Device for differential and quantitative assessment of pulmonary edema and pneumothorax based on multi-modal flexible ultrasound

Abstract

The application discloses a kind of lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound, the method is by synchronously obtaining the morphological ultrasound image and the elastic mechanics image of the same anatomical position, B line quantitative index, pleural line sliding index and the average strain value of region of interest are extracted from morphological ultrasound image and elastic mechanics image;B line quantitative index, pleural line sliding index, average strain value are combined into feature vector, feature vector is input into multi-parameter fusion decision model, obtains lung edema index and pneumothorax risk probability;Lung edema index and pneumothorax risk probability are displayed in display terminal, and when lung edema index exceeds preset index threshold or pneumothorax risk probability exceeds preset probability threshold, alarm is triggered;It can reduce the misdiagnosis rate under emergency scene, reduce end-to-end response time, provide objective, real-time, quantifiable decision support for clinician, significantly reduce the risk of missed diagnosis and improve emergency treatment efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent medical treatment and medical image analysis, and in particular to a lung edema and pneumothorax differential and quantitative evaluation device based on multi-modal flexible ultrasound. BACKGROUND

[0002] Flexible ultrasound sensors bring a glimmer of hope for long-term continuous monitoring of lung diseases; the current technical focus is mostly on how to realize the flexibility, miniaturization of the sensor and the stability of image acquisition; however, a similarly critical but often overlooked challenge is: how to effectively and automatically interpret the vast amount of ultrasound image data generated?

[0003] In clinical practice, lung ultrasound diagnosis relies on doctors identifying specific imaging markers, such as:

[0004] 1. Pulmonary edema: mainly judged by observing and counting the number and characteristics of B-lines (comet tail artifacts produced by ultrasound waves in water-filled interlobular septa).

[0005] 2. Pneumothorax: mainly judged by observing whether "lung sliding sign" (relative movement of visceral and parietal pleura with respiration) exists or "sandwich sign" disappears.

[0006] However, this diagnostic method has significant limitations:

[0007] 3. High subjectivity: different doctors have different judgments on B-line counting and lung sliding sign strength.

[0008] 4. Low efficiency: it is unrealistic to manually analyze the large amount of image data produced by continuous monitoring.

[0009] 5. Lack of quantitative standards: unable to provide traceable, comparable objective numerical indicators, which is not conducive to accurate assessment of disease progression.

[0010] 6. Insufficient information utilization: existing methods mainly use morphological imaging (B-mode), while the elastographic information provided by flexible ultrasound is not fully utilized for differential diagnosis. SUMMARY

[0011] The main purpose of the present application is to provide a lung edema and pneumothorax differential and quantitative evaluation device based on multi-modal flexible ultrasound, which aims to solve the technical problems of low clinical diagnosis accuracy and high misdiagnosis rate in the prior art due to the difficulty in accurately differentiating lung edema and pneumothorax, as both of them show increased B-lines and similar pleural sliding signs on ultrasound images.

[0012] The present application provides a lung edema and pneumothorax differential and quantitative evaluation device based on multi-modal flexible ultrasound, which comprises:

[0013] a data extraction module, configured to synchronously acquire a morphological ultrasound image and an elastomechanical image of a same anatomical position, and extract a B-line quantification index, a pleural line sliding index, and an average strain value of a region of interest from the morphological ultrasound image and the elastomechanical image;

[0014] a discriminant quantification evaluation module, configured to combine the B-line quantification index, the pleural line sliding index, and the average strain value into a feature vector, input the feature vector into a multi-parameter fusion decision model, and obtain a pulmonary edema index and a pneumothorax risk probability;

[0015] a display warning module, configured to display the pulmonary edema index and the pneumothorax risk probability on a display terminal, and trigger an alarm when the pulmonary edema index exceeds a preset index threshold or the pneumothorax risk probability exceeds a preset probability threshold.

[0016] Optionally, the data extraction module is further configured to synchronously acquire, by a flexible ultrasound sensor, the morphological ultrasound image and the elastomechanical image of the same anatomical position in real time;

[0017] The data extraction module is further configured to automatically identify the B-line in the morphological ultrasound image based on a convolutional neural network, and output the B-line quantification index.

[0018] The data extraction module is further configured to analyze a motion velocity and a motion amplitude of a pleural line in the morphological ultrasound image, and generate the pleural line sliding index according to the motion velocity and the motion amplitude.

[0019] The data extraction module is further configured to analyze the elastomechanical image, and calculate the average strain value of the region of interest.

[0020] Optionally, the data extraction module is further configured to perform pixel-level segmentation on the preprocessed morphological ultrasound image by using a convolutional neural network, and obtain a B-line pixel-level segmentation mask.

[0021] The data extraction module is further configured to calculate, according to the B-line pixel-level segmentation mask, an absolute number of B-lines in a B-line region, an average brightness value of the B-line region, and a B-line fusion region proportion, and take the absolute number, the average brightness value, and the B-line fusion region proportion as the B-line quantification index.

[0022] Optionally, the data extraction module is further configured to perform contrast enhancement and standardization processing on the morphological ultrasound image, and obtain the preprocessed morphological ultrasound image.

[0023] The data extraction module is further configured to perform noise reduction and connection breakpoint processing on the B-line pixel-level segmentation mask, and locate the B-line region by using a multi-scale convolutional feature extraction and an attention mechanism.

[0024] Optionally, the data extraction module is further configured to perform time-series dynamic analysis on the morphological ultrasound image to obtain a pleural line boundary in each frame of image, and generate a spatial coordinate sequence according to the pleural line boundary;

[0025] The data extraction module is further configured to calculate a motion speed and a motion amplitude of the pleural line in respiratory motion according to the spatial coordinate sequence, and perform weighted fusion on the motion speed and the motion amplitude according to motion periodicity and motion direction according to a preset fusion weight to obtain a dimensionless pleural line sliding index.

[0026] Optionally, the data extraction module is further configured to input a sequence of continuous B-mode images in the morphological ultrasound image as data input, and perform automatic detection and tracking on a pleural line of the sequence of continuous B-mode images based on image gradient or machine learning.

[0027] The data extraction module is further configured to, when the analysis model is selected as M-type analysis, generate an M-type image of time and depth, analyze a sand beach sign or a barcode sign of the pleural line in M mode according to the M-type image, and calculate a motion amplitude and a motion periodicity of a motion signal according to the sand beach sign or the barcode sign.

[0028] The data extraction module is further configured to, when the analysis model is selected as optical flow analysis, calculate a motion vector of the pleural line of continuous frames, and estimate a motion speed and a motion direction according to the motion vector.

[0029] Optionally, the data extraction module is further configured to automatically segment a region of interest from the elastomechanical image through a preset lung anatomical feature.

[0030] The data extraction module is further configured to perform displacement estimation on adjacent frames in the elastomechanical image by using a cross-correlation algorithm to obtain a local strain value of each pixel point in the region of interest.

[0031] The data extraction module is further configured to generate a strain distribution map according to the local strain value, and perform spatial weighted averaging on strain values of all effective pixels in the strain distribution map to obtain an average strain value.

[0032] Optionally, the discriminant quantification evaluation module is further configured to combine the B-line quantification index, the pleural line sliding index and the average strain value into a feature vector according to a preset dimension order.

[0033] The discriminant quantification evaluation module is further configured to input the feature vector into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree to obtain a pulmonary edema index and a pneumothorax risk probability.

[0034] Optionally, the differential quantification evaluation module is further configured to acquire a heart rate and a respiration rate of the to-be-detected object through the biosignal sensor, and take the heart rate and the respiration rate as basic physiological parameters;

[0035] The differential quantification evaluation module is further configured to take the basic physiological parameters as additional input features, input the additional input features and the feature vector into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree, and establish a nonlinear pathological correlation among the B-line quantification index, the pleural line sliding index, and the average strain value.

[0036] The differential quantification evaluation module is further configured to perform dynamic risk calibration according to the nonlinear pathological correlation, and output a pulmonary edema index and a pneumothorax risk probability.

[0037] Optionally, the display warning module is further configured to display the pulmonary edema index in real time on a display terminal by using a pointer-type instrument panel, and display the pneumothorax risk probability in real time on the display terminal by using a percentage progress bar.

[0038] The display warning module is further configured to compare the pulmonary edema index with a preset index threshold value, and compare the pneumothorax risk probability with a preset probability threshold value.

[0039] The display warning module is further configured to trigger an alarm when the pulmonary edema index exceeds the preset index threshold value or the pneumothorax risk probability exceeds the preset probability threshold value.

[0040] This invention proposes a device for the differentiation and quantitative assessment of pulmonary edema and pneumothorax based on multimodal flexible ultrasound. The device comprises: a data extraction module for simultaneously acquiring morphological ultrasound images and elastography images at the same anatomical location, and extracting B-line quantitative indicators, pleural line slippage index, and average strain values ​​of the region of interest from the morphological ultrasound images and the elastography images; a differentiation and quantitative assessment module for combining the B-line quantitative indicators, the pleural line slippage index, and the average strain values ​​into a feature vector, and inputting the feature vector into a multi-parameter fusion decision model to obtain a pulmonary edema index and a pneumothorax risk probability; and a display and warning module for displaying the pulmonary edema index and the pneumothorax risk probability on a display terminal, and warning the system when the pulmonary edema index exceeds a preset threshold. An alarm is triggered when the number of thresholds or the probability of pneumothorax risk exceeds a preset probability threshold; this can reduce the misdiagnosis rate in emergency scenarios, reduce end-to-end response time, provide clinicians with objective, real-time, and quantifiable decision support, significantly reduce the risk of missed diagnosis and improve the efficiency of emergency treatment, and transform subjective imaging judgments into objective digital indicators, eliminating differences in judgment among doctors and providing a precise scale for disease tracking; it achieves automated monitoring, automatically analyzes data and provides diagnostic opinions at any time, freeing up medical staff and making large-scale, long-term continuous monitoring possible; multi-parameter cross-validation greatly improves the specificity and accuracy of diagnosis, surpassing single-modality analysis; through continuous trend analysis of quantitative indicators, it can detect signs of worsening pulmonary edema or pneumothorax before clinical symptoms appear, achieving prospective early warning. Attached Figure Description

[0041] Figure 1 This is a functional block diagram of the first embodiment of the device for differentiating and quantitatively assessing pulmonary edema and pneumothorax based on multimodal flexible ultrasound of the present invention.

[0042] Figure 2 This is a schematic diagram of the overall structure and data flow of the pulmonary edema and pneumothorax differentiation and quantitative assessment system based on multimodal flexible ultrasound of the present invention;

[0043] Figure 3 This is a flowchart of the intelligent identification and quantification process of B-line in the pulmonary edema and pneumothorax differentiation and quantitative assessment device based on multimodal flexible ultrasound of the present invention;

[0044] Figure 4 This is a schematic diagram of the dynamic analysis of lung sliding sign in the pulmonary edema and pneumothorax differentiation and quantitative assessment device based on multimodal flexible ultrasound of the present invention.

[0045] Figure 5 This is a schematic diagram of the workflow of the multi-parameter fusion decision engine in the pulmonary edema and pneumothorax identification and quantitative assessment device based on multimodal flexible ultrasound of the present invention.

[0046] Figure 6 The figure is a man-machine interaction interface schematic diagram of the lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound of the present application.

[0047] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0048] It should be understood that the specific embodiments described herein are merely intended to explain the present application and not to limit the present application.

[0049] The solution of the embodiment of the present application is mainly: the lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound comprises: a data extraction module, which is used for synchronously acquiring a morphological ultrasound image and an elastomechanical image of the same anatomical position, and extracting a B-line quantitative index, a pleural line sliding index and an average strain value of a region of interest from the morphological ultrasound image and the elastomechanical image; an identification and quantitative evaluation module, which is used for combining the B-line quantitative index, the pleural line sliding index and the average strain value into a feature vector, inputting the feature vector into a multi-parameter fusion decision model, and obtaining a lung edema index and a pneumothorax risk probability; and a display and early warning module, which is used for displaying the lung edema index and the pneumothorax risk probability on a display terminal, and triggering an alarm when the lung edema index exceeds a preset index threshold or the pneumothorax risk probability exceeds a preset probability threshold; the device can reduce the misdiagnosis rate in the emergency scene, reduce the end-to-end response time, provide objective, real-time and quantifiable decision support for clinicians, significantly reduce the risk of missed diagnosis and improve the efficiency of emergency treatment, can convert subjective imaging judgment into objective digital indicators, eliminate the judgment differences among doctors, provide accurate scales for disease tracking, realize automatic monitoring, automatically analyze data and give diagnostic opinions at any time, liberate medical staffs, and make large-scale and long-term continuous monitoring possible; multi-parameter cross-validation greatly improves the specificity and accuracy of diagnosis, and surpasses single modal analysis; through continuous trend analysis of the quantitative index, signs of lung edema aggravation or pneumothorax occurrence can be found early before the clinical symptoms are obvious, prospective early warning is realized, and the technical problems that it is difficult to accurately identify lung edema and pneumothorax in the prior art, the B-line increases and the pleural sliding sign is similar in the ultrasound images of the two, and the clinical diagnosis accuracy is low and the misdiagnosis rate is high are solved.

[0050] Reference Figure 1 , Figure 1 The figure is a function module diagram of the first embodiment of the lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound of the present application.

[0051] The application is based on the first embodiment of the lung edema and pneumothorax differential and quantitative evaluation device based on multi-modal flexible ultrasound, which comprises a data extraction module 10, a differential quantitative evaluation module 20 and a display warning module 30; wherein,

[0052] The data extraction module 10 is used for synchronously acquiring morphological ultrasound images and elastomechanical images of the same anatomical position, and extracting B-line quantitative indicators, pleural line sliding indexes and average strain values of the region of interest from the morphological ultrasound images and the elastomechanical images.

[0053] The differential quantitative evaluation module 20 is used for combining the B-line quantitative indicators, the pleural line sliding indexes and the average strain values into a feature vector, inputting the feature vector into a multi-parameter fusion decision model, and obtaining a lung edema index and a pneumothorax risk probability.

[0054] The display warning module 30 is used for displaying the lung edema index and the pneumothorax risk probability on a display terminal, and triggering an alarm when the lung edema index exceeds a preset index threshold or the pneumothorax risk probability exceeds a preset probability threshold.

[0055] It should be noted that the data extraction module can synchronously collect the same anatomical position of the patient in real time, ensure that the morphological ultrasound images and the elastomechanical images are strictly aligned in the time dimension and completely consistent in the spatial position, avoid feature distortion caused by image offset, extract B-line quantitative indicators, pleural line sliding indexes and average strain values of the region of interest from the morphological ultrasound images and the elastomechanical images, thereby providing high-precision, multi-modal pathological feature input for multi-parameter fusion decision, and completely solving the lung edema and pneumothorax differential difficulty problem caused by single ultrasound mode in the prior art.

[0056] It can be understood that the differential quantitative evaluation module realizes accurate quantitative differentiation of lung edema and pneumothorax by systematically fusing multi-modal pathological features, integrates B-line quantitative indicators, pleural line sliding indexes and average strain values into a feature vector, ensures that the features are dimensionless and dimensionless, and eliminates dimensional interference; then, the feature vector is input into a multi-parameter fusion decision model in real time, finally, the model outputs a lung edema index and a pneumothorax risk probability, completely solves the lung edema and pneumothorax misjudgment problem caused by single ultrasound feature in the prior art, and provides objective and quantifiable decision basis for clinical practice.

[0057] It should be understood that the pulmonary edema index and the pneumothorax risk probability are displayed on the display terminal, the pulmonary edema index and the preset threshold and the pneumothorax risk probability and the preset threshold are compared in real time, and once any index exceeds the threshold, an alarm is activated, end-to-end response delay is realized in an emergency scene, the risk of misdiagnosis and missed diagnosis is reduced, and immediate decision support is provided for a clinician, so that the efficiency of treatment of a critical patient and the reliability of diagnosis are significantly improved.

[0058] Further, the data extraction module 10 is further configured to obtain a morphological ultrasound image and an elastomechanical image by synchronously collecting the same anatomical position in real time through the flexible ultrasound sensor.

[0059] The data extraction module 10 is further configured to automatically identify B lines in the morphological ultrasound image based on a convolutional neural network and output a B line quantitative index.

[0060] The data extraction module 10 is further configured to analyze a motion speed and a motion amplitude of a pleural line in the morphological ultrasound image, and generate a pleural line sliding index according to the motion speed and the motion amplitude.

[0061] The data extraction module 10 is further configured to analyze the elastomechanical image and calculate an average strain value of a region of interest.

[0062] It should be noted that the same anatomical position of a patient is synchronously collected in real time through the flexible ultrasound probe, so as to ensure that the morphological ultrasound image sequence and the elastomechanical image sequence are strictly aligned in the time dimension and consistent in the spatial position. Based on the morphological ultrasound image, a lightweight convolutional neural network is used to automatically identify a B line region, accurately calculate the number of B lines, the average brightness value of B lines and the proportion of B line fusion regions in a unit field of view, and form core indexes for quantitatively evaluating the degree of pulmonary interstitial edema. At the same time, by analyzing the continuous displacement trajectory of the pleural line in the respiratory cycle in the morphological ultrasound image sequence, the motion speed and amplitude parameters of the pleural line are calculated, and a sliding index is generated to dynamically reflect the pleural sliding sign. In addition, the average strain value is extracted as a key elastomechanical parameter for evaluating the elasticity and pathological state of lung tissue by selecting a lung region of interest from the elastomechanical image and based on a strain distribution calculation method.

[0063] Further, the data extraction module 10 is further configured to perform pixel-level segmentation on the preprocessed morphological ultrasound image by using a convolutional neural network to obtain a B line pixel-level segmentation mask.

[0064] The data extraction module 10 is further configured to calculate the absolute number of B lines in the B line region, the average brightness value of the B line region and the proportion of the B line fusion region according to the B line pixel-level segmentation mask, and take the absolute number, the average brightness value and the proportion of the B line fusion region as the B line quantitative index.

[0065] In a specific implementation, a lightweight convolutional neural network (such as MobileNetV3-Small) can be used for end-to-end deep learning analysis of morphological ultrasound images, precise positioning of B-line regions through multi-scale convolution feature extraction and attention mechanism, effective suppression of noise interference and differentiation of B-lines and normal lung tissue structures; in the network processing process, a B-line pixel-level segmentation mask can be automatically generated, and then the absolute number of B-lines in a unit field of view, the average brightness value of the B-line region and the B-line fusion region proportion can be quantitatively calculated; the finally output B-line quantification index is used as a standardized parameter, which is directly input into a pulmonary edema quantification evaluation module, provides high-precision and repeatable pathological feature input for subsequent multi-modal data fusion, and significantly improves the objectivity and sensitivity of early identification of pulmonary edema.

[0066] Further, the data extraction module 10 is further configured to perform contrast enhancement and standardization processing on the morphological ultrasound images to obtain preprocessed morphological ultrasound images.

[0067] The data extraction module 10 is further configured to perform noise removal and connection breakpoint processing on the B-line pixel-level segmentation mask, and locate the B-line region through multi-scale convolution feature extraction and attention mechanism.

[0068] It can be understood that performing adaptive contrast enhancement (such as a (Contrast Limited Adaptive Histogram Equalization, CLAHE) algorithm of contrast limited adaptive histogram equalization) and global normalization processing (normalizing pixel values to the [0, 1] range and eliminating device differences) on the original morphological ultrasound image ensures that the image remains consistent in features under different acquisition conditions, providing high-quality input for subsequent analysis; secondly, for the B-line pixel-level segmentation mask (generated by a lightweight Convolutional Neural Network, CNN), the module implements two-stage optimization: morphological opening operation and median filtering are used for denoising processing to effectively eliminate small-scale noise points (such as isolated pixels caused by artifacts or motion blur), and broken B-lines (such as repairing B-line breakpoints caused by respiratory motion) are connected through connected region analysis and edge growth algorithm, making the mask continuous and complete in space; finally, multi-scale convolution feature extraction (capturing multi-level structural features of B-lines such as small B-lines and fusion regions through convolution kernels of different scales in MobileNetV3-Small network) and channel attention mechanism (such as Squeeze-and-Excitation Block, SEBlock dynamically weighting key feature channels, focusing on B-line high response areas) are used to realize precise positioning and semantic segmentation of B-line regions, ensuring that the calculation error of B-line quantitative indicators (number per field of view, average brightness, fusion ratio) is reduced to <3%, thereby providing high-precision, anti-interference pathological feature input for real-time identification of pulmonary edema and pneumothorax, and completely solving the quantitative deviation problem caused by image noise or B-line breakage in traditional methods.

[0069] Further, the data extraction module 10 is further configured to perform time sequence dynamic analysis on the morphological ultrasound image to obtain a pleural line boundary in each frame of image, and generate a spatial coordinate sequence according to the pleural line boundary.

[0070] The data extraction module 10 is further configured to calculate a motion speed and a motion amplitude of the pleural line in respiratory motion according to the spatial coordinate sequence, and perform weighted fusion on the motion speed and the motion amplitude according to motion periodicity and motion direction according to a preset fusion weight to obtain a dimensionless pleural line sliding index.

[0071] It should be noted that in each frame of image, the edge detection algorithm (such as Canny edge detection operator) or lightweight convolutional neural network (MobileNetV3-Small) is used to automatically identify the pixel-level boundary of the pleural line, and the spatial coordinate sequence of the pleural line key points of the continuous frames (such as the trajectory of the center point x, y coordinates changing with time) is generated; then, based on the coordinate sequence, the system strictly calculates the dynamic parameters of the pleural line in the respiratory motion - the motion speed is obtained by dividing the displacement between adjacent frames by the time interval (unit: mm / s), which reflects the real-time speed of the pleural line sliding; the motion amplitude is calculated by the maximum displacement difference of the pleural line in the respiratory cycle (unit: mm), which represents the integrity of the sliding range; on this basis, the module is weighted and fused according to the periodicity of the respiratory motion (such as the symmetry of the inspiration / expiration phase) and the direction of the pleural line motion (such as the radial displacement perpendicular to the pleural line) according to the preset fusion weight (for example: speed weight 0.6, amplitude weight 0.4). Generate dimensionless pleural line sliding index (Sliding Index, SI), which dynamically captures the continuity and stability of pleural sliding, realizes the accurate quantification of pathological state: in patients with pulmonary edema, the sliding index is significantly reduced (such as <50) because the pleural sliding is limited; in patients with pneumothorax, the index tends to 0 (such as <5) because the pleural sliding completely disappears, thereby providing key dynamic feature basis for multi-parameter fusion decision, and completely solving the misjudgment problem of pulmonary edema and pneumothorax caused by similar pleural sliding signs in traditional methods.

[0072] Further, the data extraction module 10 is also used to input a sequence of continuous B-mode images in the morphological ultrasound image as data, and automatically detect and track the pleural line of the continuous B-mode image sequence based on image gradient or machine learning;

[0073] The data extraction module 10 is also used to generate an M-mode image of time and depth when the analysis model is selected as M-type analysis, analyze the sand beach sign or bar code sign of the pleural line in the M mode according to the M-mode image, and calculate the motion amplitude and motion periodicity of the motion signal according to the sand beach sign or the bar code sign.

[0074] The data extraction module 10 is also used to calculate the motion vector of the continuous frame pleural line when the analysis model is selected as optical flow analysis, and estimate the motion speed and motion direction according to the motion vector.

[0075] It can be understood that in the B mode analysis mode, a sequence of continuous B mode ultrasound images is taken as input data, an image gradient operator (such as a Sobel operator) or a lightweight machine learning model (such as MobileNetV3-Small) is used to automatically detect the pleural line boundary and perform time series tracking, and a sequence of spatial coordinates of key points of the pleural line in each frame of image (such as the trajectory of the center point x, y coordinates changing with time) is generated, so as to ensure that the boundary positioning error is less than 5 pixels; in the M type analysis mode, the module automatically generates a time-depth graph (M type image), intelligently identifies the characteristic signs of the pleural line in the M mode, i.e. the sand beach sign (showing irregular and broken waveforms, indicating the interruption of pleural sliding abnormality, which is common in pulmonary edema) or the barcode sign (showing regular and continuous stripe waveforms, indicating stable pleural sliding, which is common in healthy state), and accurately calculates the motion amplitude (the maximum displacement difference of the pleural line in the respiratory cycle, unit: mm) and the motion periodicity (such as respiratory rate, unit: times / minute), to provide visual basis for pathological state; in the optical flow analysis mode, the system calculates the motion vector field of the pleural line based on the sequence of continuous frames (using the TV-L1 optical flow algorithm), estimates the motion velocity (unit: mm / s) by the vector size and analyzes the motion direction (such as radial or tangential), to realize the millisecond-level capture of the dynamic characteristics of the pleural sliding; the three analysis modes can work independently or cooperatively, to ensure that in different clinical scenarios (such as irregular patient breathing or image noise interference), the system can adaptively select the optimal algorithm, output high-fidelity motion parameters, provide accurate input for the calculation of the pleural line sliding index (generated by fusing the velocity and amplitude with a weight of 0.6:0.4), and reduce the sliding index calculation error to less than 3%, completely solving the misjudgment problem of pulmonary edema and pneumothorax caused by similar pleural sliding signs in traditional methods, and significantly improving the reliability of real-time differential diagnosis.

[0076] Further, the data extraction module 10 is further configured to automatically segment a region of interest from the elastomechanical image based on preset lung anatomical features.

[0077] The data extraction module 10 is further configured to estimate the displacement of adjacent frames in the elastomechanical image by using a cross-correlation algorithm to obtain a local strain value of each pixel point in the region of interest.

[0078] The data extraction module 10 is further configured to generate a strain distribution map based on the local strain value, and perform spatial weighted averaging on the strain values of all effective pixels in the strain distribution map to obtain an average strain value.

[0079] It should be understood that the high-precision elastomechanical quantitative parameters are provided for the differentiation between pulmonary edema and pneumothorax through the intelligent elastomechanical image processing flow: first, based on the preset lung anatomical features (such as the position of the pleural line, the lung parenchyma density threshold, and the blood vessel distribution boundary), the region of interest (ROI) in the elastomechanical image is automatically segmented, ensuring that the ROI accurately covers the key areas of the lung parenchyma and subpleural tissue (coverage range > 95%), effectively excluding blood vessels, artifacts and other interference sources; then, the TV-L1 cross-correlation algorithm is used to estimate the sub-pixel level displacement of adjacent frames in the elastomechanical image sequence, accurately calculate the local strain value of each pixel point in the ROI (strain = pixel displacement / original tissue length, spatial resolution up to 100 μm), and generate a high-precision strain distribution map (dynamic range 0-50%); on this basis, the spatial weighted average processing (weight function dynamically allocates the distance between pixels and the pleural line, the closer the distance, the higher the weight) is performed on all effective pixels in the strain distribution map (effective rate > 98% after removing noise points), and the dimensionless average strain value is output, which is significantly reduced in pulmonary edema patients (typical value < 5%) and is strongly negatively correlated with the degree of pulmonary interstitial edema, and is locally abnormally high or fluctuating in pneumothorax patients (typical value > 15%), thereby providing key mechanical basis for the multi-parameter fusion decision model, and controlling the quantization error of elastomechanical characteristics to < 2%, completely solving the problem of insufficient differentiation accuracy caused by the lack of tissue elasticity parameters in traditional methods.

[0080] Further, the differentiation quantitative evaluation module 20 is further configured to combine the B-line quantitative indicators, the pleural line sliding index and the average strain value into a feature vector according to a preset dimension order.

[0081] The differentiation quantitative evaluation module 20 is further configured to input the feature vector into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree, to obtain a pulmonary edema index and a pneumothorax risk probability.

[0082] It can be understood that the B-line quantitative indicators (including the number of B-lines in a unit field of view, the average brightness of B-lines, and the proportion of B-line fusion area), the pleural line sliding index (a dimensionless parameter generated by dynamically weighting the motion speed and amplitude of the pleural line in a respiratory cycle with a weight of 0.6:0.4), and the average strain value (a tissue strain distribution value calculated by spatially weighted averaging in the elastic mechanics image of the lung region of interest) are strictly integrated into a standardized five-dimensional feature vector, and the dimension order is preset as [B-line number, B-line brightness, B-line fusion proportion, sliding index, average strain], which ensures the dimensionless alignment between features and eliminates dimensional interference; then, the feature vector is input in real time into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree (GBDT), which is trained by fusing 50,000+ clinical lung ultrasound data (including pulmonary edema, pneumothorax, and normal control group), can deeply mine the nonlinear pathological correlation between B-line indicators, pleural sliding sign, and elastic mechanics parameters (such as the interaction between B-line fusion proportion and average strain), and automatically associate basic physiological parameters such as respiratory rate and heart rate for dynamic risk calibration; finally, the model outputs a pulmonary edema index (0-100 score continuous quantitative value, index>50 indicates moderate to severe pulmonary edema) and a pneumothorax risk probability (0%-100% dynamic probability value, probability>30% indicates high risk), generates a comprehensive diagnostic report through a weighted fusion algorithm (pulmonary edema weight 0.7, pneumothorax weight 0.3), realizes a 92.3% differential accuracy (22.3% higher than traditional methods) and <200ms real-time response in emergency scenarios, completely solves the misjudgment problem of pulmonary edema and pneumothorax caused by single ultrasound feature in the prior art, and provides precise and quantifiable decision support for clinicians.

[0083] Further, the differential quantitative evaluation module 20 is further configured to acquire a heart rate and a respiratory rate of the to-be-detected object through a biological signal sensor, and take the heart rate and the respiratory rate as basic physiological parameters.

[0084] The differential quantitative evaluation module 20 is further configured to take the basic physiological parameters as additional input features, input the additional input features and the feature vector into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree, and establish a nonlinear pathological correlation between the B-line quantitative indicators, the pleural line sliding index, and the average strain value.

[0085] The differential quantitative evaluation module 20 is further configured to perform dynamic risk calibration according to the nonlinear pathological correlation, and output a pulmonary edema index and a pneumothorax risk probability.

[0086] It should be understood that the heart rate (unit: times / minute) and the respiratory rate (unit: times / minute) of the object to be detected are collected in real time by a non-invasive biosignal sensor such as a patch-type electrocardio electrode or a respiratory impedance sensor, and these basic physiological parameters are taken as key auxiliary features, which are integrated with the B-line quantification index (unit: number of B-lines in the field of view, average brightness, and fusion area ratio), the pleural line sliding index (a dimensionless value generated by fusing the respiratory cycle movement speed and amplitude with a weight of 0.6:0.4), and the average strain value (tissue strain calculated based on elastic mechanics image space weighted average) into a six-dimensional extended feature vector; the feature vector is input in real time into a multi-parameter fusion decision model based on gradient boosting decision tree (GBDT), which is trained by fusing 50,000+ clinical data, deeply mines the nonlinear pathological correlation between the B-line index, the pleural sliding sign, the elastic mechanics parameter, and the basic physiological parameter (for example, the positive correlation between the increased heart rate and the pulmonary edema index, and the interaction between the abnormal respiratory rate and the pneumothorax risk probability), and dynamically calibrates the risk assessment model; finally, the module performs real-time risk calibration based on the learned nonlinear pathological correlation, and outputs the pulmonary edema index (0-100 score continuous quantification value, index>50 indicates moderate to severe pulmonary edema) and the pneumothorax risk probability (0%-100% dynamic probability value, probability>30% indicates high risk), wherein the calibration process automatically fuses the dynamic changes of the basic physiological parameters (such as adjusting the pulmonary edema index weight when the heart rate fluctuation is>10%), so that the diagnostic accuracy is improved to 92.3% (increased by 22.3% compared with the model relying only on ultrasound features), and the misdiagnosis rate is reduced to<5%, providing objective and adaptive real-time decision support for emergency clinics, and completely solving the misjudgment problem caused by ignoring the dynamic changes of physiological state in the prior art.

[0087] Further, the display warning module 30 is also used for displaying the pulmonary edema index in real time on the display terminal by using a pointer-type instrument panel, and displaying the pneumothorax risk probability in real time on the display terminal by using a percentage progress bar.

[0088] The display warning module 30 is also used for comparing the pulmonary edema index with a preset index threshold value, and comparing the pneumothorax risk probability with a preset probability threshold value.

[0089] The display warning module 30 is also used for triggering an alarm when the pulmonary edema index exceeds the preset index threshold value or the pneumothorax risk probability exceeds the preset probability threshold value.

[0090] It should be noted that on the display terminal (such as the high-definition touch screen of the portable ultrasonic equipment or the interface of the hospital monitoring system), the module adopts a pointer type instrument panel to dynamically display the pulmonary edema index (0-100 point continuous quantitative value) in real time, the pointer moves accurately on the scale disc, and an intelligent color coding strategy (green 0-50 points represent a safe range, yellow 50-75 points indicate that mild risk needs close monitoring, and red >75 points identify moderate to severe pulmonary edema that needs emergency intervention), so that the clinician can instantly perceive the severity of pulmonary interstitial edema; at the same time, the module presents the risk probability of pneumothorax (0%-100%) in the form of a dynamic percentage progress bar, the filling length of the progress bar is strictly matched with the numerical value (such as 35% display as a 35% filled area, and superimposed with a digital label "pneumothorax risk: 35%"), and the risk level is intuitively reflected through the gradient color (green 0-30%, yellow 30-70%, and red >70%), ensuring that the user can quickly identify key information in a complex emergency environment; the module continuously performs real-time comparison in two dimensions: the pulmonary edema index is continuously and dynamically compared with the preset threshold (50 points), and the risk probability of pneumothorax is compared at the millisecond level with the preset threshold (30%), establishing an uninterrupted monitoring flow; when either condition of the real-time value of the pulmonary edema index exceeding 50 points or the real-time value of the risk probability of pneumothorax exceeding 30% is triggered, the system immediately activates a three-level linkage alarm mechanism - first, a full-screen warning box pops up on the display interface with a 5-second pulsing red flash (to avoid false touch interference), second, a high-intensity buzzer sound (85dB, effectively penetrating the noise of the emergency environment) is started, and finally, an emergency notification containing specific numerical values (such as "pulmonary edema index: 62", "pneumothorax risk: 38%"), patient location, and priority identification is pushed to the designated medical staff's mobile terminal through an encrypted channel, achieving an end-to-end response delay <500ms, reducing the risk of misdiagnosis to <5%, providing clinicians with actionable and immediate decision support, and significantly improving the efficiency and reliability of diagnosis for critically ill patients.

[0091] In a specific implementation, referring to Figure 2 , Figure 2 The overall structure and data flow diagram of the pulmonary edema and pneumothorax identification and quantitative evaluation system based on multi-modal flexible ultrasound of the present application is shown in Figure 2 The system comprises:

[0092] 1. Multi-modal data acquisition module: used to obtain morphological ultrasound images (B-mode) and elastomechanical images of the same anatomical position from the flexible ultrasound sensor.

[0093] 2. Pathological feature quantification extraction module: comprising:

[0094] 3. B-line intelligent recognition and quantification unit: Based on a lightweight convolutional neural network (such as MobileNetV3-Small), automatically recognize B-lines in B-mode images and output quantitative indicators, such as: the number of B-lines per field of view, the average brightness of B-lines, and the proportion of B-line fusion area.

[0095] 4. Dynamic analysis unit for lung sliding sign: Based on M-mode ultrasound or continuous B-mode image sequences, calculate the motion velocity and amplitude of the pleural line through cross-correlation algorithm or optical flow method, generate a continuous sliding index (0-1), and quantify the strength of lung sliding sign.

[0096] 5. Elasticity parameter analysis unit: Analyze elastography data to calculate the average strain value or strain ratio (compared to the surrounding normal tissue) of the region of interest (such as the area corresponding to the dense B-line area or suspected pneumothorax area).

[0097] 6. Multi-parameter fusion decision engine: Receive quantitative features (B-line number, sliding index, strain value) from the above units and optional basic physiological parameters (such as respiratory rate, heart rate); this engine uses machine learning models such as gradient boosting decision tree (GBDT) or support vector machine (SVM) to output two core diagnostic indicators:

[0098] 1. Pulmonary edema index (0-100): A continuous score that comprehensively reflects the severity of pulmonary edema.

[0099] 2. Pneumothorax risk probability (0-100%): Indicates the likelihood of pneumothorax.

[0100] 1. Human-computer interaction and alarm module: Dynamically display the above quantitative indices and probabilities on the display terminal (such as mobile APP), and set thresholds, automatically trigger audio-visual alarms when thresholds are exceeded.

[0101] This system diagram clearly outlines the architecture and innovative workflow of the entire intelligent diagnosis system:

[0102] 1. Clear four-layer architecture: The system is divided into data acquisition layer (A), feature quantification extraction layer (B), intelligent decision-making layer (C), and application presentation layer (D), reflecting the gradual abstraction and information extraction process from raw data to clinical decision-making, with very clear logical levels.

[0103] 2. Highlight multi-modal data fusion:

[0104] 1. The left side of the diagram clearly shows two ultrasound modalities (B-mode, elastography) and one physiological signal as input, emphasizing the diversity of information sources.

[0105] 2. In the feature layer (B), four feature extraction units work in parallel, and finally all the quantized features (structured data) are collected into the decision engine (C), which intuitively shows the core innovative path from multi-modal raw data to fusion decision.

[0106] 3. Embedding of algorithms and models is clear:

[0107] 1. In the feature layer (B), the key technologies used are clearly labeled, such as lightweight CNN models and M-mode / image sequence-based analysis methods, reflecting the depth of technology.

[0108] 2. In the decision layer (C), the core multi-parameter fusion decision engine is highlighted, and possible models (GBDT, SVM) are exemplified, showing the intelligent mapping from features to diagnostic results.

[0109] 3. Closed-loop workflow: data flow runs from left to right throughout the system, finally outputting two key quantitative results (pulmonary edema index, pneumothorax risk probability), and pointing to two terminals: a display interface for providing decision support for clinicians and an alarm module for realizing automatic monitoring; this forms a complete "perception-analysis-decision-action" closed loop.

[0110] 4. Visualization of innovative value: this figure most powerfully shows that the present embodiment is no longer just an image viewer, but an intelligent system that can output direct diagnostic conclusions; it encapsulates complex imaging knowledge in algorithms and delivers to users not difficult-to-interpret images, but intuitive indexes and risk probabilities, which is the embodiment of its core creativity and practicality.

[0111] In specific implementation, see Figure 3 , Figure 3 The flow chart of the B-line intelligent recognition and quantification process in the lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound of the present application is as follows: Figure 3As shown, taking B-mode ultrasound images as the input source, first, image preprocessing is performed, through adaptive contrast enhancement (CLAHE algorithm is adopted to dynamically adjust the local contrast) and global standardization processing (pixel values are normalized to the range of [0, 1] and the device difference is eliminated), so that the image remains consistent in different acquisition conditions; subsequently, based on the pleural line position, the area below the pleural line is preset as the region of interest (ROI), focusing on the lung parenchyma-pleural junction zone (covering range > 95%) where B lines are prone to occur, effectively excluding interference sources such as blood vessels and artifacts; then, the preprocessed ROI image is input into the lightweight convolutional neural network MobileNetV3-Small, which realizes pixel-level segmentation of B lines through multi-scale convolution feature extraction (capturing the fine structure of B lines and fusing the region) and channel attention mechanism (dynamically focusing on key feature channels), generating a high-precision B line region mask; the mask is subjected to morphological post-processing, small-scale noise points (such as isolated pixels caused by motion artifacts) are removed by median filtering and morphological opening operation, and broken B lines are connected through connected region analysis and edge growth algorithm (repairing B line breakpoints caused by respiratory motion), ensuring that the mask is continuous and complete in space; in the B line detection and classification stage, the system identifies the B line region based on the mask and performs classification (distinguishing between isolated B lines and fused B lines), realizing accurate positioning of B lines; then, the B line region is labeled and measured, and the geometric features (such as length, area) and pixel-level attributes of each B line are extracted; in the quantitative index calculation link, the system accurately counts the absolute number of B lines in the unit field of view (B line count), calculates the average brightness value of the B line region (B line feature statistics, unit: gray value), and obtains the area ratio of the fusion region through connected region analysis (B line fusion degree calculation, reflecting the connectivity and overlap degree of B lines); finally, the structured B line quantitative indicators (including B line number, average brightness and fusion region ratio) are output, providing core input for pulmonary edema quantitative evaluation, and the B line recognition error is controlled to be <3%, significantly improving the accuracy of pulmonary edema and pneumothorax differentiation.

[0112] In specific implementation, referring to Figure 4 , Figure 4 The dynamic analysis schematic diagram of the lung sliding sign in the lung edema and pneumothorax differentiation and quantitative evaluation device based on multi-modal flexible ultrasound of the application is shown as Figure 4 Figure 4 ​The dynamic analysis schematic diagram of lung sliding sign is displayed in detail. B-mode ultrasound image sequence is taken as data input source. Through intelligent analysis mode selection mechanism (M-mode analysis or optical flow analysis), the motion characteristics of pleural line are accurately quantified: in the M-mode analysis mode, the system first generates a time-depth graph (M-mode image), automatically detects and tracks the pleural line boundary (based on image gradient operator or lightweight CNN model, such as MobileNetV3-Small), and captures the dynamic trajectory of the pleural line in the respiratory cycle in the continuous frame sequence (frames Ft-1, Ft, Ft+1); for the M-mode image, the system intelligently identifies the pleural line waveform characteristics: normal lung presents "beach sign" (irregular and broken waveform, reflecting good continuity of pleural sliding), and pneumothorax patients present "barcode sign" (regular and continuous stripe waveform, reflecting complete disappearance of pleural sliding); based on waveform analysis, the amplitude (maximum displacement difference of pleural line in respiratory cycle, unit: mm) and periodicity (respiratory frequency, unit: times / minute) of the motion signal are accurately calculated, providing basic parameters for sliding index calculation. In the optical flow analysis mode, the system calculates the motion vector field of the pleural line between continuous frames based on B sequence (TV-L1 optical flow algorithm is adopted), estimates the motion velocity (unit: mm / s) through the vector size and analyzes the motion direction (radial or tangential), and realizes millisecond-level motion parameter capture. Finally, the system weights and fuses the motion velocity and amplitude according to the preset weight (velocity weight 0.6, amplitude weight 0.4), and normalizes to the range of 0-1, outputs the dimensionless sliding index SI: normal lung SI is high (typical value >0.5, corresponding to "beach sign"), and pneumothorax patient SI is low / zero (typical value <0.1, corresponding to "barcode sign"). The index dynamically quantifies the continuity and stability of pleural sliding, provides key dynamic basis for real-time identification of pulmonary edema and pneumothorax, and effectively solves the misjudgment problem in the traditional method due to the similarity of pleural sliding sign.

[0113] In specific implementation, referring to Figure 5 , Figure 5 The working process schematic diagram of the multi-parameter fusion decision engine in the lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound of the application is shown as Figure 5 , Figure 5The workflow of the multi-parameter fusion decision engine is detailed. The multi-dimensional quantitative feature vector is used as the input source, covering B-line related features (number of B-lines in a unit field of view, average brightness, fusion area ratio), lung sliding sign features (pleural line sliding index SI), elastic parameter features (local strain value, strain ratio), and basic physiological parameters (respiratory rate, heart rate). First, feature preprocessing is performed. Missing values are processed by the mean filling algorithm (such as using historical mean to replace the missing respiratory rate) and Z-score normalization is used to eliminate dimension differences, ensuring that the features are analyzed on a unified scale. Then, feature selection and dimension reduction are performed. Based on the feature importance ranking of the GBDT model (such as B-line fusion ratio weight 0.32, sliding index weight 0.28), redundant features (such as strain ratio unrelated to diagnosis) are removed, and the five core dimensions most relevant to the differentiation of pulmonary edema and pneumothorax are retained. The decision engine uses gradient boosting decision tree (GBDT) as the core model (SVM as auxiliary verification). This model is trained by integrating 50,000+ clinical lung ultrasound data (including pulmonary edema, pneumothorax, and normal control group), deeply mining the nonlinear pathological correlation between multiple parameters (such as the interaction between B-line fusion ratio and strain value). In the reasoning stage, the engine simultaneously runs the pulmonary edema discrimination branch and the pneumothorax discrimination branch: the pulmonary edema discrimination branch outputs a continuous quantitative index on a 0-100 scale (index > 50 indicates moderate to severe pulmonary edema, such as index 72 corresponding to moderate interstitial edema), and the pneumothorax discrimination branch outputs a dynamic risk probability of 0-100% (probability > 30% indicates high risk, such as probability 38% indicating high likelihood of pneumothorax). The output includes structured index / probability and decision basis explanation. For example, when high B-line number (> 15 lines / field of view), high fusion (fusion ratio > 60%), and sliding index SI > 0.5 are detected, the system determines pulmonary edema. When low / zero sliding index (SI < 0.1), abnormal strain (strain value > 15%), and B-line number < 5 lines / field of view are detected, the system determines pneumothorax. The dynamically generated decision logic diagram intuitively shows the mapping relationship between pathological features and diagnosis results (such as the weight distribution diagram of the pulmonary edema feature combination), ensuring that clinicians can quickly understand the diagnosis basis and achieve a 92.3% differentiation accuracy and <200ms real-time response in emergency scenarios, completely solving the misdiagnosis problem caused by single feature in the prior art.

[0114] In a specific implementation, referring to Figure 6 , Figure 6 The human-computer interaction interface diagram of the pulmonary edema and pneumothorax differentiation and quantitative evaluation device based on multi-modal flexible ultrasound of the present application is shown in Figure 6 , Figure 6The intelligent diagnosis interface for continuous lung monitoring is demonstrated in detail. The interface adopts a highly integrated multi-modal visualization design to provide real-time and intuitive assessment support for lung edema and pneumothorax for clinicians. The patient information area clearly presents the patient's name (e.g., "Zhang San"), ID (e.g., "001"), and real-time monitoring time (e.g., "14:30"), ensuring that the patient's identity and monitoring status are immediately apparent. The real-time image area dynamically synchronously displays B-mode ultrasound images and elastography images (supporting dual-screen split-screen or intelligent switching), real-time capturing lung morphological and elastic characteristic changes, and assisting doctors in accurately positioning the examination area. The quantitative indicator dashboard innovatively integrates four core parameters: the lung edema index is displayed in a pointer-type instrument panel (0-100 point system, green 0-50, yellow 50-75, and red >75) to dynamically display the degree of interstitial edema, the pneumothorax risk probability is visually presented by a percentage progress bar (0-100%) with a gradient color (green 0-30%, yellow 30-70%, and red >70%) to indicate the risk level, and the sliding index is displayed as a numerical value (e.g., "0.65") combined with a column chart (highly reflecting sliding continuity) to quantify pleural motion, and the B-line density is displayed as a numerical value (e.g., "12 lines / field") superimposed with a trend arrow (↑ / ↓ / →) to indicate the B-line trend in real time. The trend analysis area provides dynamic curve graphs of the lung edema index and the pneumothorax risk probability at multiple time scales (supporting 1-hour, 6-hour, and 24-hour rolling analysis), visually displaying the disease evolution trajectory through smooth curves to assist doctors in identifying early warning signals. The alarm and status area displays the current state ("normal," "observation," or "alarm") in real time, and the alarm log list records event details (e.g., "14:25 - lung edema index 62, high risk"), event type, and level (e.g., "urgent") in chronological order, and is equipped with one-key mute and confirmation buttons to achieve rapid response, ensuring that the end-to-end interaction delay is <500 ms in emergency scenarios, the misdiagnosis rate is controlled at <5%, and objective and operable real-time support is provided for clinical decision-making, significantly improving the efficiency of critical patient treatment.

[0115] It should be noted that the multi-modal data synchronous acquisition and preprocessing process is as follows:

[0116] 1. Synchronously acquire B-mode images and elastic strain images of a specific lung region.

[0117] 2. Standardize and preprocess the images (e.g., contrast enhancement, noise reduction), and define a unified region of interest for subsequent analysis.

[0118] The process of parallel pathological feature quantification is as follows:

[0119] 3. S2a: Input the B-mode image into the pre-trained B-line recognition model to output structured B-line quantitative indicators.

[0120] 4. S2b: Analyze the image sequence of the same region to calculate the motion information of the pleural line and output the sliding index.

[0121] 5. S2c: Extract the average strain value of the region of interest from the elastic strain map.

[0122] The process of multi-parameter fusion and intelligent decision-making is as follows:

[0123] 6. All the extracted quantitative features (B-line number, sliding index, strain value) are combined into a feature vector.

[0124] 7. The feature vector is input into the pre-trained multi-parameter fusion decision-making model.

[0125] 8. The decision-making model outputs the final pulmonary edema index and pneumothorax risk probability.

[0126] The process of result visualization and alarm is as follows:

[0127] 9. The quantitative diagnostic results are presented to the user in the form of dashboards, trend charts, etc.

[0128] 10. If the pulmonary edema index exceeds the preset threshold (e.g., > 50), or the pneumothorax risk probability exceeds the preset threshold (e.g., > 30%), an alarm is triggered immediately.

[0129] The model training implementation process is as follows:

[0130] 1. Data preparation: Collect a large amount of lung ultrasound video and image data annotated by clinical experts, including B-line position and number, presence of lung sliding sign, pneumothorax diagnosis results, elastography data, etc. Build a high-quality database.

[0131] 2. Model training:

[0132] 1. B-line identification model: Use a lightweight network such as MobileNetV3 as the backbone, and perform transfer learning and training on image classification and target detection datasets.

[0133] 2. Decision engine model: Use the collected quantitative feature vectors and corresponding gold standard diagnosis results (such as "severe pulmonary edema" and "confirmed pneumothorax") as training data to train a GBDT classifier, so that it can learn the complex mapping relationship from the feature vector to the final diagnosis.

[0134] 3. System deployment implementation:

[0135] Deploy the trained lightweight model on edge computing devices, such as intelligent terminals (mobile phones / tablets) connected to flexible sensors or embedded processors (such as JetsonNano).

[0136] The software flow is as follows:

[0137] 1. Terminal APP receives a frame of B-mode image and a frame of elastogram.

[0138] 2. Call local AI inference engine, execute S2a, S2b, S2c steps in parallel.

[0139] 3. Send the extracted feature values to the decision engine (S3) to get pulmonary edema index and pneumothorax risk probability.

[0140] 4. Update the result display on the UI and determine whether an alarm is needed (S4).

[0141] Example effects:

[0142] In a retrospective validation of 200 patients, the diagnostic performance of the system of the present embodiment is as follows:

[0143] 5. For the detection of moderate to severe pulmonary edema, the sensitivity reaches 96%, and the specificity reaches 92%.

[0144] 6. For the detection of pneumothorax, the sensitivity reaches 94%, and the specificity reaches 89%.

[0145] 7. Compared with the subjective judgment of three experienced ultrasound doctors, the consistency (Kappa value) of the system is more than 0.85, proving that the quantitative indicators of the output have high clinical reliability.

[0146] Those skilled in the art can understand that all or part of the steps in the above-mentioned implementation methods can be completed by programs instructing related hardware, the programs are stored in a storage medium, and include a plurality of instructions for causing a device (which can be a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application; and the aforementioned storage medium is a computer-readable storage medium, including: a U disk, a mobile hard disk, a read-only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0147] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article, or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article, or device including the element.

[0148] The above-mentioned serial numbers of the embodiments of the application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0149] The above are only preferred embodiments of the application, and do not limit the patent scope of the application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, which is made by using the content of the specification and drawings of the application, is also included in the patent protection scope of the application.

Claims

1. A device for differential and quantitative assessment of pulmonary edema and pneumothorax based on multi-modal flexible ultrasound, characterized in that, The lung edema and pneumothorax identification and quantitative evaluation device based on multi-modal flexible ultrasound comprises: A data extraction module is configured to synchronously acquire morphological ultrasound images and elastomechanical images of the same anatomical position, and extract a B-line quantitative index, a pleural line sliding index, and an average strain value of a region of interest from the morphological ultrasound images and the elastomechanical images; An identification and quantitative evaluation module is configured to combine the B-line quantitative index, the pleural line sliding index, and the average strain value into a feature vector, input the feature vector into a multi-parameter fusion decision model, and obtain a lung edema index and a pneumothorax risk probability; A display and warning module is configured to display the lung edema index and the pneumothorax risk probability on a display terminal, and trigger an alarm when the lung edema index exceeds a preset index threshold or the pneumothorax risk probability exceeds a preset probability threshold; The data extraction module is further configured to synchronously acquire real-time morphological ultrasound images and elastomechanical images of the same anatomical position by using a flexible ultrasound sensor; The data extraction module is further configured to automatically identify B-lines in the morphological ultrasound images based on a convolutional neural network, and output a B-line quantitative index; The data extraction module is further configured to analyze the motion speed and amplitude of the pleural line in the morphological ultrasound images, and generate a pleural line sliding index according to the motion speed and amplitude; The data extraction module is further configured to analyze the elastomechanical images, and calculate an average strain value of a region of interest.

2. The multi-modal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus of claim 1, wherein, The data extraction module is further configured to perform pixel-level segmentation on the preprocessed morphological ultrasound images by using a convolutional neural network, and obtain a B-line pixel-level segmentation mask; The data extraction module is further configured to calculate the absolute number of B-lines in a B-line region, the average brightness value of the B-line region, and the B-line fusion region proportion according to the B-line pixel-level segmentation mask, and use the absolute number, the average brightness value, and the B-line fusion region proportion as the B-line quantitative index.

3. The multi-modal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus of claim 2, wherein, The data extraction module is further configured to perform contrast enhancement and standardization processing on the morphological ultrasound images, and obtain preprocessed morphological ultrasound images; The data extraction module is further configured to perform noise reduction and connection breakpoint processing on the B-line pixel-level segmentation mask, and locate a B-line region by using multi-scale convolution feature extraction and an attention mechanism.

4. The multi-modal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus of claim 1, wherein, The data extraction module is further configured to perform time-series dynamic analysis on the morphological ultrasound images, obtain a pleural line boundary in each frame of image, and generate a spatial coordinate sequence according to the pleural line boundary; The data extraction module is further configured to calculate the motion speed and amplitude of the pleural line in the respiratory motion according to the spatial coordinate sequence, and perform weighted fusion of the motion speed and the motion amplitude according to motion periodicity and motion direction according to a preset fusion weight, to obtain a dimensionless pleural line sliding index.

5. The multi-modal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus as claimed in claim 4, wherein, The data extraction module is further configured to input a sequence of continuous B-mode images in the morphological ultrasound images as data, and automatically detect and track the pleural line of the sequence of continuous B-mode images based on image gradients or machine learning. The data extraction module is further configured to generate an M-mode image of the time-depth graph when the analysis model is selected as M-mode analysis, analyze a beach sign or a barcode sign of the pleural line in the M mode according to the M-mode image, and calculate a motion amplitude and a motion periodicity of the motion signal according to the beach sign or the barcode sign. The data extraction module is further configured to calculate a motion vector of the pleural line in consecutive frames when the analysis model is selected as optical flow analysis, and estimate a motion speed and a motion direction according to the motion vector.

6. The multimodal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus of claim 1, wherein, The data extraction module is further configured to automatically segment a region of interest from the elastomechanical image according to a preset lung anatomical feature. The data extraction module is further configured to perform displacement estimation on adjacent frames in the elastomechanical image by using a cross-correlation algorithm, and obtain a local strain value of each pixel point in the region of interest. The data extraction module is further configured to generate a strain distribution graph according to the local strain value, and perform spatial weighted averaging on strain values of all effective pixels in the strain distribution graph to obtain an average strain value.

7. The multimodal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus of claim 1, wherein, The differential quantitative evaluation module is further configured to combine the B-line quantitative index, the pleural line sliding index, and the average strain value into a feature vector according to a preset dimension order. The differential quantitative evaluation module is further configured to input the feature vector into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree to obtain a pulmonary edema index and a pneumothorax risk probability.

8. The multi-modal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus as claimed in claim 7, wherein, The differential quantitative evaluation module is further configured to collect a heart rate and a respiration rate of a to-be-detected object by using a biological signal sensor, and take the heart rate and the respiration rate as basic physiological parameters. The differential quantitative evaluation module is further configured to take the basic physiological parameters as additional input features, input the additional input features and the feature vector into a multi-parameter fusion decision model constructed based on a gradient boosting decision tree, and establish a nonlinear pathological correlation among the B-line quantitative index, the pleural line sliding index, and the average strain value. The differential quantitative evaluation module is further configured to perform dynamic risk calibration according to the nonlinear pathological correlation, and output a pulmonary edema index and a pneumothorax risk probability.

9. The multimodal flexible ultrasound based lung edema versus pneumothorax differentiation and quantification assessment apparatus of claim 1, wherein, The display warning module is further configured to display the pulmonary edema index in real time on a display terminal by using a pointer-type instrument panel, and display the pneumothorax risk probability in real time on the display terminal by using a percentage progress bar. The display warning module is further configured to compare the pulmonary edema index with a preset index threshold value, and compare the pneumothorax risk probability with a preset probability threshold value. The display warning module is further configured to trigger an alarm when the pulmonary edema index exceeds the preset index threshold value or the pneumothorax risk probability exceeds the preset probability threshold value.

Images