Digital archiving and management methods for historical radiology images

Abstract

This invention relates to the field of medical information and discloses a method for the digital archiving and management of historical radiology images. The method includes the following steps: acquiring and digitizing historical radiology film images and associated text information; generating enhanced digital archives with standardized metadata and establishing patient associations with existing image archiving and communication systems; adaptively generating spatially standardized mapping representations at different levels based on evaluation results; performing context-aware depth comparison and quantitative analysis of historical and current images during clinical retrieval; and generating structured reports based on the results of the depth comparison and quantitative analysis. By using intelligent graph matching and Bayesian inference with the deep semantic features of historical images and information from the existing system, high-precision and automated association between historical images and current patient records is achieved. Furthermore, by constructing hierarchical spatial mapping representations, the transformation of historical images from static archives into computable and quantifiable intelligent assets is realized.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a method for digital archiving and management of historical radiological images. Background Technology

[0002] Radiological imaging data, especially historical images on silver halide film, are valuable data assets for recording the progression of a patient's disease, evaluating treatment effectiveness, and conducting medical research. With the advancement of hospital information technology, modern image management systems centered on image archiving and communication systems have become standard practice in clinical work. However, the potential clinical and research value of a large number of historical film archives generated before the widespread adoption of PACS has not been effectively realized due to issues such as physical storage, fragility and aging, and information silos.

[0003] Currently, the common technical solution for the digital management of historical film images is a linear process of "scanning-archiving-retrieval." This involves converting the images into digital images using a professional scanner and storing them. Then, basic identifiers such as patient names and examination dates are added manually or using optical character recognition technology to create an electronic catalog for retrieval. However, this only achieves static data storage. When retrieval is needed, operators search the catalog using keywords such as patient name and ID to find the corresponding digital file and view it. This lacks the ability to deeply semantically associate and match images with patients based on the image content itself. When the historical film identification information is vague, incorrect, or incomplete, it can easily lead to association failures or errors, making it impossible to accurately link the images to the patient records in the current PACS. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for the digital archiving and management of historical radiology images, solving the problem that existing digital archiving and management methods for historical radiology images are prone to association failures or errors, and cannot be accurately linked to patient records in the current PACS.

[0005] To achieve the above objectives, the present invention provides a method for the digital archiving and management of historical radiological images, comprising the following steps:

[0006] Collect historical radiology film images and associated text information and digitize them to generate primary digital archives;

[0007] The primary digital archives are intelligently semantically annotated and associated to generate enhanced digital archives with standardized metadata and established patient association with existing image archiving and communication systems;

[0008] The enhanced digital archives are subjected to hierarchical quality assessment, and spatially normalized mapping representations of different levels are adaptively generated based on the assessment results.

[0009] Based on the aforementioned spatial standardized mapping representation, context-aware depth comparison and quantitative analysis of historical images and current images are performed during clinical retrieval.

[0010] Structured reports are generated based on the results of in-depth comparison and quantitative analysis, and iterative optimization is carried out using clinical feedback.

[0011] By adopting the above technical solution, intelligent graph matching and Bayesian inference are performed between the deep semantic features of historical images and current system information, thereby achieving high-precision and automated association between historical images and current patient records. Furthermore, by constructing a hierarchical spatial mapping representation, a unified spatiotemporal coordinate system is established for the comparison of images across periods. This realizes the transformation of historical images from static archives to calculable and quantifiable intelligent assets, solving the problem that the existing digital archiving and management of radiology historical image archives is prone to association failures or errors, and cannot be accurately linked to patient records in the current PACS.

[0012] Preferably, the intelligent semantic annotation and association of the primary digital archives includes the following steps:

[0013] High-level semantic feature vectors of historical radiology films were extracted using a deep convolutional neural network.

[0014] Integrate the text information in the primary digital archives with the high-level semantic feature vectors to construct a metadata knowledge graph;

[0015] Based on the high-level semantic feature vector, a graph matching algorithm is used to find matching patients for historical radiology films in the existing image archiving and communication system, thereby obtaining effective patient associations.

[0016] Based on effective patient associations and metadata knowledge graphs, the metadata of historical radiology films is completed and standardized through Bayesian inference to generate standardized semantic tags.

[0017] Preferably, the step of using a graph matching algorithm to find matching patients for historical radiology films in an existing image archiving and communication system includes the following steps:

[0018] Based on the high-level semantic feature vector, a bipartite graph is constructed with historical radiology film images as the first node and patient image sets in the current image archiving and communication system as the second node.

[0019] Calculate the edge weights between the first node and each second node, whereby the edge weights are calculated based on the feature vector similarity of the corresponding images and the time compatibility of the inspection.

[0020] The weighted bipartite graph maximum matching algorithm is used to solve the matching relationship based on the edge weights, and the matching with weights exceeding a preset threshold is confirmed as a valid patient association.

[0021] Preferably, the step of completing and standardizing the metadata of historical radiology film images through Bayesian inference includes the following steps:

[0022] Based on the metadata knowledge graph and the effective patient associations, the prior probability distribution of standardized labels for historical radiology films is calculated.

[0023] The likelihood probability of each standardized label under the condition of historical radiology film image features was calculated using a classifier model.

[0024] The posterior probability of the standardized labels of the historical radiology films is calculated by combining the prior probability distribution, the likelihood probability, and the transition probability obtained from the current image labels associated with the effective patients.

[0025] The standardized label with the highest posterior probability is selected as the annotation result to obtain the standardized semantic label.

[0026] Preferably, the hierarchical quality assessment of the enhanced digital archive includes the following steps:

[0027] The clarity, noise level, and structural integrity of historical radiology films in enhanced digital archives are evaluated using a multi-task learning model.

[0028] A comprehensive quality score is calculated based on the assessment results of the clarity, noise level, and structural integrity.

[0029] Historical radiology films are classified into three quality levels: first, second, and third, based on preset quality scoring thresholds.

[0030] Preferably, the step of adaptively generating spatial normalization mapping representations at different levels based on the evaluation results includes the following steps:

[0031] For historical radiographic films of the first quality grade, the deformation field to the standard anatomical space is calculated using the large deformation differential homeomorphic registration algorithm as the first mapping characterization.

[0032] For historical radiographic films of the second quality level, a sparse similarity transformation matrix to the standard anatomical space is calculated using a feature point matching algorithm as the second mapping representation.

[0033] For historical radiographic films of the third quality level, the global semantic feature vector is extracted by an encoder network as the third mapping representation.

[0034] The first mapping representation, the second mapping representation, or the third mapping representation is stored as the spatially standardized mapping representation of the corresponding historical radiology film image.

[0035] Preferably, in the context-aware depth comparison and quantitative analysis of historical images and current images during clinical retrieval, the context-aware depth comparison includes the following steps:

[0036] When reviewing current images in clinical settings, the contextual relevance score of historical radiology films is calculated based on effective patient association, standardized semantic labels, temporal relevance, and feature similarity.

[0037] The associated historical radiology films are sorted and recommended based on the context relevance score.

[0038] Preferably, the context-aware depth comparison further includes the following steps:

[0039] For historical radiology films with a first or second mapping representation, the pre-stored mapping representation is called and combined with the mapping of the current image to the standard space to perform initial registration and obtain the registration result.

[0040] A deep learning-based deformation registration network is used to adjust the registration results, and the Monte Carlo Dropout method is used to output the displacement field mean and uncertainty map of the registration results.

[0041] Preferably, in the context-aware depth comparison and quantitative analysis of historical images and current images during clinical retrieval, the quantitative analysis includes the following steps:

[0042] Multi-dimensional feature extraction is performed on the same regions of interest in historical radiology films and current images after registration adjustment.

[0043] Calculate the relative rate of change of the multidimensional features between historical radiology films and current images;

[0044] The statistical significance of the relative rate of change was calculated using statistical hypothesis testing methods, and its confidence interval was calculated.

[0045] Preferably, the iterative optimization using clinical feedback includes the following steps:

[0046] Collect and record correction data for semantic annotation, registration results, or quantitative analysis results;

[0047] A feedback dataset is constructed based on the collected correction data;

[0048] The parameters of the deep convolutional neural network and classifier model for semantic annotation, the multi-task learning model for hierarchical quality assessment, or the deformation registration network for registration fine-tuning are optimized and trained using the feedback dataset at preset times.

[0049] This invention provides a method for digital archiving and management of historical radiological images. It offers the following advantages:

[0050] 1. This invention achieves high-precision and automated association between historical images and current patient records by intelligent graph matching and Bayesian inference using the deep semantic features of historical images and current system information. Furthermore, by constructing a hierarchical spatial mapping representation, a unified spatiotemporal coordinate system is established for the comparison of images across different periods. This realizes the transformation of historical images from static archives to calculable and quantifiable intelligent assets, solving the problem that the existing digital archiving and management of radiology historical image archives is prone to association failures or errors, and cannot be accurately linked to patient records in the current PACS.

[0051] 2. This invention extracts deep semantic features of images through deep convolutional neural networks and uses graph matching algorithms and Bayesian inference to achieve high-confidence intelligent association and metadata standardization at the patient level, enabling massive historical archives to be accurately and efficiently integrated into the modern medical information ecosystem. By introducing an intelligent semantic annotation and association mechanism based on deep features and graph matching, the accuracy and automation level of the fusion of historical images with the current PACS system are improved.

[0052] 3. This invention performs multi-dimensional quality assessment of digital images, including sharpness, noise, and structural integrity, and adaptively selects the most suitable processing path based on the scoring results: high-precision registration is performed on high-quality images to generate a dense deformation field, feature point matching is performed on medium-quality images to generate a sparse transformation, and global semantic features are extracted from low-quality images. This ensures the comparison accuracy of high-quality images and achieves a balance between processing efficiency and result reliability. By implementing hierarchical quality assessment and adaptive spatial mapping strategies, resource utilization is optimized and a reliable foundation is laid for accurate comparison of images across different periods.

[0053] 4. This invention enhances the real-time support capability and quantitative analysis value of historical images in clinical decision-making by constructing a context-aware deep comparison process based on pre-computed mapping representations. When doctors access current images, the most relevant historical images can be intelligently recommended based on multi-dimensional context such as patient, time, and semantic content. By calling pre-generated spatially standardized mapping representations, the initial alignment of anatomical structures between historical and current images can be quickly completed. After fine-tuning by a deep learning network, sub-pixel-level accurate registration can be achieved. On this basis, multi-dimensional feature extraction and quantitative analysis can be performed on specific regions, and statistical significance assessment of changes can be provided, thereby elevating subjective experience comparison to objective data-driven accurate diagnosis. Attached Figure Description

[0054] Figure 1 This is a flowchart of the digital archiving and management method for historical radiological images proposed in this invention;

[0055] Figure 2 This is an architecture diagram of the digital archiving and management system for historical radiological images proposed in an embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1:

[0058] In a first embodiment of the present invention, the present invention provides a method for digital archiving and management of historical radiological images, such as... Figure 1 As shown, it includes the following steps:

[0059] Collect historical radiology film images and associated text information and digitize them to generate primary digital archives;

[0060] Specifically, the process involves collecting and digitizing historical radiology film images and associated text information to generate preliminary digital archives. The aim is to transform physical media such as silver halide film and paper records into digital objects that can be processed by computers.

[0061] First, the film is physically pretreated in a clean environment. Surface stains are removed with a professional lint-free cloth and a mild detergent. Curled or deformed film is reshaped with a flatbed press to avoid geometric distortion during scanning.

[0062] Next, the pre-processed film is placed into a calibrated medical film scanner. The scanner has an optical resolution of at least 400 dpi and a grayscale depth of 16 bits. Automatic calibration is enabled to ensure parameter consistency. The scanner linearly converts the optical density values ​​on the film to digital grayscale values ​​according to the specifications for secondary image capture in standard digital imaging and communication protocols. This conversion process follows a predetermined mapping relationship between optical density and pixel values. The core algorithm is as follows: ,in, To output image pixel values, This is the scanner's maximum pixel value. This is the optical density value of the film. The calibration coefficients are used to ultimately output a digital image matrix with high dynamic range.

[0063] At the same time, the accompanying paper documents are scanned using a document scanner with a resolution of no less than 600 dpi, and text content such as patient information and examination information is extracted through optical character recognition.

[0064] Finally, the digitized image data and extracted text information are encapsulated to create a digital image and communication object with a unique identifier for each inspection. This is written into the image data, identifier metadata, and key text fields to form a structured primary digital archive.

[0065] Intelligent semantic annotation and association are performed on primary digital archives to generate enhanced digital archives with standardized metadata and patient association with existing image archiving and communication systems;

[0066] Furthermore, intelligent semantic annotation and association are performed on the primary digital archives, including the following steps:

[0067] High-level semantic feature vectors of historical radiology films were extracted using a deep convolutional neural network.

[0068] Integrate textual information from primary digital archives with high-level semantic feature vectors to construct a metadata knowledge graph;

[0069] Based on high-level semantic feature vectors, a graph matching algorithm is used to find matching patients for historical radiology films in the existing image archiving and communication system, thus obtaining effective patient associations.

[0070] Based on effective patient associations and metadata knowledge graphs, the metadata of historical radiology films is completed and standardized through Bayesian inference to generate standardized semantic tags.

[0071] Furthermore, using graph matching algorithms to find matching patients for historical radiology films in existing image archiving and communication systems includes the following steps:

[0072] Based on high-level semantic feature vectors, a bipartite graph is constructed with historical radiology film images as the first node and patient image sets in the current image archiving and communication system as the second node.

[0073] Calculate the edge weights between the first node and each second node. The edge weights are calculated based on the feature vector similarity of the corresponding images and the time compatibility of the inspection.

[0074] The weighted bipartite graph maximum matching algorithm is used to solve the matching relationship based on edge weights, and the matching with weights exceeding a preset threshold is identified as a valid patient association.

[0075] Furthermore, the metadata of historical radiology film images is supplemented and standardized through Bayesian inference, including the following steps:

[0076] Based on metadata knowledge graphs and effective patient associations, the prior probability distribution of standardized labels for historical radiology films is calculated.

[0077] The likelihood probability of each standardized label under the condition of historical radiology film image features was calculated using a classifier model.

[0078] By combining the prior probability distribution, the likelihood probability, and the transition probability obtained from the current image labels associated with effective patients, the posterior probability of the standardized labels of historical radiology films is calculated.

[0079] The standardized label with the highest posterior probability is selected as the annotation result to obtain the standardized semantic label.

[0080] Specifically, intelligent semantic annotation and association are carried out on primary digital archives to generate enhanced digital archives with standardized metadata and patient association with existing image archiving and communication systems. The aim is to give machine-understandable semantics to the original digital data and establish accurate cross-temporal and spatial associations.

[0081] First, high-level semantic feature vectors of historical radiology films are extracted using a deep convolutional neural network. A convolutional neural network model pre-trained on a large medical image dataset is employed. The top-level classifier is removed, and the output of the final fully connected layer is used as the feature vector, which encodes the anatomical structure and potential pathological features of the image.

[0082] Next, the textual information of the primary digital archives is integrated with the aforementioned high-level semantic feature vectors to construct an initial metadata knowledge graph. Each historical examination is treated as an entity, and the initially identified patient identifier, examination date, and extracted text keywords are entered as attributes. These attributes are then embedded using feature vectors to build a preliminary network of relationships between entities.

[0083] Subsequently, based on high-level semantic feature vectors, a graph matching algorithm is used to match historical images to assigned patients within the existing system. A bipartite graph matching problem is constructed and solved, defining the bipartite graph. Contains a set of vertices and , Historical image nodes to be associated , This is a subset of image nodes grouped by patient in the current system. Historical image nodes are calculated. Compared with existing image nodes Edge weights between: ,in The cosine similarity function is used. and To check the timestamp, and The scale parameter is used. A weighted bipartite graph maximum matching algorithm is applied to each historical image node. Find in the set The corresponding node with the highest total matching weight is selected. If the highest weight exceeds a preset threshold, then a valid patient association is confirmed between this historical image and the corresponding current patient.

[0084] Then, based on the effective patient association and metadata knowledge graph, historical image metadata is completed and standardized using a Bayesian inference framework. Standardized tags are obtained. Prior probability distribution Using a classifier model Calculate the likelihood probability Obtain the label transfer probability from the existing image standard labels. Substituting into Bayes' formula: The standardized label that maximizes the posterior probability is selected as the final semantic label, thus completing high-quality semantic annotation of historical images and accurate association with patients, generating enhanced digital archives that can be integrated into the modern medical information ecosystem.

[0085] A hierarchical quality assessment is performed on the enhanced digital archives, and spatially standardized mapping representations at different levels are adaptively generated based on the assessment results.

[0086] Furthermore, a tiered quality assessment of the enhanced digital archives is conducted, including the following steps:

[0087] The clarity, noise level, and structural integrity of historical radiology films in enhanced digital archives are evaluated using a multi-task learning model.

[0088] A comprehensive quality score is calculated based on the assessment results of sharpness, noise level, and structural integrity.

[0089] Historical radiology films are classified into three quality levels: first, second, and third, based on preset quality scoring thresholds.

[0090] Furthermore, based on the evaluation results, spatial normalization mapping representations at different levels are adaptively generated, including the following steps:

[0091] For historical radiographic films of the first quality grade, the deformation field to the standard anatomical space is calculated using the large deformation differential homeomorphic registration algorithm as the first mapping characterization.

[0092] For historical radiographic films of the second quality level, a sparse similarity transformation matrix to the standard anatomical space is calculated using a feature point matching algorithm as the second mapping representation.

[0093] For historical radiographic films of the third quality level, the global semantic feature vector is extracted by an encoder network as the third mapping representation.

[0094] The first mapping representation, the second mapping representation, or the third mapping representation is stored as the spatially normalized mapping representation of the corresponding historical radiology film image.

[0095] Specifically, conducting tiered quality assessments of enhanced digital archives and adaptively generating spatially standardized mapping representations at different levels based on the assessment results involves matching corresponding computing resources and processing strategies according to the physical and informational quality of historical images, thus building an efficient and robust standardized foundation for subsequent accurate cross-period comparisons.

[0096] First, a multi-task learning model is used to quantitatively evaluate the quality of historical images. The model shares a backbone feature extraction network and outputs three evaluation branches in parallel: the first branch assesses sharpness using statistical features of gradient magnitudes in local image regions; the second branch analyzes the energy of high-frequency components or estimates noise variance to assess noise levels; and the third branch, relying on a lightweight pre-trained anatomical structure segmentation network, evaluates the sharpness and coherence of key anatomical contours to determine structural integrity. A comprehensive quality score is then calculated based on the results of these three branches. The formula is: ,in For normalized sharpness scoring, For normalized noise level scoring, To normalize the structural integrity score, , , The preset weighting coefficients sum to 1 are used to balance the three indicators, and the final output is a comprehensive quality score between 0 and 1. .

[0097] The images are divided into three quality levels according to a preset scoring threshold. The thresholds are set as T1 = 0.7 and T2 = 0.4. Images with Q ≥ T1 are high-quality images of the first level, images with T1 > Q ≥ T2 are medium-quality images of the second level, and images with Q < T2 are low-quality images of the third level.

[0098] Subsequently, a spatially normalized mapping representation is adaptively generated according to the quality level: for high-quality images, the large deformation diffeomorphic registration algorithm is used to calculate the deformation field to the standard anatomical space, and the following energy function is optimized to achieve: , where represents the diffeomorphic transformation from the image space to the standard template space, which is generated by integrating the time-dependent velocity field over the time interval from zero to one. Here, the symbol represents function composition operation, which is the result of how one function acts on another function. That is, it means that the original image is spatially resampled through the inverse transformation of the deformation field to obtain the image. The first term is the regularization term of the velocity field, ensuring the smoothness and rationality of the transformation. The second term is the image similarity term, which measures the difference between the transformed image and the standard template image . is the weighting parameter, and the smooth and reversible deformation field is obtained by solving as the first mapping representation; for medium-quality images, the feature point matching algorithm is used to extract stable key points and match them with the standard template anatomical landmark points, and the similarity transformation matrix is estimated through the random sample consensus algorithm as the second mapping representation. This method is computationally efficient and insensitive to local quality degradation; for low-quality images, since the geometric structure information is severely damaged, the geometric space mapping is abandoned, and the image is encoded into a low-dimensional global semantic feature vector through an encoder network as the third mapping representation, retaining the macroscopic semantic information.

[0099] Finally, the mapping representations corresponding to each quality level, the deformation field, the transformation matrix or the semantic feature vector, are associated and stored with the image data and metadata. This hierarchical adaptive mechanism matches and adapts the standardized interfaces for historical images of different qualities, and overall optimizes the resource utilization efficiency and processing effect of the system.

[0100] Based on the spatially normalized mapping representation, a context-aware deep comparison and quantitative analysis of historical images and current images are performed during clinical review;

[0101] Furthermore, in the context-aware deep comparison and quantitative analysis of historical images and current images during clinical review, the context-aware deep comparison includes the following steps:

[0102] When reviewing current images in clinical settings, the contextual relevance score of historical radiology films is calculated based on effective patient association, standardized semantic labels, temporal relevance, and feature similarity.

[0103] Historical radiology films are ranked and recommended based on context relevance scores.

[0104] Furthermore, context-aware deep comparison also includes the following steps:

[0105] For historical radiology films with a first or second mapping representation, the pre-stored mapping representation is called and combined with the mapping of the current image to the standard space to perform initial registration and obtain the registration result.

[0106] A deep learning-based deformation registration network is used to adjust the registration results, and the Monte Carlo Dropout method is used to output the displacement field mean and uncertainty map of the registration results.

[0107] Furthermore, during clinical image retrieval, context-aware depth comparison and quantitative analysis of historical and current images are performed. The quantitative analysis includes the following steps:

[0108] Multi-dimensional feature extraction is performed on the same regions of interest in historical radiology films and current images after registration adjustment.

[0109] Calculate the relative rate of change of multidimensional features between historical radiology film images and current images;

[0110] The statistical significance of the relative rate of change was calculated using statistical hypothesis testing methods, and its confidence interval was calculated.

[0111] Specifically, based on spatially standardized mapping representation, context-aware in-depth comparison and quantitative analysis of historical and current images are carried out during the clinical retrieval phase. The aim is to accurately and interpretably apply standardized and structured historical image data assets to clinical diagnosis and decision support workflows.

[0112] When clinicians access a patient's current images on the image archiving and communication system workstation, context-aware intelligent recommendation is automatically triggered: first, based on valid patient associations, all related historical images of the patient are retrieved; then, a multi-dimensional context relevance score is calculated for each historical image, using the following formula: ,in, For the final relevance score, It is a binary function that reflects patient consistency. For semantic tag matching degree, , For historical and current image examination times, It is a time constant. For high-level semantic features, cosine similarity. , , , The preset weighting coefficients sum to 1. (Press...) Historical images are sorted in descending order of value to generate a priority recommendation list, improving the efficiency of clinical search and selection.

[0113] For historical images selected by the physician and bearing first or second mapping representations, a precise registration process is initiated: Pre-stored mapping parameters from historical images are invoked, and the mapping parameters from the current image to the same standard anatomical space are calculated in real time. Using the spatial transformation chain rule, an initial transformation from the historical image to the current image space is synthesized. Then, a deep learning deformation registration network is used to perform pixel-level nonlinear fine-tuning of the initial results. Simultaneously, Monte Carlo Dropout technology is employed to quantify registration reliability, maintaining Dropout layer activation and performing multiple forward propagations to obtain displacement field samples. Calculate the final displacement field and uncertainty diagram:

[0114] ;

[0115] ;

[0116] in To optimize the registered displacement field, To estimate variance pixel by pixel, regions with high registration uncertainty can be identified, providing quality references for clinical practice.

[0117] After high-precision registration, the system performs quantitative analysis on regions of interest specified by doctors or automatically detected: It extracts multi-dimensional features such as average grayscale value, texture contrast, and morphological area from the same region in the aligned images, calculating the relative rate of change of each feature between historical and current images; it uses paired-samples t-tests for significance analysis and calculates p-values, combining this with bootstrapping to estimate confidence intervals, determining the statistical significance of the changes, and eliminating the influence of noise or residual registration errors. Finally, it integrates the quantitative results, significance levels, and confidence intervals to form objective and verifiable follow-up analysis conclusions, providing data support for clinical diagnosis and decision-making.

[0118] Structured reports are generated based on the results of in-depth comparison and quantitative analysis, and iterative optimization is carried out using clinical feedback.

[0119] Furthermore, iterative optimization is carried out using clinical feedback, including the following steps:

[0120] Collect and record correction data for semantic annotation, registration results, or quantitative analysis results;

[0121] A feedback dataset is constructed based on the collected correction data;

[0122] The parameters of a deep convolutional neural network and classifier model for semantic annotation, a multi-task learning model for hierarchical quality assessment, or a deformation registration network for registration fine-tuning are optimized and trained using the feedback dataset at preset times.

[0123] Specifically, generating structured reports based on in-depth comparison and quantitative analysis results, and combining clinical feedback to complete iterative optimization, transforms the output into standardized clinical documents. At the same time, relying on clinical correction information, the accuracy and robustness of each model are continuously improved.

[0124] First, based on the results of in-depth comparison and quantitative analysis, a structured follow-up report is automatically generated. The report template predefines structured fields such as patient information, examination comparisons, imaging findings, quantitative analysis, and impression suggestions, and automatically populates them: the imaging findings field describes the location and size changes of lesions based on the registration and segmentation results; the quantitative analysis field presents historical values, current values, rates of change, significance p-values, and confidence intervals of key features in tabular form. Report generation strictly adheres to medical reporting standards, ensuring that the information is clear, complete, and traceable.

[0125] At the same time, a continuous feedback learning mechanism has been established to record doctors' interactions and corrections to the automatically generated content during clinical use. These correction data are securely stored after being anonymized, forming a feedback dataset.

[0126] Based on the collected feedback data, key models are periodically iteratively optimized in an isolated training environment: when optimizing the deep convolutional neural network and classifier model for intelligent semantic annotation, the model prediction results are matched with the labels corrected by the doctor by minimizing the feedback loss function; when modifying the registration network for micro-plastic surgery, a training set consisting of image pairs adjusted by the doctor and corresponding displacement fields is constructed to make the displacement fields predicted by the model closely resemble the real spatial correspondence.

[0127] To integrate multi-source feedback and efficiently update multiple models, a comprehensive online or incremental learning framework is adopted. The core of this framework is to minimize the overall objective function, which includes multi-task loss terms, as follows: ,in, The set of all parameters of the model to be optimized. This represents the number of model tasks that need to be optimized. For the first The loss weight of each task, For the first The loss function corresponding to each task For the first The feedback training dataset for each task. For the first This involves optimizing the objective function iteratively and updating the model parameters using accumulated clinical feedback datasets. This enables the model to continuously learn and improve from clinical practice, gradually adapting it to the diagnostic habits and data characteristics of specific medical institutions, and providing more accurate and personalized clinical services.

[0128] Example 2:

[0129] In a second embodiment of the present invention, the present invention provides a digital archiving and management system for historical radiological images, such as... Figure 2 As shown, it includes the following modules:

[0130] Acquisition and Generation Module: Used to acquire historical radiology film images and associated text information, digitize them, and generate primary digital archives;

[0131] The annotation and association module is used to perform intelligent semantic annotation and association on primary digital archives, generating enhanced digital archives with standardized metadata and establishing patient associations with existing image archiving and communication systems.

[0132] Evaluation generation module: used to perform hierarchical quality assessment of enhanced digital archives and adaptively generate spatially normalized mapping representations at different levels based on the assessment results;

[0133] Comparison Analysis Module: Used for context-aware depth comparison and quantitative analysis of historical and current images during clinical retrieval based on spatially standardized mapping representation;

[0134] The generation optimization module is used to generate structured reports based on the results of in-depth comparison and quantitative analysis, and to perform iterative optimization using clinical feedback.

[0135] In the radiology department of a large comprehensive hospital, doctors frequently need to review historical baseline films from several years or even decades ago to accurately assess changes in the morphology, density, and size of lesions during long-term follow-up evaluations of cancer patients. Traditionally, doctors manually search for films in a physical archive and subjectively compare them using a viewing lamp. This method is not only inefficient but also fails to quantitatively measure subtle changes and makes it difficult to precisely align two-dimensional films with current three-dimensional CT or MR images at the anatomical level. Consequently, evaluation results rely on personal experience and lack objective, quantifiable data support, affecting the accurate formulation of treatment plans. To address these issues, the digital archiving and management system for historical radiological images provided by this invention was adopted. Its architecture is as follows: Figure 2 As shown. The specific implementation process of this system is as follows:

[0136] First, the acquisition and generation module coordinates a high-resolution medical film scanner and a high-speed document scanner to batch convert the patient's historical films and corresponding paper reports into primary digital archives in standard DICOM format.

[0137] Subsequently, the annotation and association module automatically processes the primary digital archives, calls a pre-trained deep convolutional neural network to extract deep features of the images, and intelligently associates them with the patient's full-cycle digital image records in the hospital's existing PACS through a graph matching algorithm to confirm the ownership of each historical film. Then, it integrates the text information identified from the report and generates standardized semantic labels such as "chest posteroanterior X-ray" for each historical film through Bayesian inference based on knowledge graphs, outputting an enhanced digital archive.

[0138] Next, the evaluation generation module automatically grades the quality of each enhanced digital archive. For high-quality historical films with clear images and complete structures, the module drives a large deformation differential homeomorphic registration algorithm to calculate the precise deformation field between the film and the standard chest template. For films of medium quality, a feature point matching algorithm is used to generate a sparse transformation matrix. All of the above outputs are stored as spatial normalization mapping representations and associated with the original images.

[0139] When a doctor accesses the patient's latest chest CT images at the workstation, the comparison and analysis module responds instantly. The module first intelligently filters and sorts the most relevant historical films from the archive based on time, examination type, and semantic tags, recommending them to the doctor. After the doctor selects a historical film, the module calls its pre-stored mapping representation in real time, performs rapid calculation and precise registration with the two-dimensional projection derived from the current CT image, and achieves accurate superposition of the two anatomical structures on the screen. It can also automatically delineate and measure specific lesions, calculate their area change rate, and provide statistically significant quantitative analysis results.

[0140] Finally, the optimization module automatically fills all the quantitative results, key image screenshots, and analysis conclusions of this comparison into a structured follow-up report template to generate a preliminary assessment report. When doctors review the report, they confirm the semantic labels of "nodules" automatically marked by the system or manually fine-tune the position of the registration box. These interactions are recorded as feedback data by the optimization module. The system periodically uses this feedback data to fine-tune the parameters of the classifier model in the labeling and association module and the registration network in the comparison and analysis module, thereby completing a performance improvement closed loop from clinical application to model optimization.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for digital archiving and management of historical radiological images, characterized in that, Includes the following steps: Collect historical radiology film images and associated text information and digitize them to generate primary digital archives; The primary digital archives are intelligently semantically annotated and associated to generate enhanced digital archives with standardized metadata and established patient association with existing image archiving and communication systems; The enhanced digital archives are subjected to hierarchical quality assessment, and spatially normalized mapping representations at different levels are adaptively generated based on the assessment results. Based on the aforementioned spatial standardized mapping representation, context-aware depth comparison and quantitative analysis of historical images and current images are performed during clinical retrieval. Structured reports are generated based on the results of in-depth comparison and quantitative analysis, and iterative optimization is carried out using clinical feedback.

2. The method for digital archiving and management of historical radiological images according to claim 1, characterized in that: The intelligent semantic annotation and association of the primary digital archives includes the following steps: High-level semantic feature vectors of historical radiology films were extracted using a deep convolutional neural network. Integrate the text information in the primary digital archives with the high-level semantic feature vectors to construct a metadata knowledge graph; Based on the high-level semantic feature vector, a graph matching algorithm is used to find matching patients for historical radiology films in the existing image archiving and communication system, thereby obtaining effective patient associations. Based on effective patient associations and metadata knowledge graphs, the metadata of historical radiology films is completed and standardized through Bayesian inference to generate standardized semantic tags.

3. The method for digital archiving and management of historical radiological images according to claim 2, characterized in that: The method of using a graph matching algorithm to find matching patients for historical radiology films in the existing image archiving and communication system includes the following steps: Based on the high-level semantic feature vector, a bipartite graph is constructed with historical radiology film images as the first node and patient image sets in the current image archiving and communication system as the second node. Calculate the edge weights between the first node and each second node, whereby the edge weights are calculated based on the feature vector similarity of the corresponding images and the time compatibility of the inspection. The weighted bipartite graph maximum matching algorithm is used to solve the matching relationship based on the edge weights, and the matching with weights exceeding a preset threshold is confirmed as a valid patient association.

4. The method for digital archiving and management of historical radiological images according to claim 2, characterized in that: The method of supplementing and standardizing metadata of historical radiology film images through Bayesian inference includes the following steps: Based on the metadata knowledge graph and the effective patient associations, the prior probability distribution of standardized labels for historical radiology films is calculated. The likelihood probability of each standardized label under the condition of historical radiology film image features was calculated using a classifier model. The posterior probability of the standardized labels of the historical radiology films is calculated by combining the prior probability distribution, the likelihood probability, and the transition probability obtained from the current image labels associated with the effective patients. The standardized label with the highest posterior probability is selected as the annotation result to obtain the standardized semantic label.

5. The method for digital archiving and management of historical radiological images according to claim 1, characterized in that: The hierarchical quality assessment of the enhanced digital archives includes the following steps: The clarity, noise level, and structural integrity of historical radiology films in enhanced digital archives are evaluated using a multi-task learning model. A comprehensive quality score is calculated based on the assessment results of the clarity, noise level, and structural integrity. Historical radiology films are classified into three quality levels: first, second, and third, based on preset quality scoring thresholds.

6. The method for digital archiving and management of historical radiological images according to claim 1, characterized in that: The process of adaptively generating spatially normalized mapping representations at different levels based on the evaluation results includes the following steps: For historical radiographic films of the first quality grade, the deformation field to the standard anatomical space is calculated using the large deformation differential homeomorphic registration algorithm as the first mapping characterization. For historical radiographic films of the second quality level, a sparse similarity transformation matrix to the standard anatomical space is calculated using a feature point matching algorithm as the second mapping representation. For historical radiographic films of the third quality level, the global semantic feature vector is extracted by an encoder network as the third mapping representation. The first mapping representation, the second mapping representation, or the third mapping representation is stored as the spatially standardized mapping representation of the corresponding historical radiology film image.

7. The method for digital archiving and management of historical radiological images according to claim 1, characterized in that: In the context-aware depth comparison and quantitative analysis of historical and current images during clinical retrieval, the context-aware depth comparison includes the following steps: When reviewing current images in clinical settings, the contextual relevance score of historical radiology films is calculated based on effective patient association, standardized semantic labels, temporal relevance, and feature similarity. The associated historical radiology films are sorted and recommended based on the context relevance score.

8. The method for digital archiving and management of historical radiological images according to claim 7, characterized in that: The context-aware depth comparison also includes the following steps: For historical radiology films with a first or second mapping representation, the pre-stored mapping representation is called and combined with the mapping of the current image to the standard space to perform initial registration and obtain the registration result. A deep learning-based deformation registration network is used to adjust the registration results, and the Monte Carlo Dropout method is used to output the displacement field mean and uncertainty map of the registration results.

9. The method for digital archiving and management of historical radiological images according to claim 1, characterized in that: In the context-aware depth comparison and quantitative analysis of historical and current images during clinical retrieval, the quantitative analysis includes the following steps: Multi-dimensional feature extraction is performed on the same regions of interest in historical radiology films and current images after registration adjustment. Calculate the relative rate of change of the multidimensional features between historical radiology films and current images; The statistical significance of the relative rate of change was calculated using statistical hypothesis testing methods, and its confidence interval was calculated.

10. The method for digital archiving and management of historical radiological images according to claim 1, characterized in that: The iterative optimization using clinical feedback includes the following steps: Collect and record correction data for semantic annotation, registration results, or quantitative analysis results; A feedback dataset is constructed based on the collected correction data; The parameters of the deep convolutional neural network and classifier model for semantic annotation, the multi-task learning model for hierarchical quality assessment, or the deformation registration network for registration fine-tuning are optimized and trained using the feedback dataset at preset times.

Images