Perioperative medical detection image accurate analysis and management system and method based on ai model

Abstract

The application discloses an AI model-based perioperative medical detection image accurate analysis and management system and method, and relates to the technical field of medical image processing. The AI model-based perioperative medical detection image accurate analysis and management system and method comprises the following steps: S1, preprocessing chest radiograph structure feature data and historical statistical data of a surgical patient in a perioperative period; S2, discriminating and segmenting a structure overlapping area by gradient field response and geometric density correction for the overlapping structure area; S3, constructing a time sequence registration criterion by adopting a structure-guided rigid and affine joint registration algorithm; S4, fusing time sequence labels by fusing response and space fluctuation suppression; and S5, constructing, training and reasoning a structure function evaluation model by adopting multi-modal feature enhancement. The technical problem that data labels are confused due to structure overlapping and time sequence changes in the prior art AI analysis of chest radiographs in a perioperative period, and the stability and accuracy of AI model training and reasoning are affected is solved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a system and method for precise analysis and management of perioperative medical examination images based on AI models. Background Technology

[0002] With the widespread application of medical imaging in clinical diagnosis, surgical decision-making, efficacy evaluation, and health management, the volume of digital medical image data has exploded. Traditional manual image interpretation methods, limited by subjective experience and workload, are no longer sufficient to meet the demands of large-scale, precise, and personalized medical image analysis. In recent years, the rapid development of artificial intelligence technologies such as computer vision, machine learning, and deep learning has provided new technological pathways for the automated processing and decision support of medical images. Existing technologies are continuously exploring the deep integration of structured image information with clinical test indicators to improve the accuracy of image diagnosis, data utilization efficiency, and medical interpretability. At the same time, in the context of big data-driven medical environments, ensuring the automation, robustness, and traceability of image analysis has gradually become an important direction for medical system research and development.

[0003] For example, invention patent CN118588245A discloses an AI-based mobile medical ultrasound contrast imaging image analysis and assistance system, including an image acquisition and processing module, an image analysis and privacy security module, an auxiliary diagnosis and report generation module, a user interface and interaction module, and a data management and storage module. The system utilizes portable ultrasound equipment and wireless communication technology to achieve real-time acquisition and transmission of ultrasound contrast imaging images, and performs denoising, enhancement, and registration processing in the cloud and on a local workstation. Through an improved joint heterogeneous image analysis method and a personalized federated learning model, the system analyzes image data and employs homomorphic encryption and privacy protection mechanisms to ensure data security. Based on CNN model training and PDPs visualization optimization algorithms, it automatically generates structured auxiliary diagnostic reports, provides an intuitive and easy-to-use user interface, and supports compliant data storage and retrieval. This invention achieves intelligent analysis, privacy protection, and automated auxiliary diagnosis of ultrasound contrast imaging images, improving the efficiency and safety of ultrasound diagnosis in mobile medical scenarios.

[0004] For example, invention patent CN120339228A discloses an AI-based cervical cell image analysis assistance method and system. The method includes: multimodal data acquisition and image preprocessing; constructing a generative adversarial network suitable for cervical cell images, improving image generation and feature extraction capabilities by fusing multimodal features and introducing an attention mechanism; fusing extracted features with clinical and genetic data, training and optimizing a cervical cell classification model based on support vector machines and random forests, and comprehensively evaluating model performance through accuracy, recall, and F1 score. The system can perform anomaly detection and type classification analysis on real cervical cell images, and output the degree of abnormality based on quantitative indicators such as cell nucleus and chromatin, ultimately achieving automated intelligent analysis and clinical auxiliary interpretation of cervical cell images, improving the efficiency and accuracy of cervical cell anomaly screening.

[0005] AI-based intelligent medical image analysis systems have been widely applied in fields such as ultrasound contrast imaging and cell imaging, significantly improving the processing efficiency and intelligence level of medical images through multimodal data fusion, deep learning modeling, and automated diagnosis. However, existing technologies still suffer from insufficient robustness, easy propagation of label noise, and limited adaptability of models to structural temporal changes in areas such as the identification of overlapping regions of complex structures, consistency correction of temporal images, dynamic fusion of labeled data, and management of multi-temporal anomalies. These limitations make it difficult to fully meet the clinical needs for high-precision, personalized, and interpretable medical image-assisted analysis.

[0006] Therefore, in order to address the above issues, there is an urgent need for a precise analysis and management system and method for perioperative medical imaging based on AI models. Summary of the Invention

[0007] To address the technical problem in existing AI analysis of chest X-rays where structural overlap and temporal variations lead to data label confusion, affecting the stability and accuracy of AI model training and inference, this invention provides a precise analysis and management system and method for perioperative medical examination images based on an AI model. The technical solution is as follows:

[0008] On the one hand, it provides a method for precise analysis and management of perioperative medical examination images based on AI models. This method includes: S1, collecting chest X-ray structural feature data and obtaining historical statistical data; preprocessing the chest X-ray structural feature data and historical statistical data; S2, based on the chest X-ray structural feature data, identifying and segmenting overlapping structural regions through gradient field response and geometric density correction; S3, using a structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction to construct temporal registration criteria, and performing spatial synchronization and consistency quantification of temporal images; S4, based on structural overlap discrimination and temporal registration, performing temporal label fusion through fusion response and spatial fluctuation suppression, and entering the modeling and inference output process; S5, under the constraints of three types of high-order criteria (structure, temporality, and labels), using multimodal feature enhancement to realize the construction, training, inference, and individualized output of the structural function assessment model.

[0009] Optionally, perioperative chest X-ray structural feature data of patients is collected, and historical statistical data is obtained. The specific process of preprocessing the chest X-ray structural feature data and historical statistical data is as follows: Chest X-ray structural feature data is collected, including: two-dimensional coordinate data of the original chest X-ray image, grayscale values ​​of the original image, pixel grayscale data at multiple time points, and original image data; historical statistical data is obtained, including: lung function test indicators, grayscale sliding windows, and the number of historical time points; spatial contour optimization and redundant point removal are performed on the chest X-ray structural feature data through edge detection algorithms and local principal curvature analysis; outlier suppression and short-term fluctuation smoothing are performed on the two-dimensional coordinate data and grayscale values ​​of the original chest X-ray image through pixel-level distribution resampling and adaptive sliding window filtering; missing data completion and trend feature extraction are performed on the historical statistical data through temporal interpolation and piecewise linear regression; and the chest X-ray structural feature data and historical statistical data are standardized and normalized through distribution standardization and linear normalization algorithms.

[0010] Optionally, based on chest X-ray structural feature data, the specific process of correcting overlapping structural regions through gradient field response and geometric density is as follows: Acquire original image data, two-dimensional coordinate data of the original chest X-ray image, pixel grayscale data at multiple time points, and grayscale values ​​of the original image; obtain structural boundary pixel set, bone width, gap pixel distance, and background region data from the original image data using an image segmentation algorithm; obtain structural prior response values ​​from the two-dimensional coordinate data of the original chest X-ray image using an automatic segmentation algorithm based on a UNet convolutional neural network; obtain spatial gradient magnitudes from the pixel neighborhood calculated by convolution of the original image grayscale values ​​using the Sobel operator; obtain local geometric density adjustment values ​​from the structural boundary pixel set, bone width, and gap pixel distance using OpenCV contour extraction and the minimum bounding rectangle method; obtain edge response intensity from the original image grayscale values ​​using the Canny edge detection algorithm; and so on. (x,y) is the center gray-level sliding window, where (x,y) represents the horizontal coordinate x and vertical coordinate y of a pixel in the original image data. The sliding window gray-level information entropy is obtained by applying the Shannon entropy algorithm to the original image gray-level values. The local background noise standard deviation is calculated by applying the Gaussian filtering algorithm and the local standard deviation to the background region data. The temporal gray-level fluctuation is obtained by applying the rigid registration algorithm to the pixel gray-level data at multiple time points. The gradient weighting term is obtained by calculating the product of the structural prior response value and the spatial gradient magnitude. The weighted response term is obtained by calculating the product of the local geometric density adjustment value and the gradient weighting term, and then adding the edge response intensity. The composite suppression term is obtained by calculating the product of the sliding window gray-level information entropy, the background noise adjustment factor, and the local background noise standard deviation, and then adding the product of the temporal gray-level fluctuation and the temporal change adjustment factor. The structural overlap perception value is obtained by dividing the weighted response term by the composite suppression term.

[0011] Optionally, the specific process for identifying and segmenting structural overlap regions is as follows: Real-time comparison of the perceived structural overlap value and the perceived structural overlap threshold. The perceived structural overlap value includes a first-level threshold and a second-level threshold. When the perceived structural overlap value is less than the second-level threshold, a normal structural label is generated and entered into the temporal label fusion and confidence assessment module, requiring no special masking or adjustment. When the perceived structural overlap value is greater than or equal to the second-level threshold and less than the first-level threshold, a suspected structural overlap label is generated, and a low-confidence label is output during inference in the structural function assessment model. When the perceived structural overlap value is greater than or equal to the first-level threshold, high-resolution reconstruction and local contrast enhancement are triggered, generating a structural overlap risk label. High-risk region patches are extracted using structural masks, and a small Patch-UNet local sub-model is used to perform feature depth extraction and discrimination of the regions.

[0012] Optionally, the specific process of constructing a temporal registration criterion using a structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction, is as follows: Obtain the original image data and structural overlap perception values; obtain the first-order registration criterion from the multi-temporal original image data using a structure-based rigid registration algorithm and a UNet convolutional neural network automatic segmentation algorithm. After time registration, the tag value and the first Label values ​​after time-series registration; spatial registration algorithms using feature point detection and affine transformation are used to align multi-time-series original image data, and the time-series labels are calculated. and The Euclidean distance between the pixel spatial coordinates is used to obtain the registration residual between adjacent time points; the temporal mean of the structural overlap sensing values ​​at all time points is obtained by the mean algorithm; the temporal variance of the registered label values ​​is obtained by the standard deviation algorithm; the calculation of the first... After time registration, the tag value is the same as the first time. The absolute value of the difference between the registered label values ​​and the registration residual exponential decay factor are multiplied and summed sequentially to obtain the registration difference weighting term. The absolute value of the difference between the perceived structural overlap value and the temporal mean of the perceived structural overlap value is calculated to obtain the structural overlap consistency term. The product of the temporal variance of the label value and the temporal variance adjustment factor is calculated to obtain the temporal fluctuation adjustment term. 1 is added to the structural overlap consistency term and the temporal fluctuation adjustment term to obtain the denominator adjustment term. The registration difference weighting term is divided by the denominator adjustment term to obtain the temporal registration consistency value.

[0013] Optionally, the specific process for spatial synchronization and consistency quantification of temporal images is as follows: real-time comparison of temporal registration consistency value and temporal registration consistency threshold: when the temporal registration consistency value is less than the temporal registration consistency threshold, it is determined to be a registration consistency area and enters the temporal label fusion and confidence assessment module without special intervention; when the temporal registration consistency value is greater than or equal to the temporal registration consistency threshold, it is determined to be a registration fluctuation area, and a local temporal correction method based on mutual information maximization and optical flow vector compensation is adopted. Dynamic registration correction of multi-time images is achieved through residual fitting and consistency smoothing. During model inference, a low-confidence label is output, and neighborhood information is fused for enhancement processing during model training and inference.

[0014] Optionally, based on structural overlap discrimination and temporal registration, the specific process of temporal label fusion through fusion response and spatial fluctuation suppression is as follows: Obtain the structural overlap sensing value, temporal registration consistency value, and the number of historical moments; obtain a multi-temporal fusion label probability set by using a temporal probability clustering algorithm and a multi-temporal weighted average method on the registered label values ​​of all historical moments; obtain the label probability spatial variance by using a standard variance algorithm on the multi-temporal fusion label probability set; calculate the absolute value of the difference between the structural overlap sensing value and the historical structural overlap sensing value, invert it, and then apply it to all historical moments. Multiply the historical time points together and raise the M-th power to obtain the structural consistency fusion term; calculate the product of the temporal registration consistency value and the temporal consistency adjustment factor, take the negative number, and perform an exponential operation to obtain the registration consistency adjustment term; multiply the structural consistency fusion term and the registration consistency adjustment term to obtain the label confidence fusion response term; calculate the product of the label probability space variance of the multi-temporal fusion label probability set and the label space variance adjustment factor, add 1 to obtain the spatial fluctuation suppression term; divide the label confidence fusion response term by the spatial fluctuation suppression term to obtain the temporal label fusion value.

[0015] Optionally, the specific process of performing time-series label fusion and entering the modeling and inference output flow is as follows: Real-time comparison of the time-series label fusion value and the time-series label fusion threshold, which includes a primary fusion threshold and a secondary fusion threshold: When the time-series label fusion value is less than or equal to the secondary fusion threshold, the label is directly removed and does not participate in model training and inference. At the same time, a fusion risk log is generated on the data side, triggering physical supplementary collection and manual review; When the time-series label fusion value is greater than the secondary fusion threshold but less than or equal to the primary fusion threshold, a structural time-series fusion questionable label is generated. The modeling process adopts multimodal label fusion technology, and low confidence labels are output during inference; When the time-series label fusion value is greater than the primary fusion threshold, it is directly included in the structural function assessment model training and enters the modeling and inference output module. High confidence labels are output during inference.

[0016] Optionally, under the constraints of three higher-order criteria—structure, temporal, and label—the specific process of constructing, training, inferring, and outputting individualized results of the structural function assessment model using multimodal feature enhancement is as follows: Using multimodal feature fusion and a hierarchical loss function, inputting original image data and lung function test indicators, dynamically selecting and hierarchically training samples based on structural overlap perception value criteria, temporal registration consistency value criteria, and temporal label fusion value criteria, strengthening supervised learning of high-consistency and high-confidence samples, and for low-confidence regions, employing neighborhood information enhancement and an attention mechanism to fuse labels, and applying robust training to handle noisy labels, jointly modeling... The probability distribution of multiple time-series labels is used to construct a structural function assessment model. During the inference phase, the structural function assessment model simultaneously outputs the graded results of structural overlap perception value, temporal registration consistency value, and temporal label fusion value. It detects and locates abnormal ventilation functional areas, boundary graded drift areas, and structurally low consistency blocks. Abnormal criterion areas are fed back to the labeling process in real time. The spatial distribution of abnormal ventilation areas is visualized in a two-dimensional heat map. A structural function correlation distribution report is generated through the joint analysis of structural feature distribution and pulmonary function test indicators. Finally, the report outputs confidence labels, abnormal ventilation area heat map, structural function correlation distribution, criterion graded results, and a full-process quality traceability report.

[0017] On the other hand, an AI-based model-based system for the precise analysis and management of perioperative medical examination images is provided. This system is applied to AI-based model-based methods for the precise analysis and management of perioperative medical examination images. The system includes: The data acquisition and preprocessing module is used to acquire perioperative chest X-ray structural feature data of patients and obtain historical statistical data on diagnosis and treatment; and to preprocess the chest X-ray structural feature data and historical statistical data. The structural overlap region identification and segmentation module is used to identify and segment structural overlap regions based on chest X-ray structural feature data by using gradient field response and geometric density correction. The multi-temporal image registration module is used to construct temporal registration criteria by employing a structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction, to perform spatial synchronization and consistency quantification of temporal images. The temporal label fusion and confidence assessment module is used to perform temporal label fusion based on structural overlap discrimination and temporal registration, through fusion response and spatial fluctuation suppression, and then enter the modeling and inference output process; The modeling and inference output module is used to construct, train, infer, and output individualized results of a structural function assessment model under the constraints of three higher-order criteria: structure, time series, and label. This is achieved through multimodal feature enhancement. The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) The gradient field response and geometric density correction algorithm is innovatively used to perform multi-level discrimination and fine segmentation of complex overlapping structural regions, so that structural interference areas can be accurately identified and diverted, thereby realizing risk classification and feature extraction of overlapping interference areas, effectively solving the problem of unclear identification of overlapping structures and the resulting decrease in model segmentation accuracy in the existing technology.

[0018] (2) By using the structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance dynamic correction methods, higher precision spatial synchronization and temporal consistency control can be achieved between multiple time-series medical images, thereby realizing accurate alignment of dynamic image sequences and effectively solving the problems of large registration error of multiple time-series images and temporal fluctuation affecting diagnostic consistency in the existing technology.

[0019] (3) Based on structural overlap perception and multiple criteria for temporal registration consistency, a dynamic temporal tag fusion and confidence layering strategy is proposed, which effectively realizes the intelligent generation and management of high-quality tag sets, thereby realizing the active control and hierarchical processing of tag data noise, effectively solving the problems of easy spread of tag noise and heavy manual correction burden in the existing technology.

[0020] (4) By adopting multimodal feature fusion and hierarchical loss function, and combining multi-layer information of structure, time series and label for deep modeling training, the model’s adaptability to complex structure and individual differences is greatly improved, thereby realizing the personalized evaluation effect of structure and function integration, effectively solving the problems of multimodal data fusion and insufficient individualized analysis capability in the existing technology. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for precise analysis and management of perioperative medical examination images based on an AI model, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the AI-based perioperative medical monitoring image precision analysis and management system provided in an embodiment of the present invention; Figure 3 This is a time-series tag fusion and registration tag distribution diagram provided in an embodiment of the present invention; Figure 4 This is a flowchart of the confidence-label-based data hierarchical processing provided in an embodiment of the present invention; Figure 5This is a thermal map showing the spatial distribution of abnormal ventilation areas provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0024] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0025] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0026] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0027] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0028] This invention provides a method for precise analysis and management of perioperative medical imaging images based on an AI model. This method can be implemented using an SSS device, which can be a terminal or a server. Figure 1The flowchart shown is for a method for precise analysis and management of perioperative medical examination images based on an AI model. The processing flow of this method can include the following steps: S1, collect the structural feature data of the patient's perioperative chest X-ray and obtain historical statistical data; preprocess the chest X-ray structural feature data and historical statistical data; S2, based on the chest X-ray structural feature data, identify and segment overlapping structural regions through gradient field response and geometric density correction; S3, adopt a structure-guided rigid and affine joint registration algorithm, combine residual weighting and variance correction to construct temporal registration criteria, and perform spatial synchronization and consistency quantification of temporal images; S4, based on structural overlap discrimination and temporal registration, perform temporal label fusion through fusion response and spatial fluctuation suppression, and enter the modeling and inference output process; S5, under the constraints of three types of high-order criteria (structure, temporality, and labels), adopt multimodal feature enhancement to realize the construction, training, inference, and individualized output of the structural function assessment model.

[0029] Optionally, chest X-ray structural feature data is collected, and historical statistical data is obtained; the specific process of preprocessing the chest X-ray structural feature data and historical statistical data is as follows: Collect chest X-ray structural feature data, which includes: two-dimensional coordinate data of the original chest X-ray image, grayscale value of the original image, pixel grayscale data at multiple time points, and original image data. The pixel grayscale data at multiple time points directly serve the subsequent time-series image registration and label fusion process. Historical statistics are obtained, including: lung function test indicators, gray-scale sliding window, and number of historical moments. The size parameter of the gray-scale sliding window is directly used to calculate the gray-scale information entropy of the sliding window, and the number of historical moments defines the time span of the temporal registration criterion and label fusion calculation. By employing edge detection algorithms and local principal curvature analysis, spatial contour optimization and redundant point removal are performed on chest X-ray structural feature data. Local principal curvature analysis is used to evaluate the surface geometry to distinguish between real structures and noise. Pixel-level distributed resampling and adaptive sliding window filtering are used, with the adaptive sliding window filtering dynamically adjusting the filtering intensity based on local variance to suppress outliers and smooth short-term fluctuations in the original chest X-ray image's two-dimensional coordinate data and grayscale values. Temporal interpolation and piecewise linear regression are used, with piecewise linear regression fitting the trends of each time period using the least squares method to complete missing data and extract trend features from historical statistical data. Distribution standardization and linear normalization algorithms are used, with linear normalization mapping the data to the [0,1] interval, to standardize and normalize the chest X-ray structural feature data and historical statistical data.

[0030] In this implementation plan, multi-layer processing ensures data quality by collecting perioperative chest X-ray structural feature data and historical statistical data of pulmonary function test indicators to construct a complete multimodal dataset. Edge detection and local principal curvature analysis are used to optimize spatial contours. Pixel-level resampling and adaptive filtering suppress data anomalies. Temporal interpolation and piecewise regression methods are used to fill in missing values ​​and extract trend features. Finally, data standardization is performed. These preprocessing operations significantly improve data quality and consistency, laying a reliable foundation for subsequent steps such as structural overlap discrimination, temporal registration, and label fusion analysis, and ensuring the stability of model training and the accuracy of inference results.

[0031] Optionally, based on the patient's perioperative chest radiograph structural feature data, the specific process of correcting overlapping structural regions using gradient field response and geometric density is as follows: The original image data, the original chest radiograph's two-dimensional coordinate data, multi-time-point pixel grayscale data, and the original image's grayscale values ​​constitute the complete input set required for calculating the perceived value of structural overlap. The original image data is processed using an image segmentation algorithm to obtain the set of structural boundary pixels, bone width, gap pixel distance, and background region data, accurately analyzing the anatomical structures and spatial relationships in the image. The original chest radiograph's two-dimensional coordinate data is processed using an automatic segmentation algorithm based on a UNet convolutional neural network to obtain the prior structural response value, quantifying the model's confidence in the existence of a specific structure. The original image's grayscale values ​​are processed using Sobel operator convolution to calculate the pixel neighborhood and obtain the spatial gradient magnitude, which is then calculated using two 3×3 convolution kernels. The algorithm calculates gradients in both horizontal and vertical directions. For the set of pixels at structural boundaries, bone width, and gap pixel distances, it uses OpenCV contour extraction and the minimum bounding rectangle method to obtain local geometric density adjustment values. It also uses the Canny edge detection algorithm to obtain edge response intensity from the original image grayscale values, accurately fitting the minimum bounding rectangle of the structural boundaries. A grayscale sliding window centered at (x,y) represents the horizontal x-coordinate and vertical y-coordinate of a pixel in the original image data. The Shannon entropy algorithm is used to obtain the sliding window grayscale information entropy, reflecting the randomness of the grayscale distribution within the window. The algorithm uses Gaussian filtering and local standard deviation calculation to obtain the local background noise standard deviation from the background region data. Finally, a rigid registration algorithm is used to obtain temporal grayscale fluctuations from multi-time-point pixel grayscale data, and image alignment is achieved through rotation and translation transformations.

[0032] The product of the prior structural response value and the spatial gradient magnitude is calculated element-wise to obtain the gradient weighting term. The product of the local geometric density adjustment value and the gradient weighting term, plus the edge response intensity, yields the weighted response term, integrating geometric features, gradient information, and edge intensity. The product of the sliding window grayscale entropy, the background noise adjustment factor, and the local background noise standard deviation, plus the product of the temporal grayscale fluctuation and the temporal variation adjustment factor, yields the composite suppression term, comprehensively considering the effects of texture complexity, background noise, and temporal fluctuation. Dividing the weighted response term by the composite suppression term yields the structural overlap perception value, calculated using the following formula: ; In the formula, The perceived value represents the structural overlap, directly reflecting the probability that there is a risk of structural overlap at that pixel location; This represents the prior structural response value, reflecting the degree of confidence that the location belongs to a specific anatomical structure; It represents the magnitude of the spatial gradient, which characterizes the degree of drastic gray-level change at a given location in the image; This represents the local geometric density adjustment value, reflecting the width and pixel distance geometry of the surrounding area; This represents the edge response strength, used to enhance the contribution weight of real edges; It represents the grayscale information entropy of the sliding window, which quantifies the texture complexity and information richness of a local image region; This represents the standard deviation of local background noise, characterizing the noise interference level in the background area. It represents the amount of temporal grayscale fluctuation, reflecting the grayscale stability of the same location at different points in time; This represents the local geometric density adjustment factor, which is obtained by using the minimum bounding rectangle algorithm and the local weighted average method to adjust the edge response intensity. The value ranges from 0 to 1. This represents the edge response adjustment factor, which is obtained by using the Canny edge detection algorithm and image gradient weighting method to adjust the edge response intensity. The value ranges from 0 to 1. This represents the background noise adjustment factor, which is obtained by using the local standard deviation algorithm and multi-channel background noise analysis on the background region data. The value ranges from 0 to 2. This represents the time-series variation adjustment factor, which is obtained by stability analysis of multiple time-series images to measure the time-series grayscale fluctuation. The value ranges from 0 to 1.

[0033] In this implementation scheme, this step effectively achieves accurate identification and quantitative evaluation of structural overlap regions in chest X-ray images by constructing a structural overlap perception value. It comprehensively utilizes multi-dimensional information including structural prior response, spatial gradient features, geometric density information, edge strength features, and temporal stability. By calculating the ratio of the weighted response term to the composite suppression term, it can accurately distinguish between real anatomical structures and overlap artifacts. This method not only considers image features in the spatial domain but also introduces grayscale stability analysis in the temporal domain, significantly improving the robustness and accuracy of overlap region identification. This provides a reliable structural discrimination basis for subsequent temporal registration and label fusion, thereby ensuring the accuracy and reliability of the entire medical image analysis process.

[0034] Optionally, the specific process for identifying and segmenting overlapping structural regions is as follows: real-time comparison of the perceived structural overlap value and the perceived structural overlap threshold, wherein the perceived structural overlap value includes a primary perception threshold and a secondary perception threshold. When the structural overlap perception value is less than the secondary perception threshold, it indicates that the regional organizational structure is clear and there is no overlap interference. The generated structural normal label enters the time-series label fusion and confidence assessment module, and will continue to be processed according to the standard procedure without special masking or adjustment.

[0035] When the perceived structural overlap value is greater than or equal to the second-level perception threshold but less than the first-level perception threshold, there is a certain degree of structural overlap in the region. A structural overlap doubt label is generated and a low-confidence label is output during the inference of the structural function assessment model. This label will serve as an important reference in the subsequent decision-making process.

[0036] When the perceived structural overlap value is greater than or equal to the first-level perception threshold, it confirms that there is a significant structural overlap phenomenon in the region, triggering high-resolution reconstruction and local contrast enhancement, which can effectively improve the accuracy of subsequent analysis. A structural overlap risk label is generated, and high-risk region patches are extracted through structural masking. The extracted patch regions will enter a dedicated analysis process. A small Patch-UNet local sub-model is used to perform feature depth extraction and discrimination of the region. The sub-model adopts an attention mechanism and multiple residual connections, and is specifically optimized for complex overlapping regions.

[0037] In this implementation plan, this step establishes a hierarchical discrimination mechanism to achieve intelligent identification and classification of structural overlap regions. Based on real-time comparison of the perceived structural overlap value with a preset threshold, it can accurately distinguish between normal regions, suspected overlap regions, and high-risk overlap regions, and adopt differentiated processing strategies for different risk levels. Normal regions maintain a standard processing procedure, suspected regions are marked with low confidence for subsequent reference, while high-risk regions are processed using an enhanced analysis process including high-resolution reconstruction and a dedicated Patch-UNet sub-model. This hierarchical processing approach ensures both processing efficiency and accurate discrimination results through a specially optimized model in high-risk situations, providing a reliable foundation for subsequent temporal registration and structural function assessment.

[0038] Optionally, the specific process of constructing a temporal registration criterion using a structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction, is as follows: Obtain the original image data and structural overlap perception values ​​to form the basic input for registration analysis; for multiple temporal original image data, obtain the first... After time registration, the tag value and the first After time-series registration, the label values ​​are used to achieve accurate structural segmentation through deep learning methods, ensuring accurate alignment of images from different time points in terms of anatomical structures. Multi-time-series raw image data are aligned using spatial registration algorithms based on feature point detection and affine transformation, and the alignment is achieved by calculating the time sequence. and The Euclidean distance between pixel spatial coordinates is used to obtain the registration residual between adjacent time steps, which quantifies the spatial error in the registration process and reflects the level of registration accuracy. The time-series mean of the structural overlap perception value at all time points is obtained by the mean algorithm. The time-series variance of the registered label value is obtained by the standard variance algorithm, which is used to evaluate the stability of the structural overlap feature in the entire time series and provide a benchmark reference for subsequent consistency analysis.

[0039] Calculate the first After time registration, the tag value is the same as the first time. The product of the absolute value of the difference between the registered label values ​​and the exponential decay factor of the registration residual is summed time-series to obtain the registration difference weighting term. This term comprehensively considers the label differences and corresponding registration accuracy between adjacent time points, and assigns lower weights to periods with larger registration residuals through the exponential decay factor. The absolute value of the difference between the perceived structural overlap value and the time-series mean of the perceived structural overlap value is calculated to obtain the structural overlap consistency term, which reflects the degree of deviation of the current structural overlap characteristics from the time-series average level and is used to assess the time-series stability of structural overlap. The product of the time-series variance of the label value and the time-series variance adjustment factor is calculated to obtain the time-series fluctuation adjustment term, which balances the influence of the time-series fluctuation of the label value on the final result through an adjustable parameter. 1 is added to the structural overlap consistency term and the time-series fluctuation adjustment term to obtain the denominator adjustment term, which incorporates the influence of structural overlap consistency and label fluctuation while maintaining computational stability. The registration difference weighting term is divided by the denominator adjustment term to obtain the time-series registration consistency value. The specific calculation formula is as follows: ; In the formula, The temporal registration consistency value reflects the quality level of multi-temporal image registration. Indicates the first The label value after time-registration represents the structural segmentation result at that time. Indicates the first The time-registered label values ​​represent the structural segmentation state at adjacent time points; This represents the registration residual between adjacent time points, reflecting the spatial alignment error between temporal images; This represents the perceived value of structural overlap, quantifying the degree of risk associated with structural overlap. The time-series mean of the perceived structural overlap value provides a benchmark reference for structural overlap characteristics; It represents the temporal variance of the label values, reflecting the stability of the labels over time; This represents the registration residual adjustment factor. The distribution characteristics of historical registration errors are statistically analyzed using the maximum likelihood estimation method for the registration residuals at adjacent time points. The registration residual adjustment factor is obtained by fitting the probability distribution parameters, and its value ranges from 0.05 to 0.5. The time series variance adjustment factor is obtained by using analysis of variance and cross-validation and the optimal split point search algorithm to adjust the time series variance adjustment factor for the registered label values. The value ranges from 0.3 to 1.0.

[0040] In this implementation plan, this step achieves a precise quantitative assessment of the registration quality of multi-temporal chest radiographs by constructing a calculation model for temporal registration consistency values. It comprehensively utilizes a structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction mechanisms, effectively integrating multi-dimensional information such as structural segmentation results, spatial registration errors, structural overlap feature stability, and label temporal volatility. Through the comprehensive calculation of registration difference weighting terms, structural overlap consistency terms, and temporal volatility adjustment terms, the criterion accurately reflects the spatial synchronization and consistency level between temporal images, providing a reliable basis for registration quality in subsequent temporal label fusion and structural function assessment, significantly improving the accuracy and reliability of multi-temporal medical image analysis.

[0041] Optionally, the specific process for spatial synchronization and consistency quantization of temporal images is as follows: real-time comparison of temporal registration consistency values ​​and temporal registration consistency thresholds, and precise quantitative evaluation at each spatial location: When the temporal registration consistency value is less than the temporal registration consistency threshold, it is determined to be a registration consistent area, indicating that the region maintains a good spatial correspondence in multiple temporal images and has a stable structural position. It then enters the temporal label fusion and confidence assessment module and will be processed according to the procedure without special intervention.

[0042] When the temporal registration consistency value is greater than or equal to the temporal registration consistency threshold, it is identified as a registration fluctuation zone, indicating that there is a significant spatial registration deviation in the region, which may affect the accuracy of subsequent analysis. A local temporal correction method based on mutual information maximization and optical flow vector compensation is adopted. By optimizing the information similarity between images and combining pixel motion vectors, the spatiotemporal consistency of local regions can be effectively improved. Dynamic registration correction of images at multiple time points is achieved through residual fitting and consistency smoothing, which can significantly improve the spatial alignment accuracy of image sequences. The model output is marked with low confidence during inference, and neighborhood information is fused for enhancement processing during model training and inference. The robustness and reliability of the results are improved by introducing spatial context information.

[0043] In this implementation plan, this step establishes a quantitative evaluation system for temporal registration consistency, enabling precise judgment and intelligent processing of the spatial synchronization quality of multi-temporal medical images. The system divides image regions into two levels: consistent registration zones and fluctuating registration zones, by comparing temporal registration consistency values ​​with preset thresholds in real time. For consistent zones maintaining good spatial correspondence, a standardized process ensures processing efficiency; while for fluctuating zones with spatial registration deviations, a local correction mechanism based on maximizing mutual information and optical flow vector compensation is activated, achieving dynamic registration correction through residual fitting and consistency smoothing algorithms. Simultaneously, a complete confidence assessment system is established, outputting low-confidence labels for fluctuating zones. Furthermore, a neighborhood information fusion strategy is introduced during model training and inference, effectively improving the reliability and robustness of the analysis results. This processing mechanism not only ensures the spatial alignment accuracy of image sequences but also provides reliable technical support for subsequent quantitative analysis and clinical diagnosis, demonstrating the advantages of intelligent medical image analysis systems in processing complex temporal data.

[0044] Optionally, based on structural overlap discrimination and temporal registration, the specific process of temporal label fusion through fusion response and spatial fluctuation suppression is as follows: Obtain the perceived value of structural overlap, the consistent value of temporal registration, and the number of historical moments. These parameters provide quantitative basis for label fusion from three dimensions: structural characteristics, temporal consistency, and data scale. After registration, the label values ​​of all historical moments are processed using a temporal probability clustering algorithm and a multi-temporal weighted average method to obtain a multi-temporal fusion label probability set. The temporal probability clustering algorithm uses a Gaussian mixture model for probability distribution analysis. The set effectively integrates label information from multiple time points through a probability fusion mechanism. The spatial variance of the label probability is obtained from the multi-temporal fusion label probability set using a standard variance algorithm, accurately reflecting the dispersion of label probabilities in the spatial dimension.

[0045] The absolute value of the difference between the perceived structural overlap value and the historical perceived structural overlap value is calculated, inverted, and multiplied over all historical time points. The M-th root is then used to obtain the structural consistency fusion term. The stability of the structural overlap feature over time is evaluated using a geometric mean algorithm. The product of the temporal registration consistency value and the temporal consistency adjustment factor is calculated, inverted, and exponentially calculated to obtain the registration consistency adjustment term. An exponential decay model is used to quantify the registration quality as an influence weight. The structural consistency fusion term is multiplied by the registration consistency adjustment term to obtain the label confidence fusion response term. A multiplicative fusion strategy integrates both structural and temporal consistency features. The product of the label probability space variance of the multi-temporal fusion label probability set and the label space variance adjustment factor is calculated, and 1 is added to obtain the spatial fluctuation suppression term. Linear weighting and Laplace smoothing are used to effectively balance the impact of spatial fluctuations on the fusion result. The label confidence fusion response term is divided by the spatial fluctuation suppression term to obtain the temporal label fusion value. The specific calculation formula is as follows: ; In the formula, This represents the time-series tag fusion value, which comprehensively reflects the reliability of the tags in terms of time sequence. This represents the perceived value of structural overlap, assessing the current risk level of structural overlap. It represents the perceived value of historical structural overlap, which records the overlapping feature state at past time points; P(x,y) represents the temporal registration consistency value, reflecting the spatial alignment quality of multi-temporal images; P(x,y) represents the multi-temporal fusion label probability set, which integrates the label probability information of historical moments. It represents the spatial variance of label probability, quantifying the volatility of label probability in the spatial dimension; This indicates the number of historical moments and defines the time range for participation in the fusion analysis; The time series consistency adjustment factor is obtained by using maximum likelihood estimation and parameter fitting to calculate the historical time series registration consistency value. The value ranges from 0.1 to 1.0. This represents the label space variance adjustment factor. It is obtained by combining the label probability space variance with the global label distribution characteristics, and through cross-validation and split point search algorithms. The value ranges from 0.1 to 2.0.

[0046] In this embodiment, Table 1 is a data table for evaluating the confidence level of temporal label fusion. It records in detail the registered label values, temporal label fusion values ​​and label confidence levels of different samples at each time point, which is used to quantify the performance of the AI ​​model in evaluating the consistency and confidence level of temporal image labels. Among them, the registered label values ​​of S1 at the five times from t1 to t5 are 0.92, 0.88, 0.90, 0.89, and 0.91, respectively, and the corresponding time-series label fusion values ​​are 0.91, 0.90, 0.93, 0.92, and 0.94, respectively, with all labels having high confidence levels; the registered label values ​​of S2 at each time are 0.70, 0.58, 0.61, 0.45, and 0.48, and the time-series label fusion values ​​are 0.68, 0.60, 0.59, 0.54, and 0.50, with confidence levels ranging from medium to low; the registered label values ​​of S3 at each time are 0.21, 0.19, 0.14, 0.10, and 0.12, and the time-series label fusion values ​​are 0.25, 0.22, 0.18, 0.12, and 0.13, with all labels having low confidence levels. It should be noted that the label residual adjustment factor, standard deviation normalization coefficient, and distribution smoothing parameter not covered in the formula have been standardized in this table to eliminate the influence of additional variables and ensure the comparability of the main evaluation parameters.

[0047] Table 1. Confidence Assessment Data for Time-Series Label Fusion like Figure 3 The figure shows the distribution map of time-series label fusion and registration. Combined with Table 1, it can be seen that there are significant differences in label consistency and confidence levels among different numberings during the time-series changes. The registered label value and time-series label fusion value of number S1 remain at a high level with minimal fluctuation, reflecting strong label consistency and high confidence in the time-series image, belonging to a typical healthy or stable partition. Number S2 shows large fluctuations at different times, with label and fusion values ​​gradually decreasing from medium to low levels, and the confidence level decreasing accordingly, reflecting the questionable characteristics of the label classification boundary area. Number S3 shows significantly low registered label and fusion values ​​at all times with little fluctuation, indicating a high-risk area, suitable for model risk removal and manual review. The distribution map of time-series label fusion and registration visually reflects the differences among different numberings in time-series discrimination, label consistency, and confidence level classification, providing a data foundation for anomaly area identification and model classification decisions.

[0048] In this implementation plan, this step achieves comprehensive evaluation and optimization of multi-temporal image label quality by constructing a computational model for temporal label fusion values. This method innovatively integrates information from multiple dimensions, including structural overlap perception, temporal registration consistency, and label spatial volatility. Through a systematic combination of structural consistency fusion terms, registration consistency adjustment terms, and spatial volatility suppression terms, a complete label confidence evaluation system is established. It not only considers the structural characteristics of a single time point but also comprehensively analyzes the structural stability at historical moments, while effectively suppressing interference from regional volatility through spatial variance adjustment. This multi-factor fusion evaluation mechanism significantly improves the accuracy and reliability of label fusion, providing a high-quality label data foundation for subsequent model training and inference, and effectively ensuring the overall performance and stability of the medical image analysis system.

[0049] Optionally, the specific process of performing time-series label fusion and entering the modeling and inference output flow is as follows: Figure 4 The diagram shows the process flow for graded data processing based on confidence labels. It compares the time-series label fusion value and the time-series label fusion threshold in real time. The time-series label fusion threshold includes a primary fusion threshold and a secondary fusion threshold to ensure the scientific validity and reliability of the grading standard. When the time-series label fusion value is less than or equal to the secondary fusion threshold, it is judged as low-quality label data, whose confidence is insufficient to support reliable model training and inference. The label is directly removed and does not participate in model training and inference. At the same time, a fusion risk log is generated on the data side to record quality assessment parameters, reasons for removal and timestamp information, triggering physical supplementary collection and manual review. Through multiple protection mechanisms, the quality of input data is ensured to meet clinical diagnostic standards.

[0050] When the time-series label fusion value is greater than the secondary fusion threshold but less than or equal to the primary fusion threshold, special processing is required while preserving its data value to generate structural time-series fusion questionable labels. The modeling process adopts multimodal label fusion technology, which improves the reliability of labels by integrating image features and multi-source information of indicators. During inference, low-confidence labels are output.

[0051] When the time sequence label fusion value is greater than the first-level fusion threshold, all its quality indicators reach the optimal standard and are directly included in the structural function assessment model training. It enters the modeling and inference output module and serves as the core training sample to participate in model parameter optimization. During inference, it outputs high-confidence labels, and the analysis results have high credibility.

[0052] In this implementation plan, this step achieves refined management and intelligent grading of multi-temporal image label quality by establishing a three-level evaluation of temporal label fusion values. Based on preset first- and second-level fusion thresholds, label data is divided into three levels: high quality, medium quality, and low quality, and differentiated processing strategies are adopted for each level. For low-quality labels, a strict rejection mechanism is implemented and a review process is triggered to ensure the reliability of the data source; for medium-quality labels, multimodal fusion technology is used to improve their usability, and their confidence levels are clearly marked; high-quality labels are directly used for model training and inference to fully realize their value. This grading mechanism ensures the quality of model training data while maximizing the retention of usable data. Confidence marking provides clear reference for clinical use, significantly improving practicality and reliability, and providing important support for accurate decision-making in medical image analysis.

[0053] Optionally, under the constraints of three higher-order criteria—structure, temporal, and label—the specific process of constructing, training, inferring, and outputting individualized results of the structural function assessment model using multimodal feature enhancement is as follows: Multimodal feature fusion and a hierarchical loss function are employed, with the loss function dynamically adjusting weights based on sample quality. Original image data and lung function test indicators are input to construct a multi-source feature input space. Samples are dynamically selected and hierarchically trained based on structural overlap perception value criteria, temporal registration consistency value criteria, and temporal label fusion value criteria. A sample quality assessment system is established, strengthening supervised learning of high-consistency, high-confidence samples to improve model learning efficiency. For low-confidence regions, neighboring samples are used... Domain information enhancement is achieved by aggregating contextual features through graph convolutional networks and fusing labels using an attention mechanism to focus features on key regions. Robust training is applied to handle noisy labels, and a symmetric cross-entropy noise-resistant loss function is used to jointly model the probability distribution of multiple temporal labels, constructing a structural function assessment model. During the inference phase, the structural function assessment model simultaneously outputs hierarchical results of structural overlap perception values, temporal registration consistency values, and temporal label fusion values, forming a multi-dimensional quality assessment report. This detects and locates abnormal ventilation functional areas, boundary hierarchical drift areas, and structurally low-consistency blocks, achieving accurate abnormal region identification. Abnormal criteria regions are fed back to labeling processing in real time, establishing a closed-loop optimization mechanism, such as... Figure 5 The image shows a heat map of the spatial distribution of abnormal ventilation areas. The spatial distribution of abnormal ventilation areas is visualized in a two-dimensional heat map format, providing an intuitive diagnostic reference. By jointly analyzing the distribution of structural features and pulmonary function test indicators, a structural-functional correlation distribution report is generated, revealing the deep correlation between imaging features and function. Finally, the report outputs confidence level labels, heat maps of abnormal ventilation areas, structural-functional correlation distribution, criterion grading results, and a full-process quality traceability report, forming a complete diagnostic decision support package.

[0054] In this implementation plan, this step achieves end-to-end optimization of the structural function assessment model from construction to output by constructing an intelligent analysis process constrained by three high-order criteria: structure, temporal sequence, and labeling. During the model construction and training phases, multimodal feature fusion and a hierarchical loss function are innovatively employed. Dynamic sample selection and quality weighting mechanisms significantly improve the accuracy and robustness of model learning. Neighborhood information enhancement, attention mechanisms, and noise-resistant training strategies are adopted to address the characteristics of different confidence levels, effectively overcoming the challenges posed by low-quality data. In the inference output phase, the model not only accurately identifies and locates various abnormal regions but also generates comprehensive reports with high clinical value through multidimensional quality assessment, two-dimensional heatmap visualization, and structural function correlation analysis. The final output diagnostic decision support package integrates confidence labeling, abnormal region visualization, functional correlation analysis, and key information for end-to-end quality traceability, providing comprehensive, reliable, and interpretable decision-making support, fully demonstrating the significant advantages of intelligent medical systems in accurate diagnosis and personalized treatment.

[0055] On the other hand, it provides a precise analysis and management system for perioperative medical imaging based on AI models. Figure 2 This is a schematic diagram illustrating the structure of a perioperative medical imaging image precision analysis and management system based on an AI model, according to an exemplary embodiment. This system is used for a method of precision analysis and management of perioperative medical imaging images based on an AI model. (Refer to...) Figure 2 The system includes an acquisition and preprocessing module, a structural overlap region identification and segmentation module, a multi-temporal image registration module, a temporal label fusion and confidence assessment module, and a modeling and inference output module. Among them:

[0056] The data acquisition and preprocessing module is used to acquire chest X-ray structural feature data and obtain historical statistical data; and to preprocess the chest X-ray structural feature data and historical statistical data. The structural overlap region identification and segmentation module is used to identify and segment structural overlap regions based on chest X-ray structural feature data by using gradient field response and geometric density correction. The multi-temporal image registration module is used to construct temporal registration criteria by employing a structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction, to perform spatial synchronization and consistency quantification of temporal images. The temporal label fusion and confidence assessment module is used to perform temporal label fusion based on structural overlap discrimination and temporal registration, through fusion response and spatial fluctuation suppression, and then enter the modeling and inference output process; The modeling and inference output module is used to construct, train, infer, and output individualized results of a structural function assessment model under the constraints of three higher-order criteria: structure, time series, and label. This is achieved through multimodal feature enhancement. In this implementation plan, the system constructs a complete intelligent medical image analysis and management system through the collaborative work of five core modules. Based on data acquisition and preprocessing, the system effectively solves the structural overlap problem in chest X-ray analysis through a structural overlap region identification and segmentation module, ensures the consistency of time-series data through a multi-temporal image registration module, improves the reliability of label data through a time-series label fusion and confidence assessment module, and finally achieves individualized structural function assessment through a modeling and inference output module. This system innovatively integrates three high-order criteria—structure, temporality, and label—throughout the entire analysis process. Through multimodal feature enhancement and intelligent algorithm fusion, it not only significantly improves the accuracy and reliability of medical image analysis but also provides a complete solution including quality traceability, confidence assessment, and visual reporting, providing strong technical support for clinical diagnostic decision-making and demonstrating the advanced nature and practicality of artificial intelligence technology in the field of medical image analysis.

[0057] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0058] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0059] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0060] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0061] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, systems, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0063] In the embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0065] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0066] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for precise analysis and management of perioperative medical examination images based on AI models, characterized in that, The method includes: S1. Collect perioperative chest X-ray structural feature data of patients and obtain historical statistical data; preprocess the chest X-ray structural feature data and historical statistical data. S2, based on chest X-ray structural feature data, uses gradient field response and geometric density correction to identify and segment overlapping structural regions; S3 employs a structure-guided rigid and affine joint registration algorithm, combining residual weighting and variance correction to construct temporal registration criteria, and performs spatial synchronization and consistency quantification of temporal images; S4, based on structural overlap discrimination and temporal registration, performs temporal label fusion through fusion response and spatial fluctuation suppression, and then enters the modeling and inference output process; S5, under the constraints of three higher-order criteria—structure, temporal, and label—employs multimodal feature enhancement to realize the construction, training, inference, and individualized output of the structural function assessment model.

2. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process of collecting chest X-ray structural feature data and obtaining historical statistical data, and preprocessing the chest X-ray structural feature data and historical statistical data is as follows: Collect chest X-ray structural feature data, which includes: two-dimensional coordinate data of the original chest X-ray image, grayscale values ​​of the original image, pixel grayscale data at multiple time points, and original image data; Obtain historical statistics, which include: lung function test indicators, grayscale sliding windows, and the number of historical moments; By employing edge detection algorithms and local principal curvature analysis, spatial contour optimization and redundant point removal are performed on chest X-ray structural feature data. Pixel-level distributed resampling and adaptive sliding window filtering are used to suppress outliers and smooth short-term fluctuations in the original chest X-ray image's two-dimensional coordinate data and grayscale values. Temporal interpolation and piecewise linear regression are used to complete missing data and extract trend features from historical statistical data. Distribution standardization and linear normalization algorithms are then used to standardize and normalize the chest X-ray structural feature data and historical statistical data.

3. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process of correcting overlapping structural regions based on chest X-ray structural feature data using gradient field response and geometric density is as follows: The process involves acquiring raw image data, 2D coordinate data of the original chest X-ray image, pixel grayscale data at multiple time points, and grayscale values ​​of the original image. From the raw image data, image segmentation algorithms are used to obtain the structural boundary pixel set, bone width, gap pixel distance, and background region data. From the 2D coordinate data of the original chest X-ray image, an automatic segmentation algorithm based on a UNet convolutional neural network is used to obtain the structural prior response values. From the grayscale values ​​of the original image, the Sobel operator is used to calculate the pixel neighborhood to obtain the spatial gradient magnitude. Finally, the structural boundary pixel set, bone width, and gap pixel distance are analyzed using OpenCV. Local geometric density adjustment values ​​are obtained through contour extraction and the minimum bounding rectangle method; edge response intensity is obtained from the original image grayscale values ​​using the Canny edge detection algorithm; a grayscale sliding window centered at (x,y) is used, where (x,y) represents the horizontal coordinate x and vertical coordinate y of a pixel in the original image data, and the sliding window grayscale information entropy is obtained from the original image grayscale values ​​using the Shannon entropy algorithm; the local background noise standard deviation is obtained from the background region data using the Gaussian filtering algorithm and local standard deviation calculation; and temporal grayscale fluctuation is obtained from the multi-time point pixel grayscale data using the rigid registration algorithm. The gradient weighting term is obtained by multiplying the structural prior response value by the spatial gradient magnitude. The weighted response term is obtained by multiplying the local geometric density adjustment value and the gradient weighting term, and then adding the edge response intensity. The composite suppression term is obtained by multiplying the sliding window grayscale information entropy, the background noise adjustment factor, and the local background noise standard deviation, and then adding the temporal grayscale fluctuation and the temporal change adjustment factor. The weighted response term is divided by the composite suppression term to obtain the structural overlap perception value.

4. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process for identifying and segmenting overlapping structural regions is as follows: Real-time comparison of structural overlap sensing values ​​and structural overlap sensing thresholds, including primary and secondary sensing thresholds: When the structural overlap perception value is less than the secondary perception threshold, a normal structural label is generated and enters the temporal label fusion and confidence evaluation module without special masking or adjustment. When the structural overlap perception value is greater than or equal to the second-level perception threshold and less than the first-level perception threshold, a structural overlap doubt label is generated and a low confidence label is output during the structural function assessment model inference. When the perceived structural overlap value is greater than or equal to the first-level perception threshold, high-resolution reconstruction and local contrast enhancement are triggered to generate structural overlap risk labels. High-risk region patches are extracted through structural masks, and a small Patch-UNet local sub-model is used to perform feature depth extraction and discrimination of the region.

5. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process of constructing the temporal registration criterion using the structure-guided rigid and affine joint registration algorithm, combined with residual weighting and variance correction, is as follows: Obtain raw image data and structural overlap perception values; for multi-temporal raw image data, use a rigid registration algorithm based on structural segmentation and an automatic segmentation algorithm based on UNet convolutional neural network to obtain the first... After time registration, the tag value and the first The registered tag value at any time; Spatial registration algorithms for multi-temporal raw image data are used to align the data through feature point detection and affine transformation, and the time sequence is calculated. and The Euclidean distance between the pixel space coordinates is used to obtain the registration residuals at adjacent time steps; The temporal mean of the structural overlap sensing values ​​at all time points is obtained by using the mean algorithm; the temporal variance of the registered label values ​​is obtained by using the standard deviation algorithm. Calculate the first After time registration, the tag value is the same as the first time. The product of the absolute value of the difference between the registered tag values ​​and the registration residual exponential decay factor is summed time-series to obtain the registration difference weighting term; The absolute value of the difference between the perceived structural overlap value and the time-series mean of the perceived structural overlap value is calculated to obtain the structural overlap consistency term; The product of the time series variance of the label value and the time series variance adjustment factor is calculated to obtain the time series fluctuation adjustment term; Add 1 to the structural overlap consistency term and the temporal fluctuation adjustment term to obtain the denominator adjustment term; divide the registration difference weighted term by the denominator adjustment term to obtain the temporal registration consistency value.

6. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process for spatial synchronization and consistency quantization of temporal images is as follows: Real-time comparison of time-series registration consistency values ​​and time-series registration consistency thresholds: When the time-series registration consistency value is less than the time-series registration consistency threshold, it is determined to be a registration consistency region and enters the time-series label fusion and confidence assessment module without special intervention. When the temporal registration consistency value is greater than or equal to the temporal registration consistency threshold, it is determined to be a registration fluctuation zone. A local temporal correction method based on mutual information maximization and optical flow vector compensation is adopted. Dynamic registration correction of multi-time images is achieved through residual fitting and consistency smoothing. The model outputs a low confidence label during inference, and neighborhood information is fused for enhancement processing during model training and inference.

7. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process of temporal tag fusion based on structural overlap discrimination and temporal registration, through fusion response and spatial fluctuation suppression, is as follows: Obtain structural overlap sensing values, temporal registration consistency values, and the number of historical moments; obtain a multi-temporal fusion label probability set by using temporal probabilistic clustering algorithm and multi-temporal weighted average method for the registered label values ​​of all historical moments; obtain the label probability space variance by using standard variance algorithm for the multi-temporal fusion label probability set. Calculate the absolute value of the difference between the perceived structural overlap value and the historical perceived structural overlap value, invert it, multiply it over all historical moments, and take the Mth root to obtain the structural consistency fusion term. Calculate the product of the temporal registration consistency value and the temporal consistency adjustment factor, take the negative number, and perform an exponential operation to obtain the registration consistency adjustment term; multiply the structural consistency fusion term and the registration consistency adjustment term to obtain the tag confidence fusion response term; Calculate the product of the spatial variance of the tag probability set and the tag spatial variance adjustment factor, add 1, and obtain the spatial fluctuation suppression term; Divide the tag confidence fusion response term by the spatial fluctuation suppression term to obtain the temporal tag fusion value.

8. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process of performing time-series label fusion and entering the modeling and inference output flow is as follows: Real-time comparison of time-series label fusion values ​​and time-series label fusion thresholds, which include primary fusion thresholds and secondary fusion thresholds: When the time sequence label fusion value is less than or equal to the secondary fusion threshold, the label is directly removed and does not participate in model training and inference. At the same time, a fusion risk log is generated on the data side, triggering physical supplementary data collection and manual review. When the time series label fusion value is greater than the second-level fusion threshold and less than or equal to the first-level fusion threshold, a structural time series fusion question label is generated. The modeling process adopts multimodal label fusion technology, and low confidence labels are output during inference. When the time sequence label fusion value is greater than the first-level fusion threshold, it is directly included in the training of the structural function assessment model and enters the modeling and inference output module. During inference, a high-confidence label is output.

9. The method for precise analysis and management of perioperative medical examination images based on an AI model according to claim 1, characterized in that, The specific process of constructing, training, inference, and outputting individualized results of the structural function assessment model under the constraints of three higher-order criteria: structure, time series, and label is as follows: This study employs multimodal feature fusion and a hierarchical loss function. Inputting raw image data and pulmonary function test indicators, it dynamically selects and hierarchically trains samples based on structural overlap perception criteria, temporal registration consistency criteria, and temporal label fusion criteria. Supervised learning of high-consistency, high-confidence samples is strengthened. For low-confidence regions, neighborhood information enhancement and an attention mechanism are used to fuse labels, and robust training is applied to handle noisy labels. The probability distribution of multiple temporal labels is jointly modeled to construct a structural function assessment model. During the inference phase, the structural function assessment model simultaneously outputs hierarchical results of structural overlap perception values, temporal registration consistency values, and temporal label fusion values. It detects and locates abnormal ventilation functional areas, boundary hierarchical drift areas, and structurally low-consistency blocks. Abnormal criterion areas are fed back to the labeling process in real time. The spatial distribution of abnormal ventilation areas is visualized using a two-dimensional heatmap. A structural function correlation distribution report is generated through joint analysis of structural feature distribution and pulmonary function test indicators. Finally, the system outputs confidence labels, a heatmap of abnormal ventilation areas, structural function correlation distribution, criterion hierarchical results, and a full-process quality traceability report.

10. A perioperative medical examination image precision analysis and management system based on an AI model, employing the perioperative medical examination image precision analysis and management method based on an AI model as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect perioperative chest X-ray structural feature data of surgical patients and obtain historical statistical data. Preprocessing of chest X-ray structural feature data and historical statistical data; The structural overlap region identification and segmentation module is used to identify and segment structural overlap regions based on chest X-ray structural feature data by using gradient field response and geometric density correction. The multi-temporal image registration module is used to construct temporal registration criteria by combining structure-guided rigid and affine joint registration algorithms with residual weighting and variance correction, and to perform spatial synchronization and consistency quantification of temporal images. The temporal label fusion and confidence assessment module is used to perform temporal label fusion based on structural overlap discrimination and temporal registration, through fusion response and spatial fluctuation suppression, and then enter the modeling and inference output process; The modeling and inference output module is used to construct, train, infer, and output individualized results of the structural function assessment model under the constraints of three types of high-order criteria: structure, time series, and label, using multimodal feature enhancement.

Images