Multi-model fusion pancreatic cystic lesion feature selection method and system
By employing a multi-model fusion approach, combining feature fusion and co-occurrence rule screening of pathology reports and biochemical indicators, and utilizing a two-stream neural network scoring method, the shortcomings of multimodal data processing in the diagnosis of pancreatic cystic lesions were addressed. This approach achieved highly reliable feature screening and coupling with pathological mechanisms, thereby improving the sensitivity and accuracy of diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN TUMOR HOSPITAL
- Filing Date
- 2026-05-13
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies suffer from multimodal data processing deficiencies in the diagnosis of pancreatic cystic lesions, especially IPMN, failing to achieve dynamic interactive fusion. This results in a disconnect between feature screening results and pathological mechanisms, low diagnostic sensitivity, and isolated feature selection, making it impossible to provide a highly reliable basis for precision diagnosis and treatment.
By employing a multi-model fusion approach, combining pathology report text and biochemical indicators for feature fusion and adjustment, and utilizing co-occurrence rules and co-occurrence intensity thresholds to screen features, a two-stream neural network is constructed for feature scoring analysis. This generates a multimodal coupled feature matrix and performs stability evaluation to ensure the clinical logical consistency of feature selection.
It significantly improves the sensitivity of diagnosis of pancreatic cystic lesion subtypes, provides a highly reliable basis for precision diagnosis and treatment, realizes multimodal deep coupling of feature screening and pathological mechanisms, and improves the accuracy and reliability of diagnosis.
Smart Images
Figure CN122175988B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for selecting features of pancreatic cystic lesions using multi-model fusion, and particularly to the field of medical feature selection technology. Background Technology
[0002] The accurate diagnosis of pancreatic cystic lesions, especially IPMN, faces serious limitations in multimodal data processing. Current technologies analyze pathology reports, imaging features, and biochemical indicators independently, creating data silos and hindering dynamic interactive fusion. Feature selection primarily relies on statistical correlation or general algorithm thresholds, lacking direct guidance from medical guidelines. This leads to a disconnect between selection results and the pathological mechanisms of IPMN, failing to reflect clinical co-occurrence patterns. For example, the synergistic relationship between imaging features and biochemical indicators has not been validated by medical rules, potentially resulting in low-correlation features and reduced diagnostic reliability. Furthermore, existing methods neglect the stability of multimodal features in clinical cases and fail to assess whether feature co-occurrence ratios conform to IPMN guidelines, leading to insufficient consistency of selected features in practical applications. These shortcomings collectively result in low sensitivity for IPMN subtype diagnosis, isolated feature selection, and an inability to provide highly reliable evidence for precise clinical diagnosis and treatment, severely limiting the improvement of diagnostic efficiency and accuracy. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for selecting features of pancreatic cystic lesions through multi-model fusion, which can significantly improve the sensitivity of subtype diagnosis and provide a highly reliable basis for precision diagnosis and treatment.
[0004] This application proposes a multi-model fusion method for feature selection of pancreatic cystic lesions, which includes:
[0005] S1: Perform fusion adjustment processing on at least one raw value of an image feature based on the pathology report text and biochemical indicators to determine the multimodal coupling feature matrix;
[0006] S2: Perform feature filtering on the multimodal coupling feature matrix according to at least one pre-determined co-occurrence rule and at least one co-occurrence intensity threshold to obtain a subset of medical guidance features;
[0007] S3: Perform feature scoring analysis and scoring calibration on the medical guidance feature subset to determine the importance score of the target feature;
[0008] S4: Based on the target feature importance score, perform feature screening and stability evaluation on the medical guidance feature subset to determine the target lesion feature set.
[0009] Preferably, S1 includes:
[0010] S11: Perform preset analysis and processing operations on the pathology report text and biochemical indicators to determine the associated instruction set;
[0011] S12: Perform fusion adjustment processing on at least one original image feature value using the associated instruction set to obtain a fused feature vector;
[0012] S13: Perform correlation calibration processing on at least one of the fused feature vectors to determine the multimodal coupling feature matrix.
[0013] Preferably, S1 includes:
[0014] S111: Perform semantic analysis and recognition on the pathology report text to obtain at least one pathology report keyword and determine the image segmentation parameters;
[0015] S112: Generate image feature weighting factors based on biochemical indicators and biochemical thresholds;
[0016] S113: Determine the associated instruction set based on the image segmentation parameters and the image feature weighting factors.
[0017] Preferably, S2 includes:
[0018] S21: Perform a preset semantic analysis and recognition process on the first tool guide document to determine at least one co-occurrence rule and at least one co-occurrence strength threshold;
[0019] S22: Perform feature filtering on the multimodal coupling feature matrix according to at least one of the co-occurrence rules and at least one of the co-occurrence intensity thresholds to obtain a subset of medical guidance features.
[0020] Preferably, S21 further includes:
[0021] S211: Perform a preset semantic analysis and recognition process on the first tool guide document to extract at least one co-occurrence rule;
[0022] S212: Determine at least one co-occurrence intensity threshold according to at least one of the co-occurrence rules.
[0023] Preferably, S22 further includes:
[0024] S221: Perform co-occurrence strength analysis on the multimodal coupling feature matrix according to at least one of the co-occurrence rules to determine at least one first co-occurrence strength;
[0025] S222: Perform feature filtering on the multimodal coupling feature matrix based on at least one of the first co-occurrence intensities and at least one of the co-occurrence intensity thresholds to obtain a subset of medical guidance features.
[0026] Preferably, S3 includes:
[0027] S31: Construct a two-stream neural network to perform feature scoring analysis on the medical guidance feature subset to determine feature importance scores;
[0028] S32: Perform a score calibration process on the feature importance score to determine the target feature importance score.
[0029] Preferably, S4 includes:
[0030] S41: Based on the target feature importance score, perform feature filtering on the medical guidance feature subset to determine the first feature subset;
[0031] S42: Perform stability evaluation processing on the first feature subset to determine the target lesion feature set.
[0032] This application also proposes a multi-model fusion-based pancreatic cystic lesion feature selection system to implement the aforementioned multi-model fusion-based pancreatic cystic lesion feature selection method, including:
[0033] Multimodal coupling feature matrix determination module: Performs fusion adjustment processing on at least one original image feature value based on pathology report text and biochemical indicators to determine the multimodal coupling feature matrix;
[0034] Medical guidance feature subset determination module: Performs feature filtering on the multimodal coupling feature matrix according to at least one pre-determined co-occurrence rule and at least one co-occurrence intensity threshold to obtain a medical guidance feature subset;
[0035] Target feature importance score determination module: Performs feature score analysis and score calibration processing on the medical guidance feature subset to determine the target feature importance score;
[0036] Target lesion feature set determination module: Based on the target feature importance score, perform feature screening and stability evaluation on the medical guidance feature subset to determine the target lesion feature set.
[0037] This invention proposes a multi-model fusion method and system for feature selection in pancreatic cystic lesions, belonging to the field of medical feature selection technology. This invention proposes a multi-modal feature fusion and screening method for the diagnosis of pancreatic cystic lesions. Based on a medical knowledge base, it dynamically fuses pathological imaging and biochemical data to generate a fully coupled feature vector. Multi-modal features are screened according to guideline co-occurrence rules to ensure clinical consistency. A two-stream neural network is used to calculate feature importance scores, combined with medical thresholds for feature screening. Furthermore, the stability of the IPMN clinical rules is statistically analyzed, and a threshold is set to determine the final feature set. Through the technical solution of this application, multi-modal deep coupling between feature screening and pathological mechanisms is achieved, significantly improving the sensitivity of subtype diagnosis and providing a highly reliable basis for precision diagnosis and treatment. Attached Figure Description
[0038] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0039] Figure 1 This is an execution flowchart of the multi-model fusion method for selecting features of pancreatic cystic lesions in this application.
[0040] Figure 2 This is a flowchart illustrating the determination of the associated instruction set in this application. Detailed Implementation
[0041] 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.
[0042] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions are only used to explain the present invention and are not intended to limit the present invention.
[0043] The following provides a detailed description of the multi-model fusion method and system for selecting features of pancreatic cystic lesions according to the present invention.
[0044] This embodiment proposes a multi-model fusion method for feature selection of pancreatic cystic lesions, the specific process of which is as follows: Figure 1 As shown.
[0045] S1: Perform fusion adjustment processing on at least one raw value of an image feature based on the pathology report text and biochemical indicators to determine the multimodal coupling feature matrix.
[0046] In this step, a medical knowledge base is constructed, and pathology report keywords are used to trigger dynamic optimization of image segmentation parameters in real time. Simultaneously, biochemical thresholds are converted into feature calculation weight factors to generate a set of associated instructions. A fused feature vector is generated through closed-loop calculation of pathological intensity, image features, and biochemical weights, and semantic relevance calibration is performed. The resulting feature matrix, dynamically coupled across the entire pathology-image-biochemistry process, can eliminate data silos to a certain extent.
[0047] S1 specifically includes the following sub-steps:
[0048] S11: Perform preset analysis and processing operations on pathology report text and biochemical indicators to determine the associated instruction set.
[0049] In this step, a medical knowledge base is constructed based on the "Guidelines for the Diagnosis and Treatment of Pancreatic Cystic Lesions (2023)". Pathology report keywords, such as "papillary", are used as instructions to trigger dynamic adjustments to image segmentation parameters. Simultaneously, biochemical thresholds are converted into image feature calculation weights, such as CA19-9 > 37 U / mL. Through these steps, real-time interaction between pathological descriptions, biochemical indicators, and image processing is achieved. Specifically, the intensity of pathological keywords determines the image segmentation sensitivity, and the size of the biochemical threshold affects the feature scaling coefficient. Both are linked through the knowledge base to generate a set of associated instructions.
[0050] The input data to be processed in this step are CT / MRI image sequences, CA19-9 / CEA biochemical values, and pathology report text.
[0051] The preset analysis and processing operations include semantic analysis and recognition, and threshold comparison analysis and processing.
[0052] Step S11 includes the following sub-steps:
[0053] S111: Perform semantic analysis and recognition on the pathology report text to obtain at least one pathology report keyword and determine the image segmentation parameters.
[0054] In this step, the pathology report text is analyzed and identified using a semantic analysis and recognition model commonly used in this field to determine at least one pathology report keyword. This keyword is then used as the basis for dynamic adjustment of subsequent image segmentation parameters.
[0055] In one example, the pathology report of patient 001 was described as "papillary protrusions seen in cyst fluid cytology". The system parsed the keyword "papillary" and triggered the image segmentation parameters. Through semantic analysis, it was found that the keyword strength was 0.9: the edge detection strength of the cyst segmentation area was enhanced, and the original parameter was optimized to 0.9.
[0056] S112: Generate image feature weighting factors based on biochemical indicators and biochemical thresholds.
[0057] In this step, biochemical indicators are judged based on pre-set biochemical thresholds to generate image feature weighting factors.
[0058] Biochemical thresholds are defined as follows:
[0059] Clinical thresholds for biochemical indicators are pre-defined in the guidelines. For example, CA19-9 > 37 U / mL is a high-risk threshold; CEA > 5 ng / mL is a high-risk threshold.
[0060] Biochemical threshold determination:
[0061] Comparison of biochemical values with biochemical thresholds:
[0062] If the biochemical value is less than or equal to the biochemical threshold, then the weighting factor is 1.0, with no scaling.
[0063] If the biochemical value is greater than the biochemical threshold, the weighting factor is 1.5, and high-risk scaling is triggered based on the clinical classification in the guidelines.
[0064] Basis for generation:
[0065] The guidelines clearly state that CA19-9 > 37 U / mL is strongly correlated with the malignancy risk of IPMN, with a correlation coefficient > 0.85, therefore a fixed scaling factor of 1.5 is set.
[0066] In another preferred embodiment, a stepped weighting factor can be determined based on different biochemical threshold ranges to clarify the differences between different risk levels.
[0067] S113: Determine the associated instruction set based on the image segmentation parameters and the image feature weighting factors.
[0068] In this step, the image segmentation parameters and the image feature weighting factors are fused to achieve real-time interactive fusion of pathological description, biochemical indicators and image processing.
[0069] Specifically, the edge detection intensity of the enhanced cyst segmentation region is adjusted in real time using the image segmentation parameters, and the real-time scaling coefficient is determined using the image feature weighting factor. Based on this set of related instructions, the adjustment and control of image data can better meet actual needs for detection and recognition.
[0070] Specific examples:
[0071] Patient 001's pathology report described "papillary protrusions seen in cyst fluid cytology," and the CA19-9 value was 42 U / mL.
[0072] The associated instruction set output is: ["Pathological trigger: papillary → segmentation parameter × 90%", "Biochemical weight: CA19-9=42 → scaling factor 1.5"].
[0073] The associated instruction set will serve as the input to S12, ensuring real-time coupling between image processing and pathological and biochemical data.
[0074] S12: Perform fusion adjustment processing on at least one original image feature value using the associated instruction set to obtain a fused feature vector.
[0075] In this step, the image segmentation parameters and the image feature weighting factors are fused into the image feature calculation using the associated instruction set.
[0076] Specifically, the image segmentation parameters are multiplied by the original image feature values, and then the image feature weighting factors are used to perform a second scaling to generate a fused feature vector, rather than simply splicing them together.
[0077] This ensures that the generated fused feature vector is an organic entity capable of expressing multimodal features.
[0078] The fusion feature vector output by this step corresponds to one patient. By repeating this step, fusion feature vectors corresponding to multiple patients can be obtained.
[0079] Specific examples:
[0080] If the original value of the image segment density index of patient 001 is 1.0, and the original value of the calcification distribution entropy index is 0.62.
[0081] The formation process of fusion features is explained below using the image segmentation density index as an example:
[0082] Step 1: Pathological intensity weighted at 0.9, 1.0 × 0.9 = 0.9.
[0083] Step 2: Scale the biochemical weights by 1.5, 0.9 × 1.5 = 1.35.
[0084] The fused feature vector output is: ["Separation density index_fusion=1.35", "Calcification point distribution entropy_fusion=0.62×0.9×1.5=0.84"].
[0085] S13: Perform correlation calibration processing on at least one of the fused feature vectors to determine the multimodal coupling feature matrix.
[0086] In this step, the UMLS medical knowledge graph is invoked to perform semantic calibration on the fusion feature vector generated in S12.
[0087] Specifically, by calling the UMLS medical knowledge graph, the correlation between each feature in the fused feature vector and the malignancy risk of IPMN is queried. For example, the correlation of "Separation Density Index_Fusion=1.35" in the UMLS medical knowledge graph is 0.88. According to the preset judgment conditions, if the correlation is >0.85, it is judged as having a high risk of malignancy, and a medical prior feature, such as "high risk of IPMN malignancy", can be generated to further ensure that the feature conforms to clinical logic.
[0088] By performing semantic calibration on each fusion feature vector in S12, a multimodal coupling feature matrix can be obtained. In this matrix, each row represents a patient, and each column represents a medical detection feature. Therefore, the multimodal coupling feature matrix can characterize the medical fusion features of multiple patients, where each feature reflects the deep coupling of imaging, biochemistry, and pathology.
[0089] Specific examples:
[0090] Patient 001's fusion feature septal density index_fusion=1.35:
[0091] The correlation between the UMLS query "Separation Density Index_Fusion" and IPMN is 0.88, which is greater than the preset 0.85.
[0092] This can generate a medical prior feature: ["IPMN has a high risk of malignancy"];
[0093] Merging all features: The patient 001 feature term in the multimodal coupling feature matrix contains a septation density index_fusion=1.35, indicating a high risk of malignancy for IPMN.
[0094] S2: Perform feature filtering on the multimodal coupling feature matrix according to at least one pre-determined co-occurrence rule and at least one co-occurrence intensity threshold to obtain a subset of medical guidance features.
[0095] In this step, the clinical co-occurrence rules of IPMN are extracted according to the "Guidelines for Pancreatic Imaging", such as "segmentation density index_fusion>1.0 and CA19-9 risk>1.0→malignancy risk". The proportion of co-occurrence of multimodal features in the multimodal coupling feature matrix is statistically analyzed and the co-occurrence intensity is calculated. Features are dynamically screened using medical rule thresholds, and only coupling feature combinations that conform to the clinical logic of IPMN are retained.
[0096] S2 specifically includes the following sub-steps:
[0097] S21: Perform a preset semantic analysis and recognition process on the first tool guide document to determine at least one co-occurrence rule and at least one co-occurrence strength threshold.
[0098] In this step, clinical co-occurrence rules need to be determined through medical guidelines as screening criteria for subsequent co-occurrence relevance features.
[0099] S21 includes the following sub-steps:
[0100] S211: Perform a preset semantic analysis and recognition process on the first tool guide document to extract at least one co-occurrence rule.
[0101] Extract at least one co-occurrence rule from the "Guidelines for Pancreatic Imaging".
[0102] Taking the IPMN-specific co-occurrence rule as an example, a separation density index_fusion > 1.0 and a CA19-9 risk_fusion > 1.0 indicate a malignant risk of IPMN.
[0103] Rule basis: The guidelines clearly state that "a high separation density index and a high CA19-9 jointly indicate a high risk of malignancy".
[0104] This step uses semantic recognition tools such as large language models to perform semantic analysis and extraction on the "Guidelines for Pancreatic Imaging" to determine at least one of the aforementioned co-occurrence rules.
[0105] S212: Determine at least one co-occurrence intensity threshold according to at least one of the co-occurrence rules.
[0106] Based on medical consensus guidelines, a co-occurrence intensity threshold is set for each of the co-occurrence rules. The co-occurrence intensity threshold refers to the percentage of cases in a patient dataset that satisfy a specific co-occurrence rule.
[0107] The co-occurrence intensity threshold can be set by expert experience or learned by a machine learning model.
[0108] The machine learning model can be a deep convolutional neural network model. In the training sample data, co-occurrence feature pairs can be used as input data, and manually labeled co-occurrence intensity thresholds can be used as output data. Then, during the prediction and setting of dynamic thresholds using the model, the corresponding dynamic threshold can be determined by the similarity between the co-occurrence rules and the input data in the sample data.
[0109] S22: Perform feature filtering on the multimodal coupling feature matrix according to at least one of the co-occurrence rules and at least one of the co-occurrence intensity thresholds to obtain a subset of medical guidance features.
[0110] In this step, the multimodal coupling feature matrix needs to be filtered according to at least one of the co-occurrence rules and at least one of the co-occurrence intensity thresholds to select features with high representativeness.
[0111] S22 includes the following sub-steps:
[0112] S221: Perform co-occurrence strength analysis on the multimodal coupling feature matrix according to at least one of the co-occurrence rules to determine at least one first co-occurrence strength.
[0113] In this step, all patient cases are traversed in the multimodal coupling feature matrix, and the proportion of cases satisfying the rules is counted. Each of the first co-occurrence intensities corresponds one-to-one with one of the co-occurrence rules.
[0114] Example as follows:
[0115] For patient 001, the septal density index_fusion = 1.35 > 1.0, and the CA19-9 risk_fusion = 1.5 > 1.0, so the judgment result is that it meets the co-occurrence rule.
[0116] For patient 002, the septal density index_fusion = 0.85 ≤ 1.0, so the result is that it does not meet the co-occurrence rule.
[0117] Assuming the dataset contains data from 500 patients, and 400 patients' data satisfy the image-biochemical co-occurrence rule, then the image-biochemical co-occurrence intensity can be calculated as: number of cases satisfying the rule / total number of cases = 400 / 500 = 0.8.
[0118] Output: Co-occurrence correlation strength matrix. For example, ["Imaging-Biochemistry co-occurrence strength=0.8", "Imaging-Pathology co-occurrence strength=0.55"].
[0119] S222: Perform feature filtering on the multimodal coupling feature matrix based on at least one of the first co-occurrence intensities and at least one of the co-occurrence intensity thresholds to obtain a subset of medical guidance features.
[0120] In this step, at least one of the co-occurrence intensity thresholds is compared with the corresponding first co-occurrence intensity to determine whether to filter the corresponding co-occurrence feature pairs.
[0121] Example as follows:
[0122] If the image-biochemical co-occurrence intensity = 0.8 > threshold 0.75, then retain the segmentation density index_fusion and CA19-9 risk_fusion;
[0123] Elimination: If the co-occurrence intensity of imaging and pathology is 0.55 < threshold 0.6, then features such as pathological risk and fusion will be eliminated.
[0124] After this step, the medical guidance feature subset can be output. The medical guidance feature subset is in matrix form, where each row represents a patient and each column represents a medical diagnostic feature.
[0125] S3: Perform feature scoring analysis and score calibration processing on the medical guidance feature subset to determine the importance score of the target feature.
[0126] In this step, a pancreas-specific dual-stream neural network is constructed to further filter the medical guidance feature subset.
[0127] S3 specifically includes the following sub-steps:
[0128] S31: Construct a two-stream neural network to perform feature scoring analysis on the medical guidance feature subset to determine feature importance scores.
[0129] In this step, the dual-stream neural network includes an imaging stream and a biochemical stream. The imaging stream uses 3D ResNet-18, specifically optimized for the 3D structure of pancreatic cysts, to handle the continuity of cyst wall segments. The biochemical stream initializes weights based on guideline thresholds and embeds pancreatic clinical rules in a shared attention layer.
[0130] The specific structure of the two-stream neural network is as follows:
[0131] Image stream: Input specific 3D features of pancreatic cystic lesions, such as cyst septum density index and calcification distribution entropy, and process them using a 3D ResNet-18 backbone network. This network is specifically optimized for the 3D spatial structure of pancreatic CT / MRI, avoiding the loss of pancreatic anatomical details in 2D processing.
[0132] Biochemical flow: Input IPMN-related biochemical indicators, such as CA19-9 risk and CEA risk, and use medical threshold-aware MLP. Its weight initialization is based on the threshold rules of the "Guidelines for the Diagnosis and Treatment of Pancreatic Tumors". For example, CA19-9 > 37 U / mL corresponds to high weight.
[0133] Shared Attention Layer: A pancreatic lesion-driven interaction mechanism that dynamically calculates image-biochemical feature weights. Specifically, when the CA19-9 risk output by the biochemical stream is >1.0, the weight of the "segmentation density index" feature in the image stream is automatically increased. This layer uses pancreatic clinical association rules as a prior, rather than a general attention mechanism.
[0134] This step will output the feature importance score, which directly reflects the IPMN-related interactions.
[0135] Specific examples:
[0136] Patient 001 data input:
[0137] Image stream input: Segmentation density index_fusion = 1.35, which is a typical feature of pancreatic IPMN.
[0138] Biochemical input: CA19-9 risk = 1.5, which is a high-risk marker for IPMN.
[0139] Shared attention layer interaction: Because CA19-9 risk = 1.5 > 1.0, the pancreatic guideline rule "increase the weight of the septal density index when CA19-9 is high" is triggered, and the weight increment is calculated by +0.2.
[0140] Feature importance score output: [Separation density index_fusion, CA19-9 risk].
[0141] S32: Perform a score calibration process on the feature importance score to determine the target feature importance score.
[0142] In this step, the UMLS knowledge graph is invoked to calibrate the feature importance scores. Specifically, the UMLS correlation score is queried for each feature. For example, the IPMN correlation score for "Separation Density Index_Fusion" is 0.88. If the correlation score is greater than 0.88, the original score is retained; otherwise, the score is reduced. For example, if the correlation score is 0.75, the score is reduced from 0.25 to 0.15, ensuring that the algorithm output is consistent with medical consensus.
[0143] S4: Based on the target feature importance score, perform feature screening and stability evaluation on the medical guidance feature subset to determine the target lesion feature set.
[0144] In this step, the medical guidance feature subset needs to be further screened based on the target feature importance score, and the stability of the screened features needs to be evaluated to determine the final target lesion feature set.
[0145] S4 specifically includes the following sub-steps:
[0146] S41: Perform feature filtering on the medical guidance feature subset based on the target feature importance score to determine the first feature subset.
[0147] In this step, the target feature importance score output by S3 is used to set a medical consensus threshold, such as based on the "Guidelines for Pancreatic Imaging": a feature importance score >0.20 represents clinical significance. The subset of medically guided features output by S2 is then filtered. Only features with a score >0.20 are retained to ensure that the screening results are deeply coupled with the pathological mechanism of IPMN, rather than relying on general statistical thresholds.
[0148] Specific examples:
[0149] Separation density index - fusion 0.25 retained (0.25>0.20)
[0150] CA19-9 Risk_Merge 0.28 (0.28>0.20)
[0151] Calcification point distribution entropy - fusion 0.15 removal (0.15 < 0.20)
[0152] Bile duct dilatation index - fusion ≥ 0.18 (0.18 < 0.20) - removed.
[0153] Filtering results: This step outputs the first feature subset, retaining only the features with high importance.
[0154] S42: Perform stability evaluation processing on the first feature subset to determine the target lesion feature set.
[0155] In this step, for the first feature subset, the proportion of cases meeting the clinical co-occurrence rules defined in the IPMN guidelines is calculated, such as "Separation Density Index_Fusion > 1.0 and CA19-9 Risk_Fusion > 1.0". Coupling stability is calculated as: Coupling stability = Number of cases meeting the rule / Total number of cases. A stability threshold is set; if stability > the threshold, the feature is retained.
[0156] This application also proposes a multi-model fusion pancreatic cystic lesion feature selection system to implement the above-mentioned multi-model fusion pancreatic cystic lesion feature selection method.
[0157] This invention proposes a multi-model fusion method and system for feature selection in pancreatic cystic lesions, belonging to the field of medical feature selection technology. This invention proposes a multi-modal feature fusion and screening method for the diagnosis of pancreatic cystic lesions. Based on a medical knowledge base, it dynamically fuses pathological imaging and biochemical data to generate a fully coupled feature vector. Multi-modal features are screened according to guideline co-occurrence rules to ensure clinical consistency. A two-stream neural network is used to calculate feature importance scores, combined with medical thresholds for feature screening. Furthermore, the stability of the IPMN clinical rules is statistically analyzed, and a threshold is set to determine the final feature set. Through the technical solution of this application, multi-modal deep coupling between feature screening and pathological mechanisms is achieved, significantly improving the sensitivity of subtype diagnosis and providing a highly reliable basis for precision diagnosis and treatment.
[0158] The above description is only a preferred embodiment of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the claims of this patent application are included in the scope of this patent application.
Claims
1. A method for selecting features of pancreatic cystic lesions using multi-model fusion, characterized in that, The method includes: S1: Perform fusion adjustment processing on at least one raw value of an image feature based on the pathology report text and biochemical indicators to determine the multimodal coupling feature matrix; S2: Perform feature filtering on the multimodal coupling feature matrix according to at least one pre-determined co-occurrence rule and at least one co-occurrence intensity threshold to obtain a subset of medical guidance features; S3: Perform feature scoring analysis and scoring calibration on the medical guidance feature subset to determine the importance score of the target feature; S4: Based on the target feature importance score, perform feature screening and stability evaluation on the medical guidance feature subset to determine the target lesion feature set; S1 includes: S11: Perform preset analysis and processing operations on the pathology report text and biochemical indicators to determine the associated instruction set; S11 includes: S111: Perform semantic analysis and recognition on the pathology report text to obtain at least one pathology report keyword and determine the image segmentation parameters; S112: Generate image feature weighting factors based on biochemical indicators and biochemical thresholds; S113: Determine the associated instruction set based on the image segmentation parameters and the image feature weighting factors; S12: Perform fusion adjustment processing on at least one original image feature value using the associated instruction set to obtain a fused feature vector; S12 includes: The image segmentation parameters are multiplied by the original image feature value, and then the image feature weighting factor is used to perform a second scaling to generate a fused feature vector. S13: Perform correlation calibration processing on at least one of the fused feature vectors to determine the multimodal coupling feature matrix.
2. The method for selecting features of pancreatic cystic lesions by multi-model fusion according to claim 1, characterized in that, S2 includes: S21: Perform a preset semantic analysis and recognition process on the first tool guide document to determine at least one co-occurrence rule and at least one co-occurrence strength threshold; S22: Perform feature filtering on the multimodal coupling feature matrix according to at least one of the co-occurrence rules and at least one of the co-occurrence intensity thresholds to obtain a subset of medical guidance features.
3. The method for selecting features of pancreatic cystic lesions by multi-model fusion according to claim 2, characterized in that, S21 further includes: S211: Perform a preset semantic analysis and recognition process on the first tool guide document to extract at least one co-occurrence rule; S212: Determine at least one co-occurrence intensity threshold according to at least one of the co-occurrence rules.
4. The method for selecting features of pancreatic cystic lesions by multi-model fusion according to claim 3, characterized in that, S22 further includes: S221: Perform co-occurrence strength analysis on the multimodal coupling feature matrix according to at least one of the co-occurrence rules to determine at least one first co-occurrence strength; S222: Perform feature filtering on the multimodal coupling feature matrix based on at least one of the first co-occurrence intensities and at least one of the co-occurrence intensity thresholds to obtain a subset of medical guidance features.
5. The method for selecting features of pancreatic cystic lesions by multi-model fusion according to claim 1, characterized in that, S3 includes: S31: Construct a two-stream neural network to perform feature scoring analysis on the medical guidance feature subset to determine feature importance scores; S32: Perform a score calibration process on the feature importance score to determine the target feature importance score; The dual-stream neural network includes an image stream and a biochemical stream. The image stream uses 3D ResNet-18 to optimize the 3D structure of pancreatic cysts to handle the continuity of cyst wall segments. The biochemical stream initializes weights based on guideline thresholds and embeds pancreatic clinical rules in a shared attention layer.
6. The method for selecting features of pancreatic cystic lesions by multi-model fusion according to claim 1, characterized in that, S4 includes: S41: Based on the target feature importance score, perform feature filtering on the medical guidance feature subset to determine the first feature subset; S42: Perform stability evaluation processing on the first feature subset to determine the target lesion feature set.
7. A multi-model fusion system for selecting features of pancreatic cystic lesions, used to implement the multi-model fusion method for selecting features of pancreatic cystic lesions according to any one of claims 1-6, comprising: Multimodal coupling feature matrix determination module: Performs fusion adjustment processing on at least one original image feature value based on pathology report text and biochemical indicators to determine the multimodal coupling feature matrix; Medical guidance feature subset determination module: Performs feature filtering on the multimodal coupling feature matrix according to at least one pre-determined co-occurrence rule and at least one co-occurrence intensity threshold to obtain a medical guidance feature subset; Target feature importance score determination module: Performs feature score analysis and score calibration processing on the medical guidance feature subset to determine the target feature importance score; Target lesion feature set determination module: Based on the target feature importance score, perform feature screening and stability evaluation on the medical guidance feature subset to determine the target lesion feature set.