A method for assisting diagnosis of child pneumonia based on lung ultrasonic image and application thereof

Abstract

The application discloses a kind of based on lung ultrasound image auxiliary diagnosis method, system, electronic equipment and storage medium of children pneumonia, belong to medical artificial intelligence and medical image processing technical field.It includes: obtaining the static image and / or dynamic video of child lung ultrasound collected according to preset scanning specification;Desensitization, quality control, standardization pretreatment are carried out to the image;The pretreated image is input into multi-task artificial intelligence model, the model at least includes feature extraction module, pneumonia sign identification module, lesion segmentation module, dynamic time sequence fusion module and severity evaluation module;Output pneumonia positive / negative judgment result, lesion region segmentation result, sign identification result and severity classification result, and generate visual heat map or lesion annotation chart to assist clinical interpretation.The application can reduce the dependence of lung ultrasound examination on operator experience, improve the consistency, objectivity and efficiency of bedside screening and auxiliary diagnosis of children pneumonia, and is suitable for pediatric outpatient and emergency department, ward, PICU and primary medical institutions.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, artificial intelligence-assisted diagnosis, and pediatric bedside ultrasound technology. Specifically, it relates to a method, system, electronic device, and storage medium for the auxiliary diagnosis of pediatric pneumonia based on lung ultrasound images. Background Technology

[0002] Pneumonia in children is a common and important respiratory disease in pediatrics. In severe cases, it can cause hypoxemia, respiratory failure, and multi-system complications. Timely identification and assessment of the condition are crucial for clinical management. While chest X-rays and CT scans can be used to aid in diagnosis, they suffer from problems such as ionizing radiation exposure, limited repeatability, and inconvenience at the bedside. Lung ultrasound, due to its advantages of being radiation-free, repeatable, and accessible at the bedside, is increasingly being used in pediatric outpatient and emergency departments, inpatient wards, and PICU settings.

[0003] However, the application of lung ultrasound in pediatric pneumonia still faces significant limitations. First, considerable differences exist among operators in the selection of scanning sections, image quality control, and identification of key signs, leading to insufficient repeatability of examination results. Second, signs such as pulmonary consolidation, abnormal pleural lines, B-line changes, air bronchograms, and pleural effusion often require comprehensive judgment by experienced physicians, which is highly subjective. Third, dynamic signs, such as lung sliding and B-line changes with respiration, are difficult to quantify consistently through manual observation alone. Existing technologies largely focus on single-classification tasks, failing to simultaneously address lesion localization, sign identification, severity assessment, and dynamic information utilization, thus failing to meet the auxiliary diagnostic needs in real-world clinical scenarios. Summary of the Invention

[0004] To address the problems of subjective interpretation, high operator dependence, insufficient utilization of dynamic signs, and inability of single-task models to provide comprehensive decision support in existing technologies for pediatric lung ultrasound image interpretation, this invention proposes a method and system for auxiliary diagnosis of pediatric pneumonia based on lung ultrasound images.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] On one hand, this invention provides an auxiliary diagnostic method for pediatric pneumonia based on lung ultrasound imaging, comprising: S101, acquiring static lung ultrasound images and / or dynamic videos of the child subject to be analyzed, wherein the images are acquired according to a preset scanning zoning specification; S102, performing desensitization, quality control screening, size normalization, brightness standardization, and data augmentation processing on the images to obtain standardized input data; S103, inputting the standardized input data into a multi-task artificial intelligence model to extract multi-scale spatial features and / or time series features; S104, outputting pneumonia positive / negative judgment results, lesion region segmentation results, sign recognition results, and severity grading results based on the multi-task artificial intelligence model; S105, performing reverse mapping on the model's internal feature responses or attention weights to generate heat maps, lesion mask maps, and / or sign marker maps, and displaying the results.

[0007] Preferably, the preset scanning zoning specifications include multi-zone scanning of the anterior chest, lateral chest, and back, covering different anatomical regions of both lungs.

[0008] Preferably, the sign recognition results include one or more of the following: lung consolidation area, abnormal pleural line area, dense B-line area, air bronchus sign, abnormal lung sliding sign, and pleural effusion.

[0009] Preferably, the severity grading results are determined comprehensively based on the lesion area, number of signs, distribution area of ​​signs, number of affected areas, and dynamic change characteristics.

[0010] On the other hand, the present invention also provides an auxiliary diagnostic system for pediatric pneumonia, including a data acquisition and standardization module, a multi-task artificial intelligence reasoning module, and a result output and interpretability interaction module.

[0011] In another aspect, the present invention also provides an electronic device and a computer-readable storage medium for performing the above-described method.

[0012] The function and effects of this invention:

[0013] The core idea of ​​this invention is to use static lung ultrasound images and / or dynamic videos obtained through standardized scanning as input, and to construct a multi-task artificial intelligence model that integrates classification, segmentation, sign recognition, dynamic temporal modeling and severity assessment to achieve automated auxiliary judgment of pneumonia in children, and output interpretable results that combine qualitative and quantitative information.

[0014] Compared with existing technologies, the present invention has at least the following beneficial effects: (1) Reduced operator dependence. By automatically analyzing lung ultrasound data under standardized scanning sections, the consistency of interpretation results among different physicians is improved. (2) Multi-task joint output. The model not only provides positive / negative results for pneumonia, but also outputs lesion areas, sign markers, and severity grades, providing more complete information. (3) Introduction of dynamic temporal information. By modeling continuous frames of dynamic video, the ability to identify abnormal lung sliding signs and signs that change with respiration is improved. (4) Enhanced clinical interpretability. The model intuitively presents the areas of interest through heat maps, lesion mask maps, and sign marker maps, facilitating physician review. (5) Applicable to bedside and primary care settings. It can be combined with ultrasound equipment or clinical workstations and is suitable for outpatient, emergency, ward, PICU, and primary care pediatric settings. Attached Figure Description

[0015] Figure 1 The flowchart illustrates the overall process of the ultrasound-assisted diagnosis method for pediatric pneumonia provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the multi-task network structure and dynamic temporal fusion provided in an embodiment of the present invention. Detailed Implementation

[0016] This embodiment provides a method for the auxiliary diagnosis of pediatric pneumonia based on lung ultrasound imaging, including:

[0017] S1. Acquire static and / or dynamic ultrasound images and / or videos of the lungs of the child subject to be analyzed, and collect the static and / or dynamic ultrasound images and videos of the lungs according to the preset scanning zoning specifications;

[0018] The pre-defined scanning zoning standards include standardized multi-zone scanning of the anterior chest, lateral chest, and back; pneumonia-related signs include at least one or more of the following: pulmonary consolidation, abnormal pleural lines, increased or fused B lines, air bronchogram sign, abnormal lung sliding sign, and pleural effusion; severity grading is determined based on the number, extent, area proportion, distribution region, and / or dynamic change characteristics of pneumonia-related signs.

[0019] S2. Desensitize, quality control, size normalize, and enhance static and / or dynamic images and videos of lung ultrasound to obtain standardized input data;

[0020] S3. Input standardized input data into a multi-task artificial intelligence model for joint reasoning. The multi-task artificial intelligence model includes at least a feature extraction module, a pneumonia sign recognition module, a lesion segmentation module, a dynamic temporal fusion module, and a severity assessment module.

[0021] The multi-task artificial intelligence model includes:

[0022] A backbone feature extraction network is used to extract multi-scale spatial features;

[0023] The classification branch is used to output positive / negative probability values ​​for pneumonia and / or pathogen type classification results;

[0024] Segmentation branches are used for pixel-level segmentation of pulmonary consolidation regions and / or pleural effusion regions;

[0025] The sign recognition branch is used to identify abnormal pleural lines, dense areas of B lines, air bronchus sign, and abnormal lung sliding sign.

[0026] The temporal fusion branch is used to perform temporal modeling on the features of consecutive frames in dynamic videos in order to extract dynamic change information within the breathing cycle;

[0027] Hierarchical branching is used to output mild, moderate, or severe hierarchical results based on classification results, segmentation results, feature recognition results, and time-series features.

[0028] The temporal fusion branch employs temporal convolutional networks, recurrent neural networks, temporal attention modules, Transformer temporal encoders, or combinations thereof, to aggregate features from consecutive frames in dynamic videos, thereby enhancing the ability to identify abnormal lung sliding signs, dynamic B-line changes, and lesion boundary changes.

[0029] S4. Based on the joint reasoning results, output the pneumonia positive / negative judgment results, lesion area segmentation results, pneumonia-related sign identification results, and severity grading results;

[0030] S5. Generate visual heatmaps, sign marker maps, and / or lesion mask maps based on the feature responses or attention weights in the joint reasoning process to assist in clinical interpretation.

[0031] In addition, this method also uses double-blind or multi-blind annotation of training samples by multiple annotation experts to form a gold standard label library; the gold standard labels include at least normal / pneumonia classification labels, lesion segmentation labels, sign labels and severity labels; during model training, a weighted combination of classification loss, segmentation loss and grading loss is used as a joint optimization objective.

[0032] Example 1: Data Acquisition and Standardized Preprocessing

[0033] This embodiment illustrates the specific implementation of the present invention in the data input stage.

[0034] (1) Data acquisition: Standardized zonal scanning was performed on pediatric subjects using lung ultrasound to acquire static images and / or dynamic videos of different areas of the anterior chest, lateral chest, and back. To ensure data consistency, information such as equipment model, probe type, center frequency, gain, depth, and scanning position can be recorded.

[0035] (2) Desensitization: Remove or replace the name, hospital number, examination number and other identification information in the original image.

[0036] (3) Quality control: Images that do not meet the analysis requirements due to crying, obvious motion artifacts, cross-sectional offset, severe image blurring, or missing key anatomical structures are excluded.

[0037] (4) Standardization processing: The retained images are normalized in size, grayscale range is normalized, noise is suppressed and contrast is adjusted; for dynamic videos, frame sampling or fixed-length cropping can be performed.

[0038] (5) Data augmentation: During the offline training phase, rotation, flipping, translation, scaling, brightness and contrast changes, and timing jitter can be used to improve the robustness of the model.

[0039] Example 2: Gold Standard Establishment and Label Construction

[0040] This embodiment illustrates how training labels and evaluation labels are generated.

[0041] (1) Gold standard definition: The final clinical diagnosis is used as the source of disease label. The final clinical diagnosis can be determined by combining medical history, physical signs, laboratory tests, chest imaging results, treatment response and follow-up results.

[0042] (2) Expert annotation: The lung ultrasound images are independently annotated by two or more senior ultrasound doctors, and in case of disagreement, they are reviewed by experts with higher seniority to form a unified label.

[0043] (3) Label content: including normal / pneumonia classification label, lung consolidation area segmentation label, pleural effusion area segmentation label, pleural line abnormality label, B line label, air bronchus sign label, lung sliding sign abnormality label and severity label.

[0044] (4) Grading rules: The severity label can be defined as mild, moderate or severe based on indicators such as the proportion of lesion area, the number of affected areas, the distribution range of both lungs, the density of B lines and whether there is pleural effusion.

[0045] Example 3: Construction and Training of Multi-Task Artificial Intelligence Models

[0046] This example illustrates the model architecture and training process.

[0047] (1) Backbone network: Multi-scale spatial features of lung ultrasound images are extracted using convolutional neural networks, visual Transformer networks, or a combination of both.

[0048] (2) Classification branch: Output the positive / negative probability of pneumonia based on the global features; when needed, it can further output the classification results of viral pneumonia, bacterial pneumonia or other categories.

[0049] (3) Branch segmentation: The encoder-decoder structure is used to segment the lung consolidation region and the pleural effusion region at the pixel level, and the corresponding mask is output.

[0050] (4) Sign recognition branch: Output detection results or probability values ​​for key signs such as abnormal pleural line, dense B-line area, air bronchus sign and abnormal lung sliding sign.

[0051] (5) Dynamic temporal branch: For dynamic video input, after extracting continuous frame features, use temporal convolution, recurrent network, temporal attention or Transformer temporal encoder to aggregate them in order to capture the dynamic sonogram representation that changes with the breathing cycle.

[0052] (6) Hierarchical branch: integrate classification results, segmentation results, sign recognition results and time series features to output severity level or severity score.

[0053] (7) Joint optimization: During the training phase, the weighted sum of classification loss, segmentation loss, feature recognition loss and hierarchical loss is used as the total loss function to achieve multi-task collaborative optimization.

[0054] Example 4: Online Reasoning and Interpretable Output

[0055] This embodiment is used to illustrate the operation process of the present invention in a real clinical scenario.

[0056] [Step S201] Image access: Acquire ultrasound images and / or videos of the child's lungs from ultrasound equipment, workstation or in-hospital system.

[0057] [Step S202] Automatic preprocessing: The system completes desensitization, quality control and standardization processes.

[0058] [Step S203] Model reasoning: Input the processed data into the multi-task artificial intelligence model to obtain the pneumonia auxiliary diagnosis results.

[0059] [Step S204] Interpretable generation: Extract the feature response or attention weight of the model during the inference process and map it back to the original image space to generate a heat map or lesion marker map.

[0060] [Step S205] Result display: The probability of pneumonia, lesion area, sign markers and severity classification are displayed simultaneously on the display interface for doctors to review.

[0061] In a specific example, if the system detects patchy hypoechoic consolidation in the lower lung field, interruption of the adjacent pleural line, and air bronchogram sign, and the local lung sliding sign is weakened in the dynamic video, the system can output a positive result for pneumonia and provide the corresponding lesion mask and moderate to severe grade indication.

[0062] Example 5: System Deployment and Equipment Example

[0063] This embodiment is used to illustrate the system and hardware deployment method corresponding to the present invention.

[0064] The system can be deployed at hospital ultrasound workstations, pediatric bedside terminals, edge computing devices, or cloud servers. The system includes a processor, memory, communication interfaces, and a display terminal.

[0065] The communication interface can establish data communication connections with ultrasound equipment, PACS, HIS, EMR or other hospital systems to automatically receive image data and transmit analysis results back.

[0066] The processor performs the aforementioned data preprocessing, multi-task inference, and result output steps when executing a computer program. To improve operating efficiency, the processor may include a CPU, GPU, NPU, or other parallel computing units.

[0067] The storage medium can store trained model parameters, rule configuration files, log information, and software programs used to implement the method of this invention.

[0068] Example 6: Clinical Application Case

[0069] This embodiment illustrates the specific application process of the present invention in the context of bedside lung ultrasound-assisted diagnosis in pediatric outpatient and emergency departments.

[0070] In a specific implementation scenario, this invention is deployed in a pediatric outpatient and emergency ultrasound workstation. The subject is a child presenting with fever, cough, and shortness of breath. After the attending physician completes a basic vital sign assessment, a high-frequency linear array probe and / or a micro-convex probe are used to perform a lung ultrasound scan according to a preset zoning standard for the anterior chest, lateral chest, and back. Static images of multiple standard sections of both lungs and several dynamic video segments are acquired. To improve comparability, each dynamic video segment can cover at least one respiratory cycle, and the video duration can be set to 3 to 5 seconds.

[0071] After acquisition, the image data is automatically transmitted to the auxiliary diagnostic system of this invention. The system first performs desensitization, quality control, and standardization processing, filtering out images that do not meet the interpretation requirements, and performing size normalization, grayscale standardization, and temporal segment extraction on the retained images. Subsequently, the standardized static images and dynamic videos are input into a multi-task artificial intelligence model to perform pneumonia positive probability calculation, lesion region segmentation, pneumonia-related sign recognition, and severity grading, respectively.

[0072] In this embodiment, the system output shows that there is a discontinuity of the pleural line and patchy hypoechoic consolidation in the posterior inferior region of the left lower lung. Air bronchograms are visible within the consolidation area, and multiple fused B-lines can be identified in the adjacent area. A weakened local lung sliding sign was also detected in the dynamic video. The system further provides a positive probability value for pneumonia, a lesion mask map, a sign marker map, and a moderate severity grading result. The results output interface can simultaneously display a heat map, the area ratio of the consolidation region, the number of affected regions, and a list of signs for rapid review by clinicians.

[0073] Clinicians, combining the child's medical history, physical examination, laboratory inflammatory markers, and the auxiliary results output by the system of this invention, determine whether the child may have pneumonia, and decide whether to conduct further chest imaging examinations, admit the child for observation, or administer anti-infective and symptomatic treatment. If the doctor believes that there is a deviation in the identification result of a certain sign, the lesion area or sign label can be manually corrected in the interactive interface. The correction results can serve as a data source for subsequent model continuous optimization.

[0074] In post-treatment follow-up scenarios, doctors can collect lung ultrasound images and / or dynamic videos again 24 to 72 hours later according to the same zoning standards, and then use the system of this invention for repeated analysis. By comparing the area of ​​consolidation, the distribution range of B-lines, the degree of pleural line abnormalities, and changes in dynamic signs between the two examinations, the system can output trend information such as lesion shrinkage, sign reduction, or decrease in severity, thereby assisting in the assessment of treatment response and disease progression.

[0075] Therefore, this invention can not only be used for the initial auxiliary diagnosis of pneumonia in children, but also for bedside dynamic follow-up, efficacy monitoring, and rapid screening in primary healthcare settings, and has good clinical application value.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, any equivalent substitutions, changes, or improvements made without departing from the concept of the present invention should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for auxiliary diagnosis of pediatric pneumonia based on lung ultrasound imaging, characterized in that, include: Acquire static and / or dynamic ultrasound images and / or videos of the lungs of the pediatric subjects to be analyzed, which are collected according to a preset scanning zoning standard. Desensitize, quality control, normalize, and enhance the static and / or dynamic ultrasound images and videos to obtain standardized input data. Input the standardized input data into a multi-task artificial intelligence model for joint inference. The multi-task artificial intelligence model includes at least a feature extraction module, a pneumonia sign recognition module, a lesion segmentation module, a dynamic temporal fusion module, and a severity assessment module. Based on the joint inference results, output pneumonia positive / negative judgment results, lesion region segmentation results, pneumonia-related sign recognition results, and severity grading results. Generate visual heatmaps, sign marker maps, and / or lesion mask maps based on the feature responses or attention weights during the joint inference process to assist clinical interpretation.

2. The method according to claim 1, characterized in that, The preset scanning zoning standard includes multi-zone standardized scanning of the anterior chest, lateral chest, and back; the pneumonia-related signs include at least one or more of the following: pulmonary consolidation, abnormal pleural lines, increased or fused B lines, air bronchogram sign, abnormal lung sliding sign, and pleural effusion; the severity grading is determined based on the number, range, area proportion, distribution area, and / or dynamic change characteristics of the pneumonia-related signs.

3. The method according to claim 1 or 2, characterized in that, The multi-task artificial intelligence model includes: a backbone feature extraction network for extracting multi-scale spatial features; a classification branch for outputting positive / negative probability values ​​for pneumonia and / or pathogen type classification results; a segmentation branch for pixel-level segmentation of pulmonary consolidation areas and / or pleural effusion areas; a sign recognition branch for identifying abnormal pleural lines, dense B-lines, air bronchogram signs, and abnormal lung sliding signs; a temporal fusion branch for temporal modeling of continuous frame features in dynamic videos to extract dynamic change information within the respiratory cycle; and a grading branch for outputting mild, moderate, or severe grading results based on classification results, segmentation results, sign recognition results, and temporal features.

4. The method according to claim 3, characterized in that, The temporal fusion branch employs temporal convolutional networks, recurrent neural networks, temporal attention modules, Transformer temporal encoders, or combinations thereof, to aggregate features from consecutive frames in dynamic videos, thereby enhancing the ability to identify abnormal lung sliding signs, dynamic B-line changes, and lesion boundary changes.

5. The method according to claim 1, characterized in that, The method further includes: forming a gold standard label library by double-blind or multi-blind labeling of training samples by multiple labeling experts; the gold standard labels include at least normal / pneumonia classification labels, lesion segmentation labels, sign labels and severity labels; and a weighted combination of classification loss, segmentation loss and grading loss is used as a joint optimization objective during model training.

6. A pediatric pneumonia auxiliary diagnostic system based on lung ultrasound imaging, characterized in that, include: The data acquisition and standardization module is used to acquire static images and / or dynamic videos of children's lung ultrasound, and to perform desensitization, quality control and standardization processing; The multi-task AI reasoning module is used to perform feature extraction, sign recognition, lesion segmentation, dynamic temporal fusion, and severity assessment on standardized image data; the result output and interpretability interaction module is used to output pneumonia auxiliary diagnosis results, lesion area annotation, sign marking, severity grading, and heat map or lesion mask map.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images