Image report generation method and system with adaptive correction of multi-source data

Abstract

This application provides a method and system for generating image reports with adaptive correction for multi-source data, aiming to solve the problem of report grading oscillations caused by baseline bias in multi-source data. The method utilizes systematic correction to perform modal correction calculations on multi-source data, generating consistent corrected lesion size measurements across channels, serving as a unified quantitative basis for lesion size descriptions in the report. The consistent corrected lesion size measurements across channels are compared with dynamic grading thresholds; when the threshold is exceeded, a preliminary report grading adjustment signal is generated. The stable grading adjustment results are written into the lesion assessment section of the diagnostic report, automatically generating or adjusting lesion description text, risk level labeling, and follow-up recommendations. This invention, through systematic bias correction and dynamic threshold adjustment, effectively filters out false grading signals, significantly improving the accuracy and stability of medical image report generation in multi-source data environments, providing reliable support for clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for generating image reports using adaptive correction of multi-source data. Background Technology

[0002] Intelligent generation and adjustment of medical imaging diagnostic reports is an important research direction in radiology-assisted diagnosis. Its core value lies in helping doctors quickly and accurately assess and describe lesions such as nodules, thereby improving diagnostic efficiency and consistency. In scenarios such as medical imaging screening, lesion size directly determines subsequent management strategies, such as whether follow-up, biopsy, or surgical intervention is needed. Therefore, the reliability of size measurement and its threshold judgment has a decisive impact on clinical decision-making. Currently, many imaging diagnostic platforms attempt to obtain target lesion size information through multi-data source fusion, including values ​​manually annotated by doctors, results automatically measured by AI algorithms, and historical measurement data directly imported from PACS systems. This multi-channel input seems to provide more comprehensive information support, but in actual operation, it exposes a prominent contradiction: the measurement benchmarks used by different data sources are not entirely consistent. Manual annotation relies on doctors' visual judgment of lesion boundaries, AI measurement is based on segmentation algorithms during model training, and PACS imported data may come from automatic measurements of equipment from different manufacturers or earlier versions of software. The systematic deviation between these benchmarks leads to significant dispersion in the size values ​​of the same lesion in different channels. When the actual size of a lesion is close to the clinical threshold—for example, a lesion may appear as 5.9 mm on one data source but 6.1 mm on another—this slight difference can cause the system to repeatedly trigger and deactivate threshold alerts. Doctors frequently see these alerts pop up and disappear during report writing, or the same image may show inconsistent alert statuses at different times. This constant fluctuation disrupts the continuity and reliability of the report, causing doctors to doubt the stability of the intelligent assistance system and thus reducing their reliance on and acceptance of it. Current technologies typically use simple arithmetic averages or directly select a single measurement value, ignoring the cumulative effect of systematic biases from different data sources over time. They also lack mechanisms to dynamically adjust decision boundaries based on data consistency, leading to false alarms of "false progression" near critical values. Therefore, ensuring the stability and consistency of the lesion size threshold alert triggering logic under the reality of baseline biases in size values ​​from multiple data sources, and preventing false, repeated alerts due to data channel switching, has become a critical problem that urgently needs to be solved in multi-source data adaptive correction image report generation systems. Summary of the Invention

[0003] This invention provides a method for generating image reports with adaptive correction of multi-source data, including: The lesion size measurement results generated by AI automatic measurement are used as the main channel data. The main channel source identifier is extracted. The current data source operation mode is identified according to the actual available channel combination. The lesion size measurement results of all available channels and their respective channel source identifiers are used as multi-source input data for report generation. The report generation engine queries the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the systematic correction amount accumulated by each channel. When a channel does not have a historical calibration record, the quality control parameters of the device to which the corresponding channel belongs are called as the initial correction benchmark. Based on the multi-source input data and the current data source operating mode, the sub-mode correction calculation is performed using the systematic correction amount of each channel to generate consistent corrected lesion size measurement results between channels; Obtain a preset lesion size grading threshold, and dynamically adjust the preset lesion size grading threshold in combination with the multi-source input data, historical calibration records and the current data source operation mode to generate a dynamic grading threshold that is adapted to the current channel combination. The corrected lesion size measurement results between the channels are compared with the dynamic grading threshold. When there is a situation that exceeds the risk level boundary, a preliminary report grading adjustment signal is generated. Simultaneously extract the lesion size measurement results after correction for each channel involved in the report grading adjustment signal, count the number of channels exceeding the dynamic grading threshold and the corresponding exceedance of each channel, identify and filter false grading adjustment signals, and obtain stable grading adjustment results; The results of the stability grading adjustment are written into the lesion assessment section of the diagnostic report, and lesion description text, risk level labeling, and follow-up recommendations are generated or adjusted.

[0004] Furthermore, the process involves obtaining the lesion size measurement results generated by AI automatic measurement as the main channel data, extracting the main channel source identifier, identifying the current data source operating mode based on the actual available channel combinations, and using the lesion size measurement results of all available channels along with their respective channel source identifiers as multi-source input data for report generation, including: Extract the target lesion from the current medical image and perform automatic contouring. Identify the long diameter and short diameter measurement values ​​of the lesion and combine them to form the lesion size measurement result generated by AI automatic measurement. Simultaneously generate the main channel source identifier and bind the lesion size measurement result and the main channel source identifier to store it in the main channel data record. Read the segmentation confidence score from the main channel data record. When the segmentation confidence score is lower than the confidence threshold, or when the lesion size measurement result falls into the grading critical range, trigger a manual annotation prompt signal. After manual annotation is completed, the lesion size measurement results and manual annotation source identifiers of the manually annotated channels are extracted and included in the current available channel set. The current data source operation mode is identified according to the number of channel types contained in the available channel set, and the corresponding mode label and multi-source input data of all available channels are output.

[0005] Furthermore, the report generation engine queries the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. These historical calibration records contain the accumulated systematic correction amounts for each channel. When a channel does not have a historical calibration record, the quality control parameters of the corresponding channel's device are used as the initial correction benchmark, including: Extract the channel source identifier of each available channel from the multi-source input data, make a historical calibration record query request for each channel source identifier, retrieve all related historical calibration entries using the channel source identifier as an index, and sum the correction deviation values ​​of all historical calibration entries under the same channel after assigning decreasing weights according to the calibration execution time from recent to distant to obtain the systematic correction amount of the channel. For each channel in the systematic correction set, it is determined whether there is a valid historical calibration record. If no historical calibration entry is found for a certain channel, the most recent phantom calibration result of the imaging equipment to which the corresponding channel belongs is extracted from the equipment archive, and the inherent measurement deviation value of the equipment in the phantom calibration result is written into the systematic correction set as the initial correction benchmark. The quality control parameters of the device include spatial resolution deviation value obtained based on standard phantom scanning, edge detection deviation value caused by noise, etc. The systemic correction set is traversed, and the systemic correction values ​​are retained for channels with historical calibration records. The initial correction reference is retained for channels with written initial correction references. The correction value information of all channels is integrated to form a channel calibration configuration table.

[0006] Furthermore, the step of performing sub-mode correction calculations based on the multi-source input data and the current data source operating mode, using the systematic correction amount of each channel, to generate consistent corrected lesion size measurement results across channels, includes: Read the current data source operating mode label from the channel calibration configuration table; When the operating mode label indicates the three-source full-function mode, the correction values ​​for the AI ​​channel and the manual channel are extracted from the channel calibration configuration table respectively. The measurement results of the AI ​​channel and the manual channel are subtracted from their respective correction values ​​to obtain two independent correction values. Historical measurement time series are extracted from the multi-source input data, and the trend prediction value is calculated using the linear regression method. The historical channel correction values ​​are subtracted from the trend prediction value to obtain the historical independent correction values. The arithmetic mean of the correction values ​​is taken to obtain the measurement result after three-source joint correction.

[0007] Furthermore, the step of obtaining a preset lesion size grading threshold, and dynamically adjusting the preset lesion size grading threshold in conjunction with the multi-source input data, historical calibration records, and the current data source operating mode to generate a dynamic grading threshold adapted to the current channel combination includes: The basic grading configuration is formed by reading the preset list of lesion size grading thresholds and the upper and lower boundary offsets associated with each grading node from the report template configuration file. The boundary offset is adjusted according to the current data source operating mode label. When the operating mode is single source mode, the upper and lower boundary offsets of each hierarchical node are multiplied by the widening coefficient. When the operating mode is dual source mode or three source full-function mode, the difference between the maximum and minimum values ​​in the measurement results after calibration of each channel is calculated as the dispersion value. The dispersion value is divided by the reference dispersion value to obtain the adjustment factor. The upper and lower boundary offsets of each hierarchical node are multiplied by the adjustment factor. For each graded node value, the adjusted upper boundary offset is added to obtain the dynamic upper boundary, and the adjusted lower boundary offset is subtracted to obtain the dynamic lower boundary. These are then arranged in order of risk level to form dynamic graded thresholds.

[0008] Furthermore, the step of comparing the corrected lesion size measurement results consistent across channels with the dynamic grading threshold, and generating a preliminary report grading adjustment signal when a situation exceeds the risk level boundary, includes: Traverse each grading node in the dynamic grading threshold, compare the corrected lesion size measurement results with the dynamic upper boundary and dynamic lower boundary of each grading node respectively, mark the boundary exceedance state, and obtain the boundary exceedance state mark set. Based on the set of boundary exceedance status markers, the classification node to which the AI ​​channel belongs before correction is compared with the classification node to which it belongs after correction to determine the change in classification position. When a boundary exceedance status exists, a preliminary report classification adjustment signal is generated. The report classification adjustment signal includes the risk level label before the change, the risk level label after the change, and the trigger classification node value.

[0009] Furthermore, the synchronous extraction of the corrected lesion size measurement results for each channel involved in the report grading adjustment signal, the statistical analysis of the number of channels exceeding the dynamic grading threshold and the corresponding exceedance magnitude for each channel, the identification and filtering of false grading adjustment signals, and the obtaining of stable grading adjustment results include: Based on the trigger grading node value in the report grading adjustment signal, the corrected lesion size measurement results of each available channel are extracted, the number of channels exceeding the dynamic grading threshold is counted, and the exceedance amplitude of each exceeding channel is calculated to form an exceedance feature set. Based on the current data source operating mode label, the corresponding misjudgment identification threshold is read. When the number of exceeding channels is lower than the lower limit of the number of channels, or when the exceeding amplitude of all exceeding channels is lower than the lower limit of the amplitude, the report grading adjustment signal is marked as a false signal and filtered; otherwise, it is retained as a stable grading adjustment result.

[0010] Furthermore, the step of writing the stability grading adjustment results into the lesion assessment section of the diagnostic report, and generating or adjusting the lesion description text, risk level labeling, and follow-up recommendations, includes: Based on the changed risk level labels in the stable grading adjustment results, the corresponding lesion description text template, risk level labeling position and follow-up suggestion template are matched from the report template configuration file. The corrected lesion size measurement results and changed risk level labels are filled into the corresponding positions to form the lesion assessment content. If it is a single-source mode, the single measurement status label and short-term re-examination suggestion label are extracted from the channel calibration configuration table and written as additional annotations into the lesion assessment content. The lesion assessment content is then written into the lesion assessment section of the diagnostic report.

[0011] On the other hand, the present invention discloses an image report generation system for adaptive correction of multi-source data, the system comprising: The data acquisition and pattern recognition module is used to acquire the lesion size measurement results generated by AI automatic measurement as the main channel data, extract the main channel source identifier, identify the current data source operation mode according to the actual available channel combination, and use the lesion size measurement results of all available channels and their respective channel source identifiers as multi-source input data for report generation. The channel calibration query module is used by the report generation engine to query the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the systematic correction amount accumulated by each channel. When a channel does not have a historical calibration record, the quality control parameters of the device to which the corresponding channel belongs are called as the initial correction benchmark. The sub-mode correction calculation module is used to perform sub-mode correction calculations based on the multi-source input data and the current data source operation mode, using the systematic correction amount of each channel, to generate consistent corrected lesion size measurement results between channels. The dynamic grading threshold generation module is used to read the preset lesion size grading threshold from the report template configuration, and dynamically adjust the preset lesion size grading threshold in combination with the multi-source input data, historical calibration records and the current data source operation mode to generate a dynamic grading threshold that is adapted to the current channel combination. The grading adjustment signal generation module is used to compare the consistent corrected lesion size measurement results between the channels with the dynamic grading threshold, and generate a preliminary report grading adjustment signal when there is a situation that exceeds the risk level boundary. The false signal filtering module is used to simultaneously extract the corrected lesion size measurement results of each channel involved in the report grading adjustment signal, count the number of channels exceeding the dynamic grading threshold and the corresponding exceedance of each channel, identify and filter false grading adjustment signals, and obtain stable grading adjustment results; The report automatic generation module is used to write the stability grading adjustment results into the lesion assessment section of the diagnostic report, and to generate or adjust the lesion description text, risk level labeling, and follow-up recommendations.

[0012] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an image report generation method based on adaptive correction of multi-source data. It proposes an integrated solution to address systematic biases in lesion size measurement, the need for dynamic grading threshold adjustment, and report generation consistency issues in multi-source data channels of medical imaging. This invention extracts multi-source input data and channel source identifiers, combines historical calibration records and device-level quality control parameters, and performs modal correction calculations to ensure consistency in lesion size measurement results across channels. Simultaneously, it dynamically adjusts the target lesion size grading threshold, generates evaluation criteria adapted to the current channel combination, and filters out false signals by comparing the correction results with the thresholds, outputting stable grading adjustment results. Finally, it automatically generates a diagnostic report containing lesion description, risk level, and follow-up recommendations. This invention significantly improves the accuracy and reliability of report generation in multi-source data environments, achieving intelligent and adaptive medical image assessment and providing precise support for clinical decision-making. Attached Figure Description

[0013] Figure 1 This is a flowchart of the image report generation method for multi-source data adaptive correction according to the present invention.

[0014] Figure 2 This is a schematic diagram of the image report generation method for multi-source data adaptive correction according to the present invention.

[0015] Figure 3 This is another schematic diagram of the image report generation method for multi-source data adaptive correction according to the present invention.

[0016] Figure 4 This is a schematic diagram illustrating the principle of image report generation for multi-source data adaptive correction in an embodiment of the present invention.

[0017] Figure 5 This is a schematic diagram of the image report generation system for multi-source data adaptive correction according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0019] like Figures 1 to 5 The image report generation method and system for multi-source data adaptive correction in this embodiment may specifically include: Step S101: Obtain the lesion size measurement results generated by AI automatic measurement as the main channel data, extract the main channel source identifier, identify the current data source operation mode according to the actual available channel combination, and use the lesion size measurement results of all available channels together with the channel source identifier as the multi-source input data for report generation.

[0020] The image segmentation module is invoked to automatically extract the contour of the target lesion in the current medical image, obtaining the major and minor diameter measurements of the lesion. These measurements are then combined to form the lesion size measurement result generated by AI automatic measurement. Simultaneously, a main channel source identifier containing a measurement timestamp and the segmentation module version number is generated. The lesion size measurement result and the main channel source identifier are bound and stored in the main channel data record. Based on the patient's unique identifier and the examination site code, a historical image retrieval request is initiated to the PACS database. If the retrieval results contain previous examination records of the same patient and the same site, the historical lesion size measurement results and their corresponding historical channel source identifiers are extracted from these records. The historical lesion size measurement results are arranged by examination date to form a historical measurement time series. If the retrieval results are empty, the historical data channel status is marked as unavailable. For the main channel data record, the segmentation confidence score attached to the AI ​​measurement output is read. If the segmentation confidence score is lower than the preset confidence threshold, or the lesion size measurement result falls within the preset grading critical interval, or the absolute difference between the lesion size measurement result and the most recent measurement value in the historical measurement time series exceeds the preset deviation limit, a manual annotation prompt signal is triggered. After annotation is completed, the lesion size measurement result and manual annotation source identifier of the manually annotated channel are extracted and included in the current available channel set. The current data source operation mode is identified according to the number of channel types included in the available channel set, and the corresponding mode label and multi-source input data of all available channels are output.

[0021] In the intelligent generation of medical image diagnostic reports, accurate measurement of lesion size and effective integration of multi-source data are crucial for clinical decision-making. This implementation method focuses on a medical image screening scenario, detailing the acquisition of lesion size measurement results, the retrieval and matching of historical data, and the specific implementation process of multi-channel data fusion.

[0022] In one possible implementation, the image segmentation module uses a deep learning-based semantic segmentation method to extract the contours of target lesions in medical images.

[0023] Specifically, the module receives a preprocessed medical image sequence as input, locates suspected lesion areas through layer-by-layer scanning, and then performs pixel-level boundary recognition on each lesion area, outputting a two-dimensional contour mask of the lesion. Based on the contour mask, the measurement module extracts the maximum line length along the main axis of the lesion contour as the major diameter measurement value, and extracts the maximum line length along the direction perpendicular to the main axis as the minor diameter measurement value, combining the two to form a complete lesion size measurement result.

[0024] For example, the manual annotation channel serves as an anomaly detection trigger mechanism rather than a routine operating procedure. Its design intention is that the system only issues a manual annotation prompt to the doctor when the AI ​​segmentation edge significantly deviates from the true contour of the lesion, the measurement value is within the critical range affecting clinical decision-making, or the current measurement result is significantly inconsistent with the historical measurement trend of the same patient. After the doctor completes the intervention, the manual channel data becomes effective, thereby triggering the system's multi-source fusion correction process. In the multi-channel architecture, the manual annotation channel is only activated when the system detects a suspicious deviation in the AI ​​measurement result. Its core function is to solve the problem of smooth transition when AI errors occur, avoiding unreasonable jumps in report grading due to occasional AI measurement anomalies. In actual clinical work, the accuracy of AI automatic measurement has reached a high level, and doctors can usually complete the diagnosis by directly referring to the AI ​​measurement results.

[0025] It should be noted that PACS historical image retrieval achieves precise matching based on the patient's unique identifier and the examination site code. The patient's unique identifier is usually an ID card number or a patient number assigned internally by the hospital, and the examination site code follows the standard coding specifications for medical imaging. After the retrieval request is sent to the PACS database, a list of all previous examination records for that patient at the same examination site is returned. For each previous examination record, the stored historical lesion size measurement results and the corresponding channel source identifier are extracted and arranged in chronological order of examination date to construct a historical measurement time series. If no matching previous record exists in the PACS database, the historical data channel status is marked as unavailable, and subsequent processing relies solely on the main channel data.

[0026] In one embodiment, the manual annotation triggering mechanism is based on three independent triggering conditions. The first triggering condition is the segmentation confidence score, which is output synchronously by the image segmentation module during contour extraction. This score reflects the degree of certainty of the segmentation algorithm in identifying the current lesion boundary, and its value is typically between zero and one. When this score is lower than a preset confidence threshold, a manual annotation prompt is triggered. The second triggering condition is the grading critical interval, which refers to the sensitive area where the lesion size measurement result falls near the clinical grading threshold. For example, when the clinical threshold is set to six millimeters, the grading critical interval can be set to the range of 5.5 millimeters to 6.5 millimeters. A lesion size falling into this interval means that measurement error may lead to changes in the grading result. The third triggering condition is the trend deviation judgment. By calculating the absolute difference between the current measurement value and the most recent measurement value in the historical measurement time series, when this difference exceeds a preset deviation limit, it indicates that the current measurement result is significantly inconsistent with the historical trend, and manual review and confirmation are required.

[0027] Specifically, when any one of the three triggering conditions mentioned above is met, the report generation interface displays a manual annotation prompt signal, prompting the doctor to manually delineate and measure the lesion boundary. After the doctor completes the annotation, the lesion size measurement result from the manual annotation channel is extracted and a corresponding manual annotation source identifier is generated. This identifier includes the annotating doctor's employee number and the annotation timestamp. Specifically, in this embodiment, the lesion includes various lesions such as pulmonary tuberculosis, cerebral hematoma, liver nodules, and breast nodules. The manual annotation channel data is then incorporated into the currently available channel set.

[0028] Understandably, the identification of the data source operating mode is based on the number of channel types included in the currently available channel set. If only the main channel data is available, it is identified as a single-source mode and the corresponding mode label is output; if both the main channel and historical data channels are available, or both the main channel and manually labeled channels are available, it is identified as a dual-source mode; if all three channels are available, it is identified as a three-source full-function mode. The final output mode label and the multi-source input data of all available channels serve as the basic input for the subsequent report generation process. It should be noted that this method does not require all three data channels to be available simultaneously to operate, but rather adaptively selects the corresponding operating mode based on the actual available channel combination. For genuine first-time patients with no external hospital data to import, the PACS system does not contain previous examination records for the same patient and the same site. In this case, the historical data channel is automatically marked as unavailable, the system degenerates into single-source mode, and runs according to the regular process, relying only on AI channel data. The report then marks the current result as based on a single measurement and suggests a short-term follow-up to establish a measurement baseline.

[0029] Specifically, the applicable scenarios for each mode are as follows: If only AI data is available, it is a single-source mode. The system uses device-level quality control parameters to perform basic correction on the AI ​​measurement results and marks the measurement confidence level as limited. If both AI data and historical data are available, it is a dual-source mode. The system initiates historical trend consistency verification, cross-validating the current measurement results with historical measurement sequences. If both AI data and doctor-manually intervened data are available, it is a dual-source mode. If all three channels of data are available, i.e., a three-source full-function mode, the system initiates full-function dynamic threshold adjustment and false signal filtering. Switching between the above four modes is completely automated and requires no manual selection.

[0030] In step S102, the report generation engine queries the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the systematic correction amount accumulated by each channel. When the system detects that a channel does not have a historical calibration record, it automatically calls the device-level quality control parameters as the initial correction benchmark for that channel.

[0031] The channel source identifiers of each available channel are extracted from the multi-source input data. For each channel source identifier, a historical calibration record query request is initiated to the channel calibration module. The channel calibration module uses the channel source identifier as an index key to retrieve all historical calibration entries associated with that channel. Each historical calibration entry includes the calibration execution time and the correction deviation value generated by that calibration. The correction deviation values ​​of all historical calibration entries under the same channel are summed after being weighted in descending order of calibration execution time, from most recent to oldest, to obtain the systematic correction amount for that channel. The systematic correction amount set corresponding to each available channel is output. For each channel in the systematic correction amount set, it is determined whether there is a valid historical calibration record for that channel. If no historical calibration entry is found for a channel, the channel is marked as having no calibration record. The most recent phantom calibration result of the imaging equipment to which the channel belongs is extracted from the equipment archive. The phantom calibration result is generated by comparing the nominal value of the phantom with the scanning measurement performed periodically by the equipment using a standard phantom, and includes the inherent measurement deviation value of the equipment. The inherent measurement deviation value of the equipment is used as the initial correction benchmark for that channel and written into the systematic correction amount set. For channels that have completed the initial correction benchmark writing, it is checked whether the current data source operating mode is single-source mode. If it is currently in single-source mode and there is no valid phantom calibration result for the device in the device archive, the current lesion size measurement result is marked as a single measurement, and a short-term review suggestion label is generated and associated with the measurement record. The short-term review suggestion label includes the suggested review time interval. The systematic correction quantity set is traversed. For channels with historical calibration records, their systematic correction quantities are retained; for channels with initial correction benchmarks, their initial correction benchmarks are retained; and for channels marked as single measurement, their status labels and the short-term review suggestion labels are retained. The correction quantity information and status labels of all the above channels are integrated to form a channel calibration configuration table.

[0032] In the channel calibration process of intelligent medical image reporting, the querying of historical calibration records and the calculation of systematic corrections constitute the foundational steps of multi-source data fusion. This implementation focuses on the scenario of measuring lesion size in medical imaging, detailing the working mechanism of the channel calibration module, the specific implementation of time-weighted accumulation, and the calling logic of device-level quality control parameters. The above design of the channel calibration module provides a unified calibration framework for various sources of systematic bias in clinical practice. The first typical scenario involves benchmark bias between different AI systems: when a patient is measured using one AI system at hospital A and then transferred to hospital B, where another AI system is used for measurement, the measurement benchmarks of the two AI systems may have inherent differences. If the doctor at hospital B imports historical measurement values ​​from hospital A through the Image Archiving and Communication System (PACS), the difference between the two values ​​may cause clinical confusion. This method distinguishes measurement data from different AI systems by channel source identification, eliminates benchmark bias by utilizing the systematic corrections of each channel, and avoids misclassification due to cross-system bias in subsequent steps through dynamic grading threshold adjustment and false signal filtering mechanisms. The second typical scenario involves AI model version iteration: When a hospital's AI system is upgraded from version V1.0 to V2.0, the measurement results for the same lesion in the new version may have a systematic bias, for example, the measurement results of version V2.0 are generally larger than those of version V1.0. For two follow-up visits of the same patient, if the current examination uses version V2.0 and the historical record comes from version V1.0, a direct comparison may lead to the erroneous conclusion that the lesion has increased in size. This method identifies the measurement data of different versions by using the segmentation module version number field in the channel source identifier, and incorporates the systematic bias between versions into the correction calculation, thereby avoiding measurement artifacts introduced by version iteration. The third typical scenario involves differences in different scanning equipment and image parameters: Different models of CT or MRI equipment have inherent differences in spatial resolution, noise level, etc. The measurement results of the same equipment under different scanning parameters (such as slice thickness, reconstruction kernel, window width and window level, sequence type) may also be biased. These device-level and parameter-level differences constitute data alignment issues. This method establishes independent correction benchmarks for each device using phantom calibration results stored in the device archive, and achieves cross-device measurement data normalization correction by combining the device information recorded in the channel source identifier. Specifically, the device-level quality control parameters are generated by the device's periodic phantom calibration; when in single-source mode and no phantom calibration parameters are available, the current measurement is marked as a single measurement and a short-term review is recommended to establish a measurement baseline.

[0033] In one possible implementation, the channel calibration module uses a key-value index structure to store the historical calibration entries for each channel.

[0034] Specifically, each channel source identifier serves as a unique index key, corresponding to a time-sorted list of calibration entries. Each calibration entry records the execution time and correction deviation value of a calibration operation, whereby the correction deviation value represents the difference between the measured channel output value and the reference standard value during that calibration. When a query request arrives, the channel calibration module locates the corresponding list of calibration entries based on the channel source identifier and returns all entries for subsequent calculations.

[0035] It should be noted that the core idea of ​​time-weighted accumulation is to give higher weight to recent calibration results.

[0036] In one embodiment, the weight allocation adopts an exponential decay method, using the calibration execution time of the most recent calibration entry as a benchmark, calculating the number of days between each historical calibration entry and the benchmark time, and the larger the number of days between the benchmark time and the benchmark time, the smaller the weight value allocated.

[0037] Specifically, if a channel has five historical calibration entries, with the most recent calibration dates being 0 days, 30 days, 90 days, 180 days, and 360 days respectively, then the corresponding weight values ​​decrease sequentially, with the most recent calibration entry having the highest weight value. The systematic correction amount for that channel is obtained by multiplying the correction deviation value of each calibration entry by its corresponding weight value, summing these values, and then dividing by the total sum of weight values. This method allows the systematic correction amount to better reflect the recent equipment status and reduces the impact of outdated calibration data on current measurement corrections.

[0038] For example, when no historical calibration entry is found in the calibration record database for a certain AI measurement channel, the channel calibration module marks the channel as having no calibration record. At this time, the device-level quality control parameter call process is triggered to extract the phantom calibration results of the imaging equipment to which the channel belongs from the device archive.

[0039] Specifically, phantom calibration is a quality control operation performed periodically by medical imaging equipment. During the operation, the equipment scans using a standard phantom with known nominal dimensions. This standard phantom is typically a resin block containing multiple standard spheres or cylinders, and the diameter of each embedding is precisely measured and recorded as the phantom's nominal value. After performing a CT scan on the phantom, the equipment measures the size of each embedding in the image and compares this measurement with the phantom's nominal value. The difference between the two is the equipment's inherent measurement deviation. This deviation reflects the equipment's measurement characteristics within the current calibration cycle and serves as the initial correction benchmark for that channel when no historical calibration records are available.

[0040] In one embodiment, the device archive stores the quality control records of each imaging device by its unique device number. Each device's quality control record includes the execution date of the most recent phantom calibration, the phantom number, the deviation values ​​at each measurement point, and the calibration validity period. When the channel calibration module initiates a quality control parameter call request, it locates the corresponding device's quality control record based on the device number field included in the channel source identifier, extracts the device's inherent measurement deviation value, and writes it into the systematic correction set. Further, for channels that have completed the initial correction benchmark writing, the current data source operating mode needs to be detected. If only the AI ​​measurement channel is currently available, and both the historical data channel and the manual annotation channel are unavailable, it is determined to be a single-source mode. In single-source mode, if the device archive does not contain a valid phantom calibration result for that device, it means that the current measurement has neither historical calibration accumulation information nor device-level quality control parameters as a correction basis.

[0041] Understandably, the measurement results in this scenario are subject to significant uncertainty. This scenario corresponds to the extreme case where a newly diagnosed patient is undergoing their first examination and the equipment lacks a valid phantom calibration record. The system marks the current lesion size measurement result as a single measurement and simultaneously generates a short-term follow-up recommendation label. This short-term follow-up recommendation label includes a suggested follow-up interval, which is determined based on the clinical grading range of the lesion size; lesions with sizes close to the grading threshold correspond to shorter follow-up intervals. This label is linked to the current measurement record and presented as a prompt in the diagnostic report, reminding clinicians to pay attention to establishing a measurement baseline.

[0042] Preferably, the integration process of the channel calibration configuration table traverses all channel entries in the systematic correction set. For channels with historical calibration records, the systematic correction calculated using time-weighted summation is retained; for channels that call device-level quality control parameters, their initial correction reference is retained; for channels marked as single measurement status, their status label and short-term review suggestion label are retained. The above information is organized according to the channel source identifier to form a structured channel calibration configuration table. This configuration table contains the correction value or status label for each channel, serving as the input basis for subsequent sub-mode calibration calculations.

[0043] Step S103: Based on the multi-source input data and the current data source operating mode, perform sub-mode correction calculation using systematic correction amount to generate consistent corrected lesion size measurement results between channels, which serve as a unified quantitative basis for lesion size description in the report.

[0044] The current data source operating mode label is read from the channel calibration configuration table. If the operating mode label indicates a single-source mode, the correction amount corresponding to the AI ​​channel is extracted from the channel calibration configuration table. The correction amount is subtracted from the lesion size measurement result of the AI ​​channel to obtain the measurement result after single-source calibration. If the operating mode label indicates a dual-source mode combining AI and manual intervention, the lesion size measurement result of the manually labeled channel is extracted from the multi-source input data as a reference benchmark value. The difference between the original measurement result of the AI ​​channel and the reference benchmark value is used as the channel deviation value. The channel deviation value and the current measurement timestamp are written as a new correction entry into the calibration record database of the channel calibration module. At the same time, the reference benchmark value and the original measurement result of the AI ​​channel are summed with a weight of 50% each to obtain the measurement result after dual-source manual calibration. If the operating mode label indicates a three-source full-function mode, the correction values ​​for the AI ​​channel and the manual channel are extracted from the channel calibration configuration table. The measurement results of the AI ​​channel and the manual channel are subtracted from their respective correction values ​​to obtain two independent correction values. At the same time, historical measurement time series are extracted from the multi-source input data, and a trend prediction value is calculated using a linear regression method. The correction values ​​for the historical channels are extracted from the channel calibration configuration table, and the trend prediction value is subtracted from the correction value to obtain historical independent correction values. The arithmetic mean of the three independent correction values ​​is taken to obtain the measurement result after three-source joint correction. This corrected measurement result is used as the unified quantitative basis for the description of lesion size in the report.

[0045] In the multi-mode correction process of intelligent medical image reports, selecting the corresponding correction path based on the current data source operating mode is a crucial step in ensuring the consistency of lesion size measurement. In single-source mode, device-level quality control parameters are applied to the lesion size measurement results of the AI ​​channel for correction. In dual-source mode combining AI and historical data, historical trend consistency verification is initiated, cross-validating the current AI measurement results with historical measurement sequences and performing correction. In dual-source mode combining AI and manual intervention, the lesion size measurement results manually annotated by the doctor are used as a reference benchmark. The channel deviation value between the original AI measurement results and the manual measurement results is calculated, and the two are merged to generate the corrected result. This channel deviation value is then written as a new correction entry into the channel correction module for subsequent calibration accumulation. In three-source full-function mode, the systematic correction amounts of all channels are comprehensively applied for joint correction. This implementation method focuses on the lung CT lesion size measurement scenario, detailing the correction calculation logic and specific implementation process of each mode in single-source, dual-source, and three-source full-function modes.

[0046] In one possible implementation, the operating mode label is obtained from the header field of the channel calibration configuration table. The operating mode label is an enumeration type, with values ​​including single-source mode, dual-source mode combining AI and historical data, dual-source mode combining AI and manual intervention, and three-source full-function mode. The calibration process enters the corresponding processing branch based on the label value, and different branches are executed mutually exclusively.

[0047] For example, in single-source mode, only the AI ​​measurement channel is available, and the correction calculation is performed using a direct subtraction method.

[0048] Specifically, the correction amount corresponding to the AI ​​channel is extracted from the channel calibration configuration table. This correction amount comes from device-level quality control parameters or historical calibration accumulation. The lesion size measurement result of the AI ​​channel is subtracted from this correction amount to obtain the measurement result after single-source calibration. This calibration method is suitable for scenarios such as the first inspection or when historical data is missing.

[0049] It should be noted that the dual-source model combining AI and historical data introduces a historical trend consistency verification mechanism. Linear regression plays a central role in this model, with its input being historical measurement time series and its output being the trend slope and trend intercept describing the changes in measured values ​​over time.

[0050] Specifically, the examination date of each record in the historical measurement time series is converted into the number of days since the first examination as the independent variable, and the corresponding lesion size measurement value is used as the dependent variable. The least squares method is used to solve for the slope and intercept values ​​that minimize the sum of squared residuals. The slope value represents the average rate of change of lesion size over time; a positive value indicates a trend of lesion enlargement, a negative value indicates a trend of lesion shrinkage, and a zero value indicates size stability. The intercept value represents the baseline size estimate at the time of the first examination. Substituting the number of days since the first examination into the slope and intercept, the trend prediction value at the current time point is calculated. This prediction value reflects the expected measurement result based on historical change patterns. Further, the current measurement result of the AI ​​channel is compared with the trend prediction value. If the absolute value of the difference between the two exceeds a preset consistency threshold, it indicates that the current AI measurement result deviates from the historical change trend, and weighted fusion correction is initiated. The consistency threshold is usually set as a multiple of the standard deviation of the historical measurement series, reflecting the fluctuation range of historical measurements. The weighted fusion method uses equal weights, summing the AI ​​measurement results and trend prediction values ​​with each weighted equally to obtain the measurement results after dual-source historical correction. Equal-weight fusion ensures that the correction results retain both the actual measured information from the current AI measurements and incorporate the constraints of historical trends. The calculation formula for the dual-source historical correction measurement results is expressed as: M corrected =0.5×M AI +0.5×M trend M trend =k×t+b Among them, M AI For the current measurement results of the AI ​​channel, M trend The trend prediction value is obtained based on linear regression of historical measurement time series, where k is the trend slope, b is the trend intercept, and t is the number of days since the first inspection.

[0051] In one embodiment, the dual-source mode combining AI and manual intervention uses manual annotation by the physician as a reference benchmark. This mode is typically triggered when the confidence level of the AI ​​measurement is low or the measurement value is within the critical range of clinical grading. In clinical practice, this mode is triggered infrequently, only activated when the physician discovers a significant deviation between the lesion outline drawn by the AI ​​and the actual boundary and actively makes corrections, thus not increasing the physician's routine workload. After the physician manually outlines the lesion boundary based on the image, the lesion size measurement result output by the manual annotation channel is considered a reliable reference value for that measurement. Subtracting the original measurement result from the AI ​​channel from this reference benchmark value yields the channel deviation value, reflecting the measurement difference between the two channels.

[0052] Specifically, the calculation and accumulation of channel deviation values ​​constitute the core mechanism for the dynamic updating of calibration records. After each manual annotation, the channel deviation value, along with the current measurement timestamp, is written into the calibration record database of the channel correction module, forming a new correction entry. This entry participates in the time-weighted cumulative calculation during subsequent queries, allowing the systematic correction amount of the AI ​​channel to continuously absorb information from manual correction. As the number of manual annotations increases, the correction amount of the AI ​​channel gradually converges to a level consistent with manual annotation, achieving adaptive accumulation of calibration. In clinical practice, doctors' intervention on AI contours is not limited to complete redrawing; a more common approach is to make local fine-tuning to the segmentation contours automatically generated by the AI, such as adjusting several control points in the edge region. The measurement results generated by this fine-tuning can be regarded as a "semi-manual" data source between fully automatic and fully manual measurement. The difference between it and the original AI measurement results is usually small but has systematic characteristics. This method also includes the fine-tuned size in the manual annotation channel for processing, recording the correction amount generated by each fine-tuning and its corresponding lesion morphology feature descriptor. Furthermore, when different doctors or the same doctor modifies the AI ​​contour of the same patient multiple times at different times, each modification may cause a jump in the measurement value, affecting the continuity of the lesion growth curve. To solve this problem, the channel correction module performs statistical analysis on the accumulated records of multiple manual corrections, clusters the correction amounts according to lesion morphology categories, and extracts the average correction direction and correction magnitude under each category as systematic correction prior knowledge for that type of lesion. When encountering lesions with similar morphological characteristics, the system can automatically apply pre-correction to the original AI measurement results based on the prior knowledge, outputting measurement results closer to clinical expectations even without manual intervention by the doctor, thereby maintaining the smoothness and continuity of the lesion size follow-up data curve. This pre-correction result is presented in the form of a suggested value, which does not replace the doctor's manual judgment, and the doctor can accept or modify the suggested value. At the same time, the reference baseline value and the original AI channel measurement results are summed with a weight of 50% each to obtain the measurement result after dual-source manual correction. The calculation formula for the measurement result after dual-source manual correction is expressed as: M manual corrected =0.5×M AI +0.5×M manual； δ manual =M AI -M manual Among them, M AI For the raw measurement results of the AI ​​channel, M manual The measurement results of the channels were manually labeled by the doctor, δ manual This represents the channel deviation value. Let δ... manual The current measurement timestamp, along with the current timestamp, is written as a new correction entry into the calibration record database of the channel calibration module.

[0053] Understandably, the three-source full-function mode comprehensively applies correction information from all available channels. This mode is triggered when the AI ​​channel, historical data channel, and manually labeled channel are all available, typically corresponding to scenarios where patients are being followed up and the doctor has performed a manual review. The correction values ​​for each of the three channels are extracted from the channel calibration configuration table, and the lesion size measurement results for each of the three channels are subtracted from their respective correction values ​​to obtain three independent correction values.

[0054] Preferably, the arithmetic mean of the three independent correction values ​​is taken to obtain the measurement result after three-source joint correction. The calculation formula for the measurement result after three-source joint correction is expressed as follows: M tri =(M AI corrected +M manual corrected +M hist corrected ) / 3; M AI corrected =M AI -B AI M manual corrected =M manual -B manual M hist corrected =M trend -B hist Where M AI corrected M is the independent correction value for the AI ​​channel. manual corrected M is the manual channel independent calibration value. hist corrected B is the independent correction value for the historical channel. AI B manual B hist These are the systematic corrections extracted from the channel calibration configuration table for the three channels, M. trend The trend prediction value is obtained based on linear regression of historical measurement time series. Arithmetic averaging assigns equal weight to the measurement information from all three channels, eliminating the excessive influence of outliers from a single channel on the final result. This three-source joint correction measurement result serves as a unified quantitative basis for lesion size description in the report, presented in a standard format in the lesion assessment section of the diagnostic report, and serves as the data foundation for clinical grading and follow-up recommendations. The core value of this multimodal correction architecture lies in establishing a robust, self-correcting report generation system to address the inevitable minor biases in measurement data from various sources, including across AI systems, algorithm versions, and scanning devices. This ensures the stability and consistency of measurement results and risk grading during long-term follow-up, eliminating fluctuations in clinical decisions caused by the accumulation of measurement errors.

[0055] In this way, the present invention effectively solves the systematic measurement bias caused by different AI systems, different algorithm versions or different imaging devices, and provides a unified and reliable quantitative basis for subsequent classification judgment.

[0056] Step S104: Obtain the preset lesion size grading threshold, and dynamically adjust the preset lesion size grading threshold by combining multi-source input data, historical calibration records and the current data source operation mode to generate a dynamic grading threshold that is adapted to the current channel combination.

[0057] The system reads a preset list of lesion size grading thresholds from the report template configuration file. The grading threshold list contains multiple clinical grading node values ​​and their corresponding risk level labels. At the same time, it extracts the upper and lower boundary offsets associated with each grading node value. The upper and lower boundary offsets define the judgment interval width of the grading node in the default state. The grading threshold list and the boundary offsets are combined to form a basic grading configuration. The boundary offsets in the basic hierarchical configuration are adjusted according to the current data source operating mode label. If the operating mode label indicates a single-source mode, the upper and lower boundary offsets of each hierarchical node are multiplied by a preset widening coefficient to obtain the widened boundary offset. The widening coefficient is read from the report template configuration file. If the operating mode label indicates a dual-source mode or a three-source full-function mode, the difference between the maximum and minimum values ​​in the measurement results after calibration of each channel is calculated as the dispersion value. The dispersion value is divided by a preset benchmark dispersion value to obtain an adjustment factor. When the dispersion is less than the benchmark, the adjustment factor is less than 1, which reduces the boundary offset and narrows the judgment interval. When the dispersion is greater than the benchmark, the adjustment factor is greater than 1, which increases the boundary offset and widens the judgment interval. The upper and lower boundary offsets of each hierarchical node are multiplied by the adjustment factor to obtain the adjusted boundary offset. For each graded node value in the graded threshold list, add the adjusted upper boundary offset to obtain the dynamic upper boundary of the node, and subtract the adjusted lower boundary offset to obtain the dynamic lower boundary of the node. Arrange the dynamic upper and lower boundaries of all graded nodes in order of risk level to form a dynamic graded threshold that adapts to the current channel combination.

[0058] In the lesion grading process of intelligent medical image reporting, the construction of dynamic grading thresholds is a core step in adapting to the characteristics of multi-channel data. This implementation focuses on the scenario of lung CT lesion size grading, describing in detail the process of reading grading thresholds, dynamically adjusting boundary offsets, and ultimately forming dynamic grading thresholds.

[0059] In one possible implementation, the report template configuration file stores a pre-defined list of lesion size grading thresholds in a structured format. This grading threshold list contains several clinical grading node values, each corresponding to a risk level label.

[0060] For example, a 6mm grading node value corresponds to the boundary between low and medium risk, while an 8mm grading node value corresponds to the boundary between medium and high risk. Each grading node value is associated with both the upper and lower boundary offsets, which together define the width of the judgment interval for that node in the default state.

[0061] It should be noted that the dynamic adjustment mechanism for boundary offsets exhibits different adjustment directions depending on the current data source operating mode. In single-source mode, due to the availability of only the AI ​​measurement channel and the lack of cross-validation information, the uncertainty of the measurement results is relatively high. In this case, a preset widening coefficient is read from the report template configuration file, and the upper and lower boundary offsets of each grading node are multiplied by this widening coefficient to obtain the widened boundary offset. The widened judgment interval covers a larger size range, ensuring that lesions near the grading boundary will not trigger grading changes due to fluctuations in a single measurement.

[0062] Specifically, in single-source mode, the range of the grading threshold boundary is widened to accommodate the uncertainty of a single measurement. In dual-source or triple-source full-function mode, the range of the grading threshold boundary is narrowed based on the confidence level of multi-channel cross-validation. Especially in dual-source or triple-source full-function mode, cross-validation of multi-channel data provides additional measurement reliability information. The dispersion value is calculated using the range method, which is the difference between the maximum and minimum values ​​of the calibrated measurement results for each channel. The smaller the dispersion value, the more consistent the measurement results of each channel, and the higher the measurement reliability. The adjustment factor is obtained by dividing the current dispersion value by a preset baseline dispersion value, which is read from the report template configuration file and represents the typical fluctuation range of the measurement results for each channel. The formula for calculating the adjustment factor is expressed as: α=D actual / D base Where α is the adjustment factor, D actual D represents the dispersion (range) of the measurement results after calibration for each channel. base The preset baseline discrete value.

[0063] The adjusted boundary offset is calculated as follows: Δ upper adj =α×Δ upper default ;Δ lower adj =α×Δ lower default Where Δ upper default and Δ lower defaultThese are the default upper and lower boundary offsets, respectively. When the actual dispersion is less than the baseline dispersion value, the adjustment factor is less than 1, the boundary offset multiplied by the adjustment factor decreases, and the judgment interval narrows, reflecting the high grading accuracy brought by the high consistency of multiple channels. When the actual dispersion is greater than the baseline dispersion value, the adjustment factor is greater than 1, the boundary offset increases, and the judgment interval widens to accommodate larger measurement discrepancies between channels. Further, for each grading node value in the grading threshold list, the adjusted upper boundary offset is added to obtain the dynamic upper boundary of that node, and the adjusted lower boundary offset is subtracted from it to obtain the dynamic lower boundary of that node. The dynamic upper and lower boundaries of all grading nodes are arranged in ascending order of risk level to form a dynamic grading threshold adapted to the current channel combination. This dynamic grading threshold serves as the grading basis in subsequent lesion size comparison and judgment.

[0064] Step S105: Compare the corrected lesion size measurement results between channels with the dynamic grading threshold, and generate a preliminary report grading adjustment signal when the risk level boundary is exceeded.

[0065] The system acquires consistent corrected lesion size measurement results and dynamic grading thresholds across channels. It iterates through each grading node within the dynamic grading threshold, reading the dynamic upper and lower boundaries of each node. The corrected lesion size measurement results are compared one by one with the dynamic upper and lower boundaries of each grading node. If the corrected lesion size measurement result is greater than the dynamic upper boundary or less than the dynamic lower boundary of a grading node, the grading node is marked as being in a boundary exceedance state, resulting in a boundary exceedance state marker set. Based on this boundary exceedance state marker set, the system identifies the positional changes of the corrected lesion size measurement results relative to each risk level. Specifically, it compares the grading node to which the AI ​​channel lesion size measurement result belongs before correction with the grading node to which the corrected lesion size measurement result belongs to, determining the grading positional change. If at least one grading node is marked as being in a boundary exceedance state, a preliminary report grading adjustment signal is generated. This signal includes a pre-change risk level label, a post-change risk level label, and the grading node value that triggered the signal. The pre-change risk level label is determined based on the comparison between the pre-corrected AI channel lesion size measurement result and the dynamic grading threshold.

[0066] In the lesion grading process of intelligent medical image reporting, the comparison between the corrected measurement result and the dynamic grading threshold is a key step in triggering grading adjustments. This implementation describes the specific process of boundary exceedance determination and grading adjustment signal output in the scenario of lung CT lesion size grading.

[0067] In one possible implementation, the comparison process is performed node-by-node. For each grading node in the dynamic grading threshold, its dynamic upper and lower boundaries are read, and the corrected lesion size measurement result is numerically compared with the two boundary values. If the measurement result is greater than the dynamic upper boundary of a grading node, it indicates that the lesion size has exceeded the upper limit of the risk level corresponding to that node; if the measurement result is less than the dynamic lower boundary of a grading node, it indicates that the lesion size has fallen below the lower limit of the risk level corresponding to that node.

[0068] Specifically, after each comparison of a tiered node is completed, if a boundary exceedance is detected, the tiered node is marked as boundary exceedance and added to the boundary exceedance status mark set. After traversing all tiered nodes, if at least one marked node exists in the boundary exceedance status mark set, the output process of the tiered adjustment signal is triggered. The reported tiered adjustment signal includes the risk level label before the change, the risk level label after the change, and the tiered node value that triggered the signal, for use in subsequent tiered determination processes.

[0069] Step S106: Simultaneously extract the lesion size measurement results after correction for each channel involved in the report grading adjustment signal, count the number of channels exceeding the dynamic grading threshold and the corresponding exceedance of each channel, identify and filter false grading adjustment signals, and obtain stable grading adjustment results.

[0070] Based on the trigger grading node value contained in the report grading adjustment signal, the corrected lesion size measurement results of each available channel are synchronously extracted from the multi-source input data. The corrected measurement result of each channel is compared with the dynamic upper boundary and dynamic lower boundary corresponding to the trigger grading node. The number of channels exceeding the dynamic grading threshold is counted. At the same time, the difference between the measurement result of each exceeding channel and the exceeded boundary is calculated as the exceedance amplitude of the channel. The number of channels and the exceedance amplitude of each channel are combined to form the exceedance feature set. The system reads the current data source operating mode label from the channel calibration configuration table. If the operating mode label indicates single-source mode, it reads the single-source misjudgment identification threshold from the report threshold configuration file. The single-source misjudgment identification threshold only includes the lower limit of amplitude. When the excess amplitude of the AI ​​channel is lower than the lower limit of amplitude, the current grade adjustment signal is marked as a false signal. If the operating mode label indicates dual-source mode, it reads the conservative misjudgment identification threshold from the report threshold configuration file. The conservative misjudgment identification threshold includes the lower limit of the number of channels and the lower limit of amplitude, and both values ​​are higher than the conventional setting. If the operating mode label indicates three-source full-function mode, it reads the conventional misjudgment identification threshold and matches the read misjudgment identification threshold with the excess feature set. For the excess feature set, if the number of excess channels is lower than the lower limit of the number of channels in the misjudgment identification threshold, or the excess amplitude of all excess channels is lower than the lower limit of the amplitude in the misjudgment identification threshold, the current reported grade adjustment signal is marked as a false signal and filtered out. Otherwise, the current reported grade adjustment signal is retained as a stable grade adjustment result.

[0071] In the grading and adjustment process of intelligent medical image reports, the identification and filtering of false grading adjustment signals are crucial to ensuring report stability. This implementation method focuses on the scenario of grading lung CT lesion size, detailing the extraction of outlier features, the selection of misjudgment identification thresholds, and the output of stable grading adjustment results.

[0072] In one possible implementation, the synchronous extraction of the calibrated measurement results for each channel is performed based on the trigger grading node value contained in the report grading adjustment signal. For each available channel, its calibrated measurement result is numerically compared with the dynamic upper and lower boundaries of the trigger grading node. If the measurement result exceeds the boundary, the channel is counted as an overshoot channel, and the numerical difference between the channel's measurement result and the overshoot boundary is recorded as the overshoot magnitude.

[0073] It should be noted that the choice of the false positive threshold is closely related to the current data source operating mode. In single-source mode, since only the AI ​​measurement channel is available, false signals are judged solely based on the amplitude exceeding the limit; when the amplitude exceeds the limit, it is marked as a false signal. In dual-source mode, due to the limited number of available channels, a conservative false positive threshold is adopted, setting relatively high values ​​for both the lower limit of the number of channels exceeding the limit and the lower limit of the amplitude exceeding the limit. In three-source full-function mode, the cross-validation of the three channels provides more comprehensive measurement reliability information; in this case, a standard false positive threshold is used, with both the lower limit of the number of channels and the lower limit of the amplitude below the conservative threshold.

[0074] Specifically, the identification of false signals is based on the matching judgment between the exceedance feature set and the false judgment threshold. If the number of exceedance channels is lower than the lower limit of the threshold, it indicates that there are too few exceedance channels, insufficient to support the reliability judgment of the grade change; if the exceedance amplitude of all exceedance channels is lower than the lower limit of the threshold, it indicates that the exceedance degree of each channel is too weak, possibly caused by measurement noise. In both of these cases, the current report grade adjustment signal is marked as a false signal and filtered. Further, when the number of exceedance channels reaches the lower limit of the channel number and the exceedance amplitude of at least one channel reaches the lower limit of the amplitude, the current report grade adjustment signal is determined to be a valid signal and retained, forming a stable grade adjustment result. This stable grade adjustment result contains risk level change information after filtering false signals, and serves as the basis for the grade of the lesion assessment section in the diagnostic report.

[0075] Step S107: Write the stability grading adjustment results into the lesion assessment section of the diagnostic report, and automatically generate or adjust the lesion description text, risk level label, and follow-up recommendations.

[0076] Based on the changed risk level labels included in the stability grading adjustment results, the corresponding lesion description text template and follow-up suggestion template are matched from the report template configuration file. The corrected lesion size measurement results are filled into the size placeholder position of the lesion description text template, and the changed risk level labels are filled into the risk level label position. The corresponding follow-up period and follow-up method are extracted from the follow-up suggestion template and filled into the follow-up suggestion position, combining to form a complete lesion assessment content. The current data source operation mode label is read from the channel calibration configuration table. If the operation mode label indicates single-source mode, the single measurement status label and short-term re-examination suggestion label associated with that channel are extracted from the channel calibration configuration table. Both are written as additional annotations at the end of the lesion assessment content, and the lesion assessment content with additional annotations is written into the lesion assessment paragraph of the diagnostic report.

[0077] In the output process of intelligent medical image reports, the final step in report presentation is to transform the stable grading adjustment results into the lesion assessment section of the diagnostic report. This implementation method focuses on the scenario of lung CT lesion size grading, describing the specific processes of template matching, content filling, and report writing.

[0078] In one possible implementation, the report template configuration file stores lesion description text templates and follow-up suggestion templates categorized by risk level. The lesion description text template contains placeholders for lesion location, morphological characteristics, and size values; the follow-up suggestion template contains preset content for follow-up periods and follow-up examination methods. Based on the changed risk level labels in the stability grading adjustment results, the corresponding template combination is located, and the corrected lesion size measurement results, risk level labels, follow-up periods, and follow-up methods are sequentially filled into the placeholder positions to form complete lesion assessment content.

[0079] Specifically, the additional annotation mechanism in single-source mode provides extra prompts for scenarios where measurement reliability is limited. When the operating mode label read from the channel calibration configuration table indicates single-source mode, two annotations are added to the end of the lesion assessment content: a single measurement prompt indicating that the current result is based on a single measurement; and a recommendation for short-term follow-up to establish a measurement baseline. The short-term follow-up recommendation suggests performing a follow-up at a preset interval to establish a measurement baseline. The lesion assessment content with additional annotations is then written into the lesion assessment section of the diagnostic report. In dual-source or triple-source full-function modes, the lesion assessment content is directly written into the diagnostic report without additional annotations.

[0080] For example: To further clarify the actual operating logic and beneficial effects of this technical solution, a specific pulmonary nodule follow-up scenario is constructed below for full-process numerical simulation, mapping the entire process from steps S101 to S107. This embodiment demonstrates how the system handles multi-source data conflicts, performs adaptive correction, and dynamically adjusts the decision threshold. Basic information of the embodiment: Male, 55 years old, solid nodule in the upper lobe of the right lung, this is the 4th follow-up examination (6 months since the last examination). Clinical decision threshold: According to guidelines, surgery or biopsy is recommended for nodule diameters >8.0mm, and continued follow-up is recommended for nodule diameters <8.0mm. Initial grading threshold setting standard threshold is T. std = 8.0mm, initial safety buffer zone ±0.5mm, reference discrete value D base = 0.4mm (representing the normal fluctuation range as perceived by the system).

[0081] First, data collection and input are performed. The system receives the data from this inspection and searches the historical database to obtain the following three sets of raw measurement data: the maximum diameter M of the nodule measured by the AI ​​automatic measurement channel. AI= 7.6mm (If only this data is considered, the conclusion is "follow-up"); the doctor manually reviewed the maximum diameter of the nodule measured via the channel. Man = 8.2mm (Based solely on this data, the conclusion would be "surgery / biopsy"); Maximum nodule diameter M obtained from historical trend prediction channel. Trend = 7.2 [The system extracts the corrected data from the patient's past three measurements (6.8mm, 7.0mm, 7.1mm), fits the growth curve using linear regression, and predicts the theoretical value for this measurement]. Relying on any single channel can lead to decision oscillations (jumping from follow-up to surgery).

[0082] Then, multi-source data adaptive correction is performed, and the system enters the three-source fusion mode, performing the following calculations: ① Calculate the systematic bias by calling the historical calibration record library. This version of the AI ​​algorithm has an average negative bias of -0.3mm on nodules of the same density (i.e., it tends to measure smaller), therefore the AI ​​measurement correction value M' is... AI = 7.6 + 0.3 = 7.9mm; This doctor's manual measurement habit has an average positive deviation of +0.4mm (i.e., tends to conservatively overestimate), therefore the manual measurement correction value M' Man = 8.2 - 0.4 = 7.8 mm; Trend value M Trend = 7.2mm (as a biological reference, with low weight). ② Calculate the measured value after three-source joint correction, and adjust the AI ​​channel correction value M'. AI = 7.9mm, manual channel calibration value M' Man = 7.8mm and historical independent correction value M' Trend =7.2mm, arithmetic mean: M Final = (7.9 + 7.8 + 7.2) / 3 ≈ 7.63mm.

[0083] Third, dynamic threshold adjustment and decision-making are performed. The system dynamically adjusts the decision boundary based on the dispersion of the data, rather than mechanically using 8.0mm. The set of calibrated measurement results for each channel is extracted: {7.9, 7.8, 7.2}, and the dispersion (range) is calculated: D actual = 7.9 - 7.2 = 0.7 mm, indicating high dispersion and suggesting measurement uncertainty (mainly due to the difference between the trend value and the measured value). Calculate the adjustment factor α = D actual / D base = 0.7 / 0.4 = 1.75, calculate the adjusted boundary offset Δ adj = Δ default × α = 0.5mm × 1.75 = 0.875mm. Determine the new decision band, dynamic threshold upper limit T. high= 8.0 + 0.875 = 8.875mm, lower limit of dynamic threshold T low = 8.0 - 0.875 = 7.125mm, grayscale observation area: [7.125mm, 8.875mm]. The final measured value M... Final The value (7.63mm) was compared with the new decision band: 7.125 < 7.63 < 8.875, and the value fell into the grayscale observation area. The final logical judgment was: although the original manual data (8.2mm) exceeded the static threshold (8mm), after correction and dynamic threshold adjustment, the system extracted the measurement results of each channel after correction and compared them with the lower limit (7.125mm) and upper limit (8.875mm) of the dynamic threshold: the AI ​​correction value of 7.9mm, the manual correction value of 7.8mm, and the trend correction value of 7.2mm all fell within the range of [7.125, 8.875]. The number of channels exceeding the upper limit of the dynamic threshold was 0, which was lower than the lower limit of the number of channels in the three-source mode (set to 2). The system marked the preliminary grade adjustment signal as a false signal and filtered it, and finally maintained the original grade conclusion. The intelligent report generated by the system includes the following: Measurement result: nodule in the upper lobe of the right lung, corrected diameter 7.6mm (original AI: 7.6mm, manual review: 8.2mm, trend prediction: 7.2mm); Grading conclusion: LU-RADS 4A. Since the current operating mode is three-source full-function mode, the system matches the corresponding follow-up suggestion template from the report template configuration file, automatically fills in the follow-up suggestion as "CT follow-up examination recommended in 3 months," and writes the corrected measurement value and risk level label into the lesion assessment paragraph. However, if the current operating mode is single-source mode, the system will additionally add "single measurement status" and "short-term follow-up recommendation" labels, and this path will not be involved. This embodiment fully demonstrates the entire process from "range calculation" to "adjustment factor division" and then to "offset multiplication," proving the logical closed loop of the technical solution. When the data discrepancy is large (D... actual >D base It can adaptively tolerate errors and automatically widen the threshold (from 0.5 to 0.875), avoiding over-medicalization (such as unnecessary surgical recommendations) caused by measurement errors. It ensures the stability of decisions. Even if the manually measured value (8.2mm) seems to trigger the critical value, the system outputs a more robust 7.6mm conclusion and follow-up recommendations through multi-source correction and dynamic threshold mechanism, eliminating the decision shock caused by "false progress".

[0084] This invention provides an image report generation system with multi-source data adaptive correction, mainly comprising: The data acquisition and pattern recognition module is used to acquire the lesion size measurement results generated by AI automatic measurement as the main channel data, extract the main channel source identifier, identify the current data source operation mode according to the actual available channel combination, and use the lesion size measurement results of all available channels and the channel source identifier together as the multi-source input data for report generation. The channel calibration query module is used by the report generation engine to query the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the systematic correction amount accumulated by each channel. When the system detects that a channel does not have a historical calibration record, it automatically calls the device-level quality control parameters as the initial correction benchmark for that channel. The sub-mode correction calculation module is used to perform sub-mode correction calculations based on multi-source input data and the current data source operation mode, using systematic correction amounts to generate consistent corrected lesion size measurement results between channels, which serve as a unified quantitative basis for lesion size descriptions in reports. The dynamic grading threshold generation module is used to read the preset lesion size grading threshold from the report template configuration, and dynamically adjust the preset lesion size grading threshold in combination with multi-source input data, historical calibration records and the current data source operation mode to generate a dynamic grading threshold that is adapted to the current channel combination. The grading adjustment signal generation module is used to compare the consistent corrected lesion size measurement results between channels with the dynamic grading threshold, and generate a preliminary report grading adjustment signal when the risk level boundary is exceeded. The false signal filtering module is used to simultaneously extract the corrected lesion size measurement results of each channel involved in the report grading adjustment signal, count the number of channels exceeding the dynamic grading threshold and the corresponding exceedance of each channel, identify false grading adjustment signals and filter them to obtain stable grading adjustment results; The automatic report generation module is used to write the stability classification adjustment results into the lesion assessment section of the diagnostic report, and automatically generate or adjust the lesion description text, risk level labeling, and follow-up recommendations.

[0085] If the technical solution of this application involves the collection, processing, or application of personal information, the relevant products have, in accordance with current laws and regulations, fully and clearly informed individuals of the processing rules and obtained their voluntary and explicit consent before implementing any personal information processing activities. If sensitive personal information is involved, the product has obtained the individual's separate consent before processing, and such consent has been given in an explicit manner.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating image reports using adaptive correction of multi-source data, characterized in that, include: The lesion size measurement results generated by AI automatic measurement are used as the main channel data. The main channel source identifier is extracted. The current data source operation mode is identified according to the actual available channel combination. The lesion size measurement results of all available channels and their respective channel source identifiers are used as multi-source input data for report generation. The report generation engine queries the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the systematic correction amount accumulated by each channel. When a channel does not have a historical calibration record, the quality control parameters of the device to which the corresponding channel belongs are called as the initial correction benchmark. Based on the multi-source input data and the current data source operating mode, the sub-mode correction calculation is performed using the systematic correction amount of each channel to generate consistent corrected lesion size measurement results between channels; Read the preset lesion size grading threshold and initial boundary offset, and dynamically adjust the preset lesion size grading threshold in combination with the multi-source input data, historical calibration records and the current data source operation mode to generate a dynamic grading threshold that is adapted to the current channel combination. The corrected lesion size measurement results between the channels are compared with the dynamic grading threshold. When there is a situation that exceeds the risk level boundary, a preliminary report grading adjustment signal is generated. Simultaneously extract the lesion size measurement results after correction for each channel involved in the report grading adjustment signal, count the number of channels exceeding the dynamic grading threshold and the corresponding exceedance of each channel, identify and filter false grading adjustment signals, and obtain stable grading adjustment results; The results of the stability grading adjustment are written into the lesion assessment section of the diagnostic report, and lesion description text, risk level labeling, and follow-up recommendations are generated or adjusted.

2. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The process involves acquiring lesion size measurement results generated by AI automatic measurement as the main channel data, extracting the main channel source identifier, identifying the current data source operating mode based on the actual available channel combinations, and using the lesion size measurement results of all available channels along with their respective channel source identifiers as multi-source input data for report generation, including: Extract the target lesion from the current medical image and perform automatic contouring. Identify the long diameter and short diameter measurement values ​​of the lesion and combine them to form the lesion size measurement result generated by AI automatic measurement. Simultaneously generate the main channel source identifier and bind the lesion size measurement result and the main channel source identifier to store it in the main channel data record. Read the segmentation confidence score from the main channel data record. When the segmentation confidence score is lower than the confidence threshold, or when the lesion size measurement result falls into the grading critical range, trigger a manual annotation prompt signal. After manual annotation is completed, the lesion size measurement results and manual annotation source identifiers of the manually annotated channels are extracted and included in the current available channel set. The current data source operation mode is identified according to the number of channel types contained in the available channel set, and the corresponding mode label and multi-source input data of all available channels are output.

3. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The report generation engine queries the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the accumulated systematic correction amounts of each channel. When a channel does not have a historical calibration record, the quality control parameters of the device to which the corresponding channel belongs are called as the initial correction benchmark, including: Extract the channel source identifier of each available channel from the multi-source input data, make a historical calibration record query request for each channel source identifier, retrieve all related historical calibration entries using the channel source identifier as an index, and sum the correction deviation values ​​of all historical calibration entries under the same channel after assigning decreasing weights according to the calibration execution time from recent to distant to obtain the systematic correction amount of the channel. For each channel in the systematic correction set, it is determined whether there is a valid historical calibration record. If no historical calibration entry is found for a certain channel, the most recent phantom calibration result of the imaging equipment to which the corresponding channel belongs is extracted from the equipment archive, and the inherent measurement deviation value of the equipment in the phantom calibration result is written into the systematic correction set as the initial correction benchmark. The quality control parameters of the device include spatial resolution deviation value obtained based on standard phantom scanning, edge detection deviation value caused by noise, etc. The systemic correction set is traversed, and the systemic correction values ​​are retained for channels with historical calibration records. The initial correction reference is retained for channels with written initial correction references. The correction value information of all channels is integrated to form a channel calibration configuration table.

4. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The step involves performing a sub-mode correction calculation based on the multi-source input data and the current data source operating mode, using the systematic correction amount of each channel, to generate consistent corrected lesion size measurement results across channels, including: Read the current data source operating mode label from the channel calibration configuration table; When the operating mode label indicates the three-source full-function mode, the correction values ​​for the AI ​​channel and the manual channel are extracted from the channel calibration configuration table respectively. The measurement results of the AI ​​channel and the manual channel are subtracted from their respective correction values ​​to obtain two independent correction values. Historical measurement time series are extracted from the multi-source input data, and the trend prediction value is calculated using the linear regression method. The historical channel correction values ​​are subtracted from the trend prediction value to obtain the historical independent correction values. The arithmetic mean of the correction values ​​is taken to obtain the measurement result after three-source joint correction.

5. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The preset lesion size grading threshold is dynamically adjusted based on the multi-source input data, historical calibration records, and the current data source operating mode to generate a dynamic grading threshold adapted to the current channel combination, including: The basic grading configuration is formed by reading the preset list of lesion size grading thresholds and the upper and lower boundary offsets associated with each grading node from the report template configuration file. The boundary offset is adjusted according to the current data source operating mode label. When the operating mode is single source mode, the upper and lower boundary offsets of each hierarchical node are multiplied by the widening coefficient. When the operating mode is dual source mode or three source full-function mode, the difference between the maximum and minimum values ​​in the measurement results after calibration of each channel is calculated as the dispersion value. The dispersion value is divided by the reference dispersion value to obtain the adjustment factor. The upper and lower boundary offsets of each hierarchical node are multiplied by the adjustment factor. For each graded node value, the adjusted upper boundary offset is added to obtain the dynamic upper boundary, and the adjusted lower boundary offset is subtracted to obtain the dynamic lower boundary. These are then arranged in order of risk level to form dynamic graded thresholds.

6. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The step of comparing the corrected lesion size measurement results consistent across channels with the dynamic grading threshold, and generating a preliminary report grading adjustment signal when a situation exceeds the risk level boundary, includes: Traverse each grading node in the dynamic grading threshold, compare the corrected lesion size measurement results with the dynamic upper boundary and dynamic lower boundary of each grading node respectively, mark the boundary exceedance state, and obtain the boundary exceedance state mark set. Based on the set of boundary exceedance status markers, the classification node to which the AI ​​channel belongs before correction is compared with the classification node to which it belongs after correction to determine the change in classification position. When a boundary exceedance status exists, a preliminary report classification adjustment signal is generated. The report classification adjustment signal includes the risk level label before the change, the risk level label after the change, and the trigger classification node value.

7. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The process involves simultaneously extracting the corrected lesion size measurement results for each channel involved in the report grading adjustment signal, statistically analyzing the number of channels exceeding the dynamic grading threshold and the corresponding exceedance magnitude for each channel, identifying and filtering false grading adjustment signals, and obtaining stable grading adjustment results, including: Based on the trigger grading node value in the report grading adjustment signal, the corrected lesion size measurement results of each available channel are extracted, the number of channels exceeding the dynamic grading threshold is counted, and the exceedance amplitude of each exceeding channel is calculated to form an exceedance feature set. Based on the current data source operating mode label, the corresponding misjudgment identification threshold is read. When the number of exceeding channels is lower than the lower limit of the number of channels, or when the exceeding amplitude of all exceeding channels is lower than the lower limit of the amplitude, the report grading adjustment signal is marked as a false signal and filtered; otherwise, it is retained as a stable grading adjustment result.

8. The image report generation method for adaptive correction of multi-source data according to claim 1, characterized in that, The step of writing the stability grading adjustment results into the lesion assessment section of the diagnostic report, and generating or adjusting the lesion description text, risk level labeling, and follow-up recommendations, includes: Based on the changed risk level labels in the stable grading adjustment results, the corresponding lesion description text template, risk level labeling position and follow-up suggestion template are matched from the report template configuration file. The corrected lesion size measurement results and changed risk level labels are filled into the corresponding positions to form the lesion assessment content. If the current operating mode is single-source mode, the single measurement status label and short-term review suggestion label are extracted from the channel calibration configuration table and written as additional annotations into the lesion assessment content. The lesion assessment content is then written into the lesion assessment section of the diagnostic report.

9. An image report generation system with multi-source data adaptive correction, characterized in that, The system includes: The data acquisition and pattern recognition module is used to acquire the lesion size measurement results generated by AI automatic measurement as the main channel data, extract the main channel source identifier, identify the current data source operation mode according to the actual available channel combination, and use the lesion size measurement results of all available channels and their respective channel source identifiers as multi-source input data for report generation. The channel calibration query module is used by the report generation engine to query the historical calibration records corresponding to each available channel from the channel calibration module based on the source identifier of each channel. The historical calibration records contain the systematic correction amount accumulated by each channel. When a channel does not have a historical calibration record, the quality control parameters of the device to which the corresponding channel belongs are called as the initial correction benchmark. The sub-mode correction calculation module is used to perform sub-mode correction calculations based on the multi-source input data and the current data source operation mode, using the systematic correction amount of each channel, to generate consistent corrected lesion size measurement results between channels. The dynamic grading threshold generation module is used to read the preset lesion size grading threshold from the report template configuration, and dynamically adjust the preset lesion size grading threshold in combination with the multi-source input data, historical calibration records and the current data source operation mode to generate a dynamic grading threshold that is adapted to the current channel combination. The grading adjustment signal generation module is used to compare the consistent corrected lesion size measurement results between the channels with the dynamic grading threshold, and generate a preliminary report grading adjustment signal when there is a situation that exceeds the risk level boundary. The false signal filtering module is used to simultaneously extract the corrected lesion size measurement results of each channel involved in the report grading adjustment signal, count the number of channels exceeding the dynamic grading threshold and the corresponding exceedance of each channel, identify and filter false grading adjustment signals, and obtain stable grading adjustment results; The report automatic generation module is used to write the stability grading adjustment results into the lesion assessment section of the diagnostic report, and to generate or adjust the lesion description text, risk level labeling, and follow-up recommendations.

Images