A medical image examination critical value automatic identification and reporting system

Abstract

The present application relates to the medical information technology field, more specifically, it relates to a kind of medical image examination critical value automatic identification report system, for solving the problem that prior art cannot accurately fuse image, gene, immunity and clinical data, it is difficult to effectively combine cause-effect reasoning and multi-modal deep learning, reduce risk prediction accuracy and decision explainability;The present application fuses image, gene, immunity and clinical data through intelligent diagnosis and treatment decision module, combines cause-effect reasoning and multi-modal deep learning, improves risk prediction accuracy and decision explainability, generates personalized diagnosis and treatment suggestions through weighted score and integrated analysis, covers risk assessment, progress prediction and optimal treatment plan, gives consideration to efficacy, safety and quality of life, outputs structured, interoperable results, significantly enhances the scientific nature, individualization level and clinical practicability of diagnosis and treatment decision.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and more specifically, to an automatic identification and reporting system for critical values ​​in medical imaging examinations. Background Technology

[0002] Traditional medical imaging critical value management faces multiple challenges: radiologists rely on manual interpretation under heavy workloads, which can easily lead to missed diagnoses or misjudgments due to fatigue; critical value identification usually lags behind imaging examinations, and alarms are often triggered only after the report is issued, delaying the timing of clinical intervention; at the same time, clinical, imaging and laboratory information systems are independent of each other, making data exchange difficult and information transmission prone to interruption, making it difficult to effectively achieve closed-loop management of critical values ​​throughout the entire process.

[0003] The patent application with publication number CN104992070A discloses an automatic identification and reporting system for critical values ​​in medical imaging examinations, which includes a critical value management module, an examination and diagnosis workstation, a registration workstation, and a clinical department workstation. It can automatically identify critical values, avoid missed reports, and through automatic linkage of the three links, it can promptly report critical values ​​to clinicians.

[0004] However, while the aforementioned reference patents achieve automatic identification and alerts for critical values ​​through medical terminology keywords, and report them after physician confirmation, effectively avoiding missed and false reports and reducing reliance on physicians' subjective judgment, they cannot accurately integrate imaging, genetic, immune, and clinical data. They also struggle to effectively combine causal reasoning with multimodal deep learning, reducing the accuracy of risk prediction and the interpretability of decisions. Furthermore, they cannot accurately integrate clinical review feedback and expert-annotated data, making it difficult to effectively combine new cases to achieve continuous model iteration. They also cannot utilize deep reinforcement learning to optimize decision-making strategies and adaptively adjust the learning rate, thus failing to achieve self-evolution and long-term performance optimization of the AI ​​system in the clinical environment.

[0005] To address these issues, we propose an automatic identification and reporting system for critical values ​​in medical imaging examinations. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic identification and reporting system for critical values ​​in medical imaging examinations. This system solves the problems of existing technologies, such as the inability to accurately integrate imaging, genetic, immune, and clinical data; the difficulty in effectively combining causal reasoning and multimodal deep learning, which reduces the accuracy of risk prediction and the interpretability of decisions; the inability to accurately integrate clinical review feedback and expert-annotated data; the difficulty in effectively combining new cases to achieve continuous model iteration; the inability to utilize deep reinforcement learning to optimize decision-making strategies and adaptively adjust the learning rate; and the inability to achieve self-evolution and long-term performance optimization of AI systems in clinical environments.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] An automatic identification and reporting system for critical values ​​in medical imaging examinations, integrated into a hospital's clinical diagnosis and treatment information platform, includes:

[0009] The data acquisition and processing module is used to acquire the raw images and their associated image acquisition parameters output by medical imaging equipment, preprocess the acquired raw images and image acquisition parameters, and output standardized three-dimensional images and structured acquisition parameters after preprocessing.

[0010] The lesion criticality prediction module, based on standardized 3D images, structured acquisition parameters and patient clinical information, uses deep convolutional neural networks, spatiotemporal convolutional networks and LSTM to generate criticality classification results;

[0011] The critical value intelligent interpretation module, based on the severity classification results, standardized 3D images and patient clinical information, uses CNN and U-Net segmentation to output structured interpretation results and standardized critical value reports;

[0012] The intelligent diagnosis and treatment decision-making module, based on structured interpretation results, gene mutation information and biochemical and immune test indicators, generates personalized disease progression predictions and treatment plan suggestions through the fusion analysis of causal reasoning graphs and multimodal deep learning models.

[0013] The intelligent optimization feedback module, based on clinical review feedback and expert annotations of structured interpretation results, combined with new case data, iterates and updates the lesion segmentation and severity judgment model through deep reinforcement learning and adaptive learning algorithms.

[0014] In a preferred embodiment of the present invention, the process by which the lesion criticality prediction module generates a severity classification result based on standardized three-dimensional images, structured acquisition parameters, and patient clinical information, using a deep convolutional neural network, a spatiotemporal convolutional network, and LSTM, includes:

[0015] After preprocessing, standardized 3D images and structured acquisition parameters are obtained, relevant clinical information of the patient is acquired, and the standardized 3D images are input into a deep convolutional neural network to extract local and global spatial features of the lesions step by step, and output high-dimensional feature maps.

[0016] For 3D images with time-series characteristics, the spatiotemporal convolutional network is input, and the network performs joint convolution in the time and spatial dimensions to output a temporal feature sequence. The structured acquisition parameters and clinical information are input into a fully connected network and transformed into a fixed-dimensional vector representation. The spatial features, spatiotemporal features and clinical information encoding vectors are fused to generate a unified comprehensive feature vector.

[0017] The fused feature vector is input into a fully connected classification layer. The softmax function is applied to the output value to convert it into a probability distribution for each category. The category with the highest probability is selected as the final classification result for the severity level.

[0018] In a preferred embodiment of the present invention, the critical value intelligent interpretation module performs image segmentation based on the criticality classification results, standardized three-dimensional images, and patient clinical information using U-Net, including the following steps:

[0019] Acquire standardized 3D medical image data, patient clinical information, and severity classification results. Perform voxel standardization on the 3D medical image data. Standardize and transform various indicators in the patient clinical information based on the mean and standard deviation statistically obtained during the training phase. Convert the severity classification results into embeddable category codes.

[0020] The standardized clinical information and the severity classification results are encoded and merged to generate a structured context vector. This vector serves as prior knowledge and is injected into the bottleneck layer of the U-Net model or used to guide feature learning through an attention mechanism.

[0021] The standardized 3D image is input into the pre-trained U-Net model. The model performs a segmentation process with an encoding-decoding structure, and the model output is a 3D probability map with the same spatial dimension as the input image.

[0022] A preset threshold is applied to the 3D probability map to binarize the probability values, generating a binary segmentation mask. Morphological post-processing is then performed, and finally, a structured lesion region segmentation result is output. The segmentation result is associated and bound with the severity classification result to generate a complete severity value interpretation report.

[0023] In a preferred embodiment of the present invention, the process by which the critical value intelligent interpretation module uses CNN to extract features from the segmented image and outputs structured interpretation results and standardized critical value reports includes:

[0024] The three-dimensional probability map output by the U-Net model, the binary segmentation mask generated after thresholding and morphological processing, and the morphological parameters of the segmentation result are obtained. The three-dimensional probability map is input into the convolutional neural network to extract the texture, boundary and internal heterogeneity features of the lesion layer by layer. The network outputs a feature vector of fixed dimension.

[0025] The extracted feature vectors are input into the pre-trained interpretation model. Based on the learned patterns, the model predicts the T stage of the lesion and maps the original T stage prediction results to a standardized severity value.

[0026] The following information is integrated to generate a structured interpretation result: standardized criticality value, original T-staging prediction result, target region volume, target region maximum diameter, three-dimensional spatial center coordinates, standardized status identifiers of image data and standardized status identifiers of patient clinical information. The structured interpretation result is then packaged into a standardized critical value report.

[0027] In a preferred embodiment of the present invention, the process by which the intelligent diagnostic decision-making module integrates and analyzes structured interpretation results, gene mutation information, and biochemical and immunological test indicators includes:

[0028] Gene mutation information, biochemical and immunological test indicators, and structured interpretation results are obtained. Based on the gene mutation information, a comprehensive gene mutation score is calculated. Based on three indicators—immunoglobulin level, cytokine concentration, and inflammatory marker concentration—a biochemical and immunological feature vector is generated. After standardization, the three indicators are used as input features of the multimodal fusion model. Four values ​​are extracted from the structured interpretation results. After normalization, the four values ​​are used as input features of the multimodal fusion model.

[0029] By merging gene mutation comprehensive assessment, biochemical and immunological feature vectors, and structured interpretation features, a pre-trained fusion model learns the correlation patterns between multi-source information to generate a unified comprehensive representation vector. Based on the comprehensive representation vector, structured diagnosis and treatment suggestions are generated, and the output includes disease progression risk level, potential treatment path suggestions, and key biomarker reminders.

[0030] In a preferred embodiment of the present invention, the process of the intelligent diagnosis and treatment decision module fusing and analyzing causal reasoning graphs and multimodal deep learning models includes:

[0031] The system acquires image features, gene mutation comprehensive scores, and biochemical and immunological comprehensive scores, maps input variables to corresponding nodes in the causal inference graph, calculates the potential impact paths of each variable on disease progression based on preset causal influence strength parameters, and generates preliminary disease risk assessment results based on causal logic.

[0032] The image features, gene mutation comprehensive score, and biochemical and immunological comprehensive score are input into a pre-trained multimodal deep learning model. The model uses a feedforward neural network structure to jointly model the nonlinear correlation pattern of multi-source information and outputs a real number to represent the disease risk score.

[0033] By integrating the results of causal reasoning analysis with those of deep learning, and combining the interpretability of causal logic with the discriminative power of deep learning, a final disease risk score and risk level assessment are generated, resulting in a structured diagnosis and treatment decision.

[0034] In a preferred embodiment of the present invention, the process by which the intelligent diagnosis and treatment decision module generates personalized disease progression predictions and treatment plan suggestions includes:

[0035] Acquire genetic data, clinical data, and historical treatment response data. Based on the input data, predict the probability of disease progression at multiple future time points. Input a set of candidate treatment plans and perform the following operations for each candidate plan: Obtain three impact values: efficacy impact value, side effect impact value, and quality of life impact value. Multiply the efficacy impact value by the efficacy assessment weight, the side effect impact value by the safety assessment weight, and the quality of life impact value by the quality of life weight. Add the three weighted values ​​to obtain a comprehensive evaluation score.

[0036] The recommended solution is selected from the candidate solution set based on the comprehensive evaluation score. The recommended solution number is output. The data is integrated to generate a unified structured output, and the final personalized diagnosis and treatment recommendations are output.

[0037] In a preferred embodiment of the present invention, the intelligent optimization feedback module, based on clinical review feedback and expert annotation of the structured interpretation results and combined with newly added case data, includes the following process:

[0038] The system acquires case data that has been automatically and structured for interpretation. The system marks the cases as pending review and adds them to the review task queue. Clinical experts review the cases pending review and examine the location, size, shape, or preliminary diagnosis of lesions in the interpretation results. They modify any content that does not meet the requirements and generate the modified interpretation results.

[0039] The modified interpretation results are bound to the original medical images to generate labeled data. The labeled data is stored in the labeled data repository. When the number of new data in the labeled data repository reaches a preset threshold, the training dataset update process is started to append the new labeled data to the historical training dataset and generate the updated training dataset.

[0040] Using the updated training dataset, the existing model is retrained to generate an updated model. The updated model is then validated on an independent test dataset. Once validation is successful, the updated model is deployed to the runtime environment, and the original model is archived.

[0041] In a preferred embodiment of the present invention, the process by which the intelligent optimization feedback module optimizes the policy and adaptively adjusts the learning rate through deep reinforcement learning includes:

[0042] The updated training dataset is obtained, which includes medical images, expert-annotated interpretation results and corresponding interpretation process records. The lesion segmentation model and the criticality judgment model are optimized by a deep reinforcement learning-based strategy. During the interpretation process, the model selects an action based on the current state. After the action is executed, the immediate reward is calculated based on the expert annotation results. The long-term value estimation is based on the state-action value function.

[0043] During training, the loss values ​​of two adjacent iterations are recorded, the direction and magnitude of the change in the loss value are calculated, and the learning rate adjustment coefficient is obtained by looking up the preset adjustment rule table based on the change. The current learning rate is multiplied by the adjustment coefficient to obtain the updated learning rate, and the updated learning rate adjustment rule table is generated.

[0044] The system outputs the following structured data: the updated interpretation strategy, the updated learning rate adjustment rule table, and statistics on the average reward value, the number of learning rate adjustments, and the direction of adjustments during the interpretation strategy optimization process.

[0045] In a preferred embodiment of the present invention, the process of the intelligent optimization feedback module iteratively updating the lesion segmentation and criticality assessment model includes:

[0046] Obtain all labeled case data from the labeled data repository, merge the newly labeled data with the historical training dataset to generate an updated training dataset, and use the updated training dataset to retrain the lesion segmentation model and the severity judgment model. The model parameters are updated using the gradient descent method.

[0047] After retraining, a lesion segmentation update model and a severity judgment update model are generated, and the lesion segmentation update model and the severity judgment update model are validated on an independent test dataset.

[0048] After successful verification, the lesion segmentation update model and the criticality assessment update model were put into operation, while the original lesion segmentation model and the original criticality assessment model were moved to the archive state.

[0049] Compared with the prior art, the advantages of this invention are:

[0050] (1) In this invention, the intelligent diagnosis and treatment decision module integrates imaging, gene, immune and clinical data, combines causal reasoning and multimodal deep learning to improve the accuracy of risk prediction and the interpretability of decision-making. It generates personalized diagnosis and treatment suggestions through weighted scoring and integrated analysis, covering risk assessment, progression prediction and optimal treatment plan, taking into account efficacy, safety and quality of life, and outputting structured and interoperable results, which significantly enhances the scientific nature, individualization level and clinical applicability of diagnosis and treatment decisions;

[0051] (2) In this invention, the clinical review feedback and expert annotation data are integrated through the intelligent optimization feedback module, and the model is continuously iterated by combining new cases. The decision-making strategy is optimized by deep reinforcement learning, the learning rate is adaptively adjusted, the training efficiency and interpretation accuracy are improved, automated verification and version management are supported, the update is safe and reliable, and the AI ​​system can achieve self-evolution and long-term performance optimization in the clinical environment. Attached Figure Description

[0052] Figure 1 This is a system block diagram of Embodiment 1 of the present invention;

[0053] Figure 2 This is a system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] Example 1: As Figure 1 As shown, the present invention proposes an automatic identification and reporting system for critical values ​​in medical imaging examinations, integrated into a hospital clinical diagnosis and treatment information platform, comprising:

[0056] The data acquisition and processing module is used to acquire the raw images and their associated image acquisition parameters output by the medical imaging equipment. The image acquisition parameters include slice thickness, pixel pitch, tube voltage, repetition time, echo time, device mode, scanning protocol name, and pulse sequence name. The module preprocesses the acquired raw images and image acquisition parameters. The preprocessing operations include parameter parsing and standardization, resolution normalization, image denoising, 3D reconstruction, and intensity normalization. After preprocessing, the module outputs standardized 3D images and structured acquisition parameters.

[0057] The data acquisition and processing module automatically acquires and standardizes multimodal medical images and their parameters (such as slice thickness, tube voltage, scanning protocol, etc.) to achieve image denoising, resolution and intensity normalization, 3D reconstruction and other preprocessing, outputting high-quality standardized 3D images and structured parameter data, effectively eliminating equipment differences, improving image quality and consistency, supporting precise analysis, AI model generalization and multi-center data fusion, and significantly enhancing data availability and clinical application value.

[0058] The lesion criticality prediction module, based on standardized 3D images, structured acquisition parameters and patient clinical information, uses deep convolutional neural networks, spatiotemporal convolutional networks and LSTM to generate criticality classification results;

[0059] The criticality prediction module for lesions, based on standardized 3D images, structured acquisition parameters, and patient clinical information, uses deep convolutional neural networks, spatiotemporal convolutional networks, and LSTM to generate severity classification results. The process includes:

[0060] After preprocessing, standardized 3D images and structured acquisition parameters are obtained, and relevant clinical information of the patient (such as age, gender, medical history, laboratory indicators, etc.) is acquired. All input data undergoes format verification and validity checks before entering the subsequent processing flow. The standardized 3D images are input into a deep convolutional neural network. The network extracts local and global spatial features of the lesion through multi-layer convolution and pooling operations, and outputs a high-dimensional feature map.

[0061] For 3D images with time-series characteristics (such as dynamically enhanced sequences), the spatiotemporal convolutional network is input. The network performs joint convolution in the temporal and spatial dimensions to capture the dynamic evolution pattern of lesions and output a temporal feature sequence. The structured acquisition parameters and clinical information are input into a fully connected network and transformed into a fixed-dimensional vector representation. The spatial features, spatiotemporal features and clinical information encoding vectors are fused using a weighted summation method. The weight coefficients of each branch feature are learnable parameters that are automatically optimized during model training, and the total weight sum is 1. After fusion, a unified comprehensive feature vector is generated, which contains multi-dimensional information of images, time series and clinical data.

[0062] The fused comprehensive feature vector is input into the fully connected classification layer. The number of output nodes of the classification layer is equal to the total number of severity categories (such as emergency, high risk, medium risk, low risk, normal). The Softmax function is applied to the output value to convert it into the probability distribution of each category. The category with the highest probability is selected as the final severity classification result.

[0063] The lesion criticality prediction module integrates standardized 3D images, structured acquisition parameters, and patient clinical information. Utilizing deep convolutional neural networks, spatiotemporal convolutional networks, and LSTM technology, it can accurately extract the spatial and temporal characteristics of lesions. Simultaneously, by combining the patient's personalized clinical background, it achieves intelligent classification of lesion severity. Its advantages lie in its high degree of automation, reducing human intervention. Through the effective fusion of multi-source heterogeneous data, it improves the model's generalization ability and diagnostic accuracy. Furthermore, it supports real-time analysis, helping to quickly identify emergency situations and providing timely and accurate support for clinical decision-making, significantly improving medical efficiency and quality.

[0064] The critical value intelligent interpretation module, based on the severity classification results, standardized 3D images and patient clinical information, uses CNN and U-Net segmentation to output structured interpretation results and standardized critical value reports;

[0065] The critical value intelligent interpretation module segmentes the images based on the severity classification results, standardized 3D images, and patient clinical information using U-Net. The process includes:

[0066] Acquire standardized 3D medical image data, patient clinical information (including age, gender, medical history, laboratory indicators, etc.), and severity classification results. Perform voxel standardization on the 3D medical image data to make the overall mean zero and standard deviation one. For each indicator in the patient clinical information, perform standardization transformation based on the mean and standard deviation statistically obtained during the training phase to ensure consistent numerical distribution. Convert the severity classification results into embeddable category codes.

[0067] The standardized clinical information and the severity classification results are encoded and merged to generate a structured context vector. This vector serves as prior knowledge and is injected into the bottleneck layer of the U-Net model or guided through the attention mechanism to enhance the model's sensitivity to identifying high-risk areas.

[0068] Standardized 3D images are input into a pre-trained U-Net model, which performs a segmentation process with an encoder-decoder structure: the encoder path extracts deep features of the image step by step, the decoder path restores spatial resolution step by step, and multi-scale features are fused through skip connections.

[0069] The model output is a three-dimensional probability map with the same spatial dimension as the input image. Each voxel value in the map represents the probability that the corresponding spatial location belongs to the target lesion area, and the value ranges from 0 to 1.

[0070] A preset threshold (e.g., 0.5) is applied to the 3D probability map to binarize the probability values, generating a binary segmentation mask. Morphological post-processing (e.g., removing small connected components and filling holes) is then performed to improve the geometric rationality of the segmentation results. Finally, a structured lesion region segmentation result is output, including spatial coordinates, volume, boundary information, etc. The segmentation result is associated and bound with the severity classification result to generate a complete severity value interpretation report, supporting inter-system transmission and clinical access.

[0071] The critical value intelligent interpretation module uses CNN to extract features from the segmented image and outputs structured interpretation results and standardized critical value reports. The process includes:

[0072] The three-dimensional probability map output by the U-Net model, the binary segmentation mask generated after thresholding and morphological processing, and the morphological parameters of the segmentation result (including the target region volume, maximum diameter, and three-dimensional spatial center coordinates) are obtained. The three-dimensional probability map is input into the convolutional neural network. The network extracts the texture, boundary and internal heterogeneity features of the lesion layer by layer through multi-level convolution and pooling operations. The network outputs a feature vector of fixed dimension.

[0073] The extracted feature vectors are input into the pre-trained interpretation model. Based on the learned patterns, the model predicts the T stage of the lesion and outputs an integer ranging from 0 to 4, corresponding to T0 to T4 stages. The original T stage prediction results are mapped to standardized severity values. The mapping method is based on the statistical distribution of T stage values ​​in the training dataset. A linear transformation is used to make the results comparable across cases. The standardized severity values ​​are continuous values, reflecting the relative severity of lesion progression.

[0074] The following information is integrated to generate structured interpretation results: standardized criticality value, original T-staging prediction results, target area volume, target area maximum diameter, three-dimensional spatial center coordinates, standardized status identifiers of image data and standardized status identifiers of patient clinical information. All data items undergo type validation and range checks. The structured interpretation results are encapsulated into a standardized critical value report. The standardized critical value report is a data object in JSON format. The field names in the report use a combination of lowercase English letters and underscores. The field set in the report is a fixed set and does not support expansion. The report uses UTF-8 encoding and supports data exchange, parsing and clinical access between systems.

[0075] By integrating imaging, clinical, and critical classification information through the critical value intelligent interpretation module, it achieves accurate lesion segmentation and feature extraction based on U-Net and CNN, enhances the identification of high-risk areas through context guidance, and outputs structured and interoperable critical value reports by combining staging prediction and standardized mapping. The entire process is automated, significantly improving the accuracy, consistency, and clinical response efficiency of interpretation.

[0076] The intelligent diagnosis and treatment decision-making module, based on structured interpretation results, gene mutation information and biochemical and immune test indicators, generates personalized disease progression predictions and treatment plan suggestions through the fusion analysis of causal reasoning graphs and multimodal deep learning models.

[0077] The intelligent diagnostic decision-making module integrates and analyzes structured interpretation results, gene mutation information, and biochemical and immunological test indicators. This process includes:

[0078] Acquire gene mutation information (including the detection status of a preset number of mutation sites), biochemical and immunological test indicators (including immunoglobulin levels, cytokine concentrations, and inflammatory marker concentrations), and structured interpretation results (including disease stage, lesion volume, maximum diameter, and three-dimensional spatial center coordinates). Based on the gene mutation information, calculate a comprehensive gene mutation score. The score is generated through a weighted summation method, and the calculation formula is as follows:

[0079] ,in This is a comprehensive score for gene mutations, reflecting the strength of a patient's disease association at the genetic level. Let be the weight of the i-th mutation site, and its value is determined during the model training phase. This represents the state value of the i-th mutation site, taking the value 0 or 1, where 0 indicates no mutation was detected and 1 indicates a mutation was detected. This is the total number of mutation sites participating in the scoring, which is a preset integer. The weights of all mutation sites are loaded during model initialization and are not updated during runtime.

[0080] Based on three indicators—immunoglobulin level, cytokine concentration, and inflammatory marker concentration—a biochemical immune feature vector is generated. After standardization, these three indicators are used as input features for the multimodal fusion model. Four values ​​are extracted from the structured interpretation results: disease stage, lesion volume, maximum diameter, and three-dimensional spatial center coordinates. After normalization, these four values ​​are used as input features for the multimodal fusion model.

[0081] By merging gene mutation comprehensive assessment, biochemical and immunological feature vectors, and structured interpretation features, a pre-trained fusion model learns the correlation patterns between multi-source information to generate a unified comprehensive representation vector to support diagnostic and treatment decision analysis. Based on the comprehensive representation vector, structured diagnostic and treatment suggestions are generated. The output includes disease progression risk level, potential treatment path suggestions, and key biomarker reminders. The results support inter-system transmission and clinical access.

[0082] The intelligent diagnosis and treatment decision-making module's process of fusing and analyzing causal reasoning graphs with multimodal deep learning models includes:

[0083] The system acquires image features (from the lesion segmentation and interpretation module), gene mutation comprehensive score (from gene mutation information analysis results), and biochemical and immunological comprehensive score (from biochemical and immunological test index analysis results). It maps input variables to corresponding nodes in the causal inference graph, which contains causal relationships between genes, immunity, images, and clinical variables. Based on preset causal influence strength parameters, it calculates the potential impact path of each variable on disease progression and generates preliminary disease risk assessment results based on causal logic. The causal influence strength parameters are loaded during system initialization and are not updated during operation.

[0084] The image features, gene mutation comprehensive score, and biochemical and immunological comprehensive score are input into the pre-trained multimodal deep learning model. The model uses a feedforward neural network structure to jointly model the nonlinear correlation pattern of multi-source information and outputs a real number to represent the disease risk score, with a value range of 0 to 1. The fusion weights of the image modality, gene modality, and clinical modality are loaded during model initialization and are not updated during operation.

[0085] The results of causal reasoning analysis and deep learning reasoning are fused together. The fusion strategy adopts a weighted integration method, which combines the interpretability of causal logic with the discriminative ability of deep learning to generate the final disease risk score and risk level judgment. The output is a structured diagnosis and treatment decision result, including: disease risk score, risk level (low, medium, high), key influencing factors, recommended follow-up and intervention path. The results support inter-system transmission and clinical access.

[0086] The process by which the intelligent diagnosis and treatment decision-making module generates personalized disease progression predictions and treatment recommendations includes:

[0087] Acquire genetic data (including patient gene mutation information), clinical data (including disease stage, lesion volume, maximum diameter, three-dimensional spatial center coordinates, and biochemical and immunological indicators), and historical treatment response data (including previous treatment plans, efficacy, and side effect records). Based on the input data, predict the probability of disease progression at multiple future time points. The prediction function is defined as follows:

[0088] ,in Let be the probability of disease progression at time t. For genetic data, For clinical data, For historical treatment response data, The prediction function is implemented by a deep neural network. The parameters are determined during the training phase. The model outputs the disease progression probability corresponding to a preset number of time points, which is used to construct the disease progression trend curve.

[0089] Input a set of candidate treatment options and perform the following operations for each option: Obtain three impact values: efficacy impact value, side effect impact value, and quality of life impact value. Multiply the efficacy impact value by the efficacy assessment weight, the side effect impact value by the safety assessment weight, and the quality of life impact value by the quality of life weight. Sum the three weighted values ​​to obtain a comprehensive assessment score. The efficacy assessment weight, safety assessment weight, and quality of life weight are loaded during system initialization and are not updated during runtime. Select the option with the lowest comprehensive assessment score from the candidate option set as the recommended option and output the recommended option number. Integrate the data to generate a unified structured output (gene mutation comprehensive score, biochemical and immunological comprehensive score, disease stage, lesion volume, maximum diameter, three-dimensional spatial center coordinates, disease risk score, and recommended treatment option number). All data items undergo type validation and consistency checks. Output the final personalized diagnosis and treatment suggestions. The suggestions include disease progression prediction results, recommended treatment option number, and key supporting data. All output items are numerical data, and Chinese terminology is used in the naming. The output format supports structured encoding and inter-system transmission.

[0090] By integrating structured interpretation results, gene mutation information, and biochemical and immunological indicators through an intelligent diagnostic and treatment decision-making module, and through causal reasoning graphs and multimodal deep learning collaborative analysis, a deep fusion of genetic, immunological, imaging, and clinical data is achieved. Weighted gene scores and standardized feature vectors are used to quantify individual patient differences, and causal models are combined to enhance decision interpretability. Simultaneously, the modeling capabilities of deep learning for nonlinear relationships are leveraged to improve the accuracy of risk prediction. An integrated strategy merges the two analysis paths to generate personalized decision-making recommendations that include disease risk scores, progression trend predictions, and optimal treatment recommendations, balancing efficacy, safety, and quality of life. Structured and verifiable diagnostic and treatment recommendations are output, supporting cross-system transmission and clinical access, significantly improving the scientific rigor, personalization, and clinical applicability of diagnostic and treatment decisions.

[0091] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that:

[0092] like Figure 2 As shown, the intelligent optimization feedback module, based on the clinical review feedback and expert annotation of the structured interpretation results, combined with the data of newly added cases, iterates and updates the lesion segmentation and criticality judgment model through deep reinforcement learning and adaptive learning algorithms;

[0093] The intelligent optimization feedback module, based on clinical review feedback and expert annotations of structured interpretation results and combined with newly added case data, includes the following process:

[0094] The system acquires case data that has been automatically and structured for interpretation. The case data includes medical images and corresponding interpretation results. The system marks the cases as pending review and adds them to the review task queue. Clinical experts review the cases pending review and examine the location, size, shape, or preliminary diagnosis of lesions in the interpretation results. They modify any content that does not meet the requirements and generate the modified interpretation results.

[0095] The modified interpretation results are bound to the original medical images to generate labeled data. The labeled data is stored in the labeled data repository. When the number of new data in the labeled data repository reaches a preset threshold, the training dataset update process is started to append the new labeled data to the historical training dataset and generate the updated training dataset.

[0096] Using the updated training dataset, the existing model is retrained while keeping the model structure unchanged and only updating the model parameters to generate an updated model. The updated model is then validated on an independent test dataset. The validation includes the output results of tasks such as lesion detection, segmentation, and classification. Once the validation is successful, the updated model is put into the runtime environment, and the original model is archived.

[0097] The intelligent optimization feedback module optimizes the policy and adaptively adjusts the learning rate through deep reinforcement learning, including the following processes:

[0098] The updated training dataset is obtained, containing medical images, expert-annotated interpretation results, and corresponding interpretation process records. This dataset is used for updating the interpretation strategy and adjusting the learning rate. Deep reinforcement learning-based strategy optimization is performed on the lesion segmentation model and the criticality assessment model. During the interpretation process, the model selects actions based on its current state, which consists of input image features, contextual information, and the current interpretation progress. Actions include adjusting the segmentation threshold, switching interpretation paths, or enabling post-processing modules. After executing an action, an immediate reward is calculated based on the expert annotation results. Long-term value estimation is based on a state-action value function.

[0099] ,in The current state of the model is represented by the features of the input image, contextual information, and the current interpretation progress. This indicates the action the model chooses to perform in the current state, such as adjusting the segmentation threshold, switching the interpretation path, or enabling the post-processing module. This represents the immediate reward obtained after performing an action, with the value based on expert correction. The design reflects the short-term effect of this action on improving interpretation accuracy. This represents a discount factor used to adjust the importance of future rewards relative to current rewards. This indicates the next state entered after performing action 'a'. Indicates the next state All possible actions to be performed. This represents the maximum long-term value that all possible actions can bring in the next state. This represents the estimated long-term cumulative reward obtained by performing action a in state s. The model optimizes... The function learns better decision-making strategies and reduces reliance on expert corrections;

[0100] During training, the loss values ​​of two adjacent iterations are recorded, the direction and magnitude of the change in the loss value are calculated, and the learning rate adjustment coefficient is obtained by looking up the preset adjustment rule table based on the change. The current learning rate is multiplied by the adjustment coefficient to obtain the updated learning rate, and the updated learning rate adjustment rule table is generated.

[0101] The system outputs the following structured data: the updated interpretation strategy, the updated learning rate adjustment rule table, and statistics on the average reward value, number of learning rate adjustments, and adjustment direction during the interpretation strategy optimization process. All fields are named using Chinese terminology and support inter-system transmission.

[0102] The intelligent optimization feedback module iteratively updates the lesion segmentation and severity assessment model, including the following processes:

[0103] Retrieve all annotated case data from the labeled data repository, merge the newly annotated data with the historical training dataset to generate an updated training dataset, and use the updated training dataset to retrain the lesion segmentation model and the severity assessment model. The model parameters are updated using the gradient descent method, and the update formula is as follows:

[0104] ,in This represents the current set of parameters for the model, including all learnable weights and biases in the neural network. This represents the updated model parameters. The learning rate controls the step size for each parameter update. This represents the gradient vector of the loss function with respect to the model parameters, indicating the direction in which the loss value increases the most. This represents the overall loss value of the model on the training samples under the current parameters, reflecting the difference between the model output and the expert annotation results. Parameter updates are performed after each training batch to gradually reduce the model error.

[0105] After retraining, a lesion segmentation update model and a severity judgment update model are generated. The lesion segmentation update model and the severity judgment update model are validated on an independent test dataset. The validation content includes the consistency between the segmentation results and the labeled data, and the consistency between the judgment results and the labeled data.

[0106] After successful verification, the lesion segmentation update model and the criticality assessment update model were put into operation, while the original lesion segmentation model and the original criticality assessment model were moved to the archive state.

[0107] By integrating clinical review feedback and expert-annotated data through an intelligent optimization feedback module, and combining new cases to achieve continuous model iteration, deep reinforcement learning is used to optimize decision-making strategies for lesion segmentation and criticality interpretation, and the learning rate is adaptively adjusted to improve training efficiency. A closed-loop feedback mechanism is adopted to transform real clinical corrections into high-quality training data, driving model parameter updates and performance optimization. Model version management and automated validation deployment are supported to ensure the safety and reliability of the update process. The reliance on manual correction is significantly reduced, enhancing the robustness, adaptability and interpretation accuracy of the model, and enabling the AI ​​system to self-evolve and maintain long-term performance in real clinical environments.

[0108] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. An automatic identification and reporting system for critical values ​​in medical imaging examinations, integrated into a hospital clinical diagnosis and treatment information platform, characterized in that, include: The data acquisition and processing module is used to acquire the raw images and their associated image acquisition parameters output by medical imaging equipment, preprocess the acquired raw images and image acquisition parameters, and output standardized three-dimensional images and structured acquisition parameters after preprocessing. The lesion criticality prediction module, based on standardized 3D images, structured acquisition parameters and patient clinical information, uses deep convolutional neural networks, spatiotemporal convolutional networks and LSTM to generate criticality classification results; The critical value intelligent interpretation module, based on the severity classification results, standardized 3D images and patient clinical information, uses CNN and U-Net segmentation to output structured interpretation results and standardized critical value reports; The intelligent diagnosis and treatment decision-making module, based on structured interpretation results, gene mutation information and biochemical and immune test indicators, generates personalized disease progression predictions and treatment plan suggestions through the fusion analysis of causal reasoning graphs and multimodal deep learning models. The intelligent optimization feedback module, based on clinical review feedback and expert annotations of structured interpretation results, combined with new case data, iterates and updates the lesion segmentation and severity judgment model through deep reinforcement learning and adaptive learning algorithms.

2. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 1, characterized in that, The criticality prediction module for lesions generates a classification result based on standardized 3D images, structured acquisition parameters, and patient clinical information, using deep convolutional neural networks, spatiotemporal convolutional networks, and LSTM. The process includes: After preprocessing, standardized 3D images and structured acquisition parameters are obtained, relevant clinical information of the patient is acquired, and the standardized 3D images are input into a deep convolutional neural network to extract local and global spatial features of the lesions step by step, and output high-dimensional feature maps. For 3D images with time-series characteristics, the spatiotemporal convolutional network is input, and the network performs joint convolution in the time and spatial dimensions to output a temporal feature sequence. The structured acquisition parameters and clinical information are input into a fully connected network and transformed into a fixed-dimensional vector representation. The spatial features, spatiotemporal features and clinical information encoding vectors are fused to generate a unified comprehensive feature vector. The fused feature vector is input into a fully connected classification layer. The softmax function is applied to the output value to convert it into a probability distribution for each category. The category with the highest probability is selected as the final classification result for the severity level.

3. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 1, characterized in that, The critical value intelligent interpretation module segments the images based on the severity classification results, standardized 3D images, and patient clinical information using U-Net, including the following process: Acquire standardized 3D medical image data, patient clinical information, and severity classification results. Perform voxel standardization on the 3D medical image data. Standardize and transform various indicators in the patient clinical information based on the mean and standard deviation statistically obtained during the training phase. Convert the severity classification results into embeddable category codes. The standardized clinical information and the severity classification results are encoded and merged to generate a structured context vector. This vector serves as prior knowledge and is injected into the bottleneck layer of the U-Net model or used to guide feature learning through an attention mechanism. The standardized 3D image is input into the pre-trained U-Net model. The model performs a segmentation process with an encoding-decoding structure, and the model output is a 3D probability map with the same spatial dimension as the input image. A preset threshold is applied to the 3D probability map to binarize the probability values, generating a binary segmentation mask. Morphological post-processing is then performed, and finally, a structured lesion region segmentation result is output. The segmentation result is associated and bound with the severity classification result to generate a complete severity value interpretation report.

4. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 3, characterized in that, The critical value intelligent interpretation module uses CNN to extract features from the segmented image and outputs structured interpretation results and standardized critical value reports. The process includes: The three-dimensional probability map output by the U-Net model, the binary segmentation mask generated after thresholding and morphological processing, and the morphological parameters of the segmentation result are obtained. The three-dimensional probability map is input into the convolutional neural network to extract the texture, boundary and internal heterogeneity features of the lesion layer by layer. The network outputs a feature vector of fixed dimension. The extracted feature vectors are input into the pre-trained interpretation model. Based on the learned patterns, the model predicts the T stage of the lesion and maps the original T stage prediction results to a standardized severity value. The following information is integrated to generate a structured interpretation result: standardized criticality value, original T-staging prediction result, target region volume, target region maximum diameter, three-dimensional spatial center coordinates, standardized status identifiers of image data and standardized status identifiers of patient clinical information. The structured interpretation result is then packaged into a standardized critical value report.

5. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 1, characterized in that, The intelligent diagnosis and treatment decision-making module integrates and analyzes structured interpretation results, gene mutation information, and biochemical and immunological test indicators, including the following processes: Gene mutation information, biochemical and immunological test indicators, and structured interpretation results are obtained. Based on the gene mutation information, a comprehensive gene mutation score is calculated. Based on three indicators—immunoglobulin level, cytokine concentration, and inflammatory marker concentration—a biochemical and immunological feature vector is generated. After standardization, the three indicators are used as input features of the multimodal fusion model. Four values ​​are extracted from the structured interpretation results. After normalization, the four values ​​are used as input features of the multimodal fusion model. By merging gene mutation comprehensive assessment, biochemical and immunological feature vectors, and structured interpretation features, a pre-trained fusion model learns the correlation patterns between multi-source information to generate a unified comprehensive representation vector. Based on the comprehensive representation vector, structured diagnosis and treatment suggestions are generated, and the output includes disease progression risk level, potential treatment path suggestions, and key biomarker reminders.

6. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 5, characterized in that, The intelligent diagnosis and treatment decision-making module's process of fusing and analyzing causal reasoning graphs with multimodal deep learning models includes: The system acquires image features, gene mutation comprehensive scores, and biochemical and immunological comprehensive scores, maps input variables to corresponding nodes in the causal inference graph, calculates the potential impact paths of each variable on disease progression based on preset causal influence strength parameters, and generates preliminary disease risk assessment results based on causal logic. The image features, gene mutation comprehensive score, and biochemical and immunological comprehensive score are input into a pre-trained multimodal deep learning model. The model uses a feedforward neural network structure to jointly model the nonlinear correlation pattern of multi-source information and outputs a real number to represent the disease risk score. By integrating the results of causal reasoning analysis with those of deep learning, and combining the interpretability of causal logic with the discriminative power of deep learning, a final disease risk score and risk level assessment are generated, resulting in a structured diagnosis and treatment decision.

7. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 6, characterized in that, The process by which the intelligent diagnosis and treatment decision-making module generates personalized disease progression predictions and treatment recommendations includes: Acquire genetic data, clinical data, and historical treatment response data. Based on the input data, predict the probability of disease progression at multiple future time points. Input a set of candidate treatment plans and perform the following operations for each candidate plan: Obtain three impact values: efficacy impact value, side effect impact value, and quality of life impact value. Multiply the efficacy impact value by the efficacy assessment weight, the side effect impact value by the safety assessment weight, and the quality of life impact value by the quality of life weight. Add the three weighted values ​​to obtain a comprehensive evaluation score. The recommended solution is selected from the candidate solution set based on the comprehensive evaluation score. The recommended solution number is output. The data is integrated to generate a unified structured output, and the final personalized diagnosis and treatment recommendations are output.

8. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 1, characterized in that, The intelligent optimization feedback module, based on clinical review feedback and expert annotations of the structured interpretation results and combined with newly added case data, includes the following process: The system acquires case data that has been automatically and structured for interpretation. The system marks the cases as pending review and adds them to the review task queue. Clinical experts review the cases pending review and examine the location, size, shape, or preliminary diagnosis of lesions in the interpretation results. They modify any content that does not meet the requirements and generate the modified interpretation results. The modified interpretation results are bound to the original medical images to generate labeled data. The labeled data is stored in the labeled data repository. When the number of new data in the labeled data repository reaches a preset threshold, the training dataset update process is started to append the new labeled data to the historical training dataset and generate the updated training dataset. Using the updated training dataset, the existing model is retrained to generate an updated model. The updated model is then validated on an independent test dataset. Once validation is successful, the updated model is deployed to the runtime environment, and the original model is archived.

9. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 8, characterized in that, The intelligent optimization feedback module optimizes the policy and adaptively adjusts the learning rate through deep reinforcement learning, including the following process: The updated training dataset is obtained, which includes medical images, expert-annotated interpretation results and corresponding interpretation process records. The lesion segmentation model and the criticality judgment model are optimized by a deep reinforcement learning-based strategy. During the interpretation process, the model selects an action based on the current state. After the action is executed, the immediate reward is calculated based on the expert annotation results. The long-term value estimation is based on the state-action value function. During training, the loss values ​​of two adjacent iterations are recorded, the direction and magnitude of the change in the loss value are calculated, and the learning rate adjustment coefficient is obtained by looking up the preset adjustment rule table based on the change. The current learning rate is multiplied by the adjustment coefficient to obtain the updated learning rate, and the updated learning rate adjustment rule table is generated. The system outputs the following structured data: the updated interpretation strategy, the updated learning rate adjustment rule table, and statistics on the average reward value, the number of learning rate adjustments, and the direction of adjustments during the interpretation strategy optimization process.

10. The automatic identification and reporting system for critical values ​​in medical imaging examinations according to claim 9, characterized in that, The process of iteratively updating the lesion segmentation and severity assessment model by the intelligent optimization feedback module includes: Obtain all labeled case data from the labeled data repository, merge the newly labeled data with the historical training dataset to generate an updated training dataset, and use the updated training dataset to retrain the lesion segmentation model and the severity judgment model. The model parameters are updated using the gradient descent method. After retraining, a lesion segmentation update model and a severity judgment update model are generated, and the lesion segmentation update model and the severity judgment update model are validated on an independent test dataset. After successful verification, the lesion segmentation update model and the criticality assessment update model were put into operation, while the original lesion segmentation model and the original criticality assessment model were moved to the archive state.

Images