Medical image analysis method

Abstract

This invention relates to the fields of medical image processing and precision medicine. The invention discloses a medical image analysis method, comprising the following steps: S1, receiving current medical image data of the target object and its time-series image data within a preset historical time window; S2, precisely aligning the current medical image data with the historical time-series image data, analyzing and calculating the changing trends of key lesion features; this medical image analysis method achieves automated analysis and report generation, significantly reducing time consumption, alleviating pressure on physician resources, and avoiding diagnostic delays; it reduces subjective bias with standardized quantitative algorithms, lowers the risk of missed or misdiagnosed cases, and improves decision reliability; it integrates multi-source data to achieve quantitative analysis and dynamic tracking, providing core evidence for precision medicine; and it generates personalized follow-up suggestions, enhancing the targeting and effectiveness of clinical interventions.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and precision medicine, specifically to methods for medical image analysis. Background Technology

[0002] Medical imaging, as a core basis for modern clinical diagnosis and treatment evaluation, visualizes the internal structure and function of the human body through various technologies such as CT, MRI, and X-rays, providing crucial support for lesion identification and treatment plan formulation. With the development of precision medicine, its analytical paradigm is deepening from qualitative observation to quantitative analysis, early warning, and prognosis prediction, placing extremely high demands on the accuracy, efficiency, and intelligence of the methods.

[0003] However, existing technologies lack a comprehensive analysis system that can integrate dynamic information from multiple time-series images with multimodal clinical data of patients, and based on this, perform automated, quantitative, and personalized risk prediction, which makes clinical practice still face severe challenges.

[0004] Currently, traditional image analysis mainly relies on physicians' manual interpretation of images. While this model is based on professional experience, it has significant limitations: First, it is inefficient. Faced with massive amounts of image data, reaching tens of TB daily, manual interpretation is time-consuming and prone to diagnostic delays, which is particularly evident in primary healthcare and public health emergencies where physician resources are scarce. Second, it is highly subjective. Different physicians, and even the same physician at different times, may make inaccurate interpretations, leading to a higher rate of missed diagnoses of early, small lesions or complex diseases. Finally, it lacks quantitative capabilities. Qualitative descriptions struggle to accurately quantify key parameters such as lesion size and density, as well as their dynamic evolution patterns. These parameters are precisely the objective foundation for assessing the benign or malignant nature of lesions and treatment effectiveness, thus hindering the practice of precision medicine. Therefore, we propose a medical image analysis method to address the aforementioned problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a medical image analysis method that solves the problems of low efficiency, diagnostic delays, high subjectivity, high rate of missed diagnoses, insufficient quantitative capabilities, difficulty in capturing the dynamic evolution of lesions, and inability to integrate multi-source data to achieve personalized risk prediction caused by the reliance on manual image reading in traditional medical image analysis.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a medical image analysis method, comprising the following steps:

[0007] S1. Receive the current medical image data of the target object and its time-series image data within a preset historical time window;

[0008] S2. Accurately align current medical imaging data with historical time-series imaging data, analyze and calculate the changing trends of key lesion features;

[0009] S3. Construct a multimodal fusion feature vector, which includes: a quantitative feature set extracted from the current image, the dynamic change trajectory of the quantitative features of the key lesions, and structured clinical indicators automatically extracted from the patient's electronic medical record.

[0010] S4. Input the fused feature vector into the pre-trained risk prediction model and output the risk probability value of the target object developing disease at a specific time point in the future;

[0011] S5. When the risk probability value exceeds a preset threshold, a structured analysis report containing risk classification and personalized follow-up recommendations is automatically generated and output.

[0012] Preferably, in S3, the trend of change of the key features of the lesion specifically refers to: calculating its growth rate by measuring changes in the tumor's volume, density, or metabolic activity in historical images.

[0013] Preferably, in step S3, the structured clinical indicators automatically extracted from the electronic medical record refer to: using computer text understanding technology to automatically identify key laboratory values, medication information, and symptom descriptions from the doctor's medical records.

[0014] Preferably, in step S3, the step of constructing a multimodal fusion feature vector employs a cross-modal alignment module based on an attention mechanism to map and align the image feature space with the clinical indicator feature space, so as to capture the implicit correlation between imaging manifestations and pathophysiological mechanisms.

[0015] Preferably, in step S4, the risk prediction model is an interpretable machine learning model that, while outputting the risk probability value, also outputs the contribution scores of the top N features that contribute the most to the prediction result.

[0016] Preferably, in S4, the specific future time point is one year after surgery, two years after diagnosis, or a specific clinical time point related to the treatment cycle.

[0017] Preferably, in step S4, the pre-trained risk prediction model adopts an incremental learning framework. When new, labeled medical image data and corresponding follow-up results are input, the model can update parameters and optimize itself while retaining the original knowledge.

[0018] Preferably, in step S5, the personalized follow-up recommendation is a dynamic programming scheme based on a Markov decision process, which recommends the optimal long-term follow-up path based on the risk probability value, false positive risk, examination cost, and patient compliance.

[0019] The present invention provides an electronic device, including a processor and a memory, wherein the memory stores a program or instructions that can run on the processor, and the program or instructions, when executed by the processor, implement the steps of the medical image analysis method.

[0020] The present invention provides a readable storage medium on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the medical image analysis method are implemented.

[0021] Beneficial effects

[0022] This invention provides a method for medical image analysis. Compared with existing technologies, it has the following advantages:

[0023] This medical image analysis method automates image alignment, lesion segmentation, feature extraction, risk prediction, and report generation, eliminating the need for physicians to perform tedious quantitative calculations and data integration manually. This significantly reduces the time required for medical image analysis and effectively addresses the processing pressure brought by massive amounts of image data. Especially in primary healthcare and public health emergencies, it can quickly output analysis results, making up for the shortage of physician resources and avoiding diagnostic delays.

[0024] By employing standardized quantitative analysis algorithms, the objective parameters and dynamic change characteristics of lesions are accurately extracted. Combined with multimodal data, risk prediction is performed, eliminating reliance on physicians' subjective experience and reducing the bias of interpretation by different physicians at different times. At the same time, the interpretable machine learning model can output key contribution features, helping physicians to quickly focus on core information, reduce the risk of missed or misdiagnosed early micro lesions and complex lesions, and improve the reliability of clinical decision-making.

[0025] It can extract rich quantitative features from images and systematically capture the dynamic evolution trajectory of key features of lesions, and integrate them with clinical indicators, realizing the transformation from qualitative description to quantitative analysis and dynamic tracking. These precise quantitative data provide core objective basis for assessing the benign or malignant nature of lesions, judging treatment effects, and predicting disease progression, effectively supporting the implementation of precision medicine.

[0026] By employing an attention-based cross-modal alignment module, multi-temporal imaging data and structured clinical indicators from electronic medical records are effectively integrated, capturing the implicit correlation between imaging manifestations and pathophysiological mechanisms, thus addressing the shortcomings of existing technologies in data integration. Simultaneously, personalized follow-up recommendations generated based on risk probability values, false-positive risk, examination costs, and patient compliance avoid a one-size-fits-all follow-up approach, providing optimal long-term management plans for each patient and enhancing the targetedness and effectiveness of clinical interventions. Attached Figure Description

[0027] Figure 1 This is a flowchart of the medical image analysis method of the present invention;

[0028] Figure 2 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0029] 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. 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.

[0030] like Figure 1-2 As shown:

[0031] Medical image analysis methods include the following steps:

[0032] S1. Receive the target object's current medical imaging data and its time-series image data within a preset historical time window. The current medical imaging data includes, but is not limited to, image data of any one or more modalities such as CT, MRI, and X-ray. The preset historical time window can be flexibly set according to clinical needs, such as 1 year or 2 years. The time-series image data consists of medical images of the target object taken at different time points within this time window, representing the same or multiple modalities, ensuring that the dynamic changes of lesions are reflected. Simultaneously, the target object's electronic medical record data is acquired, providing a foundation for subsequent multimodal feature fusion.

[0033] S2. Accurately align current medical imaging data with historical time-series imaging data to eliminate image shifts caused by differences in patient positioning, equipment parameters, etc., ensuring consistency in lesion location. Based on the aligned images, analyze and calculate the changing trends of key lesion features. Specifically, by measuring changes in tumor volume, density, or metabolic activity in historical images, calculate dynamic indicators such as growth rate to accurately capture the development trend of lesions.

[0034] S3. Construct a multimodal fusion feature vector, which contains three core dimensions:

[0035] The quantitative feature set extracted from the current image includes, but is not limited to, objective quantitative parameters such as the size, shape, density distribution, and edge features of the lesion;

[0036] The dynamic change trajectory of the key quantitative features of the lesions obtained in step S2, such as the time-series features of volume growth rate and density change curve;

[0037] Structured clinical indicators are automatically extracted from patients' electronic medical records. Using computer text understanding technology, key laboratory values ​​such as blood routine indicators and tumor marker levels, medication information such as drug type, dosage, and duration of medication, and symptom descriptions such as pain and fever are automatically identified and extracted from doctors' medical records.

[0038] During feature fusion, an attention-based cross-modal alignment module is used to map and align the image feature space with the clinical indicator feature space, effectively capturing the implicit relationship between imaging manifestations and pathophysiological mechanisms and improving the representational ability of feature vectors.

[0039] S4. Input the multimodal fusion feature vector constructed in step S3 into the pre-trained risk prediction model, and output the risk probability value of disease progression of the target object at a specific time point in the future, where:

[0040] In the future, specific time points can be flexibly set according to clinical scenarios, including but not limited to one year after surgery, two years after diagnosis, or specific clinical time points related to the treatment cycle;

[0041] The risk prediction model is an interpretable machine learning model. In addition to outputting the risk probability value, it also outputs the contribution scores of the top N features that contribute the most to the prediction result. N is a preset positive integer that can be adjusted according to clinical needs, so that physicians can clearly understand the core basis of risk prediction and improve clinical trust.

[0042] The pre-trained risk prediction model adopts an incremental learning framework. When new, labeled medical image data and corresponding follow-up results are input, the model can update parameters and optimize itself while retaining the original knowledge, continuously improving prediction accuracy and generalization ability.

[0043] S5. When the risk probability value output in step S4 exceeds the preset threshold, a structured analysis report containing risk classification and personalized follow-up recommendations is automatically generated and output. The personalized follow-up recommendations are based on a dynamic programming scheme of Markov decision process. This scheme comprehensively considers risk probability value, false positive risk of examination, examination cost, such as medical expenses and radiation dose, and patient compliance, such as follow-up convenience and patient willingness to cooperate, to recommend the optimal long-term follow-up path for the target subject, including specific suggestions such as follow-up frequency, examination item selection, and re-examination time nodes.

[0044] This invention provides an electronic device, including a processor and a memory. The memory stores programs or instructions that can run on the processor. When the program or instructions are executed by the processor, they implement the steps of a medical image analysis method, ensuring the engineering implementation and practical application of the analysis method.

[0045] This invention provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements the steps of a medical image analysis method, providing a carrier support for the storage, dissemination and reuse of the method.

[0046] In this implementation plan: medical imaging equipment, such as CT scanners and MRI machines, is used to collect the current medical imaging data of the target object, and the time-series imaging data of the object within a preset historical time window is retrieved from the hospital's imaging archiving and communication system to ensure the integrity and availability of the imaging data. The electronic medical record data of the target object, including text information such as laboratory reports, medication records, and medical records, is extracted from the hospital's electronic medical record system to provide a data source for the extraction of structured clinical indicators.

[0047] An algorithm combining rigid and flexible registration is used to accurately align the current image with historical time-series images. First, rigid registration corrects rigid deformations such as body translation and rotation. Then, flexible registration eliminates non-rigid differences such as organ peristalsis and tissue deformation, ensuring that the position of the lesion area is accurately matched in different images. Based on the aligned images, image segmentation algorithms such as U-Net and Mask R-CNN are used to automatically segment the lesion area. Then, quantitative parameters such as lesion volume, mean density, and metabolic activity value (for PET images) at each time point are calculated by pixel counting, density threshold analysis, and other methods. Finally, the lesion growth rate is calculated by linear regression, exponential fitting, and other methods to obtain the dynamic change trend of key features.

[0048] From the currently segmented lesion region, quantitative feature sets such as morphological features (e.g., roundness, irregularity), texture features (e.g., gray-level co-occurrence matrix, local binary pattern), and density distribution features (e.g., density standard deviation, peak density) are extracted to form an image feature vector. Natural language processing techniques (e.g., BERT model, keyword extraction algorithm) are used to process the electronic medical record text, automatically identifying and extracting laboratory values ​​(e.g., carcinoembryonic antigen CEA concentration, white blood cell count), medication information (e.g., paclitaxel dosage, treatment cycle), and symptom descriptive words (e.g., "persistent cough", "weight loss of 5kg"), which are then standardized into structured clinical indicator vectors. A cross-modal alignment module based on attention mechanism is used to fuse the image feature vector, the dynamic change trajectory vector of the lesion, and the clinical indicator vector: through attention weight calculation, the feature dimensions that contribute highly to the prediction of disease progression are highlighted, realizing the mapping and alignment of different modal feature spaces, and finally generating a multimodal fusion feature vector with unified dimensions and strong representation capabilities.

[0049] Pre-training of the risk prediction model: A large amount of medical imaging data, time-series image dynamic features, and corresponding electronic medical record data with disease progression results (such as whether recurrence occurred 1 year after surgery and whether metastasis occurred 2 years after diagnosis) and time-series image dynamic features were collected to construct a training dataset. Interpretable machine learning algorithms (such as Gradient Boosting Decision Tree (GBDT) and logistic regression combined with feature importance analysis) were used to train the risk prediction model, while a regularization mechanism was introduced to prevent overfitting. During training, model hyperparameters were adjusted through cross-validation to ensure the model's prediction accuracy and generalization ability. When new labeled data was acquired, incremental learning algorithms (such as momentum-based parameter update strategies) were used to fine-tune the parameters of the pre-trained model. While retaining the original model knowledge, effective information from the new data was absorbed to continuously optimize model performance. The constructed multimodal fusion feature vector was input into the trained risk prediction model. The model outputs the probability value of disease progression risk for the target object at a preset future time point, and outputs the top N key contributing features (such as a 30% increase in lesion volume and a 2-fold increase in CEA concentration) and their contribution scores, providing a basis for clinical interpretation.

[0050] A preset risk probability threshold is set. When the risk probability value output by the model exceeds this threshold, the structured report generation process is initiated. Risk levels are divided according to the risk probability value, such as low risk: <30%, medium risk: 30%-60%, and high risk: >60%. The disease progression risk level of the target subjects is clearly marked. A dynamic programming model is constructed based on the Markov decision process. The input includes the risk probability value, the false positive rate of different examination items, the examination cost, and the patient compliance assessment results, such as compliance scores based on the patient's age, residence, and previous follow-up records. Through state transition probability calculation and benefit maximization analysis, the optimal long-term follow-up path is generated, including follow-up frequency (e.g., follow-up once every 3 months for high-risk patients and once every 6 months for medium-risk patients), recommended examination items (e.g., PET-CT follow-up for high-risk patients and enhanced CT follow-up for medium-risk patients), and personalized follow-up suggestions such as follow-up time nodes.

[0051] It integrates risk levels, key contribution characteristics, and personalized follow-up recommendations to generate standardized structured analysis reports, which can be exported to PDF, docx, and other formats for easy viewing, archiving, and communication with patients by physicians.

[0052] This solution automates image alignment, lesion segmentation, feature extraction, risk prediction, and report generation, eliminating the need for physicians to perform tedious quantitative calculations and data integration manually. This significantly reduces the time required for medical image analysis and effectively addresses the processing pressure brought by massive amounts of image data. Especially in primary healthcare and public health emergencies, it can quickly output analysis results, making up for the shortage of physician resources and avoiding diagnostic delays.

[0053] By employing standardized quantitative analysis algorithms, the objective parameters and dynamic change characteristics of lesions are accurately extracted. Combined with multimodal data, risk prediction is performed, eliminating reliance on physicians' subjective experience and reducing the bias of interpretation by different physicians at different times. At the same time, the interpretable machine learning model can output key contribution features, helping physicians to quickly focus on core information, reduce the risk of missed or misdiagnosed early micro lesions and complex lesions, and improve the reliability of clinical decision-making.

[0054] It can extract rich quantitative features from images and systematically capture the dynamic evolution trajectory of key features of lesions, and integrate them with clinical indicators, realizing the transformation from qualitative description to quantitative analysis and dynamic tracking. These precise quantitative data provide core objective basis for assessing the benign or malignant nature of lesions, judging treatment effects, and predicting disease progression, effectively supporting the implementation of precision medicine.

[0055] By employing an attention-based cross-modal alignment module, multi-temporal imaging data and structured clinical indicators from electronic medical records are effectively integrated, capturing the implicit correlation between imaging manifestations and pathophysiological mechanisms, thus addressing the shortcomings of existing technologies in data integration. Simultaneously, personalized follow-up recommendations generated based on risk probability values, false-positive risk, examination costs, and patient compliance avoid a one-size-fits-all follow-up approach, providing optimal long-term management plans for each patient and enhancing the targetedness and effectiveness of clinical interventions.

[0056] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A medical image analysis method, characterized in that: Includes the following steps: S1. Receive the current medical image data of the target object and its time-series image data within a preset historical time window; S2. Accurately align current medical imaging data with historical time-series imaging data, analyze and calculate the changing trends of key lesion features; S3. Construct a multimodal fusion feature vector, which includes: a quantitative feature set extracted from the current image, the dynamic change trajectory of the quantitative features of the key lesions, and structured clinical indicators automatically extracted from the patient's electronic medical record. S4. Input the fused feature vector into the pre-trained risk prediction model and output the risk probability value of the target object developing disease at a specific time point in the future; S5. When the risk probability value exceeds a preset threshold, a structured analysis report containing risk classification and personalized follow-up recommendations is automatically generated and output.

2. The medical image analysis method according to claim 1, characterized in that: In S3, the trend of change of the key features of the lesion specifically refers to: calculating its growth rate by measuring the changes in the tumor's volume, density, or metabolic activity in historical images.

3. The medical image analysis method according to claim 1, characterized in that: In S3, the structured clinical indicators automatically extracted from the electronic medical record refer to: using computer text understanding technology to automatically find key laboratory values, medication information, and symptom descriptions from the doctor's medical record.

4. The medical image analysis method according to claim 1, characterized in that: In step S3, the step of constructing a multimodal fusion feature vector employs a cross-modal alignment module based on an attention mechanism to map and align the image feature space with the clinical indicator feature space in order to capture the implicit correlation between imaging manifestations and pathophysiological mechanisms.

5. The medical image analysis method according to claim 1, characterized in that: In step S4, the risk prediction model is an interpretable machine learning model that outputs the risk probability value and the contribution scores of the top N features that contribute the most to the prediction result.

6. The medical image analysis method according to claim 1, characterized in that: In S4, the specific future time point is one year after surgery, two years after diagnosis, or a specific clinical time point related to the treatment cycle.

7. The medical image analysis method according to claim 1, characterized in that: In S4, the pre-trained risk prediction model adopts an incremental learning framework. When new, labeled medical image data and corresponding follow-up results are input, the model can update parameters and optimize itself while retaining the original knowledge.

8. The medical image analysis method according to claim 1, characterized in that: In S5, the personalized follow-up recommendation is a dynamic programming scheme based on a Markov decision process. This scheme recommends the optimal long-term follow-up path based on the risk probability value, false positive risk, examination cost, and patient compliance.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the medical image analysis method as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the medical image analysis method as described in any one of claims 1-8.

Images