uncertainty, then mode s is retained. ij The full contribution is used to reflect the consistency between modes to the greatest extent possible;

[0029]

[0030] Through this mechanism, the conflict matrix S is weighted and adjusted based on uncertainty to construct the conflict adjustment matrix D, such that... in, Let S represent the reweighted conflict matrix, i.e., for the conflict matrix S

[0031] The reweighting mechanism, which multiplies element-wise with the conflict adjustment matrix D, ensures that more reliable modes have a greater influence in the fusion process, while the weights of unreliable modes are appropriately reduced. Meanwhile, the overall conflict score for each mode is obtained by the following formula:

[0032]

[0033] in, This represents the difference between modes m1 and m2 after reweighting. This represents the final conflict score after modality m1 is adjusted; to dynamically adjust the contribution of each modality, a method based on... The reward and punishment functions; if If the conflict exceeds the average conflict threshold θ, a penalty function is applied; otherwise, a reward function is applied, with α controlling the degree of the reward / penalty function. The reward / penalty function is defined as follows:

[0034]

[0035] Where e is the natural constant, approximately equal to 2.71828. The final conflict score after modality m1 adjustment is represented by the above dynamic fusion mechanism. This invention ensures that more reliable modalities play a crucial role in multimodal fusion, while the impact of highly conflicting or unreliable modalities on the overall prediction is effectively suppressed. Finally, the modal features x are... i According to its weight w i Perform weighted aggregation to generate the final multimodal fusion representation x. f For more accurate prognostic prediction;

[0036]

[0037] This method effectively improves the reliability and accuracy of multimodal data in predicting the prognosis of liver cancer through a dynamic fusion mechanism, and also provides a flexible approach to address modal conflicts.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] Compared to traditional fixed or heuristic methods, this invention features strong dynamic adaptability and robustness, effectively addressing the complexity of multimodal data in open clinical settings, thereby achieving high precision and dynamism in liver cancer survival analysis. This method provides crucial support for the development of personalized treatment plans and contributes to the advancement of precision medicine for liver cancer. Attached Figure Description

[0040] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0041] The present invention will be further described below with reference to embodiments.

[0042] Example 1

[0043] A multimodal ultrasound fusion system based on dynamic conflict perception includes a B-mode ultrasound image acquisition module, a contrast-enhanced ultrasound image acquisition module, a clinical data acquisition module, and further includes:

[0044] The self-calibrating uncertainty estimation module is used to propose a cross-modal uncertainty calibration regularizer to align modal uncertainty with its prediction performance;

[0045] The dynamic conflict fusion module is used to propose a dynamic conflict-aware fusion mechanism that dynamically adjusts the weights of different modes based on cross-modal consistency and modal uncertainty.

[0046] The individual survival risk prediction module is used to predict individual survival risk and to stratify the risk of liver cancer patients.

[0047] The personalized treatment plan recommendation module is used to recommend treatment plans for liver cancer patients.

[0048] Example 2

[0049] A method for a multimodal ultrasound fusion system based on dynamic conflict perception includes the following steps:

[0050] S1: Perform B-mode ultrasound imaging, contrast-enhanced ultrasound imaging, and clinical data acquisition.

[0051] S2: Self-calibrating uncertainty estimation: A cross-modal uncertainty calibration regularizer is proposed to align modal uncertainty with its prediction performance. Step S2, by introducing the cross-modal uncertainty calibration regularizer, utilizes normalized distribution matching to align the inherent uncertainty of each mode with the prediction error. This calibration ensures that the fusion process prioritizes modes with reliable predictions, thereby significantly improving the overall reliability and accuracy of the prognosis.

[0052] Furthermore, the method for estimating the self-calibration uncertainty in step S2 is as follows:

[0053] To address the characteristics of relative risk estimation in the Cox proportional hazards model, this invention employs a partial likelihood loss function instead of the commonly used mean squared error (MSE). This loss function quantifies prediction accuracy by calculating the relative risk between an individual and the risk set; the formula is defined as follows:

[0054]

[0055] Where: i represents the i-th individual, j represents the j-th individual, m represents the modality, and u i,m For uncertainty constraints, E i,m This is the prediction error, r. i,m Let r be the risk function for predicting the i-th individual. j,mThe j-th individual predicts the risk function, and the risk set R(·) refers to the risk set at time t when the event occurs. i The still existing individuals, δ i When = 1, it means that the event occurred for the i-th individual. In the formula, ∝ represents direct proportion, -log(·) is the logarithmic function, and ∑ is the summation of all individuals whose events occurred. This is used to calculate the sum of partial likelihood losses for all individuals where the event occurred, exp(r i,m ) and exp(r j,m ) represents the exponential form of the predicted risk scores of the i-th and j-th individuals under mode m; this alignment design ensures that in the subsequent mode fusion process, the mode with larger error is given higher uncertainty, thereby reducing its negative impact on the overall prediction results;

[0056] In batch data processing, this invention addresses modal uncertainty u. i,m The constraint is related to the prediction error E. i,m They are directly proportional, and after normalization, they are transformed into discrete distributions U and E, defined as follows:

[0057]

[0058] Where: B is the batch size, and M is the number of modes. This represents the uncertainty value of the last element among all samples and modes. The prediction error represents the last element in all samples and modes. The normalized uncertainty vector U and error vector E can eliminate the interference of the number of modes and batch size on uncertainty estimation, ensuring the consistency and stability of the distribution.

[0059] S3: Dynamic Conflict Fusion: A dynamic conflict-aware fusion mechanism is proposed, which dynamically adjusts the weights of different modalities based on cross-modal consistency and modal uncertainty. During the fusion phase, this invention dynamically adjusts modal weights based on cross-modal consistency and modal uncertainty. When conflicts exist between multiple modalities, the degree of conflict is first quantified, then adjusted using calibration uncertainty, and finally, modal weights are reweighted through a reward-penalty mechanism to ensure that more stable and reliable modal inputs are prioritized. Compared with traditional fixed or heuristic methods, this invention has strong dynamic adaptability and high robustness, effectively addressing the complexity of multimodal data in open clinical environments, thereby achieving high precision and dynamism in liver cancer survival analysis. This method provides important support for the development of personalized treatment plans and contributes to the advancement of precision medicine for liver cancer.

[0060] Furthermore, the method for dynamic conflict fusion in step S3 is as follows:

[0061] To quantify the conflict between different modal risk predictions, this invention defines a multimodal conflict matrix S, where each element of S represents the prognostic difference s between modal m1 and modal m1. ij s ij =|r i -r j |, where: r i Let r represent the prediction risk function for mode m1. j The predictive risk function for mode m2 is represented by the conflict matrix S, which quantifies cross-modal inconsistency by calculating the difference in prognostic outcomes between modes.

[0062] To adjust the differences between modes based on the confidence level of the modal predictions. ij This invention further introduces a conflict adjustment factor d. ij If the uncertainty of mode m1 is u i Uncertainty u below mode m2 j If mode m1 is more reliable, then the fusion is optimized by reducing the difference contribution from the less reliable mode m2; conversely, if mode m2 has lower uncertainty, then mode s is retained. ij The full contribution is used to reflect the consistency between modes to the greatest extent possible;

[0063]

[0064] Through this mechanism, the conflict matrix S is weighted and adjusted based on uncertainty to construct the conflict adjustment matrix D, such that... in, Let S represent the reweighted conflict matrix, i.e., for the conflict matrix S

[0065] The reweighting mechanism, which multiplies element-wise with the conflict adjustment matrix D, ensures that more reliable modes have a greater influence in the fusion process, while the weights of unreliable modes are appropriately reduced. Meanwhile, the overall conflict score for each mode is obtained by the following formula:

[0066]

[0067] in, This represents the difference between modes m1 and m2 after reweighting. This represents the final conflict score after modality m1 is adjusted; to dynamically adjust the contribution of each modality, a method based on... The reward and punishment functions; if If the conflict exceeds the average conflict threshold θ, a penalty function is applied; otherwise, a reward function is applied, with α controlling the degree of the reward / penalty function. The reward / penalty function is defined as follows:

[0068]

[0069] Where e is the natural constant, approximately equal to 2.71828. The final conflict score after modality m1 adjustment is represented by the above dynamic fusion mechanism. This invention ensures that more reliable modalities play a crucial role in multimodal fusion, while the impact of highly conflicting or unreliable modalities on the overall prediction is effectively suppressed. Finally, the modal features x are... i According to its weight w i Perform weighted aggregation to generate the final multimodal fusion representation x. f For more accurate prognostic prediction;

[0070]

[0071] This method effectively improves the reliability and accuracy of multimodal data in predicting the prognosis of liver cancer through a dynamic fusion mechanism, and also provides a flexible approach to address modal conflicts.

[0072] S4: Individual survival risk prediction, risk stratification and treatment recommendations for liver cancer patients;

[0073] The recommended approach for step S4 personalized treatment is as follows:

[0074] This invention proposes a scientific and comprehensive method for recommending personalized treatment plans for patients with hepatocellular carcinoma (HCC) based on a postoperative risk prediction model. This method combines multidimensional data analysis and risk stratification technology to provide patients with precise treatment recommendations. The specific steps are as follows:

[0075] First, for new HCC cases, the postoperative risk prediction model h is used. SR (t|x) and h RFA (t|x) was used to assess the postoperative risk of patients undergoing surgical resection (SR) and radiofrequency ablation (RFA), respectively. Based on the prediction results, patients undergoing SR were divided into a high-risk group and a low-risk group, and patients undergoing RFA were divided into a high-risk group and a low-risk group. The high-risk group represented patients with a poor prognosis, and the low-risk group represented patients with a good prognosis. Next, the patients in the original SR group were input into the RFA risk prediction model λ. RFA (t|x) were reclassified into a new high-risk SR group and a new low-risk SR group; at the same time, patients from the original RFA group were input into the SR risk prediction model λ. SR(t|x) is reclassified into a new high-risk RFA group and a new low-risk RFA group. Through this two-way risk assessment method, this invention can identify patients with potential prognostic improvement for different treatment options. The focus is on the following two non-overlapping patient categories: Non-overlapping patient type I: Patients who were originally in the high-risk RFA group but are now in the low-risk RFA group in the new risk grouping. These patients have a poor prognosis with RFA treatment, but may have a significantly improved prognosis if they choose SR treatment. Non-overlapping patient type II: Patients who were originally in the high-risk SR group but are now in the low-risk SR group in the new risk grouping. These patients have a poor prognosis with SR treatment, but may have a significantly improved prognosis if they choose RFA treatment. Through the precise identification of non-overlapping patients, this invention provides a clear decision-making basis for personalized treatment.

[0076] Finally, based on the patients' risk grouping and non-overlapping characteristics, this invention proposes the following treatment recommendation principles: Type I patients: RFA is recommended as the first-line treatment; Type II patients: SR is recommended as the first-line treatment; For patients with low risk in both SR and RFA: Decisions for these patients can be made by comprehensively considering various factors such as clinical preference, patient wishes, or treatment costs, providing flexible and personalized treatment options; For patients with high risk in both SR and RFA: For these patients, careful decision-making is required, combining existing prognostic models and other available treatment options to ensure optimal treatment outcomes under controllable risks. Through the above steps, this invention achieves a comprehensive assessment of postoperative risks and precise recommendations for treatment plans for HCC patients, providing patients with better personalized treatment options and significantly improving the scientific rigor and effectiveness of clinical treatment.

[0077] In summary, this invention addresses the shortcomings of existing multimodal fusion methods in complex clinical scenarios with diverse data by proposing a dynamic conflict-aware multimodal ultrasound fusion method. This method achieves dynamic optimization of modal weights and uncertainty calibration through two core steps: self-calibration uncertainty estimation and dynamic conflict-aware fusion, thereby improving the accuracy and reliability of multimodal data fusion. Furthermore, based on survival risk prediction results, this invention further develops personalized treatment plans, providing scientific and precise treatment recommendations for hepatocellular carcinoma patients. This not only significantly improves patients' clinical prognosis but also provides important technical support for promoting the development of precision medicine and personalized treatment for liver cancer.

[0078] Example 3

[0079] A professional medical team in the ultrasound diagnostic department of a top-tier hospital in Nanjing has long been engaged in the diagnosis of liver cancer and ultrasound-guided ablation therapy for liver cancer. They possess extensive experience in ultrasound image analysis and clinical data processing, providing solid technical support and data assurance for the research of this invention. In previous studies, the team retrospectively collected case data from over 400 patients with primary solitary hepatocellular carcinoma, including over 250 patients who underwent surgical resection (SR) and over 150 patients who underwent radiofrequency ablation (RFA). During the implementation of this invention, it is planned to further collect preoperative imaging data and clinical information of liver cancer patients who underwent surgical resection or RFA at a top-tier hospital in Nanjing to support the ongoing research.

[0080] Regarding data collection, this invention covers preoperative clinical data, liver ultrasound contrast imaging, postoperative follow-up, and inclusion and exclusion criteria. Preoperative clinical data includes the patient's gender, age, comorbidities (such as hepatitis B, hepatitis C, fatty liver), cirrhosis status, biochemical indicators (alpha-fetoprotein, alanine aminotransferase, total bilirubin, albumin, platelet count), and tumor characteristics (lesion size, location, and whether it is located around blood vessels or at the liver margin). This clinical information provides a comprehensive reference for the stratification of liver cancer patients and the development of treatment plans.

[0081] Liver contrast-enhanced ultrasound examinations were performed by sonographers with over five years of experience, using the second-generation contrast agent SonoVue in a low-mechanical-index imaging mode for contrast agent-specific imaging. During the procedure, 2.4 ml of contrast agent was injected into the patient via the antecubital vein, followed by flushing with 5 ml of normal saline. Contrast images were simultaneously acquired, recording the spatial distribution and flow path of the contrast agent in the lesion and surrounding tissues until the contrast agent was completely removed. All images were stored in DICOM or AVI format to provide standardized imaging resources for subsequent data analysis.

[0082] Postoperative follow-up involves a two-year systematic observation of patients who have received SR or RFA treatment. The follow-up plan includes regular assessments at 1, 3, 6, 9, and 12 months postoperatively, and follow-up examinations every 3-6 months thereafter. During follow-up, enhanced CT or contrast-enhanced MRI is used as the gold standard for postoperative assessment, focusing on tumor progression characteristics (such as local tumor progression, vascular invasion, distant intrahepatic recurrence, and extrahepatic metastases). The time to the first appearance of postoperative tumor progression characteristics is defined as progression-free survival (PFS); if there is no tumor progression, PFS is defined as the time of death from any cause or the last follow-up visit.

[0083] Regarding patient selection, this invention clearly defines the inclusion and exclusion criteria. Inclusion criteria include patients with histologically or cytologically confirmed primary solitary hepatocellular carcinoma, lesion diameter ≤5.0 cm, good liver function (Child-Pugh A), no severe dysfunction of vital organs such as the heart, lungs, and kidneys, and the ability to adhere to the study protocol and follow-up procedures. Exclusion criteria cover patients who have received other treatments, those whose lesions are not observable under CEUS or whose image quality is poor, and those unable to complete the study due to drug abuse or other factors.

[0084] In summary, this invention features detailed design at every stage of data acquisition, processing, and follow-up, coupled with stringent inclusion and exclusion criteria, ensuring the scientific rigor and reliability of the research data. This implementation method provides a comprehensive and standardized operational framework for subsequent research, laying a solid foundation for the clinical application of precision diagnosis and treatment of liver cancer.

[0085] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic conflict-aware based multi-modality ultrasound fusion system comprising a B-mode ultrasound image acquisition module, a contrast-enhanced ultrasound image acquisition module, a clinical data acquisition module, characterized in that, Also comprising: a self-calibration uncertainty estimation module for proposing a cross-modal uncertainty calibration regularizer to align the modal uncertainty with its prediction effect; a dynamic conflict fusion module for proposing a dynamic conflict-aware fusion mechanism to dynamically adjust the weights of different modalities according to cross-modal consistency and modal uncertainty; an individual survival risk prediction module for individual survival risk prediction to stratify liver cancer patients by risk; a personalized treatment recommendation module for recommending treatment plans for liver cancer patients; The method of self-calibration uncertainty estimation is as follows: A partial likelihood loss function is used instead of the commonly used mean square error MSE; the loss function quantifies the prediction accuracy by calculating the relative risk between individuals and risk sets; the formula is defined as follows: where: i represents the ith individual, j represents the jth individual, m represents the modality, u i,m is the uncertainty constraint, E i,m is the prediction error, r i,m is the prediction risk function of the ith individual, r j,m is the prediction risk function of the jth individual, the risk set R(·) refers to the individuals who have not experienced the event at the time t i when the event occurs, δ i = 1 indicates that the ith individual has experienced the event, the ∝ in the formula represents direct proportion, -log(·) is the logarithmic function, and ∑ is the summation of all the individuals who have experienced the event, is used to calculate the sum of the partial likelihood loss of all the individuals who have experienced the event, exp(r i,m ) and exp(r j,m ) represent the exponential form of the prediction risk score of the ith individual and the jth individual in the modality m; The method of dynamic conflict fusion is as follows: A multi-modal conflict matrix S is defined, each element in S represents the difference in prognosis between modality m1 and modality m2, s ij , s ij = |r i -r j |, where: r i represents the predictive risk function of modality m1, r j represents the predictive risk function of modality m2, the conflict matrix S quantifies the inconsistency across modalities by calculating the difference in prognostic outcomes between modalities; Adjust the differences between modes based on the confidence level of the modal predictions. ij Introducing a conflict adjustment factor d ij If the uncertainty of mode m1 is u i Uncertainty u below mode m2 j This indicates that mode m1 is more reliable. In this case, the fusion is optimized by reducing the difference contribution from the less reliable mode m2. If the uncertainty of mode m2 is lower, then s is retained. ij The full contribution is used to reflect the consistency between modes to the greatest extent possible; An uncertainty-based weighted adjustment is made to the conflict matrix S to construct a conflict adjustment matrix D, such that wherein, represents the re-weighted conflict matrix S, and the overall conflict score of each modality is obtained by the following formula: wherein, denotes the difference in the modalities m1, m2 after re-weighting, denotes the final conflict score after adjustment of the modality m1.

2. A method based on the system of claim 1, characterized by The method comprises the following steps: S1: B-mode ultrasound image, contrast-enhanced ultrasound image and clinical data acquisition, S2: Self-calibration uncertainty estimation: propose a cross-modal uncertainty calibration regularizer to align the modal uncertainty with its prediction effect; S3: Dynamic conflict fusion: propose a dynamic conflict-aware fusion mechanism to dynamically adjust the weights of different modalities according to cross-modal consistency and modal uncertainty.

3. The method of claim 2, wherein, In step S2, the following steps are further included: In batch data processing, the modal uncertainty u i,m is constrained to be proportional to the prediction error E i,m and is transformed into a discrete distribution U and E by normalization, defined as follows: where B is batch size, M is the number of modalities, represents the uncertainty value of the last element in all samples and modalities, represents the prediction error of the last element in all samples and modalities, the normalized uncertainty vector U and the error vector E can eliminate the interference of the number of modalities and the batch size on the uncertainty estimation, and ensure the consistency and stability of the distribution.

4. The method of claim 3, wherein, In order to dynamically adjust the contribution of each modality in step S3, a reward and penalty function based on is introduced again; if the average conflict exceeds the threshold θ, the penalty function is applied; otherwise, the reward function is applied, and α controls the degree of the reward and penalty function; the reward and penalty functions are defined as follows: wherein e is a natural constant, denotes the final conflict score of the modality m1 after adjustment; finally, the modal feature x i is weighted aggregated according to its weight w i to generate the final multi-modal fusion representation x f for more accurate prognosis prediction;

Images