E-FAST ultrasonic intelligent navigation and decision-making system

Abstract

The invention relates to an E-FAST ultrasonic intelligent navigation and decision-making system, which comprises a portable ultrasonic device used for collecting ultrasonic image data of a patient in pitching, rotating and parallel moving processes by taking a standard position as a reference; the computer processing part is used for inputting the ultrasonic image data into a navigation decision model to obtain a reference correction operation and an effusion diagnosis result; wherein the navigation decision model comprises a scanning optimal tangent plane navigation part used for calculating and outputting a reference correction operation and a reference value according to received ultrasonic image data; the scanning effusion segmentation part is used for segmenting effusion in the image in real time according to the received ultrasonic image data to obtain an effusion image; and the effusion diagnosis part is used for performing diagnosis processing on the effusion image to obtain a feature map with effusion discrimination. According to the invention, a non-professional user can be guided to use the portable ultrasonic equipment to carry out preliminary examination and qualitative diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of ultrasonic trauma assessment, and in particular to an E-FAST ultrasonic intelligent navigation and decision-making system. Background Art

[0002] E-FAST ultrasound uses the simple, fast, sensitive and accurate characteristics of portable ultrasound diagnostic equipment to help clinicians evaluate life-threatening trauma situations, such as trauma to the chest, pericardium and abdomen. Therefore, it is widely used in emergency, battlefield and large-scale disaster scenarios.

[0003] Currently, E-FAST ultrasound examinations require experienced experts and have expensive labor costs, especially at emergency accident sites, intensive care units, and areas with a lack of medical services. Due to the scarcity and high costs of professional talent, it is particularly difficult to meet the demand for experts, which may result in some patients being unable to receive appropriate treatment in a timely manner.

[0004] Medical imaging is one of the main application areas of artificial intelligence. The development of artificial intelligence in ultrasound medicine is gradually entering the clinic, involving the automatic identification of free peritoneal effusion by U-net deep learning models and deep learning visual classification of children's FAST ultrasound images. However, the above studies are all based on stored ultrasound images or videos, with the aim of obtaining intelligent and rapid diagnosis and shortening the time spent on image recognition. In complex emergency environments, this technology is accompanied by a large amount of computing cost and reasoning time, which makes it difficult to meet the needs of the actual scene. More importantly, the lack of medical personnel with professional ultrasound skills on site and the inability to obtain standard cross-sectional images are often the primary factors limiting the application of FAST / E-FAST technology. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an E-FAST ultrasonic intelligent navigation and decision-making system, which can guide non-professional users to use portable ultrasonic equipment for preliminary examination, make qualitative and quantitative diagnosis, and triage in time.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide an E-FAST ultrasonic intelligent navigation and decision-making system, including:

[0007] Portable ultrasound equipment, used to collect ultrasound image data of the patient during pitch, rotation and parallel movement based on the standard position;

[0008] The computer processing part is used to input the ultrasound image data into the navigation decision model to obtain the reference correction operation and the effusion diagnosis result;

[0009] Wherein, the navigation decision model includes:

[0010] A scanning optimal cutting plane navigation part is used to calculate and output a reference correction operation and a reference value according to the received ultrasound image data to assist the operator in positioning the ultrasound probe of the portable ultrasound device to the optimal cutting plane;

[0011] The scanning effusion segmentation part is used to perform real-time segmentation of the effusion in the image according to the received ultrasound image data to obtain an effusion image;

[0012] The effusion diagnosis part is used to perform diagnostic processing on the effusion image to obtain a feature map with effusion discrimination.

[0013] The scanning optimal cutting plane navigation part combines the image understanding ability contained in the two-dimensional pre-trained weights in the SAM model and the advanced performance of the three-dimensional architecture medical model SAMMed3D in the field of medical image analysis. At the same time, by training the three-dimensional architecture model, it is possible to use existing knowledge to analyze ultrasound image data.

[0014] The scanning effusion segmentation part adopts the UNet architecture, including an encoder and a decoder. The encoder and the decoder have four stages, and the output of each stage is obtained from the output result of the previous stage. Jump connections are used between each stage of the encoder and the decoder; the encoder adopts a 2D state space model as the basic processing unit, uses a cross scanning method to perform four context modeling on the image, and calculates the corresponding outputs respectively, and finally uses the average of the four outputs as the input of the next stage.

[0015] The effusion diagnosis section includes:

[0016] A filtering module based on diagnostic experience weights, used for filtering the effusion image using a spatial weight filter;

[0017] A mask module based on irrelevant information shielding is used to shield irrelevant information of the effusion image after filtering to obtain the effusion image to be tested;

[0018] The multi-directional displacement window attention structure module is used to extract and infer the discriminative features of the effusion image to be tested to obtain a feature map with effusion discriminative properties.

[0019] The weights of the spatial weight filter are expressed as: Among them, w is the weight of the pixel with a height h from the upper edge of the image, H represents the height of the image size, and H w A hyperparameter used to control the degree of preservation of the original image information, w0 represents the minimum weight of the spatial weight filter.

[0020] The multi-directional displacement window attention structure module includes a horizontal-vertical boundary membrane feature extraction structure, a 1×1 Bottleneck layer and a Shortcut connection layer; wherein, the horizontal-vertical boundary membrane feature extraction structure includes a horizontal feature extraction branch, a vertical feature extraction branch and an oblique feature extraction branch, each branch is provided with a W-MSA module and a SW-MSA module, the W-MSA module is used for information calculation and interaction within the window, the SW-MSA module is used for information interaction function between windows, and the SW-MSA modules of different branches have different structures to obtain various effusion discriminative features; the W-MSA module and the SW-MSA module both perform self-attention calculation based on SWIN TRANSFORMER.

[0021] The effusion diagnosis part also uses a data enhancement module based on body fat mass simulation to enhance the effusion image during training. The data enhancement module based on body fat mass simulation uses a Gaussian blur method and a random Gaussian kernel size to simulate effusion images with different body fat amounts.

[0022] When training the navigation decision model, automatic batch annotation is performed based on a large segmentation model, specifically: by selecting the video frames at the beginning, middle and end of each video data, clear and representative video frames are selected, and annotated by ultrasound experts to form a precision labeled data set; by using an image encoder model, the obtained precision labeled data set is encoded into a unified image feature representation; the image encoder model is used to encode the input image, and the obtained encoding is searched for similarity with the encoding of the precision labeled data set to obtain the most similar N input images, and the N input images are input into the SegGPT large model together with the image to be annotated, and the segmentation result is obtained after credibility evaluation and smoothing and denoising of the output, and the batch annotation results are reviewed and refined by ultrasound experts to obtain the final segmentation data set.

[0023] When training the navigation decision model, the navigation offset based on optical flow is used for automatic labeling, specifically: the positive direction and the negative direction are set as the label of each frame in a certain motion dimension; based on the data collected by the physician, the offset label of the ultrasound probe of the portable ultrasound device from the standard position is automatically calculated through the optical flow direction; the offset labels are screened, the offset labels that meet the standards are retained and numerical mapping processing is performed to obtain the final navigation data set for use in navigation model training.

[0024] The navigation decision model also includes a scanning quality assessment part, which includes a preliminary fluid effusion detection module, an image quality assessment module and a calculation and judgment module; the preliminary fluid effusion detection module is based on the ultrasonic prior characteristics of the effusion, uses image registration technology, performs image registration of a healthy human body and a patient's human body, and obtains a difference score by comparing the differences in the registered images, and screens whether the current position is an abnormal fluid accumulation position based on the difference score; the image quality assessment module uses the quality measurement index recommended by the PCP quality assessment protocol to evaluate the image and obtain an image quality score; the calculation and judgment module combines the difference score and the image quality score to obtain a regression score through weighted calculation, and performs fluid effusion diagnosis processing when the regression score exceeds a threshold.

[0025] Beneficial Effects

[0026] Due to the adoption of the above technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention can significantly lower the threshold for the use of ultrasonic operation, has no requirements for the basic knowledge of ultrasonic operation, and general practitioners, nurses or inexperienced users at emergency on-site rescue stations can complete the operation under the guidance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the E-FAST ultrasound chest scanning experimental paradigm;

[0028] Figure 2 This is a flow chart of the E-FAST ultrasound chest scan process;

[0029] Figure 3 It is a schematic diagram of the best cutting plane navigation part of the scan;

[0030] Figure 4 It is a schematic diagram of the scan quality assessment part;

[0031] Figure 5 It is a schematic diagram of the segmented part of the scan for effusion;

[0032] Figure 6 It is a schematic diagram of the effusion diagnosis part. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0034] The embodiment of the present invention relates to an E-FAST ultrasonic intelligent navigation and decision system, the hardware part of which includes: a portable ultrasonic device equipped with a high-frequency probe, a video acquisition card, a rk3399pro edge computing device (or a computing device supporting model reasoning such as a desktop computer or a laptop computer), a display device, a mouse and a keyboard. Among them, a navigation decision model is set in the rk3399pro edge computing device (or a computing device supporting model reasoning such as a desktop computer or a laptop computer), which can guide inexperienced users to complete the E-FAST examination to find the best cutting plane and perform quantitative analysis on the ultrasonic image.

[0035] The navigation decision model consists of the following aspects: experimental paradigm and scanning process, automatic batch labeling of segmentation based on the segmentation large model, automatic labeling of navigation offset based on optical flow, E-FAST scanning optimal cutting plane navigation part, E-FAST scanning quality assessment part, E-FAST scanning effusion segmentation part, E-FAST effusion diagnosis part, E-FAST scanning process control module (Pipeline), GUI interface and continuous learning module.

[0036] Experimental paradigm and scanning process. In this step, in the ultrasound image acquisition experiment of pleural effusion, a standard experimental paradigm and scanning process were adopted to ensure efficient acquisition of high-quality ultrasound images for model use, such as Figure 1 As shown. Taking chest scanning as an example, in this embodiment, ultrasound images of 50 subjects are collected in sequence. For each subject, data of 4 intercostal spaces are collected in sequence, including the 6th intercostal space, the 7th intercostal space, the 8th intercostal space, and the 9th intercostal space. In each intercostal space, ultrasound image data of the subject's pitch, rotation, and parallel movement process are collected based on the standard position, as shown in FIG. Figure 2 As shown in the figure. In the experiment, Mindray M9 (a portable color Doppler ultrasound system) was used to acquire ultrasound video. In this experimental paradigm, the standard position refers to the image with the most discriminative disease in this intercostal space. Selection of the standard position: For patients with pleural effusion, select the image position with the clearest and largest area of ​​effusion; for healthy subjects, select the intercostal area with clear and artifact-free. The process data of these movement modes can record the ultrasound images of pleural effusion in different states. More specifically, the pitch movement refers to the ultrasound probe moving from the standard position to an elevation angle of 40°, and then moving in the opposite direction to -40°. The rotation movement refers to the ultrasound probe rotating from the standard position to 30° clockwise, and then moving in the opposite direction to -30°. The forward movement refers to the ultrasound probe moving 4cm from the standard position to the direction in which the patient is facing, and then moving in the opposite direction to -4cm. The ultrasound probe needs to be kept in the intercostal gap during translation scanning to avoid being blocked by the ribs. Keep a slow and uniform motion during the movement to avoid dynamic blur and artifacts.

[0037] In the data annotation stage, this embodiment adopts automatic batch annotation of segmentation based on the large segmentation model. This method can use a small amount of precise segmentation prompts with a large model to achieve batch annotation, which greatly reduces labor costs. Specifically: First, by selecting the video frames at the beginning, middle and end of each video, clear and representative video frames are selected, and ultrasound experts perform fine annotation to form a precise dataset. Then, the image encoder model (CLIP, etc.) is used to encode the acquired precise dataset into a unified image feature representation. Next, the input image is encoded by the same image encoder model, and a similarity search is performed with the encoding of the precise dataset to obtain the most similar N input images. The N input images are input into the SegGPT large model together with the image to be annotated, and the output is subjected to credibility assessment and smoothing denoising to obtain the segmentation results. Finally, the ultrasound expert reviews and refines the batch annotation results to obtain the final segmentation dataset for segmentation model training.

[0038] In the data labeling stage, this embodiment also adopts automatic labeling of navigation offset based on optical flow, which realizes accurate labeling of navigation offset by analyzing quantifiable optical flow movement data. Specifically: First, set the positive direction and negative direction as the label of each frame in a certain motion dimension. Then, based on the data collected by the physician, the offset label of the ultrasound probe from the standard position is automatically calculated by the optical flow direction. Finally, the offset labels are screened, the offset labels that meet the standards are retained and numerical mapping processing is performed to obtain the final navigation data set for navigation model training.

[0039] The E-FAST scan optimal cutting plane navigation part in this embodiment adopts an innovative method of combining two-dimensional pre-trained weights with three-dimensional architecture during the training phase by comprehensively learning the statistical distribution characteristics of ultrasound data and its navigation information, so as to achieve a deep understanding of the target ultrasound image and its correction strategy. Specifically, this navigation part is based on the two-dimensional pre-trained weights of SAM and optimizes the three-dimensional architecture medical model SAMMed3D fine-tuning architecture, and introduces three-dimensional ultrasound data modality for training to further improve the model performance. This method combines the image understanding capabilities contained in the two-dimensional pre-trained weights in the SAM model with the advanced performance of SAMMed3D in the field of medical image analysis. At the same time, by training the three-dimensional architecture model, it also gives the model the ability to use existing knowledge to parse three-dimensional ultrasound data. Figure 3 As shown, during the inference stage, the navigation part receives real-time ultrasound image input, calculates and outputs corresponding reference correction operations (including translation, rotation and pitch) and reference values, thereby effectively assisting the operator to accurately position the ultrasound probe to the optimal cutting plane to achieve the best image acquisition effect.

[0040] The E-FAST scan quality assessment part in this embodiment includes a preliminary effusion detection module, an image quality assessment module, and a calculation and judgment module. Figure 4 As shown, first, the preliminary detection module of effusion uses image registration technology based on the ultrasound prior features of effusion, registers the images of healthy human body and patient human body, and obtains the difference (DR) score by comparing the difference of the registered images, and screens whether the current position is an abnormal effusion position based on the difference score. Subsequently, the image quality assessment module uses the quality metrics (e.g., signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), artifact voxel ratio (qi1), etc.) recommended by the PCP quality assessment protocol (see http: / / preprocessed-connectomes-project.org / quality-assessment-protocol) to evaluate the image and obtain the image quality (IQA) score. Finally, the calculation and judgment module combines the difference score and the image quality score through weighted calculation to obtain a regression (RS) score in the range of 1 to 9, and performs further diagnostic analysis when the regression score exceeds 6 points, that is, the effusion diagnosis treatment.

[0041] The E-FAST scanning and segmentation of the effusion in this embodiment has the characteristics of high accuracy, small model parameters, and fast segmentation speed, and can achieve real-time segmentation of the effusion and form visual feedback to prompt the operator. Figure 5 As shown in the figure, this part adopts the classic UNet architecture, which consists of an encoder, a decoder and a jump connection. In the encoder and decoder, the model has four stages respectively, and each stage is obtained by downsampling / upsampling the output result of the previous stage. Among them, the encoder uses a 2D state space model as the basic processing unit, uses a cross scanning method to model the context of the image four times, and calculates the corresponding outputs respectively, and finally uses the mean of the four outputs as the input of the next stage. The jump connection merges the output of the encoder with the input of the decoder in the same layer to make better use of the detailed information in the image.

[0042] The effusion diagnosis part of this embodiment is used to identify the effusion in patients with pleural effusion. This part can obtain a characteristic graph with effusion discrimination. Figure 6 As shown, the effusion diagnosis part includes: a filtering module based on diagnostic experience weights, which is used to filter the effusion image using a spatial weight filter; a mask module based on irrelevant information shielding, which is used to shield irrelevant information of the effusion image after filtering to obtain the effusion image to be tested; a multi-directional displacement window attention structure module, which is used to extract and infer the discriminative features of the effusion image to be tested to obtain a feature map with effusion discriminability.

[0043] In order to ensure the generalization ability of the effusion diagnosis part, the effusion diagnosis part also includes a data enhancement module based on body fat simulation. Specifically, after preprocessing, the ultrasound image is cropped into a 224×224 square image Is. The data enhancement module based on body fat simulation is input to simulate different body masses based on the Gaussian blur method and random Gaussian kernel size. For two-dimensional image data, the weight function of Gaussian blur (also called probability density function) is defined as:

[0044]

[0045] Among them, d x , d y They correspond to the distance from the horizontal and vertical coordinates of the current pixel to the target, and σ represents different distribution styles of the Gaussian function. In data enhancement, a random Gaussian kernel size and σ are set for the samples of each epoch to generate data enhancement. Specifically, for image data of size H×W, the Gaussian kernel size is randomly selected from [1,3,5,7,9]. The size of the Gaussian kernel determines the degree and range of blur. When the size of the Gaussian kernel is 1, it can be considered that the image is not blurred. As the kernel size increases, the blurring effect will gradually increase, and simulate ultrasound data with higher fat content, improving the generalization ability of the model.

[0046] After patch embedding, the data-enhanced image is multiplied by the filter module based on the diagnostic experience weight to introduce expert experience. Specifically, the diagnostic experience from ultrasound experts has certain discriminative information about effusion. This discriminative information can be introduced into the model by constructing a special spatial weight filter as an attention module. Therefore, a spatial weight filter is proposed for the diagnosis of pleural effusion. This filter can provide an attention module based on the prior weights. By setting the initial weights, the regional information (red area) where pleural effusion is most likely to occur is retained, and other areas are filtered and information compressed according to the degree of possibility. This transition process of weight W based on medical experience can be expressed as:

[0047]

[0048] Among them, w represents the weight of the pixel with a height h from the upper edge of the image, H represents the height of the image size, and the variable hyperparameter H w ∈[1,H) is used to control the degree of preservation of the original image information. w=1 indicates that the experience weight is not effective, and w0 indicates the minimum filtering weight. Since the data distribution of effusion is diverse, the filter cannot simply ignore some low-probability events. Therefore, the minimum filtering weight cannot usually be set to 0. It is worth noting that human experience is not always the optimal solution for model learning. Fixed filters may affect the upper limit of subsequent model learning. To solve this problem, the weight of the filter is allowed to be updated with the back-propagation learning process of the model to obtain the optimal filtering weight. This method provides a way for the model to autonomously select the data to pay attention to, so that the model can adjust the degree of attention to data in different regions based on its own learning and judgment. This flexibility enables the model to better adapt to different data distributions and challenges, thereby improving the accuracy and reliability of diagnosis.

[0049] After filtering by empirical weights, the feature map is added to the mask module based on irrelevant information shielding to shield irrelevant information. The window self-attention of the effusion diagnosis part performs a softmax operation in the calculation process. Too many feature maps will cause compression and loss of discriminative information. Therefore, this module shares a softmax with the mask of the window attention to remove irrelevant information, that is, it is only necessary to operate with the MASK template before each window self-attention to shield irrelevant information.

[0050] After irrelevant information is shielded, the feature map enters the multi-directional shift window attention structure module to extract and infer discriminative features. Specifically, the multi-directional shift window attention structure module is mainly composed of the extraction structure of the horizontal-vertical boundary membrane feature, the 1×1 Bottleneck layer and the Shortcut connection layer. In the extraction structure of the horizontal-vertical boundary membrane feature, three feature extraction branches, namely horizontal, vertical and oblique, are sampled. There are W-MSA (window based multi-head self-attention) modules and SW-MSA (shifted window based multi-head self-attention) modules on each branch. Both modules need to perform self-attention calculation based on SWIN TRANSFORMER. The W-MSA module is responsible for information calculation and interaction within the window. In different branches, the SW-MSA module has different structures and functions to obtain various discriminative features (horizontal-SW-MSA, vertical-SW-MSA, SW-MSA). At the same time, SW-MSA also has the function of information interaction between windows. Through this design, the effusion diagnosis part can better extract local information and realize the interaction between information, thereby improving the performance of the model in pleural effusion diagnosis.

[0051] The E-FAST scanning process control module (Pipeline) in this embodiment is responsible for scheduling the scanning process, GUI, E-FAST scanning optimal cutting plane navigation part, E-FAST scanning quality assessment part, E-FAST scanning effusion segmentation part, and E-FAST effusion diagnosis part at different stages, and independently or in parallel controls the calling and prompting of each part to guide the user to complete the E-FAST scan. The GUI interface is controlled by the E-FAST scanning process control module, and displays real-time scanning images, real-time output of each part, and scanning prompts. In different scanning processes, the corresponding scanning position, scanning method, reference image and real-time results are displayed on the GUI interface to assist inexperienced users to complete the E-FAST scan.

[0052] It is worth mentioning that the E-FAST ultrasonic intelligent navigation and decision-making system of this embodiment also includes a continuous learning module, which is mainly responsible for continuously optimizing the model performance of each part by using actively collected and automatically recorded ultrasonic data.

[0053] The use process of the E-FAST ultrasonic intelligent navigation and decision-making system in this embodiment is as follows:

[0054] (1) When in use, the ultrasound physician first turns on the portable ultrasound device, connects the high-frequency ultrasound probe, and then connects the HDMI of the portable ultrasound device to the E-FAST ultrasound navigation system.

[0055] (2) Turn on the edge computing device, wait for the system to automatically run the corresponding script, and start the E-FAST navigation system software (i.e., the navigation decision model).

[0056] (3) The user clicks the start button according to the system prompts to enter patient information and other information to enter the scanning GUI interface.

[0057] (4) The user refers to the system prompts, holds the probe and places it in the sample position as prompted, and clicks Continue.

[0058] (5) The user scans according to the scanning prompts, refers to the target lesion in the example and the real-time prompts (segmentation and navigation prompts) to find the best lesion, or clicks Continue to scan the next location.

[0059] (6) After the user follows the system prompts and completes the scan of all parts of the patient, the system displays the overall diagnosis results and conducts a qualitative and quantitative assessment of the patient's fluid accumulation.

[0060] (7) The user can click on the interface to continue the next scan, then enter the next patient information, repeat (3) to (6), and complete the scan of the new patient. Or click the exit button to return to the system standby interface and wait for the operator to perform three steps. Perform an ultrasonic scan of the patient's muscle injury area, use a high-frequency probe to collect ultrasonic two-dimensional images, and transmit the ultrasonic images to the mobile terminal in real time.

[0061] It is not difficult to find that the present invention can significantly lower the threshold for using ultrasonic operation, and has no requirement for basic ultrasonic operation skills. General practitioners, nurses or inexperienced users at emergency on-site rescue stations can complete the operation under the guidance of the system.

Claims

1. An E-FAST ultrasonic intelligent navigation and decision-making system, characterized in that: include: Portable ultrasound equipment, used to collect ultrasound image data of the patient during pitch, rotation and parallel movement based on the standard position; The computer processing part is used to input the ultrasound image data into the navigation decision model to obtain the reference correction operation and the effusion diagnosis result; Wherein, the navigation decision model includes: A scanning optimal cutting plane navigation part is used to calculate and output a reference correction operation and a reference value according to the received ultrasound image data to assist the operator in positioning the ultrasound probe of the portable ultrasound device to the optimal cutting plane; The scanning effusion segmentation part is used to perform real-time segmentation of the effusion in the image according to the received ultrasound image data to obtain an effusion image; The effusion diagnosis part is used to perform diagnostic processing on the effusion image to obtain a feature map with effusion discrimination.

2. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 1 is characterized in that: The scanning optimal cutting plane navigation part combines the image understanding ability contained in the two-dimensional pre-trained weights in the SAM model and the advanced performance of the three-dimensional architecture medical model SAMMed3D in the field of medical image analysis. At the same time, by training the three-dimensional architecture model, it is possible to use existing knowledge to analyze ultrasound image data.

3. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 1 is characterized in that: The scanning effusion segmentation part adopts the UNet architecture, including an encoder and a decoder. The encoder and the decoder have four stages, and the output of each stage is obtained from the output result of the previous stage. Jump connections are used between each stage of the encoder and the decoder; the encoder adopts a 2D state space model as the basic processing unit, uses a cross scanning method to perform four context modeling on the image, and calculates the corresponding outputs respectively, and finally uses the average of the four outputs as the input of the next stage.

4. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 1, characterized in that: The effusion diagnosis section includes: A filtering module based on diagnostic experience weights, used for filtering the effusion image using a spatial weight filter; A mask module based on irrelevant information shielding is used to shield irrelevant information of the effusion image after filtering to obtain the effusion image to be tested; The multi-directional displacement window attention structure module is used to extract and infer the discriminative features of the effusion image to be tested to obtain a feature map with effusion discriminative properties.

5. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 4 is characterized in that: The weights of the spatial weight filter are expressed as: Among them, w is the weight of the pixel with a height h from the upper edge of the image, H represents the height of the image size, and H w A hyperparameter used to control the degree of preservation of the original image information, w0 represents the minimum weight of the spatial weight filter.

6. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 4, characterized in that: The multi-directional displacement window attention structure module includes a horizontal-vertical boundary membrane feature extraction structure, a 1×1 Bottleneck layer and a Shortcut connection layer; wherein, the horizontal-vertical boundary membrane feature extraction structure includes a horizontal feature extraction branch, a vertical feature extraction branch and an oblique feature extraction branch, each branch is provided with a W-MSA module and a SW-MSA module, the W-MSA module is used for information calculation and interaction within the window, the SW-MSA module is used for information interaction function between windows, and the SW-MSA modules of different branches have different structures to obtain various effusion discriminative features; the W-MSA module and the SW-MSA module both perform self-attention calculation based on SWIN TRANSFORMER.

7. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 4, characterized in that: The effusion diagnosis part also uses a data enhancement module based on body fat mass simulation to enhance the effusion image during training. The data enhancement module based on body fat mass simulation uses a Gaussian blur method and a random Gaussian kernel size to simulate effusion images with different body fat amounts.

8. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 1, characterized in that: When training the navigation decision model, automatic batch annotation is performed based on a large segmentation model, specifically: by selecting the video frames at the beginning, middle and end of each video data, clear and representative video frames are selected, and annotated by ultrasound experts to form a precision labeled data set; by using an image encoder model, the obtained precision labeled data set is encoded into a unified image feature representation; the image encoder model is used to encode the input image, and the obtained encoding is searched for similarity with the encoding of the precision labeled data set to obtain the most similar N input images, and the N input images are input into the SegGPT large model together with the image to be annotated, and the segmentation result is obtained after credibility evaluation and smoothing and denoising of the output, and the batch annotation results are reviewed and refined by ultrasound experts to obtain the final segmentation data set.

9. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 1, characterized in that: When training the navigation decision model, the navigation offset based on optical flow is used for automatic labeling, specifically: the positive direction and the negative direction are set as the label of each frame in a certain motion dimension; based on the data collected by the physician, the offset label of the ultrasound probe of the portable ultrasound device from the standard position is automatically calculated through the optical flow direction; the offset labels are screened, the offset labels that meet the standards are retained and numerical mapping processing is performed to obtain the final navigation data set for use in navigation model training.

10. The E-FAST ultrasonic intelligent navigation and decision-making system according to claim 1, characterized in that: The navigation decision model also includes a scanning quality assessment part, which includes a preliminary fluid effusion detection module, an image quality assessment module and a calculation and judgment module; the preliminary fluid effusion detection module is based on the ultrasonic prior characteristics of the effusion, uses image registration technology, performs image registration of a healthy human body and a patient's human body, and obtains a difference score by comparing the differences in the registered images, and screens whether the current position is an abnormal fluid accumulation position based on the difference score; the image quality assessment module uses the quality measurement index recommended by the PCP quality assessment protocol to evaluate the image and obtain an image quality score; the calculation and judgment module combines the difference score and the image quality score to obtain a regression score through weighted calculation, and performs fluid effusion diagnosis processing when the regression score exceeds a threshold.

Images