Deep Learning-Based Medical Image Diagnostic Assistance System

Abstract

This invention discloses a deep learning-based medical image diagnostic assistance system, belonging to the field of medical image processing technology. This invention employs integrity verification, specific noise suppression, and spatial standardization processing. It captures the dynamic changes of lesion features through a temporal attention mechanism, while simultaneously utilizing a medical pathology knowledge graph to achieve precise matching of temporal lesion features with anatomical structures and pathological type nodes, significantly reducing the risk of missed or misdiagnosed diagnoses. By assessing the confidence of diagnostic assistance conclusions through multi-dimensional factors, it automatically prompts for manual review of low-confidence results, forming a dual control system of system assistance and human oversight. This system provides standardized lesion localization, pathological type candidates, and diagnostic criteria for primary care physicians or those with limited experience, while reducing repetitive data processing work for experienced physicians, helping them focus on complex case judgments. Overall, it improves the diagnostic and treatment capabilities of medical teams, effectively reducing medical risks and enhancing the quality and efficiency of diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a medical image diagnostic assistance system based on deep learning. Background Technology

[0002] With the rapid development of medical imaging technology, multimodal equipment such as CT, MRI, and ultrasound have been widely used in clinical diagnosis. However, the imaging principles and parameter settings of different equipment differ, resulting in problems such as inconsistent resolution, noise interference, and inconsistent cross-modal data distribution in the acquired medical images, which increases the difficulty of diagnosis for doctors. At the same time, the amount of medical image data has surged, and doctors need to spend a lot of time on image analysis and lesion identification, which can easily lead to missed diagnoses or misdiagnoses due to fatigue or differences in experience.

[0003] Furthermore, existing diagnostic support systems often lack the ability to capture temporal features and effectively fuse cross-modal features, resulting in insufficient accuracy and reliability of diagnostic conclusions. Additionally, inconsistent data formats hinder efficient integration with electronic medical record systems. Therefore, there is an urgent need for an intelligent diagnostic support system capable of standardized medical image processing, precise fusion of multi-dimensional features, temporal correlation analysis, and standardized report output to improve diagnostic efficiency and accuracy and support clinical decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based medical image diagnostic assistance system. Through feature fusion and diagnostic reasoning modules, it improves the accuracy of lesion identification, reduces the risk of missed or misdiagnosed cases, generates standard reports and integrates with electronic medical records, adapts to multiple scenarios, and improves the level of medical teams by relying on confidence assessment and manual review to control quality, thereby effectively reducing medical risks and improving the quality and efficiency of diagnosis and treatment, thus solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a medical image diagnostic assistance system based on deep learning.

[0006] Deep learning-based medical image diagnostic assistance systems include: The medical image data processing module is used to acquire medical image data and corresponding equipment parameters collected by various medical imaging devices, perform image processing on the acquired medical image data, and output standardized medical image data. The medical image feature fusion module is used to extract multi-dimensional features and generate calibration factors based on image categories of standardized medical image data, calculate the attention weights of each dimension feature based on a deep learning model, optimize the fusion by combining the calibration factors and attention weights, enhance the fused lesion features, and output an enhanced lesion feature map. The diagnostic reasoning mapping module is used to extract dynamic temporal data and corresponding physiological signals from standardized medical image data. Based on the periodic features of physiological signals, a temporal attention mechanism is constructed to capture the temporal change pattern of lesion features in the enhanced lesion feature map and generate temporally associated lesion features. The diagnostic reasoning mapping module is also used to call the medical pathology knowledge graph, match the temporal associated lesion features with the anatomical structure nodes and pathological type nodes in the knowledge graph, establish the association between lesion features and anatomical location and pathological type, and output the temporal lesion feature set and the list of lesion and pathological information association. The diagnostic result output module is used to generate diagnostic auxiliary conclusions from the temporal lesion feature set and the lesion-pathology information association list. It extracts lesion location coordinates based on the enhanced lesion feature map, marks the diagnostic results by combining the confidence of the diagnostic auxiliary conclusions, integrates the diagnostic auxiliary conclusions, lesion location coordinates and confidence, and outputs a standardized diagnostic report.

[0007] Furthermore, the process by which the medical image data processing module outputs standardized medical image data includes: Acquire medical image data and equipment parameters collected by various medical imaging devices, establish a mapping relationship between medical imaging devices and medical image data and equipment parameters, verify data integrity based on hash value, and remove redundant data; Based on the scanning resolution and exposure intensity in the equipment parameters, combined with the preset standard exposure intensity and standard resolution, the equipment difference calibration coefficient is calculated. Based on the equipment difference calibration coefficient, the gray value of each pixel in the medical image data is corrected to generate a medical image with gray balance adjustment. Extract target feature points from the medical image after grayscale balance adjustment, determine the edge contour coefficients of the target region and background region based on the target feature points, determine the noise type, and perform adaptive filtering denoising on the target region and background region respectively based on the noise type to generate a denoised medical image. The gray-level difference between the target region and the background region in the denoised medical image is obtained, the contrast adjustment coefficient is obtained, and the comprehensive contrast adjustment coefficient is obtained by combining the center gray value of the target region. Based on the comprehensive contrast adjustment coefficient, the distinction between lesions and normal tissues is enhanced, and a contrast-optimized medical image is generated. Based on the retrieved preset medical image standard resolution, the size of the contrast-optimized medical image is adjusted to generate a medical image with standardized resolution. The quality quantification index of the medical image after resolution standardization is calculated, the quality quantification index is compared with the preset quality threshold, the image anomaly coefficient is calculated based on the comparison result, and the quality verification result is generated. Based on the quality verification results, qualified images are converted into a standard format, embedded with equipment calibration parameters and processing logs, and output standardized medical image data.

[0008] Furthermore, data integrity is verified based on hash values, including: Establish a verification rule base, clarify the pass / fail thresholds for each indicator, and based on the verification rule base, verify the integrity of the data format and metadata, and determine the validity of the pixel value range. Calculate the target hash value of the medical image data, obtain the original hash value during transmission, and compare the target hash value with the original hash value bit by bit to verify the integrity of data transmission. Based on the label information of medical image data, the medical image data is classified, each subset is obtained, the consistency of data within each subset is calculated, and the confidence score of each subset is generated. Invalid data with a confidence level lower than the preset sub-data confidence level threshold is removed, and missing data is supplemented based on the similarity algorithm of adjacent data. The integrity of the supplemented data is verified, and a data integrity verification report is generated.

[0009] Furthermore, the medical image feature fusion module includes: The calibration factor generation unit is used to match the corresponding feature extraction strategy based on the image category of standardized medical image data, extract the modal statistical features, data distribution features and anatomical structure features of each modality image based on the matching results, calculate the difference value of the same feature under different modalities, and generate the corresponding calibration factor according to the magnitude of the difference value and the influence weight of the feature on the diagnosis. The attention weight calculation unit is used to build a deep learning model. It uses the deep learning model to quantify the importance of each dimension feature in the lesion identification task and calculates the attention weight corresponding to each dimension feature based on the quantification evaluation results. The feature space alignment unit is used to combine calibration factors and attention weights to perform feature space calibration and alignment on multi-dimensional features across modalities, eliminate the distribution differences of cross-modal features, and obtain optimized features. The feature enhancement operator construction unit is used to load a preset lesion prior knowledge base, which includes the anatomical features, gray-scale distribution features and morphological features of the lesion. Based on the lesion prior knowledge base, key feature parameters of the lesion are extracted, a feature enhancement objective function is established, and a corresponding feature enhancement operator is constructed. The lesion feature enhancement and output unit is used to input the optimized features into the feature enhancement operator, enhance the expression intensity of the lesion region features, suppress noise interference in non-lesion regions, calculate the distinguishability between the enhanced lesion features and normal tissue features, and output the enhanced lesion feature map.

[0010] Furthermore, feature space calibration and alignment processing is performed on the multi-dimensional features across modalities, specifically including: Obtain the original distribution range of multi-dimensional features of each modality, and normalize and calibrate the distribution range of multi-dimensional features of each modality based on the calibration factor; Obtain the attention weights of each dimension of the calibrated features, and then assign weighted distributions to each dimension of the calibrated features based on the attention weights. Retrieve preset unified feature space parameters and map the weighted multi-dimensional features of each modality to the unified feature space; Calculate the spatial position deviation and semantic information matching degree of the multi-dimensional features of each modality after mapping. If the position deviation is less than the preset deviation threshold and the semantic matching degree is greater than the preset similarity value, then spatial alignment is completed. Calculate the cross-modal consistency index of the aligned features. If the index meets the preset index standard, the optimized features are obtained.

[0011] Furthermore, a temporal attention mechanism is constructed based on the periodic characteristics of physiological signals, specifically including: Extract physiological signals from dynamic time-series data, analyze the periodic characteristics of physiological signals, and determine the signal period parameters; A temporal attention weight matrix is ​​constructed based on signal period parameters, and differentiated attention weights are assigned to lesion features at different period stages to enhance lesion features with significant temporal changes. The temporal attention weight matrix is ​​subjected to matrix operation with the enhanced lesion feature map to calculate the temporal variation coefficient of each feature channel, highlighting the temporal dynamic differences of lesion features and generating temporally associated lesion features.

[0012] Furthermore, the step of performing similarity matching between temporally associated lesion features and nodes in the knowledge graph includes: The feature vectors of target nodes in the medical pathology knowledge graph are extracted, the features of temporally related lesions are processed by vector transformation, and the similarity value between the feature vectors of temporally related lesions and the feature vectors of each target node is calculated. Target nodes with similarity greater than a preset similarity threshold are selected, and a mapping relationship between temporally related lesion features and the selected target nodes is established to generate a list of lesion-pathology information associations.

[0013] Furthermore, the confidence level of the diagnostic auxiliary conclusion is obtained based on a multi-dimensional influencing factor system, which includes feature reliability factor, matching validity factor, data quality factor, and model output consistency factor. The comprehensive confidence level is calculated based on the weight coefficients of each influencing factor and serves as the confidence level of the diagnostic auxiliary conclusion.

[0014] Furthermore, the construction of the temporal attention mechanism based on the periodic features of physiological signals specifically includes the following steps: The normalized temporal physiological signal intensity, the average gray-scale curve of the lesion area, and the calculated first derivative of the physiological signal, the second derivative of the gray-scale curve, the hydrodynamic lag time, and the instantaneous phase angle of the main frequency were obtained. Based on the normalized temporal physiological signal intensity, the average gray-level curve of the lesion region, and the calculated first derivative of the physiological signal, the second derivative of the gray-level curve, the hydrodynamic lag time, and the instantaneous phase angle of the dominant frequency, the hemodynamic-lesion perfusion derivative coupling efficiency index is calculated. The calculation formula is as follows: in, The derivative coupling efficiency index is dimensionless, and its value range is [range missing]. This characterizes the causal consistency between physiological driving forces and lesion responses; The normalized temporal physiological signal intensity, It is the first derivative of the physiological signal, representing the instantaneous rate of change of physiological driving pressure; To introduce fluid dynamic lag time The average grayscale curve of the lesion area after treatment. The second derivative of the average grayscale curve of the lesion area represents the perfusion acceleration of the contrast agent in the lesion tissue. This represents the total duration of the sampling period for time-series data. To prevent the regularization of small constants with a denominator of zero; and These represent the instantaneous phase angles of the dominant frequency of the physiological signal and the grayscale curve of the lesion in the Hilbert transform domain, respectively. For phase-locked constraint coefficients, The preset physiological cycle phase jitter tolerance variance; exponential term Used to forcibly suppress coupling index when two waveforms are similar but phase logic is completely unlocked; This refers to the time derivative; Based on the calculated hemodynamic-lesion perfusion derivative coupling efficiency index Adaptive gain adjustment is performed on the temporal attention weight matrix to suppress non-causal artifact interference and generate temporally associated lesion features that incorporate physiological causal constraints.

[0015] Furthermore, after establishing the correlation between lesion characteristics and anatomical location and pathological type, the process also includes performing a logical self-consistency verification step based on counterfactual topological residual analysis: Based on the current matched pathology type node as an assumption, retrieve the corresponding standard normal tissue anatomical manifold base from the medical pathology professional knowledge atlas; Using a spatial deformation mapping algorithm, the standard normal tissue anatomical manifold base is projected onto the coordinate system of the current medical image data to generate a virtual counterfactual health image; Pixel-level difference operations are performed on standardized medical image data and virtual counterfactual health images to extract the residual feature field of lesions; The Betti number of the residual feature field of the lesion is calculated using a topological data analysis algorithm to obtain the 0-dimensional Betti number. Characterization of connected components and 1-dimensional Betti numbers Characterized features of the porous structure; Determine whether the calculated Betty number feature falls within the predefined topological invariant interval of the pathological type node in the knowledge graph; if it falls within the interval, confirm the diagnostic logic is consistent and increase the confidence; if it does not fall within the interval, determine that the current feature is a false match caused by unstructured noise, and trigger the re-evaluation process of the suboptimal node.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention provides high-quality, standardized medical image data for diagnosis by preprocessing medical image data of different modalities. It extracts multi-dimensional features for different image categories and optimizes the fusion by combining attention weights and calibration factors calculated by deep learning to enhance the expression of lesion features, highlight lesion information that is key to diagnosis, and captures the dynamic changes of lesion features through a temporal attention mechanism, breaking through the limitations of traditional static image diagnosis and better conforming to the time pattern of lesion development. 2. This invention uses a medical pathology knowledge map to match temporal lesion features with anatomical structures and pathological type nodes, allowing the reasoning process to combine professional pathological knowledge, which greatly reduces the risk of missed diagnosis and misdiagnosis, and improves the accuracy of lesion localization and pathological type matching; 3. This invention establishes a dual quality control system of system assistance and human oversight through confidence assessment, avoiding misdiagnosis caused by system errors, reducing medical risks, and providing standardized lesion localization, pathological type candidates, and diagnostic criteria for doctors at the grassroots level or with insufficient experience, helping them to quickly grasp the key to diagnosis; for experienced doctors, it can reduce repetitive work, allowing them to focus on the judgment of complex cases, and improving the overall diagnosis and treatment level of the medical team. Attached Figure Description

[0017] Figure 1 This is a flowchart of the deep learning-based medical image diagnostic assistance system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 A deep learning-based medical image diagnostic assistance system includes: The medical image data processing module is used to acquire medical image data and corresponding device parameters collected by various medical imaging devices, perform image processing on the acquired medical image data, perform integrity verification based on data compliance standards, suppress specific noise by combining the characteristics of each modality, perform spatial standardization based on anatomical structure landmarks, and output standardized medical image data. The medical image feature fusion module is used to extract multi-dimensional features and generate calibration factors based on standardized medical image data, such as anatomical site categories like chest, brain, and abdomen, or disease categories like tumors and inflammation. These features include modal statistical features, data distribution features, and anatomical structure features. The module calculates the attention weights of each dimension feature based on a deep learning model, combines the calibration factors and attention weights for optimization and fusion, and enhances the fused lesion features to output an enhanced lesion feature map. The diagnostic reasoning mapping module is used to extract dynamic temporal data and corresponding physiological signals, such as blood flow signals and metabolic signals, from standardized medical image data. Based on the periodic features of physiological signals, a temporal attention mechanism is constructed to capture the temporal change pattern of lesion features in the enhanced lesion feature map and generate temporally associated lesion features. The diagnostic reasoning mapping module is also used to call the medical pathology knowledge graph, match the temporal associated lesion features with the anatomical structure nodes and pathological type nodes in the knowledge graph, establish the association between lesion features and anatomical location and pathological type, and output the temporal lesion feature set and the list of lesion and pathological information association. The diagnostic result output module is used to generate auxiliary diagnostic conclusions based on a weighted fusion algorithm by combining the temporal lesion feature set and the list of associations between lesions and pathological information. It extracts the lesion location coordinates based on the enhanced lesion feature map, and marks the diagnostic results in conjunction with the confidence level of the auxiliary diagnostic conclusions. If the confidence level is lower than the preset confidence threshold, the marking requires manual review and prompting. It integrates the auxiliary diagnostic conclusions, lesion location coordinates, and confidence level to output a standardized diagnostic report. The report format conforms to the clinical data exchange standard and supports the association display of the diagnostic report with the original medical images and standardized medical images. At the same time, it outputs structured data for integration with the electronic medical record system.

[0020] In this embodiment, medical image data from different modalities is preprocessed to specifically suppress noise specific to different modal devices and unify the spatial specifications of images acquired by different devices, providing high-quality, standardized medical image data for diagnosis. Multi-dimensional features are extracted for different image categories, and the fusion is optimized using deep learning-calculated attention weights and calibration factors to enhance lesion feature expression, highlighting key diagnostic information, reducing redundant feature interference, and making lesion identification more accurate. A temporal attention mechanism is used to capture the dynamic changes in lesion features, overcoming the limitations of traditional static image diagnosis and better aligning with the temporal patterns of lesion development. Medical pathology experts are also invoked. The professional knowledge graph matches temporal lesion characteristics with anatomical structures and pathological type nodes, allowing the reasoning process to combine professional pathological knowledge, significantly reducing the risk of missed diagnoses and misdiagnoses, and improving the accuracy of lesion localization and pathological type matching. Through confidence assessment, a dual quality control system of system assistance and human oversight is formed to avoid misdiagnosis caused by system errors, reduce medical risks, and provide standardized lesion localization, pathological type candidates and diagnostic criteria for primary care physicians or those with insufficient experience, helping them to quickly grasp the key to diagnosis. For experienced physicians, it can reduce repetitive work, allowing them to focus on the judgment of complex cases, and improve the overall diagnosis and treatment level of the medical team.

[0021] In this embodiment, the process by which the medical image data processing module outputs standardized medical image data includes: Acquire DICOM format medical image data and equipment parameters from various medical imaging devices such as CT, MRI, ultrasound, and X-ray. Equipment parameters include scanning resolution, exposure intensity, detector sensitivity, slice thickness, TE / TR value, etc. Establish a mapping relationship between medical imaging devices and medical image data and equipment parameters, verify data integrity based on hash value, and remove redundant data. Based on the scanning resolution and exposure intensity in the equipment parameters, combined with the preset standard exposure intensity and standard resolution, the equipment difference calibration coefficient is calculated. Based on the equipment difference calibration coefficient, the gray value of each pixel in the medical image data is corrected to generate a medical image with gray balance adjustment. After grayscale balance adjustment, target feature points such as organ edges and lesion areas are extracted from medical images. Based on the target feature points, the edge contour coefficients of the target area and background area are determined. The noise type is judged, including CT X-ray hardening artifacts, MRI motion artifacts, and ultrasound speckle noise. Based on the noise type matching and adaptive filtering algorithm, the target area and background area are subjected to adaptive filtering and denoising processing respectively to generate denoised medical images. The gray-level difference between the target region and the background region in the denoised medical image is obtained. The contrast adjustment coefficient is obtained based on the gray-level balance index. The comprehensive contrast adjustment coefficient is obtained by combining the center gray value of the target region. The distinction between lesions and normal tissues is enhanced based on the comprehensive contrast adjustment coefficient, and a contrast-optimized medical image is generated. Based on the preset standard resolution of medical images, the size of the contrast-optimized medical images is adjusted according to the interpolation algorithm to ensure that the pixel density of medical images acquired by different devices is uniform and to generate medical images with standardized resolution. The system calculates quality quantification indicators such as image signal-to-noise ratio, contrast-to-noise ratio, and edge sharpness of medical images after resolution standardization. It then compares these quality quantification indicators with preset quality thresholds, calculates image anomaly coefficients based on the comparison results, determines whether the image quality meets the standards, and generates quality verification results. Based on the quality verification results, qualified images are converted into standard DICOM 3.0 format, and device calibration parameters and processing logs are embedded, including metadata such as filter algorithm type, contrast adjustment coefficient, and quality verification results, and standardized medical image data is output.

[0022] In this embodiment, a secondary optimization process is triggered for unqualified images, and the adaptive filtering and noise reduction and contrast enhancement steps are re-executed; for images that are still unqualified after secondary optimization, a re-acquisition prompt is generated and fed back to the corresponding medical imaging device terminal. In this embodiment, data integrity verification based on hash value includes: Retrieve integrity indicators from data compliance standards, including data format integrity, pixel value range validity, and metadata integrity; establish a verification rule base; clarify the pass / fail threshold for each indicator; and verify the integrity of data format and metadata and determine the validity of pixel value range based on the verification rule base. Calculate the target hash value of the medical image data, obtain the original hash value during transmission, compare the target hash value with the original hash value bit by bit to verify the integrity of data transmission. If they are completely consistent, the data transmission is determined to be complete; if there are differences, the data transmission is marked as abnormal. Based on the label information of medical image data, such as device type, scanning site, and acquisition time, the medical image data is classified to obtain each subset, the consistency of data within each subset is calculated, and the confidence score of the subset is generated. Invalid data with a confidence level lower than the preset sub-data confidence level threshold is removed, and missing data is supplemented based on the similarity algorithm of adjacent data. The integrity of the supplemented data is verified, a data integrity verification report is generated, and the amount of data that passed the verification, the amount of invalid data, the status of missing data supplementation, and the hash value comparison results are recorded.

[0023] In this embodiment, the original storage path of the medical image data is obtained, the header information and body content of the data file are read, and key information such as data format identifier, pixel value range, and metadata fields are extracted. It is determined whether the data format identifier conforms to the preset format standard, and the integrity of the data format is verified. It is checked whether the pixel value range is within the valid range, such as the pixel value of CT image which is usually between -1000 and 4000 HU, and the validity of the pixel value range is determined. It is verified whether the metadata fields contain required fields such as device model, scan time, and patient basic information, and the integrity of the metadata is determined. In this embodiment, the medical image feature fusion module includes: The calibration factor generation unit is used to match the corresponding feature extraction strategy to the image category based on standardized medical image data. Based on the matching results, it extracts the modal statistical features of each modality image, such as pixel mean, variance, gray-level histogram peak value, data distribution features such as feature probability density distribution, cross-modal feature correlation coefficient, and anatomical structural features such as organ contour coordinates and relative position parameters of lesions and organs. It calculates the difference value of the same feature under different modalities and generates the corresponding calibration factor according to the magnitude of the difference value and the influence weight of the feature on the diagnosis. The calibration factor value range is mapped to the interval [0.1, 1.0] to ensure that the calibration strength is adapted to the degree of feature difference. The attention weight calculation unit is used to construct a deep learning model, such as a feature attention model based on Transformer. The deep learning model is used to quantify the importance of each dimension feature in the lesion identification task, such as calculating the mutual information value between the feature and the lesion label. Based on the quantitative evaluation results, the attention weight corresponding to each dimension feature is calculated. The attention weight is used to characterize the contribution of each dimension feature to lesion identification. The feature space alignment unit is used to combine calibration factors and attention weights to perform feature space calibration and alignment on multi-dimensional features across modalities, eliminate the distribution differences of cross-modal features, and obtain optimized features. The feature enhancement operator construction unit is used to load a preset lesion prior knowledge base. The lesion prior knowledge includes the anatomical features of the lesion, such as the typical growth location of the tumor, its proximity to blood vessels, gray-level distribution features, such as the gray-level difference range between the inflamed area and normal tissue, and morphological features, such as the boundary smoothness and shape irregularity coefficient of the lesion. Based on the lesion prior knowledge base, key feature parameters of the lesion are extracted, a feature enhancement objective function is established, and a corresponding feature enhancement operator is constructed. The operator parameters are adapted to the feature enhancement requirements of different lesion types. The lesion feature enhancement and output unit is used to input the optimized features into the feature enhancement operator, enhance the expression intensity of the lesion region features, suppress noise interference in non-lesion regions, calculate the distinguishability between the enhanced lesion features and normal tissue features, and output the enhanced lesion feature map.

[0024] In this embodiment, it is determined whether a preset discrimination threshold has been reached; if the threshold is reached, an enhanced lesion feature map is output, and the feature enhancement process parameters, such as the convolution kernel size and denoising intensity, are recorded; if the threshold is not reached, the feature enhancement operator parameters are readjusted, and the enhancement steps are repeated until an enhanced lesion feature map that meets the requirements is output.

[0025] In this embodiment, feature space calibration and alignment processing is performed on multi-dimensional features across modalities, specifically including: The original distribution range of multi-dimensional features of each modality is obtained, such as the feature value range of CT modality [0,1000] and the feature value range of MRI modality [0,2000]. The distribution range of multi-dimensional features of each modality is normalized and calibrated based on the calibration factor so that all modal features are mapped to the [0,1] interval, eliminating cross-modal domain shifts caused by differences in modal acquisition equipment, such as CT detector sensitivity, MRI TE / TR value, and different imaging principles. Obtain the attention weights of each dimension of features after calibration, and then assign weighted distributions to each dimension of features after calibration based on the attention weights. For example, the weight of high contribution features is multiplied by a factor of 1.2, and the weight of low contribution features is multiplied by a factor of 0.8. For example, increase the weight of gray-scale gradient features at the edge of lesions and reduce the influence of uniform gray-scale features in the background area. Retrieve preset unified feature space parameters, such as spatial dimension, coordinate scale, and semantic label system. Through feature space coordinate mapping algorithms, such as thin plate spline interpolation mapping and affine transformation mapping, map the weighted multi-dimensional features of each modality to the unified feature space. Calculate the spatial positional deviation of the multi-dimensional features of each modality after mapping, such as feature point coordinate error, and the semantic information matching degree, such as the association probability between features and semantic labels. If the positional deviation is less than a preset deviation threshold and the semantic matching degree is greater than a preset similarity value, spatial alignment is completed. If the criteria are not met, re-optimize the mapping algorithm parameters and repeat the mapping steps. Calculate the cross-modal consistency index of the aligned features, such as the trace value of the feature covariance matrix. If the index meets the preset index standard, such as the trace value being greater than 0.7, the optimized features with both consistency and effectiveness are obtained. If the standard is not met, the calibration, weighting, and mapping process is re-executed until the features meet the requirements.

[0026] In this embodiment, normalization calibration based on calibration factors maps features from different modalities such as CT and MRI to a unified range, effectively eliminating cross-modal domain shifts caused by differences in equipment parameters and imaging principles, laying a unified scale foundation for multimodal feature fusion. Furthermore, weighted allocation based on attention weights specifically enhances the weights of high-contribution features such as gray-scale gradients at lesion edges, while reducing the influence of redundant features such as uniform gray-scale in the background region, allowing feature representation to focus more on key diagnostic information and reducing invalid interference. Dual verification of positional deviation and semantic matching degree ensures spatial alignment accuracy, resolving the problem of spatial misalignment of multimodal features and guaranteeing the consistency of feature semantics and position during fusion. Consistency indicators such as the trace value of the feature covariance matrix are verified, and features that do not meet the standards are repeatedly subjected to calibration, weighting, and mapping processes to ensure that the final optimized features output possess both cross-modal consistency and diagnostic effectiveness.

[0027] In this embodiment, a temporal attention mechanism is constructed based on the periodic characteristics of physiological signals, specifically including: Physiological signals are extracted from dynamic time-series data, and the periodic characteristics of the physiological signals are analyzed by Fourier transform to determine the signal period parameters, such as period duration and peak interval. A temporal attention weight matrix is ​​constructed based on signal period parameters, and differentiated attention weights are assigned to lesion features at different period stages to enhance lesion features with significant temporal changes. The temporal attention weight matrix is ​​subjected to matrix operation with the enhanced lesion feature map to calculate the temporal variation coefficient of each feature channel, highlighting the temporal dynamic differences of lesion features and generating temporally associated lesion features.

[0028] In this embodiment, for physiological signals without obvious periodic characteristics, such as sudden abnormal metabolic signals, a pseudo-period is constructed using a sliding time window, and attention weights are assigned based on the local change trend of the signal within the window. At the same time, the construction process parameters of the temporal attention mechanism are recorded, including the type of filtering algorithm, the calculation results of period parameters, and the weight matrix optimization log. Metadata of temporally associated lesion features is embedded to facilitate subsequent diagnostic traceability and mechanism optimization. In this embodiment, the step of performing similarity matching between temporally associated lesion features and nodes in the knowledge graph includes: Feature vectors of target nodes in the medical pathology knowledge graph are extracted, such as feature vectors of anatomical structure nodes and feature vectors of pathology type nodes, to ensure that the extracted feature vectors can represent the core attributes of the nodes. Vector transformation is performed on the temporally related lesion features to make the transformed lesion feature vectors match the dimensions and data format of the knowledge graph node feature vectors. The similarity calculation method is adopted to calculate the similarity value between the temporally related lesion feature vectors and the feature vectors of each target node. Target nodes with similarity greater than a preset similarity threshold are selected, and the association mapping relationship between temporal associated lesion features and the selected target nodes is established. The correspondence between lesion features and anatomical location and pathological type is clarified, and a list of lesion and pathological information association is generated, including lesion temporal feature identifier, corresponding anatomical location, candidate pathological type and matching similarity.

[0029] In this embodiment, by determining the periodic parameters, the temporal variation patterns of lesion features are specifically captured. Combined with differentiated attention weight matrix allocation, lesion features with significant temporal changes are enhanced, and dynamic differences are calculated to effectively overcome the limitations of traditional static image diagnosis, making the expression of lesion features more consistent with the temporal evolution patterns of physiological signals. The anatomical structure and pathological type node feature vectors that can represent core attributes are extracted from the medical pathology professional knowledge atlas. The dimension and format of the node vector are adapted by transforming the lesion feature vectors to ensure the feasibility and accuracy of similarity calculation. An association list containing lesion temporal feature identifiers, anatomical locations, candidate pathological types, and matching similarities is generated. This makes the association between lesion features and anatomical and pathological information more targeted and reliable, and provides a clear and traceable basis for diagnostic reasoning, avoiding subjective judgment bias.

[0030] In this embodiment, the confidence level of the diagnostic auxiliary conclusion is obtained based on a multi-dimensional influencing factor system, wherein the multi-dimensional influencing factors include feature reliability factor, matching validity factor, data quality factor and model output consistency factor. The comprehensive confidence level is calculated based on the weight coefficient of each influencing factor and is used as the confidence level of the diagnostic auxiliary conclusion. In this embodiment, based on the temporal stability of each feature in the temporal lesion feature set and the inter-class distinguishability between lesions and normal tissues, a weighted summation algorithm is used to calculate the feature reliability factor; the matching similarity of each candidate pathological type in the lesion and pathological information association list is extracted, and the weighted value of the maximum and mean of the matching similarity is used as the matching validity factor; the quality quantification indicators of standardized medical image data are retrieved, including signal-to-noise ratio, contrast-to-noise ratio, and edge sharpness, and each indicator is compared with a preset optimal threshold to obtain the normalized pass rate of each indicator, and the mean algorithm is used to calculate the data quality factor; based on the multiple prediction results of the deep learning model for the same case, the proportion of the number of times the prediction conclusions are consistent is statistically analyzed, and combined with the normalized value of the model's accuracy in the test set of similar cases, a product algorithm is used to obtain the model output consistency factor; In this embodiment, the analytic hierarchy process (AHP) is used to construct a judgment matrix. Combined with the priority requirements of clinical diagnosis, the weights of each factor are obtained through eigenvalue decomposition. The weighted product algorithm is used to fuse the influencing factors, and the calculation results are mapped to the [0,1] interval as the final confidence level of the diagnostic auxiliary conclusion. In this embodiment, a multi-dimensional influencing factor system is used to achieve a comprehensive assessment of confidence level, ensuring that the factors accurately reflect the quality of the lesion characteristics themselves, objectively reflect the reliability of pathological associations, control the basic impact on diagnostic conclusions from the data source, effectively avoid the random errors of single model predictions, and fully meet actual clinical needs, making the confidence level results more consistent with the doctor's diagnostic logic, achieving effective synergy of factors of different dimensions, giving the confidence level results standardized interpretation, making it easier for doctors to quickly judge the reliability of conclusions, and improving the clinical trustworthiness of the system's assisted diagnosis.

[0031] In this embodiment, the construction of the temporal attention mechanism based on the periodic features of physiological signals specifically includes the following steps: The normalized temporal physiological signal intensity, the average gray-scale curve of the lesion area, and the calculated first derivative of the physiological signal, the second derivative of the gray-scale curve, the hydrodynamic lag time, and the instantaneous phase angle of the main frequency were obtained. Based on the normalized temporal physiological signal intensity, the average gray-level curve of the lesion region, and the calculated first derivative of the physiological signal, the second derivative of the gray-level curve, the hydrodynamic lag time, and the instantaneous phase angle of the dominant frequency, the hemodynamic-lesion perfusion derivative coupling efficiency index is calculated. The calculation formula is as follows: in, The derivative coupling efficiency index is dimensionless, and its value range is [range missing]. This characterizes the causal consistency between physiological driving forces and lesion responses; The normalized temporal physiological signal intensity, It is the first derivative of the physiological signal, representing the instantaneous rate of change of physiological driving pressure; To introduce fluid dynamic lag time The average grayscale curve of the lesion area after treatment. The second derivative of the average grayscale curve of the lesion area represents the perfusion acceleration of the contrast agent in the lesion tissue. This represents the total duration of the sampling period for time-series data. To prevent the regularization of small constants with a denominator of zero; and These represent the instantaneous phase angles of the dominant frequency of the physiological signal and the grayscale curve of the lesion in the Hilbert transform domain, respectively. For phase-locked constraint coefficients, The preset physiological cycle phase jitter tolerance variance; exponential term Used to forcibly suppress coupling index when two waveforms are similar but phase logic is completely unlocked; This refers to the time derivative; Based on the calculated hemodynamic-lesion perfusion derivative coupling efficiency index Adaptive gain adjustment is performed on the temporal attention weight matrix to suppress non-causal artifact interference and generate temporally associated lesion features that incorporate physiological causal constraints.

[0032] When processing time-varying medical imaging data such as dynamic contrast-enhanced CT, dynamic contrast-enhanced MRI (DCE-MRI), or ultrasound imaging, the system first initiates a multimodal data synchronous acquisition and high-dimensional preprocessing process. Traditional image analysis often overlooks the physiological driving force behind lesion enhancement, namely the patient's systemic hemodynamic state. This system synchronously receives high-dimensional dynamic image sequence data streams from medical imaging equipment and continuous time-series physiological signal data streams from vital sign monitoring equipment via a hardware interface. These physiological signals typically include, but are not limited to, photoplethysmography (PPG), invasive or non-invasive arterial pressure (ABP), and electrocardiogram (ECG) signals, which characterize the original driving pressure wave exerted by the heart's pumping function on the systemic vascular system.

[0033] Because data from different sources vary significantly in sampling frequency, physical dimensions, and noise characteristics, the system processor first performs data cleaning and standardization. For physiological signals, the system invokes the digital signal processing unit, employing cascaded bandpass filters (such as a fourth-order Butterworth filter) to remove power frequency interference, high-frequency electromyographic noise, and low-frequency baseline drift caused by respiratory movements. Although the processed signal retains its waveform characteristics, its absolute amplitude is significantly affected by environmental factors such as sensor contact impedance and circuit gain. Therefore, the system further performs Z-score standardization or max-min normalization to calculate the statistical distribution parameters of the signal within the current sampling period, mapping its amplitude to a standard dimensionless interval. This step eliminates the limitations of physical units, allowing the fluctuation pattern of physiological signals to be coupled with the grayscale changes of medical images at a purely mathematical level. Simultaneously, for medical image data, the system, based on pre-segmented regions of interest (ROIs), uses a weighted averaging algorithm to extract the average pixel grayscale value within each frame, constructing an average grayscale curve of the lesion region evolving over time. To suppress quantum noise in the imaging device, the system performs a smoothing filter on the curve while preserving its steep, enhanced rising edge characteristics.

[0034] After preparing the basic data, the system enters the core dynamic characteristic transformation stage. Static pressure values ​​are not the direct factor driving microcirculation perfusion; the rate of pressure change and the resulting acceleration are key physical quantities representing the blood supply characteristics of the lesion. Therefore, the system does not directly compare the original waveforms but instead extracts features in the higher-order derivative domain. The system's computational unit uses a high-precision numerical differentiation algorithm to calculate the first derivative of the normalized physiological signal with respect to time, and the second derivative of the average grayscale curve of the lesion region with respect to time. The first derivative of the physiological signal characterizes the instantaneous rate of change of the physiological driving pressure, i.e., the intensity of the driving pulse generated by cardiac ejection; while the second derivative of the lesion grayscale curve, under the premise that the grayscale value and contrast agent concentration have a linear relationship, characterizes the acceleration of contrast agent accumulation in the interstitial space of the lesion tissue. In highly vascularized lesions such as malignant tumors, the lack of a smooth muscle layer and an intact basement membrane in their neovascular network results in low vascular resistance and poor compliance. Therefore, the arrival of an external driving pulse often triggers an increase in contrast agent concentration, manifesting as a large second-derivative response. In contrast, normal tissues or benign lesions, with their well-developed vascular regulatory mechanisms, buffer against blood flow shocks, resulting in a relatively gentle acceleration response. Through this analysis, the system can distinguish between actively enhanced lesions and passively leaking backgrounds.

[0035] However, the physical transport time for blood pumped from the heart through large blood vessels to the microcirculation of distal lesions is significant. To address this causal time lag issue, the system incorporates a module for optimizing fluid dynamic lag time. Within a preset physiologically reasonable time window, the system constructs a sliding window cross-correlation analyzer to perform time-shifted scanning of the first derivative sequence of the physiological signal and the second derivative sequence of the lesion's grayscale. The system calculates the cross-correlation coefficient at different time shifts, and uses the time offset corresponding to the global peak of the cross-correlation function as the unique fluid dynamic lag time for that lesion. Subsequently, the system compensates for the time axis shift of the lesion's grayscale and its derivative sequences based on this parameter, achieving precise alignment of the driving and responding ends in the logical time phase, ensuring that subsequent coupling analysis is performed at the corresponding moment of the same physical event.

[0036] Based on this, the system performs a deep computation of the coupling efficiency index between hemodynamics and lesion perfusion derivative. This computation is designed as a composite mathematical model including energy integration, normalization, and phase locking. First, the system calculates the causal resonance energy numerator. The processor calculates the product of the first derivative of the physiological signal and the second derivative of the lesion grayscale after hysteresis compensation point by point, and integrates and accumulates the product value over the entire scan cycle. When the driving force pulse (peak of the first derivative) and the lesion perfusion acceleration (peak of the second derivative) occur synchronously in time and in the same direction, this integral term will obtain a significant positive gain, representing a strong causal driving relationship. Second, the system calculates the energy normalization denominator. To eliminate the influence of absolute signal intensity, the system calculates the self-energy (i.e., the square integral of the amplitude) of the driving signal and the response signal respectively, and uses the square root of their product as a normalization factor. This step ensures that the calculation result is a relative coefficient reflecting waveform similarity, rather than an absolute value limited by the contrast agent injection dose or cardiac output.

[0037] More importantly, to eliminate coincidental artifacts with similar waveforms but logical errors (such as respiratory artifacts or mechanical vibrations with frequencies close to heart rate), the system introduces a phase-locking constraint mechanism based on the Hilbert transform. The system transforms the original real-domain signal to the complex domain, extracting the instantaneous phase angle of the dominant frequency at each moment. In a real physically coupled system, a relatively constant phase difference (i.e., phase locking) should be maintained between the drive and the response. The system calculates the square of the difference in their phase angles and substitutes it into a Gaussian negative exponential decay function. This function includes a preset phase jitter tolerance variance parameter. When the phase difference fluctuates within the allowable range, the exponential term approaches 1, preserving the coupling coefficient; once the phase difference drifts drastically or loses lock, the exponential term rapidly approaches 0, forcibly suppressing the final coupling efficiency exponent.

[0038] The calculated coupling efficiency index is used as an adaptive gain coefficient by the system and directly applied to the temporal attention module of the deep learning model. For regions with high coupling indices, the system identifies them as high-confidence true blood supply lesions, amplifies the weights of their feature channels, and focuses the network on these areas. For regions with low coupling indices, the system identifies them as noise or artifacts and suppresses their weights. In this way, the system effectively injects causal prior knowledge from biophysics into the deep learning model, achieving a deep integration of data-driven and mechanistic models.

[0039] In this embodiment, after establishing the association between lesion features and anatomical location and pathological type, the method further includes performing a logical self-consistency verification step based on counterfactual topological residual analysis: Based on the current matched pathology type node as an assumption, retrieve the corresponding standard normal tissue anatomical manifold base from the medical pathology professional knowledge atlas; Using a spatial deformation mapping algorithm, the standard normal tissue anatomical manifold base is projected onto the coordinate system of the current medical image data to generate a virtual counterfactual health image; Pixel-level difference operations are performed on standardized medical image data and virtual counterfactual health images to extract the residual feature field of lesions; The Betti number of the residual feature field of the lesion is calculated using a topological data analysis algorithm to obtain the 0-dimensional Betti number. Characterization of connected components and 1-dimensional Betti numbers Characterized features of the porous structure; Determine whether the calculated Betty number feature falls within the predefined topological invariant interval of the pathological type node in the knowledge graph; if it falls within the interval, confirm the diagnostic logic is consistent and increase the confidence; if it does not fall within the interval, determine that the current feature is a false match caused by unstructured noise, and trigger the re-evaluation process of the suboptimal node.

[0040] After the diagnostic reasoning mapping module completes the initial feature matching and pathology type inference, in order to solve the problems of the deep learning model being uninterpretable and easily misled by texture features, the system introduces a logical self-consistency verification process based on counterfactual deduction and algebraic topology.

[0041] First, the system generates a preliminary diagnostic conclusion. Assuming the system initially determines the lesion to be of a specific pathological type (e.g., "invasive ductal carcinoma"), this conclusion is used as a hypothesis to be verified, and the system accesses a medical pathology knowledge graph. This graph stores a standard normal tissue anatomical manifold base constructed based on large-scale healthy population image data. It is a parameterized high-dimensional mathematical model describing the geometry, texture distribution, and topological features of the anatomical site in a healthy state. The system retrieves the corresponding manifold base based on the anatomical location information of the current image.

[0042] Subsequently, the system utilizes non-rigid registration algorithms such as Large Deformation Differential Homeomorphic Metric Mapping (LDDMM) to construct a deformation field from the standard manifold base to the patient's actual anatomical space. This algorithm uses normal anatomical landmarks in the patient's image as anchor points to drive elastic deformation of the standard base, causing it to coincide with the patient's organs in macroscopic contours, generating a virtual counterfactual healthy image. This image represents how the patient's anatomical structure would normally appear if they did not have the disease. Because the deformation algorithm is subject to topological preservation constraints, it does not generate abnormal structures of lesions out of thin air; therefore, the lesion area will be filled with a smooth, healthy texture in the virtual image.

[0043] The system performs pixel-level difference operations, subtracting the patient's real standardized medical image from a virtual counterfactual healthy image to extract the lesion residual feature field. This residual field eliminates interference from the background normal tissue while preserving the morphological differences caused by the lesion. To quantitatively analyze the structural characteristics of this residual field, the system abandons traditional Euclidean geometric metrics and instead applies topological data analysis (TDA) technology.

[0044] The system treats the residual field as a high-dimensional scalar field and performs persistent cohomology analysis by constructing simple complex sequences (such as Vietoris-Rips complexes) to calculate the Betti numbers of the residual field. The system focuses on calculating the 0-dimensional Betti numbers (representing the number of connected components, reflecting whether the lesion is solitary or multiple and scattered) and the 1-dimensional Betti numbers (representing the number of pores or ring structures, reflecting the presence of necrotic cavities, cystic changes, or ring enhancement within the lesion). The system records the generation and extinction processes of these topological features to generate a persistence graph.

[0045] Finally, the system compares the calculated real-time Betty number features with the predefined topological invariant intervals of the pathological type nodes in the knowledge graph. If the Betty number falls within the interval, it indicates that the lesion not only resembles the disease in texture but also conforms to the growth logic of the disease in deep topological structure. The system confirms the diagnostic consistency and increases the confidence level. If the Betty number falls outside the interval (for example, the system considers it a cavitary lesion but the calculated number of cavities is zero), it indicates a logical paradox. The system determines that the initial match is a false response, triggers a negative feedback mechanism, reduces the confidence level, and activates the re-evaluation process for suboptimal nodes. This closed-loop mechanism ensures the rigor and uniqueness of the diagnostic conclusion in terms of pathological logic.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based medical image diagnostic assistance system, characterized in that, include: The medical image data processing module is used to acquire medical image data and corresponding equipment parameters collected by various medical imaging devices, perform image processing on the acquired medical image data, and output standardized medical image data. The medical image feature fusion module is used to extract multi-dimensional features and generate calibration factors based on image categories of standardized medical image data, calculate the attention weights of each dimension feature based on a deep learning model, optimize the fusion by combining the calibration factors and attention weights, enhance the fused lesion features, and output an enhanced lesion feature map. The diagnostic reasoning mapping module is used to extract dynamic temporal data and corresponding physiological signals from standardized medical image data. Based on the periodic features of physiological signals, a temporal attention mechanism is constructed to capture the temporal change pattern of lesion features in the enhanced lesion feature map and generate temporally associated lesion features. The diagnostic reasoning mapping module is also used to call the medical pathology knowledge graph, match the temporal associated lesion features with the anatomical structure nodes and pathological type nodes in the knowledge graph, establish the association between lesion features and anatomical location and pathological type, and output the temporal lesion feature set and the list of lesion and pathological information association. The diagnostic result output module is used to generate diagnostic auxiliary conclusions from the temporal lesion feature set and the lesion-pathological information association list, extract lesion location coordinates based on the enhanced lesion feature map, mark the diagnostic results in combination with the confidence of the diagnostic auxiliary conclusions, integrate the diagnostic auxiliary conclusions, lesion location coordinates and confidence, and output a standardized diagnostic report. The step of performing similarity matching between temporally associated lesion features and anatomical structure nodes and pathological type nodes in the knowledge graph includes: The feature vectors of target nodes in the medical pathology knowledge graph are extracted, the features of temporally related lesions are processed by vector transformation, and the similarity value between the feature vectors of temporally related lesions and the feature vectors of each target node is calculated. Target nodes with similarity greater than a preset similarity threshold are selected, and a mapping relationship between temporally related lesion features and the selected target nodes is established to generate a list of lesion-pathological information associations. After establishing the correlation between lesion characteristics and anatomical location and pathological type, the process also includes a logical self-consistency verification step based on counterfactual topological residual analysis: Based on the current matched pathology type node as an assumption, retrieve the corresponding standard normal tissue anatomical manifold base from the medical pathology professional knowledge atlas; Using a spatial deformation mapping algorithm, the standard normal tissue anatomical manifold base is projected onto the coordinate system of the current medical image data to generate a virtual counterfactual health image; Pixel-level difference operations are performed on standardized medical image data and virtual counterfactual health images to extract the residual feature field of lesions; The Betti number of the residual feature field of the lesion is calculated using a topological data analysis algorithm to obtain the 0-dimensional Betti number. Characterization of connected components and 1-dimensional Betti numbers Characterized features of the porous structure; Determine whether the calculated Betty number feature falls within the predefined topological invariant interval of the pathological type node in the knowledge graph; if it falls within the interval, confirm the diagnostic logic is consistent and increase the confidence; if it does not fall within the interval, determine that the current feature is a false match caused by unstructured noise, and trigger the re-evaluation process of the suboptimal node.

2. The deep learning-based medical image diagnostic assistance system as described in claim 1, characterized in that, The process by which the medical image data processing module outputs standardized medical image data includes: Acquire medical image data and equipment parameters collected by various medical imaging devices, establish a mapping relationship between medical imaging devices and medical image data and equipment parameters, verify data integrity based on hash value, and remove redundant data; Based on the scanning resolution and exposure intensity in the equipment parameters, combined with the preset standard exposure intensity and standard resolution, the equipment difference calibration coefficient is calculated. Based on the equipment difference calibration coefficient, the gray value of each pixel in the medical image data is corrected to generate a medical image with gray balance adjustment. Extract target feature points from the medical image after grayscale balance adjustment, determine the edge contour coefficients of the target region and background region based on the target feature points, determine the noise type, and perform adaptive filtering denoising on the target region and background region respectively based on the noise type to generate a denoised medical image. The gray-level difference between the target region and the background region in the denoised medical image is obtained, the contrast adjustment coefficient is obtained, and the comprehensive contrast adjustment coefficient is obtained by combining the center gray value of the target region. Based on the comprehensive contrast adjustment coefficient, the distinction between lesions and normal tissues is enhanced, and a contrast-optimized medical image is generated. Based on the retrieved preset medical image standard resolution, the size of the contrast-optimized medical image is adjusted to generate a medical image with standardized resolution. The quality quantification index of the medical image after resolution standardization is calculated, the quality quantification index is compared with the preset quality threshold, the image anomaly coefficient is calculated based on the comparison result, and the quality verification result is generated. Based on the quality verification results, qualified images are converted into a standard format, embedded with equipment calibration parameters and processing logs, and output standardized medical image data.

3. The deep learning-based medical image diagnostic assistance system as described in claim 2, characterized in that, Data integrity is verified based on hash values, including: Establish a verification rule base, clarify the pass / fail thresholds for each indicator, and based on the verification rule base, verify the integrity of the data format and metadata, and determine the validity of the pixel value range. Calculate the target hash value of the medical image data, obtain the original hash value during transmission, and compare the target hash value with the original hash value bit by bit to verify the integrity of data transmission. Based on the label information of medical image data, the medical image data is classified, each subset is obtained, the consistency of data within each subset is calculated, and the confidence score of each subset is generated. Invalid data with a confidence level lower than the preset sub-data confidence level threshold is removed, and missing data is supplemented based on the similarity algorithm of adjacent data. The integrity of the supplemented data is verified, and a data integrity verification report is generated.

4. The deep learning-based medical image diagnostic assistance system as described in claim 1, characterized in that, The medical image feature fusion module includes: The calibration factor generation unit is used to match the corresponding feature extraction strategy based on the image category of standardized medical image data, extract the modal statistical features, data distribution features and anatomical structure features of each modality image based on the matching results, calculate the difference value of the same feature under different modalities, and generate the corresponding calibration factor according to the magnitude of the difference value and the influence weight of the feature on the diagnosis. The attention weight calculation unit is used to build a deep learning model. It uses the deep learning model to quantify the importance of each dimension feature in the lesion identification task and calculates the attention weight corresponding to each dimension feature based on the quantification evaluation results. The feature space alignment unit is used to combine calibration factors and attention weights to perform feature space calibration and alignment on multi-dimensional features across modalities, eliminate the distribution differences of cross-modal features, and obtain optimized features. The feature enhancement operator construction unit is used to load a preset lesion prior knowledge base, which includes the anatomical features, gray-scale distribution features and morphological features of the lesion. Based on the lesion prior knowledge base, key feature parameters of the lesion are extracted, a feature enhancement objective function is established, and a corresponding feature enhancement operator is constructed. The lesion feature enhancement and output unit is used to input the optimized features into the feature enhancement operator, enhance the expression intensity of the lesion region features, suppress noise interference in non-lesion regions, calculate the distinguishability between the enhanced lesion features and normal tissue features, and output the enhanced lesion feature map.

5. The deep learning-based medical image diagnostic assistance system as described in claim 4, characterized in that, Feature space calibration and alignment are performed on multi-dimensional features across modalities, specifically including: Obtain the original distribution range of multi-dimensional features of each modality, and normalize and calibrate the distribution range of multi-dimensional features of each modality based on the calibration factor; Obtain the attention weights of each dimension of the calibrated features, and then assign weighted distributions to each dimension of the calibrated features based on the attention weights. Retrieve preset unified feature space parameters and map the weighted multi-dimensional features of each modality to the unified feature space; calculate the spatial position deviation and semantic information matching degree of the multi-dimensional features of each modality after mapping. If the position deviation is less than the preset deviation threshold and the semantic matching degree is greater than the preset similarity value, then spatial alignment is completed. Calculate the cross-modal consistency index of the aligned features to obtain the optimized features.

6. The deep learning-based medical image diagnostic assistance system as described in claim 1, characterized in that, A temporal attention mechanism is constructed based on the periodic characteristics of physiological signals, specifically including: Extract physiological signals from dynamic time-series data, analyze the periodic characteristics of physiological signals, and determine the signal period parameters; A temporal attention weight matrix is ​​constructed based on signal period parameters, and differentiated attention weights are assigned to lesion features at different period stages to enhance lesion features with significant temporal changes. The temporal attention weight matrix is ​​subjected to matrix operation with the enhanced lesion feature map to calculate the temporal variation coefficient of each feature channel, highlighting the temporal dynamic differences of lesion features and generating temporally associated lesion features.

7. The deep learning-based medical image diagnostic assistance system as described in claim 1, characterized in that, The confidence level of the diagnostic auxiliary conclusion is obtained based on a multi-dimensional influencing factor system, which includes feature reliability factor, matching effectiveness factor, data quality factor, and model output consistency factor. The comprehensive confidence level is calculated based on the weight coefficient of each influencing factor and is used as the confidence level of the diagnostic auxiliary conclusion.

8. The deep learning-based medical image diagnostic assistance system as described in claim 6, characterized in that, The temporal attention mechanism constructed based on the periodic features of physiological signals specifically includes the following steps: The normalized temporal physiological signal intensity, the average gray-scale curve of the lesion area, and the calculated first derivative of the physiological signal, the second derivative of the gray-scale curve, the hydrodynamic lag time, and the instantaneous phase angle of the main frequency were obtained. Based on the normalized temporal physiological signal intensity, the average gray-level curve of the lesion region, and the calculated first derivative of the physiological signal, the second derivative of the gray-level curve, the hydrodynamic lag time, and the instantaneous phase angle of the dominant frequency, the hemodynamic-lesion perfusion derivative coupling efficiency index is calculated. The calculation formula is as follows: in, The derivative coupling efficiency index has a range of values. This characterizes the causal consistency between physiological driving forces and lesion responses; The normalized temporal physiological signal intensity, It is the first derivative of the physiological signal, representing the instantaneous rate of change of physiological driving pressure; To introduce fluid dynamic lag time The average grayscale curve of the lesion area after treatment. The second derivative of the average grayscale curve of the lesion area represents the perfusion acceleration of the contrast agent in the lesion tissue. This represents the total duration of the sampling period for time-series data. It is a small constant for regularization; and These represent the instantaneous phase angles of the dominant frequency of the physiological signal and the grayscale curve of the lesion in the Hilbert transform domain, respectively. For phase-locked constraint coefficients, The preset physiological cycle phase jitter tolerance variance; exponential term Used to forcibly suppress coupling index when two waveforms are similar but phase logic is completely unlocked; This refers to the time derivative; Based on the calculated hemodynamic-lesion perfusion derivative coupling efficiency index Adaptive gain adjustment is performed on the temporal attention weight matrix to suppress non-causal artifact interference and generate temporally associated lesion features that incorporate physiological causal constraints.

Images