An ultrasonic heart dynamic monitoring system and method

Abstract

This application provides an ultrasound cardiac dynamic monitoring system and method, relating to the field of ultrasound cardiac function dynamic monitoring technology. The system includes: a patch-type ultrasound probe, a hydrogel patch, a wearable hardware module, and a cardiac function analysis module. The cardiac function analysis module is configured to: receive the ultrasound image; determine a target sampling line in the ultrasound image based on a reinforcement learning model; control the wearable hardware module to acquire M-mode images along the target sampling line; input the M-mode images into a deep learning model to output the key point positions of the left ventricular septum and the left ventricular posterior wall at end-diastole and end-systole based on the deep learning model; calculate the left ventricular ejection fraction (LVEF) based on the LVEF and the key point positions; the LVEF represents the dynamic monitoring and detection results of cardiac function. This system can improve the effectiveness and accuracy of dynamic cardiac monitoring.

Description

Technical Field

[0001] This application relates to the field of ultrasound dynamic cardiac function monitoring technology, specifically to an ultrasound dynamic cardiac monitoring system and method. Background Technology

[0002] The management of cardiovascular disease relies on continuous and accurate assessment of cardiac function, but current technologies struggle to achieve long-term, continuous, and reliable dynamic monitoring under normal living conditions. While traditional echocardiography provides rich structural and functional information, its bulky size and highly specialized operation limit its use for continuous monitoring during daily life or exercise. Current mainstream wearable health devices (such as watches based on photoplethysmography) primarily collect surface physiological signals and cannot directly image and quantitatively analyze the deep structures and pumping function of the heart, resulting in significant limitations in monitoring accuracy.

[0003] Although wearable ultrasound technology has made some progress, enabling continuous imaging, current solutions mostly focus on providing anatomical images using B-mode. This type of imaging mode has insufficient temporal resolution, making it difficult to clearly capture the instantaneous high-speed motion details of structures such as the ventricular walls and valves. Therefore, it still has significant shortcomings in dynamically and accurately quantifying cardiac systolic and diastolic function, resulting in low effectiveness and accuracy of dynamic cardiac monitoring. Summary of the Invention

[0004] Based on this, this application provides an ultrasound cardiac dynamic monitoring system and method, which can improve the effectiveness and accuracy of cardiac dynamic monitoring and realize continuous, objective, and quantitative monitoring of cardiac contractile function.

[0005] In a first aspect, this application provides an ultrasound cardiac dynamic monitoring system, comprising:

[0006] A patch-type ultrasound probe; the patch-type ultrasound probe is used to emit ultrasound waves and receive echo signals from the human body. Hydrogel patch; the hydrogel patch is attached to the radiating surface of the patch-type ultrasound probe to fill the air gap between the radiating surface of the patch-type ultrasound probe and human skin; Wearable hardware module; the wearable hardware module is connected to the patch-type ultrasound probe and is used to generate an ultrasound image based on the echo signal; wherein the ultrasound image is a B-mode ultrasound image of the long axis of the left ventricle of the heart; Cardiac function analysis module; the cardiac function analysis module is connected to the wearable hardware module; the cardiac function analysis module is configured as follows: The system receives the ultrasound image and determines a target sampling line in the ultrasound image based on a reinforcement learning model. It then controls the wearable hardware module to acquire M-mode images along the target sampling line. The M-mode images are input into a deep learning model to output the key point positions of the left ventricular septum and the left ventricular posterior wall at end-diastole and end-systole based on the deep learning model. The left ventricular ejection fraction is calculated based on the left ventricular septum and the key point positions. The left ventricular ejection fraction represents the dynamic monitoring and detection results of cardiac function.

[0007] Optionally, the wearable hardware module includes an analog switch unit, an analog front-end unit, a main control unit, a transmission unit, and a power management unit; The analog switch unit is connected to the patch-type ultrasound probe and is used to switch between multiple ultrasound transducer channels; the analog front-end unit is connected to the analog switch unit and is used to generate an ultrasonic wave transmission signal to excite the patch-type ultrasound probe, and to amplify and perform analog-to-digital conversion on the echo signal received by the patch-type ultrasound probe; the main control unit is connected to the analog front-end unit and is used to control the timing of ultrasonic wave transmission and reception, and to generate the ultrasound image based on the echo signal; the transmission unit is connected to the main control unit and is used to wirelessly transmit the ultrasound image to the cardiac function analysis module; the power management unit is connected to the analog switch unit, the analog front-end unit, the main control unit, and the transmission unit respectively, and is used to provide a stable operating voltage.

[0008] Optionally, the cardiac function analysis module is specifically used for: In the end-diastolic and end-systolic key point prediction heatmaps, the smallest bounding rectangle of each predicted key point region is found, and the coordinates of the center point of each rectangle are obtained. In the end-diastolic key point prediction heatmap, the two center points with the closest horizontal coordinates are paired together to form a diastolic group, and in the end-systolic key point prediction heatmap, the two center points with the closest horizontal coordinates are paired together to form a systolic group. Each diastolic group is paired with the systolic group with the closest horizontal distance behind it to construct a cardiac cycle. For each paired cardiac cycle, the left ventricular ejection fraction is calculated based on the left ventricular interventricular septum and the key point position.

[0009] Optionally, the cardiac function analysis module is specifically used for: The left ventricular end-diastolic diameter is calculated based on the key points of the left ventricular interventricular septum and the left ventricular posterior wall at end-diastole, and the left ventricular end-systolic diameter is calculated based on the key points of the left ventricular interventricular septum and the left ventricular posterior wall at end-systole; the left ventricular ejection fraction is calculated based on the left ventricular end-diastolic diameter and the left ventricular end-systolic diameter.

[0010] Optionally, the formula for calculating the left ventricular ejection fraction includes:

[0011] Where EF is the left ventricular ejection fraction, N is the total number of effective cardiac cycles identified within the preset monitoring period, and i represents the i-th cardiac cycle. Left ventricular end-diastolic diameter The angle between the vertical direction and the vertical direction. Left ventricular end-diastolic diameter The angle between the vertical direction and the vertical direction.

[0012] Optionally, the patch-type ultrasonic probe includes multiple ultrasonic transducer elements arranged in a linear array.

[0013] Alternatively, the hydrogel patch is manufactured using a biocompatible silicone integral molding process.

[0014] Secondly, this application provides a method for dynamic cardiac ultrasound monitoring, which can be applied to the aforementioned dynamic cardiac ultrasound monitoring system. The method includes: The system receives the ultrasound image and determines a target sampling line in the ultrasound image based on a reinforcement learning model. It then controls the wearable hardware module of the ultrasound cardiac dynamic monitoring system to acquire M-mode images along the target sampling line. The M-mode images are input into a deep learning model to output the key point positions of the left ventricular septum and the left ventricular posterior wall at end-diastole and end-systole based on the deep learning model. The left ventricular ejection fraction is calculated based on the left ventricular septum and the key point positions. The left ventricular ejection fraction represents the dynamic monitoring and detection results of cardiac function.

[0015] Optionally, the method for dynamic cardiac ultrasound monitoring further includes: acquiring a training set of historical B-mode ultrasound images containing long-axis sections of the left ventricle, wherein each historical B-mode ultrasound image is marked with a gold standard sampling line position; constructing a reward function based on the angle between the sampling line position and the gold standard sampling line position of the agent in the initial reinforcement learning model, and the scanning angle corresponding to the historical B-mode ultrasound image; repeatedly performing actions on the training set by the agent and obtaining feedback according to the reward function until the initial reinforcement learning model automatically outputs a target sampling line position that matches the gold standard sampling line position based on the input B-mode ultrasound image, thereby obtaining the reinforcement learning model.

[0016] Optionally, the ultrasound cardiac dynamic monitoring method further includes: receiving M-mode images labeled with key points of the interventricular septum at end-diastole, key points of the left ventricular posterior wall at end-diastole, key points of the interventricular septum at end-systole, and key points of the left ventricular posterior wall at end-systole as training data; generating a Gaussian distribution heatmap centered on the key point locations as training labels; and training an initial deep learning model by minimizing the loss between the model prediction heatmap and the training labels to obtain the deep learning model.

[0017] Compared with existing technologies, the beneficial effects of this application are as follows: by integrating a patch-type ultrasound probe, a wearable hardware module, and an intelligent analysis module, a monitoring scheme that can operate continuously in a natural state is formed. The ultrasound cardiac dynamic monitoring system acquires B-mode ultrasound images of the left ventricle's long axis based on wearable hardware. A reinforcement learning model automatically locates the optimal sampling line in this image and guides the system to switch to M-mode imaging to obtain high temporal resolution data required for observing high-speed motion of the ventricular wall. Subsequently, a deep learning model analyzes the M-mode images to automatically identify the positions of the left ventricular septum and the left ventricular posterior wall at key time phases, and calculates the left ventricular ejection fraction, a core clinical indicator. This technical approach combines the portability of wearable ultrasound, the ability of M-mode imaging to capture motion details, and the automated analysis advantages of artificial intelligence algorithms. It can improve the effectiveness and accuracy of dynamic cardiac monitoring, achieving continuous, objective, and quantitative monitoring of cardiac systolic function, and providing a practical solution for cardiac function assessment in scenarios where traditional ultrasound equipment is difficult to cover. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the architecture of the ultrasound cardiac dynamic monitoring system provided in the embodiments of this application.

[0019] Figure 2 This is a schematic diagram of a patch-type ultrasonic probe provided in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of a wearable hardware module provided in an embodiment of this application.

[0021] Figure 4 This is a schematic diagram of the workflow of the ultrasound cardiac dynamic monitoring system provided in the embodiments of this application.

[0022] Figure 5 This is a schematic diagram illustrating the sampling line setting training process of the reinforcement learning model provided in the embodiments of this application.

[0023] Figure 6 This is a schematic diagram of the M-mode image provided in an embodiment of this application.

[0024] Figure 7This is a schematic diagram of M-mode left ventricular key point detection provided in an embodiment of this application.

[0025] Figure 8 Key prediction heatmaps for end-diastolic and end-systolic phases provided in the embodiments of this application.

[0026] Figure 9 This is a schematic diagram showing the result of key point grouping and pairing provided in the embodiments of this application.

[0027] Figure labels: 10-Ultrasound cardiac dynamic monitoring system; 11-Patch ultrasound probe; 12-Hydrogel patch; 13-Wearable hardware module; 14-Cardiac function analysis module. Detailed Implementation

[0028] The present application will now be described in further detail with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the subject matter of the present application to the following embodiments. All technologies implemented based on the content of the present application fall within the scope of protection of the present application.

[0029] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0030] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the architecture of the ultrasound cardiac dynamic monitoring system provided in the embodiments of this application. The ultrasound cardiac dynamic monitoring system 10 may include a patch-type ultrasound probe 11, a hydrogel patch 12, a wearable hardware module 13, and a cardiac function analysis module 14.

[0031] Please refer to Figure 2 , Figure 2This is a schematic diagram of a patch-type ultrasonic probe provided in an embodiment of this application. The patch-type ultrasonic probe 11 has multiple ultrasonic transducer elements arranged in a linear array for emitting ultrasonic waves and receiving echo signals from the human body. It has a concave curvature that conforms to the curve of the human chest wall to ensure stable acoustic contact. The patch-type ultrasonic probe 11 can be made of piezoelectric composite material, with a probe crystal diameter of less than 13 mm, a probe weight of less than 40 g, and can operate at 1 MHz to 12.5 MHz, with center frequencies of 1.8 MHz and 7.1 MHz. The deviation between the acoustic operating frequency and the fundamental center frequency is within ±15%. The patch-type ultrasonic probe 11 internally contains an ultrasonic transducer based on a piezoelectric crystal, with 128 elements, an element spacing of 0.3 mm, a lateral aperture of 1.4 mm, and a focusing depth of 30 mm. The hydrogel patch 12 is attached to the radiating surface of the patch-type ultrasound probe 11 to fill the air gap between the radiating surface of the patch-type ultrasound probe 11 and the human skin. The hydrogel patch 12 is made of biocompatible silicone through a one-piece molding process, which can fit well with the probe and human tissue, has good sound transmission and has no obvious inflammatory effect on the human body.

[0032] Please refer to Figure 3 , Figure 3 This is a schematic diagram of the wearable hardware module provided in an embodiment of this application. The wearable hardware module is connected to the patch-type ultrasound probe 11 and supports B-mode (Brightness Mode) and M-mode (Motion Mode) imaging. The wearable hardware module is a miniaturized, low-power electronic device integrating ultrasound transmitting / receiving circuits, a main controller, and a wireless communication unit. It is used to drive the probe to work and perform signal processing and image reconstruction based on the received echo signals to generate a B-mode ultrasound image of the long axis of the left ventricle of the heart. The B-mode image is a two-dimensional anatomical cross-sectional view that uses brightness to represent tissue structure and is used to display the real-time spatial morphology of the heart.

[0033] The wearable hardware module 13 is a 32-channel integrated ultrasound device. Its core lies in the use of an integrated analog front-end chip that combines transceiver capabilities. Utilizing a high-density 12-layer PCB layout, the overall circuit board area is reduced to approximately the size of a mobile phone, thus enabling wearable functionality. Structurally, this module mainly includes an analog switch unit, an analog front-end unit, a main control unit, a wireless transmission unit, and a power management unit. The analog switch unit is connected to the patch probe and is responsible for switching between multiple ultrasound transducer channels. The analog front-end unit is connected to the analog switch unit and is used to generate the ultrasonic transmission signal that excites the probe, and to amplify and convert the echo signal received by the probe. The main control unit is further connected to the analog front-end unit and is responsible for controlling the timing of ultrasonic transmission and reception, and reconstructing ultrasound images based on the processed echo signals. The wireless transmission unit is connected to the main control unit and is used to wirelessly transmit image data to the cardiac function analysis module 14. The power management unit provides a stable operating voltage for all the above units.

[0034] To achieve the low power consumption and long battery life required for wearable devices, this module adopts a dynamic power management strategy that balances working and sleeping times at the system level. By optimizing the overall power consumption of the analog front-end, main control unit (such as FPGA), and wireless communication module, the peak power consumption of the system is controlled to within 13W, which is significantly lower than the power consumption level of hundreds of watts in traditional ultrasound instruments, thereby effectively extending the continuous working time of the built-in lithium battery.

[0035] The cardiac function analysis module 14 is connected to the wearable hardware module 13 and can typically be deployed on a mobile terminal or server. It is configured to execute an automated analysis process. Specifically, the cardiac function analysis module 14 can be configured as follows: The system receives the ultrasound image and determines a target sampling line in the ultrasound image based on a reinforcement learning model; it controls the wearable hardware module 13 to acquire M-mode images along the target sampling line; it inputs the M-mode images into a deep learning model to output the key point positions of the left ventricular septum and the left ventricular posterior wall at end-diastole and end-systole based on the deep learning model; and it calculates the left ventricular ejection fraction based on the left ventricular septum and the key point positions.

[0036] The cardiac function analysis module 14 may include three sub-modules: a reinforcement learning-based sampling line position setting module, a deep learning-based ventricular wall key point monitoring module, and a cardiac function calculation and analysis module. The reinforcement learning-based sampling line position setting module is an intelligent decision-making system that treats the ultrasound B-mode image as the environment and the adjustment of the sampling line position as an action. This module evaluates the effect of each action through a preset reward function and, through iterative training on a large amount of labeled historical image data, enables the model to automatically locate the optimal path, i.e., the target sampling line, in the long-axis B-mode image of the left ventricle for subsequent M-mode imaging. Its core function is to replace manual operation, achieving accurate, automatic, and reproducible positioning of the sampling line.

[0037] The deep learning-based cardiac ventricular wall key point monitoring module employs a deep neural network architecture. This module receives M-mode images acquired along the aforementioned target sampling line as input. During its training phase, the end-diastolic and end-systolic key point locations, annotated by professional physicians and represented in Gaussian heatmap form, serve as the learning target. The trained model can infer from the input M-mode images and output corresponding predicted heatmaps, thereby accurately identifying the pixel-level coordinates of the left ventricular septum and left ventricular posterior wall at end-diastole and end-systole. Its core function is to achieve automated, high-precision tracking and localization of high-speed cardiac ventricular wall motion.

[0038] The cardiac function calculation and analysis module refers to an algorithm system that performs quantitative calculations based on the aforementioned key point locations. This module first performs post-processing on the predicted key points, such as noise reduction, grouping, and cycle pairing. Then, based on the paired key points, it calculates the left ventricular end-diastolic diameter (LVDd) and left ventricular end-systolic diameter (LVDs) for each cardiac cycle. Finally, based on the diameter data, it calculates and outputs the left ventricular ejection fraction (LVEF). The function of this module is to transform image data into quantitative functional parameters with clear clinical significance, enabling continuous and dynamic monitoring of cardiac function. In clinical practice, left ventricular ejection fraction is a standard with clear diagnostic and grading significance, which can be directly used to assess the heart's pumping function and directly characterize the dynamic monitoring results of cardiac systolic function.

[0039] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the workflow of the ultrasound cardiac dynamic monitoring system provided in the embodiments of this application.

[0040] After the hydrogel patch 12 is applied to the patch-type ultrasound probe 11, it is connected to the wearable hardware module 13 via a flexible connecting cable. This assembly is placed next to the subject's ribs to accurately scan the long axis section of the left ventricle. The wearable hardware module 13 first acquires a B-mode ultrasound image of the long axis of the left ventricle and transmits it via WiFi to the cardiac function analysis module 14 deployed on a terminal smart device.

[0041] In the cardiac function analysis module 14, the reinforcement learning-based sampling line position setting module automatically optimizes the sampling line position through the following steps: First, the B-mode ultrasound image is considered as the environment, and the sampling line position is considered as an adjustable action. A reward function is set to provide feedback on the accuracy of cardiac function assessment based on the sampling line position. The higher the accuracy, the greater the reward. The algorithm continuously tries different sampling line positions, using reward feedback to adjust the strategy to gradually improve the accuracy of cardiac function assessment. Through multiple iterations, the reinforcement learning algorithm gradually converges to the optimal sampling line position, ensuring the accuracy of cardiac function measurement.

[0042] Overall, during the offline training phase, or learning phase, a large amount of historical echocardiogram image data with pre-labeled optimal sampling line positions is used to train the reinforcement learning model. The agent practices on these images, gradually learning the optimal sampling line positions corresponding to image features by analyzing actions and rewards, thereby optimizing its decision-making strategy. In the online application phase, or inference phase, when the trained model is deployed to a real-world cardiac function assessment system, for a new left ventricular long-axis B-mode image, the trained agent can automatically calculate and locate the optimal M-mode sampling line position in a very short time.

[0043] Please refer to Figure 5 , Figure 5 This diagram illustrates the sampling line setting training process for the reinforcement learning model provided in this embodiment. The reward function of the reinforcement learning model is:

[0044] Where R is the reward value, Angular deviation represents the angle between the position of the GT scan line (i.e., the gold standard scan line set by the ultrasound physician) and the scan line position attempted by the model. This refers to the scanning angle of the B-mode image.

[0045] Once the reinforcement learning model correctly sets the sampling line position, the cardiac function analysis module 14 will automatically scan the M-mode image at the sampling line position. Please refer to [link / reference needed]. Figure 6 , Figure 6 This is a schematic diagram of the M-mode image provided in an embodiment of this application.

[0046] The wearable hardware module 13 can acquire M-images every 5 seconds, which are then input into a pre-trained deep learning model for M-mode left ventricular key point detection. It outputs two key points p1 and p2 of the interventricular septum (IVS) and left ventricular posterior wall (PW) at end-diastole of the left ventricle, and two key points p3 and p4 at end-systole of the left ventricle in each cardiac cycle. Please refer to [link / reference]. Figure 7 , Figure 7 This is a schematic diagram of M-mode left ventricular key point detection provided in an embodiment of this application. Figure 7 The left side shows the location of the left ventricular key in the M-mode image, and the right side shows the detection results output by the deep learning model.

[0047] The deep learning model was trained using labeled data from physicians. Each M-mode image corresponds to two labeled images: Mask_p1, which shows the key points of the interventricular septum and left ventricular posterior wall at end-diastole, and Mask_p2, which shows the key points of the interventricular septum and left ventricular posterior wall at end-systole. Within each labeled image, key points are marked by a professional ultrasound physician at the designated locations (…). , Generate 2D Gaussian spherical keypoint annotations at the coordinates. G (x, y), its calculation formula is:

[0048] in,( , The location () represents the center of the Gaussian distribution, which is the key location for labeling. The standard deviation of the Gaussian distribution controls the width of the distribution; in this application, it is set to 3 pixel values. During training, the model's input is an M-mode grayscale image, and the output is two keypoint prediction images: Pred_p1, a prediction image of the interventricular septum keypoint and the left ventricular posterior wall keypoint at end-diastole, and Pred_p2, a prediction image of the interventricular septum keypoint and the left ventricular posterior wall keypoint at end-systole. The loss function L used for training the deep learning model is:

[0049] in, Mean squared error is used to measure the difference between the predicted value and the actual value.

[0050] Optionally, the cardiac function calculation and analysis module can be specifically used to: find the smallest bounding rectangle of each predicted key point region in the key point prediction heatmaps at end-diastole and end-systole respectively, and obtain the coordinates of the center point of each rectangle; in the key point prediction heatmap at end-diastole, pair the two center points with the closest horizontal coordinates into a group to form a diastolic group; in the key point prediction heatmap at end-systole, pair the two center points with the closest horizontal coordinates into a group to form a systolic group; pair each diastolic group with the systolic group with the closest horizontal distance behind it to construct a cardiac cycle; for each paired cardiac cycle, calculate the left ventricular ejection fraction based on the left ventricular interventricular septum and the key point position.

[0051] For example, after the deep learning model outputs two keypoint prediction images from the M-mode image in 5 seconds, the cardiac function calculation and analysis module performs denoising processing on the keypoint prediction images Pred_p1 and Pred_p2 to remove small isolated point noise. Then, the minimum bounding rectangle is searched for the keypoint prediction results in the Pred_p1 and Pred_p2 prediction images, and the center point coordinates of each minimum bounding rectangle are obtained.

[0052] Please refer to Figure 8 , Figure 8 Key prediction heatmaps for end-diastolic and end-systolic phases provided in the embodiments of this application. Figure 8 The left side shows the key points at end-diastole, and the right side shows the key points at end-systole. Key points at end-diastole and end-systole were grouped and paired, respectively, with the pairing principle being that the two key points closest in the horizontal X-axis form a group. Please refer to [link / reference]. Figure 9 , Figure 9 This diagram illustrates the results of key point grouping and pairing provided in the embodiments of this application. Each pair represents the location of a key point on the interventricular septum and the posterior wall of the left ventricle during the end-diastolic and end-systolic phases of a cardiac cycle. The center point of the smallest bounding rectangle connecting the two key points in the diastolic group is the left ventricular end-diastolic diameter (LVDd), and the center point of the smallest bounding rectangle connecting the two key points in the systolic group is the left ventricular end-systolic diameter (LVDs).

[0053] In some embodiments, the cardiac function analysis module 14 may calculate the left ventricular ejection fraction by: calculating the left ventricular end-diastolic diameter based on the key point positions of the left ventricular interventricular septum and the left ventricular posterior wall at end-diastole; calculating the left ventricular end-systolic diameter based on the key point positions of the left ventricular interventricular septum and the left ventricular posterior wall at end-systole; and calculating the left ventricular ejection fraction based on the left ventricular end-diastolic diameter and the left ventricular end-systolic diameter.

[0054] For example, assuming N key pairs of end-diastolic and end-systolic phases are detected in a 5-second M-image, the formula for calculating the left ventricular ejection fraction (EF) is:

[0055] Where EF is the left ventricular ejection fraction, N is the total number of effective cardiac cycles identified within the preset monitoring period, and i represents the i-th cardiac cycle. Left ventricular end-diastolic diameter The angle between the vertical direction and the vertical direction. Left ventricular end-diastolic diameter The angle between the vertical direction and the vertical direction.

[0056] The wearable hardware module 13 collects M-mode images every 5 seconds and sends them to the cardiac function analysis module 14. The left ventricular ejection fraction is continuously calculated through the above cardiac function analysis process, thus completing the real-time monitoring of left ventricular cardiac function based on M-mode images.

[0057] The ultrasound-guided dynamic cardiac monitoring system provided in this application integrates a patch-type ultrasound probe, a wearable hardware module, and an intelligent analysis module to form a monitoring solution that can operate continuously in a natural state. The system acquires B-mode ultrasound images of the left ventricle's long axis using wearable hardware. A reinforcement learning model automatically locates the optimal sampling line within these images and guides the system to switch to M-mode imaging to obtain high temporal resolution data required for observing high-speed motion of the ventricular wall. Subsequently, a deep learning model analyzes the M-mode images, automatically identifying the positions of the left ventricular septum and posterior wall at key time phases, and calculating the left ventricular ejection fraction (LVEF), a core clinical indicator. This approach combines the portability of wearable ultrasound, the ability of M-mode imaging to capture motion details, and the automated analysis advantages of artificial intelligence algorithms. It improves the effectiveness and accuracy of dynamic cardiac monitoring, enabling continuous, objective, and quantitative monitoring of cardiac systolic function, and provides a practical solution for assessing cardiac function in scenarios where traditional ultrasound equipment is difficult to cover.

[0058] It should be understood that when the various modules of the system provided in the above embodiments are working, the division of each functional module in the above description is only used as an example. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0059] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0060] Based on the same concept, this application also provides a method for dynamic echocardiographic monitoring, which can be applied to the aforementioned dynamic echocardiographic monitoring system 10. The method includes: The system receives the ultrasound image and determines a target sampling line in the ultrasound image based on a reinforcement learning model. It then controls the wearable hardware module of the ultrasound cardiac dynamic monitoring system to acquire M-mode images along the target sampling line. The M-mode images are input into a deep learning model to output the key point positions of the left ventricular septum and the left ventricular posterior wall at end-diastole and end-systole based on the deep learning model. The left ventricular ejection fraction is calculated based on the left ventricular septum and the key point positions. The left ventricular ejection fraction represents the dynamic monitoring and detection results of cardiac function.

[0061] In some embodiments, the method may further include a step of training a reinforcement learning model. This step of training the reinforcement learning model can be applied to a mobile terminal or server deployed with the cardiac function analysis module 14, specifically including: A training set of historical B-mode ultrasound images containing long-axis sections of the left ventricle of the heart is obtained, wherein the position of the gold standard sampling line is marked in each historical B-mode ultrasound image. A reward function is constructed based on the angle between the sampling line position and the gold standard sampling line position of the agent in the initial reinforcement learning model, and the scanning angle corresponding to the historical B-mode ultrasound image. The agent repeatedly performs actions on the training set and receives feedback according to the reward function until the initial reinforcement model automatically outputs a target sampling line position that matches the gold standard sampling line position based on the input B-mode ultrasound image, thus obtaining the reinforcement learning model.

[0062] The historical B-mode ultrasound image training set refers to a database consisting of multiple pre-acquired and professionally labeled long-axis ultrasound images of the left ventricle of the heart, from which the model learns patterns. In this embodiment, the gold standard sampling line position refers to the ideal M-mode sampling line, which is considered most suitable for assessing cardiac function and is manually marked by experienced ultrasound physicians in each historical image. This position serves as the target true value pursued by the model for learning.

[0063] At the start of training, the agent in the initial reinforcement learning model is placed in a virtual environment constructed from these historical images. In this embodiment, the agent refers to the algorithmic entity in the reinforcement learning model responsible for perceiving the environmental state, making decisions, and receiving feedback; that is, the decision-making core of the model. After each decision, the environment provides a quantified reward function value as feedback. The reward function can be found in the reward function described above. For example, if the scanning angle α of an image is 60 degrees, and the angle θ between the sampled line attempted by the agent and the gold standard line is 12 degrees, then the reward value R obtained for this action is 0.2. The design of this function ensures that the smaller the angle, i.e., the closer the attempted line is to the gold standard, the smaller the penalty (negative feedback) represented by the reward value, thereby guiding the agent to optimize its strategy in the direction of reducing the angle.

[0064] The training process involves the agent repeatedly performing actions (trying different sampling line positions) and receiving feedback (receiving a reward value R calculated based on the above formula) in the image environment of the training set. It then continuously adjusts its internal decision parameters using this feedback, ultimately enabling the model to converge to a stable and optimal policy. In this embodiment, convergence or reaching the training completion state means that after sufficient learning, the agent of the initial reinforcement learning model, when faced with a new, unseen B-mode ultrasound image, can automatically output a target sampling line that functionally matches the labeled gold standard sampling line position. The model obtained at this point is the reinforcement learning model that can be used for actual monitoring. The purpose of this training process is to transform professional experience into automatically reproducible algorithmic capabilities.

[0065] In some embodiments, the method may further include a step of training a deep learning model. This step of training the deep learning model can also be applied to a mobile terminal or server deployed with the cardiac function analysis module 14, and may specifically include: The system receives M-mode images labeled with key points of the interventricular septum at end-diastole, key points of the left ventricular posterior wall at end-diastole, key points of the interventricular septum at end-systole, and key points of the left ventricular posterior wall at end-systole as training data; Gaussian distribution heatmaps are generated centered on the key point locations as training labels; and the initial deep learning model is trained by minimizing the loss between the model prediction heatmap and the training labels to obtain the deep learning model.

[0066] Among them, the M-mode images labeled with key points of the interventricular septum at end-diastole, key points of the left ventricular posterior wall at end-diastole, key points of the interventricular septum at end-systole, and key points of the left ventricular posterior wall at end-systole refer to the original image data used as training samples. The positions of key anatomical structures at specific time phases have been pre-marked. For example, on an M-mode image, the positions of the interventricular septum and the left ventricular posterior wall at the end of diastole and at the end of systole can be marked in advance. These points together constitute the ground truth for supervised model learning.

[0067] To transform discrete keypoint coordinates into a continuous format with spatial context information more suitable for neural network learning, a Gaussian distribution heatmap needs to be generated as training labels. In this embodiment, the Gaussian distribution heatmap refers to a two-dimensional image whose pixel values ​​are distributed according to a two-dimensional Gaussian function centered on the labeled keypoint coordinates. For example, using the labeled points (… , In the heatmap generated centered at the center, the pixel values ​​closer to the center are higher (close to 1), and the pixel values ​​decay according to a Gaussian curve to close to 0 as the distance increases. The training labels are a pair of heatmaps corresponding to the original M-mode image, one representing two key points at the end of diastole and the other representing two key points at the end of systole.

[0068] The goal of training is to enable an initial deep learning model (referring to an untrained or randomly initialized neural network) to learn the mapping from an input M-pattern image to an output predicted heatmap. During training, the model outputs a predicted heatmap and compares it with the aforementioned training labels (i.e., ground truth heatmaps). In this embodiment, the loss is a scalar value used to quantify the difference between the predicted heatmap and the ground truth heatmap, typically calculated using a function such as mean squared error. This loss value is minimized using optimization methods such as backpropagation, i.e., continuously adjusting the model's internal parameters so that the predicted heatmap output by the model increasingly approximates the ground truth heatmap in terms of distribution. When the loss converges to a low level, it indicates that the model has been able to stably and accurately locate key points from the M-pattern image, and the resulting model is the deep learning model that can be used for practical monitoring. The core of this process is to distill the knowledge of key point location in the M-pattern image into the neural network parameters through heatmap distillation.

[0069] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An ultrasound cardiac dynamic monitoring system, characterized in that, include: A patch-type ultrasound probe; the patch-type ultrasound probe is used to emit ultrasound waves and receive echo signals from the human body. Hydrogel patch; the hydrogel patch is attached to the radiating surface of the patch-type ultrasound probe to fill the air gap between the radiating surface of the patch-type ultrasound probe and human skin; Wearable hardware module; the wearable hardware module is connected to the patch-type ultrasound probe and is used to generate an ultrasound image based on the echo signal; wherein the ultrasound image is a B-mode ultrasound image of the long axis of the left ventricle of the heart; Cardiac function analysis module; the cardiac function analysis module is connected to the wearable hardware module; the cardiac function analysis module is configured as follows: The system receives the ultrasound image and determines a target sampling line in the ultrasound image based on a reinforcement learning model. It then controls the wearable hardware module to acquire M-mode images along the target sampling line. The M-mode images are input into a deep learning model to output the key point positions of the left ventricular septum and the left ventricular posterior wall at end-diastole and end-systole based on the deep learning model. The left ventricular ejection fraction is calculated based on the left ventricular septum and the key point positions. The left ventricular ejection fraction represents the dynamic monitoring and detection results of cardiac function.

2. The ambulatory ultrasonic cardiac monitoring system of claim 1, wherein, The wearable hardware module includes an analog switch unit, an analog front-end unit, a main control unit, a transmission unit, and a power management unit; The analog switch unit is connected to the patch-type ultrasonic probe and is used to switch between multiple ultrasonic transducer channels; The analog front-end unit is connected to the analog switch unit and is used to generate an ultrasonic emission signal that excites the patch-type ultrasonic probe, and to amplify and perform analog-to-digital conversion on the echo signal received by the patch-type ultrasonic probe. The main control unit is connected to the analog front-end unit and is used to control the timing of ultrasonic wave transmission and reception, and to generate the ultrasonic image based on the echo signal. The transmission unit is connected to the main control unit and is used to wirelessly transmit the ultrasound image to the cardiac function analysis module. The power management unit is connected to the analog switch unit, the analog front-end unit, the main control unit, and the transmission unit respectively, and is used to provide a stable operating voltage.

3. The ambulatory ultrasonic cardiac monitoring system of claim 1, wherein, The cardiac function analysis module is specifically used for: In the key point prediction heatmaps at the end of diastole and the end of systole, respectively, find the minimum bounding rectangle of each predicted key point region and obtain the coordinates of the center point of each rectangle; In the key point prediction heatmap at the end of diastole, the two nearest center points on the horizontal coordinate are paired together to form a diastolic group; in the key point prediction heatmap at the end of systole, the two nearest center points on the horizontal coordinate are paired together to form a systolic group. Each diastolic group is paired with the systolic group that is horizontally closest to it to construct a cardiac cycle; For each paired cardiac cycle, the left ventricular ejection fraction is calculated based on the left ventricular septum and the location of the key point.

4. The ultrasound cardiac dynamic monitoring system according to claim 3, characterized in that, The cardiac function analysis module is specifically used for: The left ventricular end-diastolic diameter is calculated based on the key points of the left ventricular interventricular septum and the left ventricular posterior wall at the end of diastole, and the left ventricular end-systolic diameter is calculated based on the key points of the left ventricular interventricular septum and the left ventricular posterior wall at the end of systole. The left ventricular ejection fraction is calculated based on the left ventricular end-diastolic diameter and the left ventricular end-systolic diameter.

5. The ambulatory ultrasonic cardiac monitoring system of claim 4, wherein, The formula for calculating the left ventricular ejection fraction includes: Where EF is the left ventricular ejection fraction, N is the total number of effective cardiac cycles identified within the preset monitoring period, and i represents the i-th cardiac cycle. Left ventricular end-diastolic diameter The angle between the vertical direction and the vertical direction. Left ventricular end-diastolic diameter The angle between the vertical direction and the vertical direction.

6. The ultrasound cardiac dynamic monitoring system according to claim 1, characterized in that, The patch-type ultrasonic probe includes multiple ultrasonic transducer elements arranged in a linear array.

7. The ultrasound cardiac dynamic monitoring system according to claim 1, characterized in that, The hydrogel patch is manufactured using a biocompatible silicone integral molding process.

8. A method for dynamic echocardiographic monitoring, characterized in that, The method, applied to the ultrasound cardiac dynamic monitoring system according to any one of claims 1-7, comprises: Receive the ultrasound image and determine the target sampling line in the ultrasound image based on a reinforcement learning model; The wearable hardware module of the ultrasound cardiac dynamic monitoring system is controlled to acquire M-mode images along the target sampling line; The M-mode image is input into a deep learning model to output the key points of the left ventricular septum and left ventricular posterior wall at end-diastole and end-systole based on the deep learning model. The left ventricular ejection fraction is calculated based on the left ventricular septum and the location of the key points; the left ventricular ejection fraction represents the dynamic monitoring and detection results of cardiac function.

9. The method for dynamic cardiac monitoring by ultrasound according to claim 8, characterized in that, The method further includes: A training set of historical B-mode ultrasound images containing long-axis sections of the left ventricle of the heart was obtained, with the gold standard sampling line position marked in each historical B-mode ultrasound image. A reward function is constructed based on the angle between the agent's position on the sampling line and the gold standard sampling line in the initial reinforcement learning model, and the scanning angle corresponding to the historical B-mode ultrasound image. The agent repeatedly performs actions on the training set and receives feedback based on the reward function until the initial reinforcement model automatically outputs a target sampling line position that matches the gold standard sampling line position based on the input B-mode ultrasound image, thus obtaining the reinforcement learning model.

10. The method for dynamic cardiac monitoring by ultrasound according to claim 8, characterized in that, The method further includes: The training data consists of M-mode images labeled with key points of the interventricular septum at end-diastole, key points of the left ventricular posterior wall at end-diastole, key points of the interventricular septum at end-systole, and key points of the left ventricular posterior wall at end-systole. A Gaussian distribution heatmap centered on the key point location is generated as the training label; The initial deep learning model is trained by minimizing the loss between the model's predicted heatmap and the training labels to obtain the deep learning model.

Images