A method and system for processing MRI data for evaluating treatment efficacy before and after treatment

By standardizing fields and using evidence association networks, the problem of linking multiple MRI examination records for the same patient was solved, enabling automated efficacy evaluation and improving the accuracy and traceability of efficacy evaluation.

CN123072771APending Publication Date: 2026-11-10XIEHE HOSPITAL ATTACHED TO TONGJI MEDICAL COLLEGE HUAZHONG SCI & TECH UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611577895.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately link multiple MRI examination records of the same patient, leading to inaccurate efficacy evaluation results that are difficult to process automatically and cannot be effectively traced.

Method used

By standardizing fields and calculating multi-field similarity, an evidence association network is constructed to transfer and project evidence strength. Combined with examination time, location, and report completeness, the efficacy category is determined, thus achieving automated efficacy evaluation.

Benefits of technology

This improved the standardization and automation of efficacy evaluation data processing, reduced the burden of manual verification, lowered the risk of misjudgment, and ensured the accuracy and traceability of efficacy evaluation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072771A_ABST
    Figure CN123072771A_ABST
Patent Text Reader

Abstract

The application provides a treatment before and after MRI treatment effect evaluation data processing method and system, relates to the treatment effect evaluation data processing field, and includes the following steps: obtaining patient identity records and MRI report records, processing, deduplicating, and correlating scores to obtain baseline-reexamination candidate pairs and their pairing scores; analyzing text content to obtain multi-report completeness; constructing an evidence correlation network and performing evidence strength transmission to obtain transmitted positive evidence strength and negative evidence strength; comparing the transmitted positive evidence strength and negative evidence strength to obtain the maximum value, and obtaining projected positive evidence strength and negative evidence strength; performing treatment effect category determination on the evidence nodes respectively, and determining a candidate treatment effect set according to a preset category priority; and obtaining strictness determination results based on entity alignment scores, pairing scores, report completeness, and the candidate treatment effect set. The application improves the standardization and automation degree of treatment effect evaluation data processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of efficacy evaluation data processing technology, and in particular to a method and system for processing MRI efficacy evaluation data before and after treatment. Background Technology

[0002] Cancer patients typically require multiple MRI scans before and after treatment. Doctors evaluate treatment efficacy by comparing the baseline examination report before treatment with the follow-up examination report after treatment, providing a basis for adjusting subsequent treatment plans. However, the patient's original records and MRI reports often come from multiple different information systems or medical institutions. The names, coding formats, and filling standards of patient identification fields differ in records from different sources. The same patient may correspond to different identification information in different systems, making it difficult to accurately link records under the same patient's name and to determine which records belong to the same patient and which reports are duplicates.

[0003] Furthermore, a single examination may generate multiple versions of the report due to different stages of the process, such as drafting, supplementation, or signing. Deduplication based solely on examination time or document name cannot guarantee that the final and complete version is retained. Multiple examinations of the same patient do not necessarily constitute a comparable pre- and post-treatment pairing; a comprehensive judgment is needed considering factors such as the order of examinations and whether the examination sites are consistent. The presentation of lesion measurement results in the report text varies, and inconsistencies or even contradictions may exist between conclusions regarding target lesions, non-target lesions, and newly formed lesions. Relying solely on manual reading and comparison of each report for efficacy judgment is inefficient and prone to omissions or misjudgments, failing to meet the standardized and automated needs of clinical efficacy evaluation data processing.

[0004] Chinese patent application No. 201910427251.0 discloses a patient identity matching method based on multi-field similarity calculation. This method preprocesses the field values ​​and calculates string similarity for multiple information items within the patient identity information framework, then weights and sums the similarities of each information item to obtain the patient record similarity, and compares it with a preset threshold to determine whether two patient records point to the same patient. However, it only processes and compares patient record similarity at a single stage, and cannot combine factors such as examination time and examination site to compare multiple examination records under the same patient's name before and after treatment, thus failing to trace the efficacy evaluation results. Summary of the Invention

[0005] In view of this, this application provides a method and system for processing MRI efficacy evaluation data before and after treatment, which solves the technical problems in the prior art of lacking a comparable relationship determination for multiple examination records under the same patient before and after treatment, and making it difficult to provide verifiable and traceable efficacy evaluation results, thereby improving the standardization and automation of efficacy evaluation data processing.

[0006] The technical solution of this application is implemented as follows: On the one hand, this application provides a method for processing MRI efficacy evaluation data before and after treatment, including: S1. Obtain patient identity records and MRI report records, perform field standardization processing and multi-field similarity calculation to obtain entity alignment scores, remove duplicate reports under the same patient name based on entity alignment scores to obtain valid MRI report records, and perform correlation scoring on valid MRI report records based on examination time, examination site and entity alignment scores to obtain baseline-re-examination candidate pairs and their pairing scores. S2. Analyze the text content of each valid MRI report in the baseline-review candidate pair to obtain multiple evidence nodes and the report completeness of each evidence node's report; S3. Using entity alignment score, pairing score and report completeness as reference nodes, construct an evidence association network containing reference nodes and evidence nodes, and transfer evidence strength on the evidence association network to obtain the positive evidence strength and negative evidence strength after transfer. S4. Compare the transmitted positive evidence strength and negative evidence strength with the rule positive evidence strength and rule negative evidence strength respectively, and take the maximum value to obtain the projected positive evidence strength and negative evidence strength. S5. Based on the projected positive and negative evidence strength, determine the efficacy category for each evidence node and determine the candidate efficacy set according to the preset category priority. S6. Based on entity alignment score, pairing score, report completeness, and candidate efficacy set, obtain the rigor judgment result; if the candidate efficacy set contains only one efficacy category and the rigor judgment result meets the preset conditions, output the efficacy category as the efficacy evaluation result of the baseline-review candidate pair; otherwise, output the candidate efficacy set as the efficacy evaluation result to be reviewed.

[0007] Based on the above technical solutions, preferably, step S1 specifically includes: Retrieve the fields of name, medical record number, patient card number, standardized name, date of birth and gender from the patient identity record, as well as the patient identification field from the MRI report record, and convert the format of the fields to obtain standardized fields; The similarity between corresponding fields in the patient identity record and the MRI report record is calculated based on the normalized fields, and the similarity of multiple fields is summed to obtain the entity alignment score. Based on the entity alignment score, determine whether there are multiple report records under the same patient name pointing to the same examination. If so, retain the one with the highest information completeness and delete the rest of the report records to obtain the valid MRI report record.

[0008] Based on the above technical solutions, preferably, step S2 specifically includes: The text content of each valid MRI report in the baseline-review candidate pair was analyzed, and the lesion descriptions obtained from the analysis were classified into target lesion evidence nodes, non-target lesion evidence nodes, and new lesion evidence nodes according to the anatomical region. To determine the completeness of a MRI report, one must determine whether both the sections seen and the conclusions are present simultaneously in the valid MRI report. Determine whether the number of located anatomical regions in valid MRI report records reaches a preset number to obtain the region integrity. Determine whether the number of evidence nodes in a valid MRI report reaches a preset number to obtain the evidence completeness. The report completeness of each evidence node is obtained by summing the chapter completeness, region completeness, and evidence completeness.

[0009] Based on the above technical solutions, preferably, step S3 specifically includes: S31. Using entity alignment score, pairing score and report completeness as reference nodes, and target lesion evidence nodes, non-target lesion evidence nodes and new lesion evidence nodes as evidence nodes, establish connection edges between reference nodes and evidence nodes, and between evidence nodes to obtain the evidence association network. S32. Evidence strength is transferred along the connecting edges in the evidence association network. The features of adjacent nodes are aggregated to obtain the positive and negative evidence strengths after transfer.

[0010] Based on the above technical solutions, preferably, step S31 specifically includes: Determine whether the anatomical regions between each pair of evidence nodes are the same to obtain the same region judgment result; Determine whether the node identifiers between any two evidence nodes point to the same lesion origin to obtain the homology judgment result; Determine whether each pair of evidence nodes belongs to different time points and has the same evidence predicate to obtain the cross-time same predicate judgment result; If any one of the following is true: same region judgment result, same source judgment result, or cross-time same predicate judgment result, then establish a connection edge between the corresponding two evidence nodes; A bidirectional connection edge is established between the reference node and the evidence node to obtain the evidence association network.

[0011] Based on the above technical solutions, preferably, step S32 specifically includes: Transform the features of the nodes at both ends of the connecting edge to obtain the structural priors; The reliability term is obtained by weighting the nodes at both ends of the connecting edge based on their reliability. Logarithmic transformation is performed based on the reliability of the connected edges to obtain the edge reliability term; The structural prior terms, reliability terms, and edge reliability terms are added together and normalized within the constraints of the connecting edges to obtain the weights of the weighted aggregation. Based on the weights of the weighted aggregation, the features of adjacent nodes are weighted and aggregated to obtain the positive and negative evidence strengths after propagation.

[0012] Based on the above technical solutions, preferably, step S4 specifically includes: The positive evidence strength after transmission is compared with the baseline positive evidence strength, and the maximum value is taken to obtain the projected positive evidence strength. The negative evidence strength after transmission is compared with the baseline negative evidence strength, and the maximum value is taken to obtain the projected negative evidence strength. An uncertainty flag is obtained based on whether the difference between the projected positive evidence strength and the projected negative evidence strength is less than a preset threshold. A conflict flag is obtained based on whether the strength of the projected positive evidence and the strength of the projected negative evidence simultaneously reach a preset threshold.

[0013] Based on the above technical solutions, preferably, step S5 specifically includes: S51. Efficacy categories include complete remission (CR), partial remission (PR), stable disease (SD) or progressive disease (PD), undefined category, and conflict category. The efficacy category of the target lesion evidence node is one of CR, PR, SD, or PD. If the uncertainty flag corresponding to the evidence node is true, the efficacy category of the evidence node is determined to be undefined. If the conflict flag corresponding to the evidence node is true, the efficacy category of the evidence node is determined to be conflict category. The efficacy category of the non-target lesion evidence node is one of A_non={SD}, A_non={CR}, or A_non={NN}, where NN represents incomplete remission and non-progression. The efficacy category of the new lesion evidence node is one of A_new={PD} and entering the disease progression locked state or no new lesion. S52. If the efficacy category of the target lesion evidence node is PD, or A_non={PD}, or A_new={PD}, and there is no negative evidence in the same anatomical region with a reliability higher than the disease progression evidence by at least a preset difference δ, then enter the disease progression coverage branch and determine PD as the only category in the candidate efficacy set; at this time, except for the node that triggers PD locking, all other evidence nodes are recorded as covered conflicts. S53. If the disease progression coverage branch is not entered, and the efficacy category of the target lesion evidence node is CR, and A_non∈{CR,NN}, and A_new=Ø, then complete remission CR is determined as the only category in the candidate efficacy set. S54. If the disease progression coverage branch and CR branch are not entered, and the efficacy category of the target lesion evidence node is PR, and A_new=Ø, and A_non≠{PD}, then PR is determined as the only category in the candidate efficacy set. S55. If the disease progression coverage branch, CR branch, and PR branch are not entered, and the disease progression coverage branch is not triggered, and the efficacy category of the target lesion evidence node is SD, or A_non={NN}, then SD will be determined as the only category in the candidate efficacy set. S56. If the conditions of any branch from S52 to S55 are not met, the candidate efficacy set is determined according to the efficacy category of each evidence node in accordance with the preset category priority rule; when the number of efficacy categories contained in the candidate efficacy set is greater than 1, the efficacy category of the corresponding baseline-review candidate pair is determined as an unclear category.

[0014] Furthermore, step S6 specifically includes: Determine whether the entity alignment score reaches the preset identity threshold and there is no conflict flag to obtain the identity strictness indicator; Determine whether the pairing score reaches the preset pairing threshold and whether the time direction conforms to the preset time constraint to obtain the pairing strictness indicator; Determine whether there are non-unknown anatomical regions among the evidence nodes corresponding to the baseline report records and review report records, and whether there is an intersection between the candidate anatomical regions to which the evidence nodes belong, to obtain the site strictness indicator; Determine whether the completeness of both the baseline report record and the review report record reaches the preset completeness threshold to obtain the completeness strictness indicator; Determine whether there are comparable target lesion treatment categories or evidence of disease progression coverage that meets preset conditions in the candidate treatment set to obtain a comparability strictness indicator; determine whether the number of treatment categories included in the candidate treatment set is equal to 1 to obtain a conflict strictness indicator. The rigor assessment result is obtained by multiplying the identity rigor indicator, pairing rigor indicator, site rigor indicator, integrity rigor indicator, comparability rigor indicator, and conflict rigor indicator. If the candidate efficacy set contains only one efficacy category and the rigor assessment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

[0015] In addition, this application also provides a data processing system for evaluating the efficacy of MRI before and after treatment, to implement the aforementioned method for processing data for evaluating the efficacy of MRI before and after treatment, including: The deduplication module is used to obtain patient identity records and MRI report records, perform field standardization processing and multi-field similarity calculation to obtain entity alignment scores, and deduplicate duplicate reports under the same patient name based on entity alignment scores to obtain valid MRI report records. Based on examination time, examination site and entity alignment scores, the valid MRI report records are correlated and scored to obtain baseline-re-examination candidate pairs and their matching scores. The evidence analysis module is used to analyze the text content of each valid MRI report record in the baseline-review candidate pair to obtain multiple evidence nodes and the report completeness of each evidence node's report. The graph propagation module is used to construct an evidence association network containing reference nodes and evidence nodes, using entity alignment score, pairing score and report completeness as reference nodes, and to transmit evidence strength on the evidence association network to obtain the transmitted positive evidence strength and negative evidence strength. The evidence projection module is used to compare the transmitted positive evidence strength and negative evidence strength with the rule-based positive evidence strength and rule-based negative evidence strength, respectively, and take the maximum value to obtain the projected positive evidence strength and negative evidence strength. The category determination module is used to determine the efficacy category for each evidence node based on the projected positive and negative evidence strength, and to determine the candidate efficacy set according to the preset category priority. The results output module is used to obtain the rigor judgment result based on entity alignment score, pairing score, report completeness and candidate efficacy set; when the candidate efficacy set contains only one efficacy category and the rigor judgment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

[0016] This application has the following advantages over the prior art: (1) By performing entity alignment and deduplication, examination pairing scoring, evidence node analysis, correlation network strength transfer, projection constraint, category determination and strictness determination on the original patient records and MRI report records from multiple sources, the burden of manual verification caused by scattered record sources and inconsistent report versions is reduced, and the standardization and automation of efficacy evaluation data processing are improved.

[0017] (2) The efficacy categories of the evidence nodes for target lesions, non-target lesions and new lesions are comprehensively judged according to the preset category priority. When any evidence node clearly points to disease progression, the conclusion is reflected in the candidate efficacy set in a timely manner, so as to avoid the disease progression being covered up by the category judgment results of other evidence nodes.

[0018] (3) Rigorous determination is made based on entity alignment score, pairing score, report completeness and the composition of candidate efficacy set, which avoids outputting a single definite result in cases where identity association is uncertain, inspection pairing is unreliable, report information is incomplete or efficacy category is conflicting, thus reducing the risk of efficacy evaluation results being misclassified. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for processing MRI efficacy evaluation data before and after treatment, as described in this application. Figure 2 This is a schematic diagram of the benefit consistency rate corresponding to different selected coverage rates on real data from an embodiment of a data processing method for evaluating the efficacy of MRI before and after treatment according to this application; Figure 3 This is an audit diagram of an embodiment of a method for processing MRI efficacy evaluation data before and after treatment according to this application. Detailed Implementation

[0021] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] like Figure 1As shown, this application provides a method for processing MRI efficacy evaluation data before and after treatment, including: S1. Obtain patient identity records and MRI report records, perform field standardization processing and multi-field similarity calculation to obtain entity alignment scores, remove duplicate reports under the same patient name based on entity alignment scores to obtain valid MRI report records, and perform correlation scoring on valid MRI report records based on examination time, examination site and entity alignment scores to obtain baseline-re-examination candidate pairs and their pairing scores. Specifically, step S1 includes: Retrieve the name, medical record number, patient card number, standardized name, date of birth and gender fields from the patient identity record, as well as the patient identification field from the MRI report record, and perform format conversion on all the above fields to obtain standardized fields; The similarity between corresponding fields in the patient identity record and the MRI report record is calculated based on the normalized fields, and the similarity of multiple fields is summed to obtain the entity alignment score. Based on the entity alignment score, determine whether there are multiple report records under the same patient name pointing to the same examination. If so, retain the one with the highest information completeness and delete the rest of the report records to obtain the valid MRI report record.

[0023] Understandably, different source systems do not have uniform naming and formatting for fields such as patient identity, examination time, examination site, report status, and observed conclusion text. Therefore, a versionable field mapping table is pre-established, as shown in Table 1. Each uniform field in the table records at least the source system, source field alias, target field name, target type, allowed value range, unit conversion method, whether missing values ​​are allowed, and the corresponding failure action. The mapping table itself is saved along with the version number so that the mapping version on which the historical records are based can still be traced after the threshold or rule is adjusted.

[0024] Table 1 Field Mapping Table When a field mapping fails, only one of three processing methods is allowed: rejecting the record, retaining the record and adding an error mark, or transferring it to the manual review queue. Silently replacing the incorrect or unmapped value with a seemingly normal default value is not allowed, thus ensuring that all fields used in subsequent steps have traceable sources and conversion basis.

[0025] The standardization of date and time fields follows similar processing principles: times with time zone information are uniformly converted to the organization's time zone; times with only hours, minutes, and seconds but no time zone information are interpreted according to the time zone pre-configured in the source system; times with only a date but no specific time are internally uniformly set to 12:00 of the current day and recorded as precision=day to avoid boundary deviations caused by defaulting to 00:00; times with only year and month but lacking a specific date are not included in the process requiring pairing by day. The mapped and standardized fields, along with the original field strings, parsing results, time zone source, and precision markers, are retained for subsequent processes and manual review and traceability.

[0026] In one embodiment of this application, multi-field similarity calculation specifically includes: Retrieve the name, medical record number, patient card number, standardized name, date of birth, and gender fields from the patient's identity record, as well as the patient identification field from the MRI report record; the patient identification field includes the ID number and the primary patient index; Standardized fields are obtained by removing leading and trailing whitespace, converting full-width characters to half-width characters, and unifying the case of letters in the patient identity record and the patient identification field in the MRI report record.

[0027] The multi-field similarity calculation specifically includes: dividing the normalized fields into two categories: strong identifying fields and weak identifying fields. ID number, master patient index, etc., are considered strong identifying fields, while medical record number, patient card number, normalized name, date of birth, and gender are considered weak identifying fields. Different weighting coefficients are assigned to the two categories when summing them, as shown in Table 2. The weights of the matching fields are accumulated, and the same weight is deducted when there is a clear conflict between strong identifying fields, date of birth, or gender. Table 2. Field Weights and Conflict Handling The similarity scores of each field are summed according to the weights mentioned above and truncated to between 0 and 1 to obtain the entity alignment score, denoted as S. id If the score is not lower than the preset automatic merging threshold and there are no unresolved conflict markers, it can be automatically identified as the same patient. If the score is lower than the threshold but higher than the preset review threshold, it will be transferred to manual review. If the score is lower than the review threshold, it will not be merged. Regardless of whether it is merged or not, the list of fields involved in the matching and the list of conflicting fields will be output for easy traceability.

[0028] In one embodiment of this application, entities with a score of 0.70 or higher and no unresolved strong conflicts are automatically merged; those with a score between 0.55 and 0.70 undergo review; and those below 0.55 are not merged. Any merge outputs a list of match and conflict fields.

[0029] In one embodiment of this application, deduplication specifically includes: The study uses the combination of patient, examination number, examination time accuracy, examination site and examination instance to form a joint key. When the examination number is missing, the combination of patient, examination date, site and report check value is used to form an approximate joint key. Multiple report records pointing to the same joint key are judged as duplicate candidates. Calculate a comprehensive score for each duplicate record: ; in, This represents the overall score. This indicates the report status item. The status is taken as final version, issued, or reviewed. ,otherwise , The information completeness item is determined by whether the findings and conclusions of the report are simultaneously present. This indicates a traceability item; if the check number exists, it is taken. , The version number field is obtained by normalizing the version number of the report; the version number is obtained by dividing by 3 and truncating to 1. .

[0030] When there are ties, the report version number, issuance time, and source record verification value are compared in sequence, with the last item ensuring stable and reproducible sorting.

[0031] In one embodiment of this application, the association scoring specifically includes: For valid MRI reports that have passed identity alignment, hard constraints are applied. These constraints include: patient physical alignment is complete; baseline examination is no later than 7 days after treatment; follow-up examination is later than baseline examination; there is a comparable overlap in the anatomical thesaurus of the two reports; and the report text can be parsed normally. Only two valid MRI reports that simultaneously meet the hard constraints constitute a candidate examination pair and are included in the paired score calculation; the paired score is calculated for each candidate examination pair: ; ; in, Indicates the pairing score. Indicates time item, The section indicating the location represents the percentage overlap between the locations examined in the two reports. This indicates a sequence item, representing the percentage of comparable imaging sequence types from the two reports. This indicates the report item, which is the average of the baseline report completeness and the review report completeness. The identity item is used to calculate the entity alignment score. This indicates the number of days between the follow-up examination and the baseline examination.

[0032] according to The calculation results yield baseline-review candidate pairs and their paired scores. If the candidate check pair is below the preset matching threshold, it does not meet the matching strictness indicator; when matching scores are tied, they are compared sequentially. , Sorting by scores for each item.

[0033] In one embodiment of this application, the preset pairing threshold is 0.45.

[0034] In one embodiment of this application, the normalized fields are serialized in a fixed field order. Each field consists of a field name, an equal sign, and a normalized value. After the fields are connected by an invisible delimiter, a summary is calculated to obtain the source check value of the valid MRI report record. ; in, This represents the source verification value. Represents a summary function. Indicates the first The field name of a normalized field. Indicates the first The values ​​of each normalized field, Indicates an invisible separator connection. Indicates the first The field name of a normalized field. Indicates the first The values ​​of each normalized field, Indicates the first The field name of a normalized field. Indicates the first The values ​​of each normalized field, This indicates the total number of normalized fields involved in the serialization; Understandably, checksums are used for the construction of composite keys for duplicate records, the identification of report version changes, the verification of data consistency before and after processing, and the verification of the integrity of output result packages.

[0035] For the same patient, the normalized master index is concatenated with the controlled salt value independently stored by the deployer to calculate the summary. The first few hexadecimal characters are extracted and prefixed with RP to obtain the study patient identifier corresponding to that patient. ; in, Indicates patient identification for the study. A fixed prefix indicating the identification of research patients; in this example, the value is RP. This indicates that the first 20 hexadecimal characters of the summary result are taken. This indicates a controlled salt value, which is stored only on the deployer's side and not written to the result data. This represents the normalized master patient index. This represents an irreversible digest algorithm performed on the input string. This indicates string concatenation. This means that the standardized master patient index will be uniformly converted to uppercase before participating in the summary calculation, in order to avoid the same patient generating different study identifiers due to differences in capitalization.

[0036] Understandably, the same primary index provides a stable and unique patient identifier within the same deployment domain. Different deployment domains cannot be reverse-engineered or directly linked due to their different salt values. This ensures that the original patient identifier is not leaked while meeting the requirements for entity alignment and deduplication.

[0037] S2. Analyze the text content of each valid MRI report in the baseline-review candidate pair to obtain multiple evidence nodes and the report completeness of each evidence node's report; Specifically, step S2 includes: The text content of each valid MRI report in the baseline-review candidate pair is parsed, and the lesion descriptions obtained are classified into target lesion evidence nodes, non-target lesion evidence nodes, and new lesion evidence nodes according to the anatomical region. Each evidence node carries at least the following attributes: node identifier, node type, anatomical region, polarity (positive or negative), reliability, source chapter, original text fragment index, measurement interval, and source fragment check value. To determine the completeness of a MRI report, one must determine whether both the sections seen and the conclusions are present simultaneously in the valid MRI report. Determine whether the number of located anatomical regions in valid MRI report records reaches a preset number to obtain the region integrity. Determine whether the number of evidence nodes in a valid MRI report reaches a preset number to obtain the evidence completeness. The report completeness of each evidence node is obtained by summing the chapter completeness, region completeness, and evidence completeness.

[0038] Understandably, the measurement interval represents the lesion measurement value in the form of upper and lower bounds, rather than taking only a single center value, in order to preserve the uncertainty carried by approximate expressions such as "approximately several centimeters" in the original report. Evidence nodes are deduplicated and sorted according to stable node identifiers to ensure that the same input yields a consistent set of nodes when processed repeatedly at different times.

[0039] In one embodiment of this application, when parsing the text content of a valid MRI report and obtaining evidence nodes and report completeness, the observed chapters (corresponding to expressions such as "image findings", "examination findings", "findings", etc.) and conclusion chapters (corresponding to expressions such as "diagnosis opinion", "image diagnosis", "conclusion", "impression", etc.) and comparison chapters (corresponding to expressions such as "compared with previous film", "compared with previous", "compared to previous", etc.) are first identified according to preset aliasing rules. Then, the identified chapter texts are segmented into sentences according to periods, semicolons and line breaks. The anatomical region entity dictionary on which the evidence node classification depends is organized according to preset anatomical groups (such as primary area, regional lymph nodes, retropharyngeal lymph nodes, distant lesions) and version management is performed.

[0040] In the process of classifying evidence nodes into target lesion evidence nodes, non-target lesion evidence nodes, and new lesion evidence nodes, a scope restriction is set for negative expressions appearing in sentences (such as "not seen," "not reached," "none," "not found," "not displayed," "not prompted"). That is, the scope of negation is limited to the preset character range before the negation trigger word and does not cross sentences. Based on this, the polarity of the evidence node is determined (for example, "no new lesions seen" corresponds to a new lesion evidence node with a negative polarity; "reduced compared to before" corresponds to a target lesion evidence node with a positive polarity). The reliability of the evidence node is differentiated according to the chapter in which it is located. That is, the reliability of evidence nodes from the conclusion chapter is converted according to a preset coefficient, and the reliability of evidence nodes from the seen chapter is converted according to another preset coefficient. Pure measurement evidence nodes are converted according to different preset coefficients in the seen chapter and other chapters. This reliability will be one of the initial characteristics of the evidence node when constructing the evidence association network.

[0041] In one embodiment of this application, the negation scope is limited to the first 12 characters of the negation trigger word and does not span across sentences. For example, "no new lesions were found" generates a new_lesion node with polarity=-1; "shrink compared to before" generates a shrink node with polarity=+1. Here, the new_lesion node represents an evidence node in the report text regarding "whether new lesions exist"; the shrink node represents an evidence node indicating a reduction in the size of the target lesion compared to the baseline or previous examination; polarity indicates the direction of evidence, with +1 indicating that the text definitely supports the event described by the node, -1 indicating that the text negates the event, and 0 or an uncertain value used in cases of insufficient evidence, semantic conflict, or unclear direction.

[0042] In one embodiment of this application, the reliability of event nodes in the conclusion section is multiplied by 0.95, and that in the seen section by 0.90; the reliability of pure measurement nodes is multiplied by 0.95 in the seen section and by 0.85 in other sections.

[0043] In one embodiment of this application, the measurement value parsing in the report text containing the target lesion evidence node includes two categories: single-diameter description and double-diameter description. The double-diameter description is 1.2×0.8 cm, 12×8 mm, and 12×8 mm, while the single-diameter description is 1.2 cm. The larger value of the double-diameter description is taken as the major diameter, and the smaller value is taken as the minor diameter, and both are uniformly converted to millimeters. When the measurement description contains approximate words such as "approximately," "about," "nearly," or "around," the corresponding value is not treated as an exact point value. Instead, an 8% measurement interval is generated centered on that value. For example, "approximately 1 cm" represents [9.2, 10.8] mm.

[0044] When multiple uncertain words appear in the same sentence, the one with the lowest reliability is taken as the reliability of the evidence node. Evidence nodes formed by uncertain statements with low reliability (such as newly emerging candidate evidence) do not directly cover negative evidence nodes with high reliability in the same area. Both are retained in the constructed evidence association network and are reflected by uncertainty and conflict indicators.

[0045] Different uncertainty terms correspond to different basic reliability levels, as shown in Table 3: Table 3. Basic Reliability of Uncertain Terms In one embodiment of this application, the three components—chapter completeness, region completeness, and evidence completeness—are included in the sum of the report completeness according to preset weights, with chapter completeness having a weight of 0.45, region completeness having a weight of 0.3, and evidence completeness having a weight of 0.25. When both the described and conclusion chapters exist, the chapter completeness is set to 1; when the number of clearly located anatomical regions reaches a preset number (e.g., at least two), the region completeness is set to the maximum value; when the number of evidence nodes reaches a preset number (e.g., at least four), the evidence completeness is set to the maximum value. A report completeness score below 0.35 is considered insufficient, i.e. =0.

[0046] Each evidence node includes at least the following: a unique identifier (node_id), a node type (node_type), the corresponding anatomical region (region), the polarity indicating affirmative or negative evidence direction, the reliability of evidence from different sections or expressions, the source section (source_section), the text segment position of the evidence in the original report (text_span), the measurement interval (measurement_interval) obtained from the measurement value and approximate words, and the irreversible verification value (source_hash) of the source report or source text. A graph-level object includes an inspection number, a section set, completeness, uncertainty quality, and a node list. Node types include target lesions, non-target lesions, new lesions, or measurement nodes. The inspection number in the graph-level object identifies the current inspection; the section set records the parsed report sections; the completeness indicates whether the report information is sufficient for evaluation; the uncertainty quality summarizes ambiguous, conflicting, or missing evidence in the report; and the node list stores all traceable evidence nodes.

[0047] In one embodiment of this application, in addition to evidence nodes, a graph-level summary object is also constructed for each valid MRI report record during the parsing process. This object records the examination number, chapter set, report completeness, overall situation of whether there is an uncertain flag for the evidence node, and a list of evidence nodes, serving as the basic data structure for subsequently constructing the evidence association network.

[0048] S3. Using entity alignment score, pairing score and report completeness as reference nodes, construct an evidence association network containing reference nodes and evidence nodes, and transfer evidence strength on the evidence association network to obtain the positive evidence strength and negative evidence strength after transfer. Specifically, step S3 includes: S31. Using entity alignment score, pairing score and report completeness as reference nodes, and target lesion evidence nodes, non-target lesion evidence nodes and new lesion evidence nodes as evidence nodes, establish connection edges between reference nodes and evidence nodes, and between evidence nodes to obtain the evidence association network. S32. Evidence strength is transferred along the connecting edges in the evidence association network. The features of adjacent nodes are aggregated to obtain the transferred positive and negative evidence strengths. The calculation formula for evidence strength transfer is as follows: ; in, Indicates the first Nodes in layer propagation To the node Attention logarithmic value, , Let be the feature representations of nodes u and v at layer l, respectively. The characteristics of the connecting edges themselves, , , This is the learnable matrix for linear transformation of the source node features, target node features, and edge features in this layer. This is a learnable attention vector; This represents vector concatenation, where SiLU is the activation function. The learnable nonnegative scaling factor. To prevent logarithmic operations from diverging to the smallest positive numbers, Indicates connecting edges Edge reliability, Indicates the source node The reliability is determined by the entity alignment score, pairing score, or report completeness when the source node is a factor node.

[0049] In this system, the first dimension of the feature vector of the reference node is written with the corresponding score of 0 to 1, and the other dimensions are distinguished by one-hot encoding of factor type. The reference node corresponding to the report completeness is constructed independently according to the baseline report and the review report, that is, it is only connected to the evidence node at the corresponding time point. The reference nodes corresponding to the entity alignment score and the pairing score are connected to the evidence nodes at both the baseline and review time points. The evidence nodes are defined as target lesion evidence nodes, non-target lesion evidence nodes and new lesion evidence nodes. The feature vector of the evidence node includes the aforementioned dimensions such as reliability, polarity, measurement attribute, time point, region and chapter source.

[0050] In one embodiment of this application, step S31 specifically includes: Determine whether the anatomical regions between each pair of evidence nodes are the same to obtain the same region judgment result; Determine whether the node identifiers between any two evidence nodes point to the same lesion origin to obtain the homology judgment result; Determine whether each pair of evidence nodes belongs to different time points and has the same evidence predicate to obtain the cross-time same predicate judgment result; If any one of the following is true: same region judgment result, same source judgment result, or cross-time same predicate judgment result, then establish a connection edge between the corresponding two evidence nodes; A bidirectional connection edge is established between the reference node and the evidence node to obtain the evidence association network.

[0051] In one embodiment of this application, the relationship mask of the connection edges in the evidence association network is: Its value is 0 or 1, indicating whether the connection edge between node u and node v is retained; each retained edge also carries the edge reliability. Its value ranges from 0.2 to 1, and it is used to characterize the prior estimate of the correlation strength between the evidence connected by the edge. The reliability carried by the node itself is denoted as . Its value ranges from greater than 0 to no more than 1. The above connection method avoids introducing connecting edges between evidence nodes that have no actual correlation, reduces mutual interference between irrelevant evidence, and also reduces the computational scale in the subsequent strength transfer process.

[0052] In one embodiment of this application, step S32 specifically includes: Transform the features of the nodes at both ends of the connecting edge to obtain the structural priors; The reliability term is obtained by weighting the nodes at both ends of the connecting edge based on their reliability. Logarithmic transformation is performed based on the reliability of the connected edges to obtain the edge reliability term; The structural prior terms, reliability terms, and edge reliability terms are added together and normalized within the constraints of the connecting edges to obtain the weights of the weighted aggregation. Based on the weights of the weighted aggregation, the features of adjacent nodes are weighted and aggregated to obtain the positive and negative evidence strengths after propagation. The formula for calculating the weights of the weighted aggregation is as follows: ; ; ; ; in, Indicates the first Nodes in layer propagation To the node The weighted aggregate weight, Indicates pointing to a node For any source node, the summation range is all nodes connected to it. There are source nodes with connecting edges between them. Indicates the target node. This represents the edge score from source node u to target node v during propagation at layer l. This represents the score of edges pointing to v from other source nodes w that share the same target node v. This represents the aggregated message received by target node v at layer l. This represents the linear transformation parameters used in message computation. This represents the linear transformation parameters used in the gating calculation. Indicates the first Layer source node eigenvectors, This represents a gating vector that controls the proportion of messages written. Indicates the first Layer target node eigenvectors, This represents the updated target node features. This indicates element-wise multiplication. This represents the activation function. Presentation layer normalization operation.

[0053] In one embodiment of this application, a node feature matrix is ​​first generated by sorting the nodes by their node identifiers (node_id), and then a node-by-node pair is generated according to the connection rules. , Edge type encoding; perform three-layer bidirectional propagation normalized to the target node, with each layer first calculating the edges for all retained edges. and Then, the messages are aggregated and the node features are updated via gated residuals to avoid inconsistencies in the update order introduced by in-situ feature overlay due to different node traversal orders within the same layer. After propagation, the positive evidence strength is transmitted by the feature outputs of the updated target lesion evidence nodes, non-target lesion evidence nodes, and newly added lesion evidence nodes. With the strength of negative evidence .

[0054] The evidence association network limits the connections between evidence nodes to three types of relationships: same region, same source, or same predicate across time, rather than establishing pairwise connections between all evidence nodes. In scenarios where the number of evidence nodes is limited and their sources are scattered, this reduces the mutual interference between irrelevant evidence nodes and also reduces the number of connection edges that need to be processed for the transmission of evidence strength.

[0055] In one embodiment of this application, the structural prior terms, reliability terms, and edge reliability terms are added together, and then, within the constraint range of the connecting edges, the target node is selected. All established edges are normalized to obtain the weights of the weighted aggregation; The target node is obtained by weighted aggregation of features of adjacent nodes based on weighted aggregation weights. In the messages aggregated by layer The target node state is then updated via a gated residual mechanism: first by the gated vector. The node features and messages of this layer are concatenated and then activated by a sigmoid function, which is then used to obtain the result. The sum of the features before the update and the gated message is obtained by layer normalization. After passing through a preset number of layers, the strength of positive and negative evidence after the transmission is output. S4. Compare the transmitted positive evidence strength and negative evidence strength with the rule positive evidence strength and rule negative evidence strength respectively, and take the maximum value to obtain the projected positive evidence strength and negative evidence strength. Specifically, step S4 includes: The strength p of the positive evidence after transmissionnet Compared with the benchmark positive evidence strength P rule The maximum value is taken from the comparison to obtain the projected positive evidence strength P; The strength of negative evidence after transmission n net Compared with the baseline negative evidence strength N rule The maximum value is taken from the comparison to obtain the negative evidence strength N after projection; An uncertainty flag is obtained based on whether the difference between the projected positive evidence strength and the projected negative evidence strength is less than a preset threshold. A conflict flag is obtained based on whether the strength of the projected positive evidence and the strength of the projected negative evidence simultaneously reach a preset threshold.

[0056] This is understandable, since the maximum value operation guarantees that P is not lower than P. rule N is not less than N rule The strength of evidence introduced in the online dissemination process can only be supplemented or modified based on the support level already confirmed by the rules, and cannot be downgraded or erased from the strength of evidence that has been clearly observed in the original report. This avoids the problem of evidence strength dilution or misrepresentation that may occur during the online dissemination of evidence, and ensures that the final judgment can always be traced back to the evidence support already observed in the original report.

[0057] This application compares the strength of positive and negative evidence after transmission with the strength of positive and negative evidence according to the rules, respectively, and takes the maximum value to obtain the projected strength of evidence. This ensures that the projected strength of evidence is not lower than the benchmark strength of evidence determined according to the rules. In this way, while supplementing and associating evidence information using the correlation network, it avoids weakening or covering the already determined strength of evidence during the network transmission process, thus ensuring the reliability of the strength of evidence on which the efficacy determination is based.

[0058] In one embodiment of this application, each region-predicate node after transmission outputs the positive evidence strength p after transmission. net and the strength of negative evidence n net Instead of directly overwriting the rule state with this output, a monotonic projection is performed: ; ; ; ; in, This represents the strength of positive evidence projected onto region r and the evidence predicate q. This indicates the strength of positive evidence obtained from the rule. This indicates the strength of positive evidence obtained through network transmission. Indicates the strength of negative evidence after projection. This indicates the strength of negative evidence for the rule after projection. Indicates the strength of negative evidence obtained through network transmission. Indicates the uncertainty after projection. Indicates the original uncertain sign, This is an indicator function that takes a value of 1 if the condition is true and 0 otherwise. Indicates the threshold for uncertain judgment. Indicates the conflict flag after projection. The table shows the threshold for the strength of positive evidence. The table shows the threshold for the strength of negative evidence. This represents the difference threshold used to trigger a judgment of uncertainty or conflict when the strengths of positive and negative evidence are close.

[0059] The projection guarantees that P* is not less than the original P and N* is not less than the original N; the network can supplement cross-node propagation support, but cannot erase the original positive, negative, or conflict history. The response conflict lattice reads P*, N*, U*, C*, and the original node chain; if network propagation is inconsistent with hard constraints, conflict flags, incorrect date direction, PD locking, and multi-value set outputs still have priority. When the node arrival order changes, the node_id sorting, softmax normalized to the target node, and summation aggregation maintain the graph-level output unchanged.

[0060] S5. Based on the projected positive and negative evidence strength, determine the efficacy category for each evidence node and determine the candidate efficacy set according to the preset category priority. Specifically, step S5 includes: S51. Based on the projected positive and negative evidence strengths, determine the efficacy category for each evidence node, including target lesion evidence nodes, non-target lesion evidence nodes, and newly diagnosed lesion evidence nodes. The efficacy categories include complete remission (CR), partial remission (PR), stable disease (SD) or progressive disease (PD), ambiguous category, and conflict category. The efficacy category for target lesion evidence nodes is one of CR, PR, SD, or PD. If the uncertainty flag corresponding to an evidence node is true, the efficacy category of that evidence node is determined to be ambiguous; if the uncertainty flag is true, the efficacy category is determined to be ambiguous. If the conflict flag corresponding to a node is true, then the efficacy category of that evidence node is determined to be a conflict category; the efficacy category of a non-target lesion evidence node is one of A_non={SD}, A_non={CR}, or A_non={NN}, where NN represents incomplete remission and no disease progression; the efficacy category of a new lesion evidence node is either A_new={PD} and it enters a disease progression locked state or there is no new lesion; where A_non is the efficacy category of a non-target lesion evidence node, and A_new is the efficacy category of a new lesion evidence node; S52. If the efficacy category of the target lesion evidence node is PD, or A_non={PD}, or A_new={PD}, and there is no negative evidence in the same anatomical region with a reliability higher than the disease progression evidence by at least a preset difference δ, then enter the disease progression coverage branch (PD_LOCKED) and determine PD as the only category in the candidate efficacy set; at this time, except for the node that triggers PD locking, all other evidence nodes are recorded as covered conflicts. S53. If the disease progression coverage branch is not entered, and the efficacy category of the target lesion evidence node is CR, and A_non∈{CR,NN} (i.e., non-target lesions are not PD), and A_new=Ø, then complete remission CR is determined as the only category in the candidate efficacy set. S54. If the disease progression coverage branch and CR branch are not included, and the efficacy category of the target lesion evidence node is PR, and A_new=Ø, and A_non≠{PD} (non-target lesions are not disease progression), then PR will be determined as the only category in the candidate efficacy set. S55. If the disease progression coverage branch, CR branch, and PR branch are not entered, and there is no evidence of disease progression coverage (i.e., PD_LOCKED is not triggered), and the efficacy category of the target lesion evidence node is SD, or A_non={NN}, then SD will be determined as the only category in the candidate efficacy set. S56. If the conditions of any branch from S52 to S55 are not met, the candidate efficacy set is determined according to the efficacy category of each evidence node and the preset category priority rule. When the number of efficacy categories contained in the candidate efficacy set is greater than 1, the efficacy category of the corresponding baseline-review candidate pair is determined as an unclear category, and the candidate efficacy set, the evidence node that caused the disagreement and the review reason are retained as part of the output content for use in the subsequent manual review stage. When it is determined as an unclear category, the SD category shall not be used to replace the output.

[0061] Understandably, entering the disease progression coverage branch means that the final conclusion of the entire baseline-review candidate pair is PD, and the results of other nodes are no longer considered. If the disease progression coverage branch is not entered (i.e., no node is determined to be PD), then the conditions for CR (complete remission) are checked: the target lesion evidence node must be CR, the non-target lesion evidence node must be CR or an intermediate state of "non-CR and non-PD", and the new lesion evidence node must not have any new lesions. All three conditions must be met simultaneously to enter the CR branch, and the final conclusion is locked as CR. If the first two branches are not entered (neither PD nor the triple conditions for CR are met), then the conditions for PR (partial remission) are checked: the target lesion evidence node is determined to be PR, the non-target lesion evidence node is not PD, and the new lesion evidence node does not have any new lesions. If these conditions are met, the final conclusion is locked as PR. When the uncertainty indicator is true, the larger value between the original uncertainty indicator being true or the positive evidence strength after transfer and the negative evidence strength after transfer is less than a preset threshold. When the conflict flag is true, the strength of the projected positive evidence is greater than or equal to the threshold for the strength of the positive evidence, and the strength of the projected negative evidence is greater than or equal to the threshold for the strength of the negative evidence, and the difference between the strength of the projected positive evidence and the strength of the projected negative evidence is less than a preset threshold.

[0062] The conditions for classifying the efficacy of the target lesion evidence node as disease progression include: the relative change interval calculated based on baseline and follow-up measurements falls entirely above a preset rise threshold interval (e.g., the relative increase percentage is not less than the preset rise threshold), or the absolute change reaches the lower bound of the preset absolute increment threshold, and there is no negative evidence with higher reliability pointing to complete remission, partial remission, or disease stability within the same anatomical region; when the relative change interval crosses the preset rise threshold, adjacent branches are retained and recorded as measurement boundary conflicts, and are not forcibly classified into the single category of disease progression. The conditions for classifying the efficacy of the new lesion evidence node as disease progression include: the intensity of positive evidence after projection reaches a preset reliability threshold, which is determined to be definite positive, and there is no negative evidence with higher reliability pointing to the absence of new lesions within the same anatomical region; when the difference between the intensity of positive evidence and the intensity of negative evidence is less than the preset threshold, it is determined to be an indeterminate marker, and is not directly determined as disease progression.

[0063] That is, disease progression is the primary trigger, but it must be based on evidence such as a clear increase in the size of the target lesion, a clear positive result for new lesions, or a clear progression of non-target lesions. In cases where the measurement interval crosses the threshold, the positive and negative evidence are close, or there is reverse evidence with higher reliability, a single disease progression category is not directly output, but adjacent candidate categories or uncertain markers are retained and entered into the review process.

[0064] This application comprehensively judges the efficacy categories of evidence nodes for target lesions, non-target lesions, and new lesions according to preset category priorities. Among them, disease progression is set as a priority category, which can promptly reflect the conclusion in the candidate efficacy set when any evidence node clearly points to disease progression, and avoid the disease progression being obscured by the category judgment results of other evidence nodes.

[0065] The criteria for classifying non-target lesion evidence nodes as disease progression include: positive evidence nodes formed by positive descriptions of lesion progression in the report text (such as "significant progression, expansion of range, increased invasion, significant increase," etc.), whose positive evidence strength reaches a preset reliability threshold, and the absence of negative evidence with a reliability higher than the positive evidence by at least a preset difference δ within the same anatomical region. When the efficacy category of a non-target lesion evidence node is in a non-CR and non-PD compatible state, this state only supports the determination of the stable disease branch and does not support the determination of the partial remission branch alone. If the uncertainty flag corresponding to the evidence node is true, the efficacy category of the evidence node is determined to be an ambiguous category; if the conflict flag corresponding to the evidence node is true, the efficacy category of the evidence node is determined to be a conflict category.

[0066] The efficacy category determination results are associated with benefit labels according to a preset mapping: complete remission (CR), partial remission (PR), and stable disease (SD) are mapped to the benefit category, and disease progression (PD) is mapped to the no-benefit category; if the candidate efficacy set contains efficacy categories corresponding to both the benefit category and the no-benefit category, the benefit label retains two values ​​at the same time, and the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

[0067] In one embodiment of this application, the formation state of each lesion region r and the evidence predicate q is... . To affirm the highest level of reliability of evidence, i.e., the strength of positive evidence in a rule, To negate the highest level of reliability of evidence, i.e., the rule of negative evidence strength, Indicates an uncertain sign. Indicates conflict. Positive and negative evidence are updated after being recalculated based on the source coefficient corresponding to their respective chapter (Conclusion Chapter, Seen Chapter, Other Chapters) and the coefficient corresponding to the uncertainty term. and Evidence from different sources is not deleted from each other, and evidence with lower reliability is still retained in the node chain.

[0068] The state update is determined using the following rules, where the positive evidence strength threshold is used. negative evidence strength threshold Difference threshold : For any set of nodes arriving in any order, the positive and negative components are merged by the maximum value of each component, and the uncertain flags are merged by a Boolean union; the conflict flags are determined by the merged set. , The sum and difference threshold δ is determined once. This merge satisfies the commutative, associative, and idempotent laws; repeated access of the same node or changes in the node order do not alter the final state, as shown in Table 4. ; ; ; ; Table 4 Status Update Table In Table 4, POSITIVE indicates that the evidence clearly supports a certain event or therapeutic branch; NEGATIVE indicates that the evidence clearly excludes the event or therapeutic branch; UNCERTAIN indicates that the strength of the evidence is insufficient or the difference between the positive and negative evidence is too small to make a definitive classification; CONFLICT indicates that both positive and negative evidence have reached the threshold and the difference is small, so it is necessary to retain both the support set and the exclusion set and proceed to review; NONE indicates that this state has not been triggered or there is no corresponding evidence.

[0069] In one embodiment of this application, the baseline measurement interval is set as follows: The re-examination measurement interval is ,and The relative range of change is then calculated using the following formula: ; in, This indicates the range of percentage changes relative to the baseline during the follow-up examination. Indicates the lower boundary of the measurement interval to be reviewed. This indicates the upper limit of the measurement interval to be reviewed. Indicates the lower boundary of the baseline measurement interval. Indicates the upper limit of the baseline measurement interval; The lower bound of D is calculated using the re-examination lower bound and the baseline upper bound, while the upper bound of D is calculated using the re-examination upper bound and the baseline lower bound. This is used to obtain the relative range of change under the most conservative and most aggressive measurement combinations. (Interval) The lower bound is determined by the ratio of the lower bound of the re-examination to the upper bound of the baseline, and the upper bound is determined by the ratio of the upper bound of the re-examination to the lower bound of the baseline. This is used to obtain the range of values ​​for the relative change amplitude even when there is uncertainty in the measurement value itself. When the relative change interval D falls into a certain preset threshold interval, a single and definite target lesion efficacy branch is formed. When the interval D crosses the preset decreasing threshold or increasing threshold (e.g., a relative decrease or increase of a certain percentage), adjacent efficacy candidates are formed and marked as measurement boundary conflicts. They are not forcibly classified into a single branch, thus preserving the uncertainty carried by the measurement itself. The absolute increment threshold is also determined by the upper and lower bounds of the interval rather than by the center value of the interval.

[0070] Each branch generates a set of allowed responses. and Exclusion Set When the evidence node for a newly emerging lesion is determined to be a clearly positive POSITIVE, it is generated. And enter the disease progression coverage branch PD_LOCKED; when the evidence node for new lesions is determined to be clearly non-existent, generate Triggered solely by new lesions Complete remission is generated based on the measurement interval of the target lesion. Partial relief ), disease stability ( ) or disease progression ( Candidate categories; generated when non-target lesion evidence nodes are determined to indicate clear progression. When the condition is determined to be a compatible state of incomplete remission and non-disease progression, it is generated. The candidate treatment set is determined using the following state machine: Initial state (INIT): Set of candidate treatments ={CR,PR,SD,PD}, reads the status of all branches.

[0071] Disease progression coverage status PD_LOCKED: Any new lesion evidence node or target lesion evidence node is determined as PD_LOCKED. Or non-target lesion evidence node is determined as If the result is positive (POSITIVE) and there is no negative evidence of high reliability (at least δ) in the same region, response_set={PD}; the result of other branches is recorded as covered_conflic, but the disease progression coverage status is not removed.

[0072] Constrained intersection state CONSTRAINED: The set of allowed branches for all non-unknown branches when the disease progression coverage state has not yet been entered. Find the intersection and remove the exclusion set of each branch. If the intersection is not empty, then the intersection is taken as... .

[0073] Conflict set state CONFLICT_SET: When the intersection of the sets is empty or there is a threshold crossing, the state is determined by... Calculate the support score for each response, where For branches The reliability of the corresponding evidence nodes, The values ​​are +1, -1, and 0 under the three orientations, respectively; responses that differ from the highest support score by no more than 0.1 and are not hard excluded by evidence with a reliability higher than 0.8 are retained to form a set value and record the evidence path that generated each candidate.

[0074] Automatic Single-Value Output Status AUTO_SINGLE: When response_set is a single element, all six Strict indicators are 1 and C≥0.70, a single efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair.

[0075] In the REVIEW_SINGLE state where the response_set is a single element but any Strict indicator is 0 or C < 0.7, the single-value suggestion is retained and the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

[0076] REVIEW_SET: When the response_set contains more than two responses, it is judged as an ambiguous category. The candidate efficacy set, conflict nodes and review reasons are output. This case is not filled with disease stability.

[0077] The results of manual review are written back as newly added review nodes, without overwriting the original evidence nodes, original response_set, or original decision_path. Therefore, the same input, mapping version, and threshold version always produce the same state transition result, and the differences before and after review can be traced node by node.

[0078] In one embodiment of this application, each evidence node is first deduplicated and sorted according to a stable node_id. The node vector consists of reliability, affirmative / negative polarity, uncertainty flag, measurement presence, measurement center and relative interval width, baseline / review time point, node type, anatomical region, and source chapter. A factor node is established for entity alignment, inspection pairing, baseline report completeness, and review report completeness. The first dimension of the factor node is written with a corresponding score of 0 to 1, and the remaining dimensions are identified by a one-hot code of the factor type.

[0079] The evidence association network does not unconditionally establish fully connected edges for any node; instead, it first generates connection edge constraint identifiers. Two evidence nodes are allowed to be considered only if they satisfy at least one of the following conditions: same origin, same region, or same predicate across time. The reference nodes corresponding to entity alignment scores and pairing scores establish bidirectional connections with the baseline and all evidence nodes in both review periods; the reference nodes corresponding to baseline report completeness only establish bidirectional connections with evidence nodes in the baseline period; the reference nodes corresponding to review report completeness only establish bidirectional connections with evidence nodes in the review period; bidirectional connections are maintained between all reference nodes; the remaining nodes are linked to... Furthermore, it is not included in the denominator of the normalization calculation. Each established connection carries an edge feature vector, which includes at least the following identifiers: whether it uses the same predicate, whether it is in the same anatomical region, whether it is from the same source, whether it spans different time points, whether it has opposite polarities, whether it is a reference node pointing to an evidence node, and the type of the target reference node.

[0080] Calculate the structural prior for each established connection edge. Homology is set to 1, same region or opposite polarity to 0.95, same predicate or factor-evidence correspondence to 0.9, cross-time relationship to 0.85, and other retention relationships to a lower limit of 0.20. Source node reliability. When the source is an evidence node, read its own reliability; when the source is a reference node, read the corresponding entity alignment score, pairing score, or report completeness.

[0081] Suppose there is a baseline-review candidate pair, after deduplication, there are a total of After adding four reference nodes—entity alignment, pairing, baseline report completeness, and review report completeness—to the evidence nodes, the node feature matrix has the following dimensions: The 29 dimensions are, in order: reliability, affirmative polarity, negative polarity, low threshold indicator, measurement presence indicator, normalized measurement center, normalized relative interval width, time point identifier, reference node identifier and its 4D type code, 7D node type code, 5D anatomical region code, and 3D source chapter code. Each row of the edge feature matrix corresponds to an established connection edge, and the 12 dimensions are, in order: same predicate identifier, same region identifier, same source identifier, cross-time identifier, opposite polarity identifier, reference node pointing to evidence node identifier, target reference node 4D type code, source node reliability, and structural prior. The edge index stores the source and target node indices of each edge in a two-row matrix.

[0082] Before being transmitted, node features undergo linear transformation, SiLU activation, and layer normalization encoding to a preset hidden dimension (e.g., 48 dimensions). This hidden dimension is maintained throughout the transmission of a preset number of layers (e.g., three layers). Each layer first calculates all established connections at once. For the same target node, the score is first subtracted from the maximum value among all incoming edges of that target node to ensure numerical stability, and then only the score is calculated for that target node. Normalization is performed within the range of incoming edges. Node pairs without established connections are not allocated storage locations and are not included in the normalization denominator with minimal values. Message computation and gating computation are implemented using preset linear transformations and activation function combinations, respectively. , Formula. This implementation allows for a variable number of evidence nodes and connecting edges for different baseline-review candidate pairs, while keeping the node feature dimension, edge feature dimension, and network parameter dimension fixed.

[0083] The transmission endpoint outputs two-dimensional values ​​of the positive and negative evidence strengths after transmission for each evidence node. Reference nodes do not participate in the grouping and merging process based on anatomical region and evidence predicate. Evidence nodes are grouped according to their time point, anatomical region, and evidence predicate combination. Within the same group, the maximum value of the positive and negative evidence strengths after transmission is taken. If there is no corresponding transmission result for the regular positive or negative evidence strength in a certain group, the transmission result for that group is 0, and the projected positive and negative evidence strengths remain consistent with the regular positive and negative evidence strengths. The above grouping criteria must be determined by the time point, anatomical region, and evidence predicate attributes carried by the evidence node itself, and not inferred from the node's position in the feature matrix.

[0084] In one embodiment of this application, when a baseline-review candidate pair with known efficacy evaluation results exists as a training sample, the training objective of the evidence association network consists of the classification loss of the candidate efficacy set and the auxiliary loss of the polarity of the evidence nodes. The polarity objective of each evidence node in the auxiliary loss directly reads the affirmative or negative flag obtained from the parsing of that evidence node, and the node-by-node binary cross-entropy is weighted according to the reliability of that evidence node. ; in, Indicates total training loss. The classification loss represents the set of candidate treatment outcomes, used to constrain the overall treatment outcome categories. This represents the set of evidence nodes that participate in auxiliary supervision. Indicates evidence node Reliability, This represents the binary cross-entropy loss. This indicates the strength of positive evidence after the evidence node has been passed on. Indicates the strength of negative evidence after the evidence node has been passed on. This indicates the positive label corresponding to the evidence node. The negative label represents the evidence node, and the reliability summation in the denominator is used to eliminate the scale difference caused by the different number of evidence nodes in different candidate pairs.

[0085] Understandably, the six strictness indicators, the disease progression coverage branch determination results, and the manual review results are not considered as supervisory targets that can be modified during training. Baseline-review candidate pairs without candidate efficacy set labels can perform evidence transfer and efficacy category determination, but do not participate in the calculation of the aforementioned classification loss; baseline-review candidate pairs without evidence nodes have both positive and negative evidence strength values ​​of 0 after transfer, and are prevented from outputting a single efficacy category result by the completeness strictness indicator and the comparability strictness indicator. The evidence association networks of training samples, validation samples, and baseline-review candidate pairs to be evaluated are indexed separately, and network parameters are frozen during the validation and inference phases.

[0086] S6. Based on entity alignment score, pairing score, report completeness, and candidate efficacy set, obtain the rigor judgment result; if the candidate efficacy set contains only one efficacy category and the rigor judgment result meets the preset conditions, output the efficacy category as the efficacy evaluation result of the baseline-review candidate pair; otherwise, output the candidate efficacy set as the efficacy evaluation result to be reviewed.

[0087] Specifically, step S6 includes: Determine whether the entity alignment score reaches the preset identity threshold and whether there is a conflict flag to obtain the identity strictness indicator. ; Determine whether the matching score reaches the preset matching threshold and whether the time direction conforms to the preset time constraint (i.e., the review time is not earlier than the baseline time and the time can be parsed normally) to obtain the matching strictness indicator. ; To determine whether there are non-unknown anatomical regions among the evidence nodes corresponding to the baseline report records and review report records, and to obtain the site strictness indicator, it is necessary to determine whether there is an overlap between the candidate anatomical regions to which the evidence nodes belong. ; Determine whether the completeness of both the baseline report record and the review report record meets the preset completeness threshold to obtain the completeness strictness indicator. ; The comparability stringency index is obtained by determining whether there are comparable target lesion treatment categories or evidence of disease progression coverage that meets pre-defined criteria within the candidate treatment set. Determine whether the number of treatment categories contained in the candidate treatment set is equal to 1 to obtain the conflict strictness indicator. ; The rigor assessment result is obtained by multiplying the identity rigor indicator, pairing rigor indicator, site rigor indicator, integrity rigor indicator, comparability rigor indicator, and conflict rigor indicator. If the candidate efficacy set contains only one efficacy category and the rigor assessment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

[0088] This application transforms the original efficacy evaluation process, which relied on manual comparison, into a structured, automated, and traceable data processing workflow by sequentially performing entity alignment and deduplication, examination pairing scoring, evidence node analysis, correlation network strength transfer, projection constraints, category determination, and rigor determination on original patient records and MRI reports from multiple sources. This reduces the burden of manual verification caused by scattered record sources and inconsistent report versions, and improves the standardization and automation of efficacy evaluation data processing.

[0089] In one embodiment of this application, the complete determination of the six indication quantities is shown in Table 5: Table 5 Determination of Indicator Quantity ; in, This indicates the result of the strictness determination.

[0090] It is understandable that six indicators, threshold, actual value, failure code and evidence node identifier should be output simultaneously, rather than just the product result.

[0091] This application makes a strictness determination based on entity alignment score, pairing score, report completeness, and the composition of the candidate efficacy set. Only when the strictness determination result meets the preset conditions and the candidate efficacy set contains only one efficacy category will the efficacy category be output as the final evaluation result. Otherwise, the candidate efficacy set will be output as a result pending review. This avoids outputting a single definite result in cases where the identity association is uncertain, the check pairing is unreliable, the report information is incomplete, or there are conflicts in the efficacy categories, thereby reducing the risk of the efficacy evaluation result being misclassified.

[0092] In one embodiment of this application, the output efficacy evaluation results or efficacy evaluation results pending review can be further organized into a unified output data structure. This structure includes at least: research_patient_id, baseline_accession, followup_accession, source_hash set, pair_score of baseline and followup examinations, response_set of candidate efficacy sets, hard_response of a single efficacy category output when strict conditions are met, benefit_set of aggregated results indicating whether it belongs to clinical benefit, confidence, conflict_state, decision_path from evidence node to efficacy category, six Strict indicators, strict_product, and review_reason. Any low-confidence output retains the aggregated values ​​and original evidence, and failures are not masked by empty strings. The source_hash set is used to store the verification values ​​of the source reports; conflict_state indicates a state of conflict, uncertainty, or insufficient evidence; the six Strict indicators represent the results of strictness checks such as entity alignment, check pairing, baseline integrity, review integrity, single candidate efficacy, and conflict status; strict_product represents the product of the above strictness indicators; and review_reason indicates the specific reason for requiring manual review.

[0093] In one embodiment of this application, when the patient's identity record or MRI report record contains a usable geographic field, the context source can be selected according to the priority of "city of work, city of birth, and city of origin," and assigned source reliability of 1, 0.8, and 0.7 respectively. The context control codes include CTX_AVAILABLE, CTX_SOURCE_RELIABILITY, CTX_YEAR_LAG, CTX_SOURCE_CODE, and CTX_CITY_HASH; each numerical code also has a "signature code __MISSING" indication. These codes, as context nodes in the evidence map, can participate in review priority or downstream outcome models, but do not change the PD coverage and CR, PR, and SD determination rules. Currently, 238 pairs of MRI data do not have a reliably connectable geographic field; therefore, this node has a mask of 0 in this embodiment and is not included in the efficacy consistency rate. Here, CTX_AVAILABLE indicates the existence of an available context field, CTX_SOURCE_RELIABILITY indicates the reliability of the context source, CTX_YEAR_LAG indicates the number of years the indicator year lags behind the review year, CTX_SOURCE_CODE indicates the type of context source, and CTX_CITY_HASH indicates an irreversible city summary. "Missing" indicates whether the value is missing; if missing, it is not replaced with a 0 value. The region name is only used for intra-institutional matching, and the output uses an irreversible city summary, not the original place name. The review year is used as the cutoff year; only city indicators from the cutoff year and the most recent year are allowed, and future years are prohibited. Each indicator outputs both a numerical code and a missing code.

[0094] The specific indicator characteristics and corresponding stable feature codes are illustrated in Table 6: Table 6 Feature groups and stable adjustment codes In Table 6, ECO_GDP represents the Gross Regional Product, ECO_GDP_PC represents the Gross Regional Product per capita, ECO_WAGE represents the average wage, ECO_RETAIL represents the total retail sales of consumer goods, ECO_DEPOSIT and ECO_SAVINGS represent indicators related to financial deposits and household savings, respectively; LAB_UNEMPLOYED represents indicators related to unemployment or underemployment, POP_REGISTERED represents the registered population; HLT_HOSPITAL, HLT_BED and HLT_DOCTOR represent indicators of health resources such as hospitals, hospital beds and doctors, respectively; EDU_COLLEGE represents indicators related to higher education, INS_MEDICAL and INS_UNEMPLOYMENT represent indicators related to medical insurance and unemployment insurance, respectively, and ENV_PM represents indicators of fine particulate matter or air particulate pollution.

[0095] In one embodiment of this application, a comprehensive confidence level can be further calculated based on the stringency determination result: ; in, Indicates the overall confidence level. Scoring is given for entity alignment. For pairing scores, , These are the baseline and the completeness of the follow-up report, respectively. The value corresponding to the clarity of evidence for newly diagnosed lesions. The penalty items corresponding to the baseline and the uncertainty mark of the review (calculated according to the preset coefficient and with an upper limit). This is a penalty calculated using a preset coefficient when the number of treatment categories in the candidate treatment set exceeds one. Penalties calculated based on preset coefficients for unresolved conflict indicators. This means truncating the calculation result to the interval between 0 and 1.

[0096] The confidence level is calculated by weighting the identity item, pairing item, baseline report completeness item, follow-up report completeness item, and the degree of known new lesions item according to preset weights, then subtracting uncertainty penalty, ambiguity penalty, and conflict penalty, and truncating it between 0 and 1. The uncertainty penalty can be obtained by multiplying the sum of the uncertainty quality on both sides of the baseline and follow-up by a certain coefficient and setting an upper limit. A fixed penalty can be applied for each additional candidate category in the candidate efficacy set, and a fixed penalty can also be applied for each unresolved conflict. When the candidate efficacy set is a single category, all six strictness indicators meet the conditions, and the above comprehensive confidence level reaches the preset threshold, the efficacy category is output as the efficacy evaluation result of the baseline-follow-up candidate pair. When the candidate efficacy set is a single category but the strictness indicator or confidence level does not meet the conditions, the single category is retained as a suggestion and the candidate efficacy set is output as the efficacy evaluation result to be reviewed. When the candidate efficacy set contains more than two efficacy categories, the candidate efficacy set, the corresponding conflict evidence path, and the review reason are output. The multi-value set is not covered by the existing judgment result, nor is this situation filled as disease stability. The aforementioned six stringency indicators verify the credibility of efficacy evaluation results from six independent dimensions: identity, pairing, location, completeness, comparability, and conflict. If any dimension fails to meet the criteria, a set of candidate efficacy outcomes is output instead of a single efficacy category. This screens out baseline-review candidate pairs with insufficient evidence or ambiguity for further review, rather than relying on a single comprehensive score to mask deficiencies in a particular dimension, thus improving the traceability and credibility of the efficacy evaluation results. The decisions made through manual review are written back as new records, without overwriting the original evidence nodes, candidate efficacy sets, or judgment processes, ensuring the recalculation and traceability of judgment results under the same input and version of parameters.

[0097] Specifically, the uncertainty penalty is the sum of the baseline and review uncertainty quality multiplied by 0.08 and capped at 0.35; the penalty is 0.13 for each additional candidate category in the response set; and the penalty is 0.10 for each unresolved conflict.

[0098] In one embodiment of this application, the efficacy evaluation results or efficacy evaluation results pending review can be further organized into a unified output data structure. This structure includes at least the study patient identifier, baseline examination number, re-examination examination number, source check value set, paired score, candidate efficacy set, efficacy category determined (if it is a single category), benefit label, six strictness indicators and their multiplication to obtain the strictness determination result, covered conflict records, and the review reason in the case of pending review. When any strictness indicator fails to meet the condition and outputs the candidate efficacy set as the efficacy evaluation result pending review, the candidate efficacy set and corresponding evidence node path are still retained, and the failure to pass the judgment is not masked by null values ​​or default values, which facilitates downstream systems or manual review links to read and trace in a structured manner.

[0099] The method flow of this application is as follows: Read the field mapping version and perform type conversion, value range validation, unit conversion, and failure handling for each source field; The source verification value is calculated for the normalized fields, and the entity alignment score is calculated based on the multi-field similarity to generate the research patient identifier; Calculate the deduplication composite score for reports under the same composite key, retain the one with the highest information completeness, and retain the association chain of other versions; Based on the time direction, inspection location, and pre-defined constraints related to the report, all baseline-review candidate pairs are generated; Calculate pairing scores for candidate pairs and determine candidate pairs according to preset winning rules; Each valid MRI report record is analyzed to extract measurement intervals, polarities, anatomical regions, and reliability to obtain evidence nodes; The evidence nodes are sorted by node identifier to remove duplicates. Four reference nodes are established for entity alignment, pairing, baseline report completeness, and review report completeness. Bidirectional connection edges with edge features are generated to obtain the evidence association network. In each transmission layer, the aggregate weights are normalized according to the connection edges of the target node, so that the reference node modulates the features of the evidence node, and the conflict information of the evidence node also updates the representation of the reference node in reverse, and the node state of the next layer is obtained by gating the residual. The positive and negative evidence strengths output by the evidence strength transmission are compared with the positive and negative evidence strengths of the rules, respectively, and the maximum value is taken to obtain the projected positive and negative evidence strengths, and the uncertainty flag and conflict flag are determined accordingly. The evidence nodes of the projected target lesions, non-target lesions and new lesions are determined according to the judgment order described in S51 to S56 to determine the candidate efficacy set. The six strictness indicators are calculated and multiplied to obtain the strictness judgment result, and the efficacy evaluation result or the efficacy evaluation result pending review is output accordingly.

[0100] To illustrate with a specific example, 238 pairs of real pre- and post-treatment MRI examination records were used to verify the evidence association network and the rigor determination process. For each pair of records, evidence nodes, identity reference nodes, pairing reference nodes, and two report completeness reference nodes for baseline and re-examination were constructed according to the above steps. Some candidate pairs have existing efficacy records available for comparison, covering four outcomes: complete remission, partial remission, stable disease, and disease progression. The comparison standard is the comparable efficacy conclusions in existing records. Samples that cannot be compared are not used in the denominator of the concordance rate but are still retained in the coverage and review load statistics. The algorithm does not use patient name or original identifier as the result field. The results are shown in Table 7. The selected coverage and benefit concordance rate data curves are shown in Table 7. Figure 2 As shown, where Figure 2 This section explains the changes in the benefit concordance rate and complete concordance rate between the output of this application and existing records under different comparable record coverage rates; the blue line represents whether the benefit is consistent in binary categories, and the orange line represents whether the benefit is completely consistent in pentatonic categories. Table 7 Results The comparison in this embodiment is based on existing recorded conclusions, not the gold standard of independent human adjudication. It does not demonstrate that CIEG-RL improves benefit consistency; the benefit consistency point estimate of the strict ensemble output is actually lower than the fixed threshold baseline. The 50% OOF result only indicates that a review ranking can be formed in 24 internally comparable cases, and the sample size is insufficient to support the performance claims. The validated technical effects are limited to: multi-valued responses are no longer backfilled by existing auto_response, and the six Strict gatings, conflict evidence, and reasons for failure can be output item by item.

[0101] Using the same 238 pairs of real baseline-review candidate pairs as validation samples, the baseline report parsing yielded 342 sets of rule-based evidence strength states, forming "region × predicate" states. The review report parsing yielded 482 sets, for a total of 824 sets. According to the merging rules, the merged results showed 774 sets of definitely dominant states, 23 sets of negatively dominant states, 24 sets of uncertain states, and 3 sets of conflicting states, including 3 pairs of candidate pairs with unresolved conflicts. Reversing the order of evidence node access for each report and recalculating, the merging results for the 824 sets of states remained unchanged, verifying the consistency of the rule-based positive and negative evidence strength merging methods when the order of evidence node access changes.

[0102] In the re-examination report, there were 6 candidate pairs whose evidence nodes for new lesions were clearly positive and 232 candidate pairs whose evidence nodes were uncertain. According to the judgment order from S51 to S56, the final candidate efficacy set was 42 candidate pairs whose efficacy set was a single efficacy category and 196 candidate pairs whose efficacy set contained two or more efficacy categories and whose efficacy set was uncertain. None of the 196 pairs used the existing single-category judgment results for backfilling.

[0103] To distinguish between the judgment results of the candidate efficacy set and the independent evaluation conclusions, this embodiment constructs a two-person independent evaluation and audit process on the same 238 candidate pairs: each candidate pair only provides the evaluators with a randomized candidate pair identifier. The candidate efficacy set described in this application, the efficacy category in the existing records, the six strictness indicators and the strictness judgment result obtained by multiplying them are not provided temporarily, and can only be compared after the two evaluators have completed their independent evaluations. The 238 candidate pairs are stratified according to the judgment status of the candidate efficacy set: 196 pairs where the candidate efficacy set is not a single category or is judged as an unclear category, 16 pairs where the output result of this application is inconsistent with the efficacy category in the existing records, 4 pairs where the strictness judgment result does not meet the preset conditions, and 22 pairs where the strictness judgment result meets the preset conditions and the candidate efficacy set is a single category. Each evaluator makes an independent judgment on the evidence nodes for target lesions, non-target lesions, new lesions, and the degree of uncertainty in measurement results. When the overall judgment is unclear, the specific reasons for insufficient evidence are recorded simultaneously. When the judgments of two evaluators are inconsistent, a third evaluator makes the final judgment. This final judgment is saved independently and is not written back to the merging thresholds for positive and negative evidence strengths, the judgment order described in S51 to S56, or the parameters of the evidence association network described in this application.

[0104] Each reader independently assigns CR, PR, SD, PD, or NE to the baseline and review data. The blinding table breaks down evidence into machine-verifiable fields: CR / PR / SD / PD / NE for target lesions, CR / NON_CR_NON_PD / PD / NE for non-target lesions, PRESENT / ABSENT / UNCERTAIN / NE for new lesions, and NONE / LOW / MODERATE / HIGH / NE for measurement uncertainty, with a separate explanation for insufficient evidence; the reason for insufficient evidence cannot be NONE if the overall response is NE. In case of discrepancy, a third reader makes the decision. The decision is independently versioned and cannot be rewritten with the evidence map threshold, conflict grid rules, or model parameters. The current embodiment has generated a real-case blinding audit queue and restricted data link, but has not yet loaded the human review results, therefore it is not presented as having achieved the gold standard for human adjudication or a verified improvement in accuracy. Among them, NE indicates that it cannot be evaluated; NON_CR_NON_PD indicates that the non-target lesions have neither achieved complete remission nor reached the intermediate state of disease progression, which can support disease stability but not support partial remission alone; PRESENT indicates the presence of new lesions, ABSENT indicates no new lesions, UNCERTAIN indicates that the evidence for new lesions is uncertain, NONE, LOW, MODERATE, HIGH, and NE indicate no measurement uncertainty, low uncertainty, moderate uncertainty, high uncertainty, and cannot be evaluated, respectively; NONE in the reason for insufficient evidence indicates that the reason for insufficient evidence has not been recorded or there is no reason for insufficient evidence.

[0105] The analysis following manual image review is pre-locked before unblinding: each reviewer is identified anonymously, their completion time is recorded, and their confidence level is 1 to 5; any discrepancies between two reviews must be resolved by a third person different from the first two reviewers, who must provide a final conclusion and justification. The analysis includes the following: five-category consistency rate of the two reviews; the proportion requiring adjudication; the five-category and benefit-based binary consistency rate of the algorithm relative to the final reference conclusion; the coverage rate of the automatic Strict output; and the stratified results of the four conflict priority layers. Confidence intervals are resampled using anonymized patient clusters, not pre- and post-treatment paired rows. After unblinding, evidence map rules, Strict thresholds, selective calibrators, or model parameters must not be refitted to improve any of the above indicators.

[0106] In this embodiment, all 238 candidate pairs have been constructed with four reference nodes: evidence node and entity alignment, pairing, baseline report completeness, and review report completeness. Among them, 104 candidate pairs have existing efficacy category labels that can be used for cross-validation, with 14, 39, 34, and 17 pairs respectively for complete remission (CR), partial remission (PR), stable disease (SD), and disease progression (PD). An evidence association network with connecting edges based on the same region, same source, and cross-time predicate is compared with a control network that does not distinguish the relationship between evidence nodes and establishes connecting edges for all evidence nodes and reference nodes in pairs. The comparison involves two networks that only change the edge set and propagation logit: fully connected evidence attention, type relationship mask plus source reliability, and structural prior. Both networks use the same node features, five-fold partitioning, 100 rounds of training per fold, number of propagation layers, and response readout, as shown in Table 8. Table 8 Comparison of Evidence Nodes like Figure 3 As shown, Figure 3 This is an out-of-sample descriptive audit of the dissemination of constrained evidence, where the color bars represent the number of cases, and CR, PR, SD, and PD represent complete remission, partial remission, stable disease, and disease progression, respectively. Figure 3A heatmap is used to show the cross-distribution between existing recorded efficacy categories and out-of-sample propagation efficacy categories. Cells on the main diagonal indicate complete consistency, while off-diagonal cells indicate category differences. This graph describes the distribution relationship between the two evaluation results and the sources of differences requiring further manual review, and is not directly used as a conclusion on accuracy improvement. It is evident that the average number of directed edges is reduced by approximately 36.93% compared to the fully connected approach. Simultaneously, the parameters of the five training folds show a change of more than 12.72 relative to the initial L2, proving that the propagation parameters on the retained edges actually participate in optimization, rather than being directly output by fixed rules. The type relationship model frozen after the last fold training traverses all 238 pairs of 824 "region × predicate" states. After merging the node polarity heads according to the time point—region—predicate ternary key, a P / N monotonically maximum projection is performed, resulting in 0 reductions in the original positive or negative support. The five-fold OOF subset covers an additional 449 states, with the same 0 reduction in support. After reversing the order of the evidence nodes, the maximum output probability difference is 5.96 × 10⁻⁶. -7 It is within the tolerance of floating-point operations.

[0107] The type mask's OOF accuracy and macro F1 on existing record labels are not higher than the fully connected control; therefore, this embodiment does not advocate its improvement in efficacy classification accuracy. Its validation function is limited to: removing message channels without structural basis according to explicit relationships; writing source reliability and structural priors into the normalized logit of each retained edge; allowing bidirectional propagation between factor nodes and evidence nodes; ensuring node rearrangement does not change the substantive output; and preventing the deletion of observed P / N support from network output. Existing record responses are not the gold standard for independent human adjudication; the final accuracy still relies on the blinded adjudication process.

[0108] This application also provides a data processing system for evaluating the efficacy of MRI before and after treatment, used to implement the aforementioned method for processing data for evaluating the efficacy of MRI before and after treatment, including: The deduplication module is used to obtain patient identity records and MRI report records, perform field standardization processing and multi-field similarity calculation to obtain entity alignment scores, and deduplicate duplicate reports under the same patient name based on entity alignment scores to obtain valid MRI report records. Based on examination time, examination site and entity alignment scores, the valid MRI report records are correlated and scored to obtain baseline-re-examination candidate pairs and their matching scores. The evidence analysis module is used to analyze the text content of each valid MRI report record in the baseline-review candidate pair to obtain multiple evidence nodes and the report completeness of each evidence node's report. The graph propagation module is used to construct an evidence association network containing reference nodes and evidence nodes, using entity alignment score, pairing score and report completeness as reference nodes, and to transmit evidence strength on the evidence association network to obtain the transmitted positive evidence strength and negative evidence strength. The evidence projection module is used to compare the transmitted positive evidence strength and negative evidence strength with the rule-based positive evidence strength and rule-based negative evidence strength, respectively, and take the maximum value to obtain the projected positive evidence strength and negative evidence strength. The category determination module is used to determine the efficacy category for each evidence node based on the projected positive and negative evidence strength, and to determine the candidate efficacy set according to the preset category priority. The results output module is used to obtain the rigor judgment result based on entity alignment score, pairing score, report completeness and candidate efficacy set; when the candidate efficacy set contains only one efficacy category and the rigor judgment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

[0109] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for processing MRI efficacy evaluation data before and after treatment, characterized in that, include: S1. Obtain patient identity records and MRI report records, perform field standardization processing and multi-field similarity calculation to obtain entity alignment scores, remove duplicate reports under the same patient name based on entity alignment scores to obtain valid MRI report records, and perform correlation scoring on valid MRI report records based on examination time, examination site and entity alignment scores to obtain baseline-re-examination candidate pairs and their pairing scores. S2. Analyze the text content of each valid MRI report in the baseline-review candidate pair to obtain multiple evidence nodes and the report completeness of each evidence node's report; S3. Using entity alignment score, pairing score and report completeness as reference nodes, construct an evidence association network containing reference nodes and evidence nodes, and transfer evidence strength on the evidence association network to obtain the positive evidence strength and negative evidence strength after transfer. S4. Compare the transmitted positive evidence strength and negative evidence strength with the rule positive evidence strength and rule negative evidence strength respectively, and take the maximum value to obtain the projected positive evidence strength and negative evidence strength. S5. Based on the projected positive and negative evidence strength, determine the efficacy category for each evidence node and determine the candidate efficacy set according to the preset category priority. S6. Based on entity alignment score, pairing score, report completeness, and candidate efficacy set, the rigor judgment result is obtained; If the candidate efficacy set contains only one efficacy category and the rigor judgment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

2. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 1, characterized in that: Step S1 specifically includes: Retrieve the fields of name, medical record number, patient card number, standardized name, date of birth and gender from the patient identity record, as well as the patient identification field from the MRI report record, and convert the format of the fields to obtain standardized fields; The similarity between corresponding fields in the patient identity record and the MRI report record is calculated based on the normalized fields, and the similarity of multiple fields is summed to obtain the entity alignment score. Based on the entity alignment score, determine whether there are multiple report records under the same patient name pointing to the same examination. If so, retain the one with the highest information completeness and delete the rest of the report records to obtain the valid MRI report record.

3. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 1, characterized in that: Step S2 specifically includes: The text content of each valid MRI report in the baseline-review candidate pair was analyzed, and the lesion descriptions obtained from the analysis were classified into target lesion evidence nodes, non-target lesion evidence nodes, and new lesion evidence nodes according to the anatomical region. To determine the completeness of a MRI report, one must determine whether both the sections seen and the conclusions are present simultaneously in the valid MRI report. Determine whether the number of located anatomical regions in valid MRI report records reaches a preset number to obtain the region integrity. Determine whether the number of evidence nodes in a valid MRI report reaches a preset number to obtain the evidence completeness. The report completeness of each evidence node is obtained by summing the chapter completeness, region completeness, and evidence completeness.

4. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 1, characterized in that: Step S3 specifically includes: S31. Using entity alignment score, pairing score and report completeness as reference nodes, and target lesion evidence nodes, non-target lesion evidence nodes and new lesion evidence nodes as evidence nodes, establish connection edges between reference nodes and evidence nodes, and between evidence nodes to obtain the evidence association network. S32. Evidence strength is transferred along the connecting edges in the evidence association network. The features of adjacent nodes are aggregated to obtain the positive and negative evidence strengths after transfer.

5. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 4, characterized in that: Step S31 specifically includes: Determine whether the anatomical regions between each pair of evidence nodes are the same to obtain the same region judgment result; Determine whether the node identifiers between any two evidence nodes point to the same lesion origin to obtain the homology judgment result; Determine whether each pair of evidence nodes belongs to different time points and has the same evidence predicate to obtain the cross-time same predicate judgment result; If any one of the following is true: same region judgment result, same source judgment result, or cross-time same predicate judgment result, then establish a connection edge between the corresponding two evidence nodes; A bidirectional connection edge is established between the reference node and the evidence node to obtain the evidence association network.

6. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 4, characterized in that: Step S32 specifically includes: Transform the features of the nodes at both ends of the connecting edge to obtain the structural priors; The reliability term is obtained by weighting the nodes at both ends of the connecting edge based on their reliability. Logarithmic transformation is performed based on the reliability of the connected edges to obtain the edge reliability term; The structural prior terms, reliability terms, and edge reliability terms are added together and normalized within the constraints of the connecting edges to obtain the weights of the weighted aggregation. Based on the weights of the weighted aggregation, the features of adjacent nodes are weighted and aggregated to obtain the positive and negative evidence strengths after propagation.

7. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 1, characterized in that: Step S4 specifically includes: The positive evidence strength after transmission is compared with the baseline positive evidence strength, and the maximum value is taken to obtain the projected positive evidence strength. The negative evidence strength after transmission is compared with the baseline negative evidence strength, and the maximum value is taken to obtain the projected negative evidence strength. An uncertainty flag is obtained based on whether the difference between the projected positive evidence strength and the projected negative evidence strength is less than a preset threshold. A conflict flag is obtained based on whether the strength of the projected positive evidence and the strength of the projected negative evidence simultaneously reach a preset threshold.

8. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 7, characterized in that: Step S5 specifically includes: S51. Efficacy categories include complete remission (CR), partial remission (PR), stable disease (SD) or progressive disease (PD), undefined category, and conflict category. The efficacy category of the target lesion evidence node is one of CR, PR, SD, or PD. If the uncertainty flag corresponding to the evidence node is true, the efficacy category of the evidence node is determined to be undefined. If the conflict flag corresponding to the evidence node is true, the efficacy category of the evidence node is determined to be conflict category. The efficacy category of the non-target lesion evidence node is one of A_non={SD}, A_non={CR}, or A_non={NN}, where NN represents incomplete remission and non-progression. The efficacy category of the new lesion evidence node is one of A_new={PD} and entering the disease progression locked state or no new lesion. S52. If the efficacy category of the target lesion evidence node is PD, or A_non={PD}, or A_new={PD}, and there is no negative evidence in the same anatomical region with a reliability higher than the disease progression evidence by at least a preset difference δ, then enter the disease progression coverage branch and determine PD as the only category in the candidate efficacy set; at this time, except for the node that triggers PD locking, all other evidence nodes are recorded as covered conflicts. S53. If the disease progression coverage branch is not entered, and the efficacy category of the target lesion evidence node is CR, and A_non∈{CR,NN}, and A_new=Ø, then complete remission CR is determined as the only category in the candidate efficacy set. S54. If the disease progression coverage branch and CR branch are not entered, and the efficacy category of the target lesion evidence node is PR, and A_new=Ø, and A_non≠{PD}, then PR is determined as the only category in the candidate efficacy set. S55. If the disease progression coverage branch, CR branch, and PR branch are not entered, and the disease progression coverage branch is not triggered, and the efficacy category of the target lesion evidence node is SD, or A_non={NN}, then SD will be determined as the only category in the candidate efficacy set. S56. If the conditions of any branch from S52 to S55 are not met, the candidate efficacy set is determined according to the efficacy category of each evidence node in accordance with the preset category priority rule; when the number of efficacy categories contained in the candidate efficacy set is greater than 1, the efficacy category of the corresponding baseline-review candidate pair is determined as an unclear category.

9. The method for processing MRI efficacy evaluation data before and after treatment as described in claim 1, characterized in that: Step S6 specifically includes: Determine whether the entity alignment score reaches the preset identity threshold and there is no conflict flag to obtain the identity strictness indicator; Determine whether the pairing score reaches the preset pairing threshold and whether the time direction conforms to the preset time constraint to obtain the pairing strictness indicator; Determine whether there are non-unknown anatomical regions among the evidence nodes corresponding to the baseline report records and review report records, and whether there is an intersection between the candidate anatomical regions to which the evidence nodes belong, to obtain the site strictness indicator; Determine whether the completeness of both the baseline report record and the review report record reaches the preset completeness threshold to obtain the completeness strictness indicator; Determine whether there are comparable target lesion treatment categories or evidence of disease progression coverage that meets preset conditions in the candidate treatment set to obtain a comparability strictness indicator; determine whether the number of treatment categories included in the candidate treatment set is equal to 1 to obtain a conflict strictness indicator. The rigor assessment result is obtained by multiplying the identity rigor indicator, pairing rigor indicator, site rigor indicator, integrity rigor indicator, comparability rigor indicator, and conflict rigor indicator. If the candidate efficacy set contains only one efficacy category and the rigor assessment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

10. A data processing system for evaluating the efficacy of MRI before and after treatment, characterized in that, The system is used to implement a method for processing MRI efficacy evaluation data before and after treatment as described in any one of claims 1-9, including: The deduplication module is used to obtain patient identity records and MRI report records, perform field standardization processing and multi-field similarity calculation to obtain entity alignment scores, and deduplicate duplicate reports under the same patient name based on entity alignment scores to obtain valid MRI report records. Based on examination time, examination site and entity alignment scores, the valid MRI report records are correlated and scored to obtain baseline-re-examination candidate pairs and their matching scores. The evidence analysis module is used to analyze the text content of each valid MRI report record in the baseline-review candidate pair to obtain multiple evidence nodes and the report completeness of each evidence node's report. The graph propagation module is used to construct an evidence association network containing reference nodes and evidence nodes, using entity alignment score, pairing score and report completeness as reference nodes, and to transmit evidence strength on the evidence association network to obtain the transmitted positive evidence strength and negative evidence strength. The evidence projection module is used to compare the transmitted positive evidence strength and negative evidence strength with the rule-based positive evidence strength and rule-based negative evidence strength, respectively, and take the maximum value to obtain the projected positive evidence strength and negative evidence strength. The category determination module is used to determine the efficacy category for each evidence node based on the projected positive and negative evidence strength, and to determine the candidate efficacy set according to the preset category priority. The results output module is used to obtain the rigor judgment result based on entity alignment score, pairing score, report completeness and candidate efficacy set; when the candidate efficacy set contains only one efficacy category and the rigor judgment result meets the preset conditions, the efficacy category is output as the efficacy evaluation result of the baseline-review candidate pair; otherwise, the candidate efficacy set is output as the efficacy evaluation result to be reviewed.

Citation Information

Patent Citations

  • Patient identity matching method based on multi-field similarity calculation

    CN110197214A