Edge segmentation method and system of ultrasonic image, terminal and medium

Abstract

The invention relates to the technical field of medical image processing, and discloses an ultrasonic image edge segmentation method and system, a terminal and a medium, and the ultrasonic image edge segmentation method comprises the steps: constructing an initial model, obtaining a training set, inputting the training set into the initial model, and obtaining a prediction probability graph; performing data preprocessing on the training set to obtain boundary supervision information; boundary sensing information is obtained according to the prediction probability graph and the boundary supervision information; obtaining a target model according to the boundary perception information, the training set and the initial model; and acquiring an echocardiogram of a user, and inputting the echocardiogram into the target model to obtain a segmentation probability graph. According to the method, the boundary monitoring information is provided by preprocessing the training set to solve the problem of boundary blur, the perceptual ability of the model to the boundary is enhanced through the boundary monitoring information, and the boundary segmentation precision of the ultrasonic image can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and particularly relates to an edge segmentation method and system of an ultrasound image, a terminal and a medium. BACKGROUND

[0002] As a non-invasive and real-time imaging examination method, echocardiography can quantitatively analyze various echocardiographic views and different heart defects and their clinical parameters, and provide accurate screening for various types of CHD (Congenital Heart Disease). Its portability, low cost and real-time capability, combined with the risk of not worrying about ionizing radiation and radioactive hazards, make it particularly suitable for infants and is widely used in the diagnosis of congenital heart disease. By using neonatal echocardiography to visualize the segmentation of atria and ventricles, doctors can accurately observe the internal structure of the heart, including the specific morphology and functional status of the left atrium, left ventricle, right atrium and right ventricle. This high-precision visualization helps to clearly distinguish the anatomical relationship between each cardiac chamber and its interaction with key structures such as heart valves. Through precise segmentation of atria and ventricles by echocardiography, not only can the development status of each part of the heart be evaluated, but also the size of the cardiac chamber can be measured, the hemodynamic parameters can be analyzed, and any possible abnormal connection or defect can be identified. For example, by calculating the left ventricular ejection fraction to determine whether the heart pumping function is normal, and by detailed examination of the atrial septum and interventricular septum to find potential defects.

[0003] From the perspective of technology for implementing segmentation, traditional cardiac structure segmentation relies on manual tracing by doctors, which is inefficient and highly subjective, and is difficult to meet the needs of consistency, efficiency and objectivity for large-scale CHD screening. Although semi-automatic segmentation methods such as algorithms based on active contour, level set or graph cuts reduce human intervention, they still require manual initialization, and the segmentation accuracy is limited when dealing with images with low contrast, blurred boundaries or artifacts. In recent years, automatic segmentation techniques based on deep learning (such as U-Net, V-Net, etc. Convolutional Neural Networks) have shown good performance in left ventricular segmentation of adult echocardiograms, but on the neonatal 4CH section, the model performance has declined due to data domain shift, small sample size and image quality fluctuations. Although existing research has attempted to introduce lightweight Transformer structures (such as Swin-UNet) to enhance the model's ability to model long-range dependencies and improve segmentation accuracy, the significant increase in parameter quantity leads to an increase in model complexity, making the training process more difficult to converge, and deployment on bedside ultrasound devices faces a bottleneck in computing power. In addition, echocardiograms have inherent image quality issues such as speckle noise, low signal-to-noise ratio, blurred boundaries, and signal attenuation, making it difficult to identify thin-walled structures such as atrial walls, further affecting the model's segmentation performance. At the same time, the large differences in image statistical properties due to different hospitals, devices, probe frequencies and gain settings limit the model's cross-center generalization ability.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] The main purpose of the present application is to provide an edge segmentation method, system, terminal and medium for an ultrasound image, aiming to solve the problem of low precision in boundary segmentation of neonatal echocardiogram images in the prior art.

[0006] The first aspect of the embodiments of the present application provides an edge segmentation method for an ultrasound image, which comprises the following steps:

[0007] An initial model is constructed, and a training set is obtained, which is input into the initial model to obtain a prediction probability map;

[0008] The training set is subjected to data preprocessing to obtain boundary supervision information;

[0009] According to the prediction probability map and the boundary supervision information, boundary perception information is obtained;

[0010] According to the boundary perception information, the training set and the initial model, a target model is obtained;

[0011] An echocardiogram of a user is obtained, and the echocardiogram is input into the target model to obtain a segmentation probability map.

[0012] Optionally, in an embodiment of the present application, the training set comprises a plurality of label images, and the boundary supervision information comprises a plurality of symbolic Euclidean distance maps.

[0013] The data preprocessing of the training set to obtain boundary supervision information specifically comprises:

[0014] obtaining a plurality of initial distance maps according to the plurality of label images, wherein the distance values in the initial distance maps comprise positive numbers and negative numbers;

[0015] resampling the plurality of initial distance maps to a uniform resolution to obtain a plurality of intermediate distance maps;

[0016] respectively performing linear mapping on the plurality of intermediate distance maps to obtain a plurality of symbolic Euclidean distance maps.

[0017] Optionally, in an embodiment of the present application, the label image comprises a first region, a second region and a third region;

[0018] The obtaining of a plurality of initial distance maps according to the plurality of label images specifically comprises:

[0019] for each label image, respectively calculating the Euclidean distances of the pixels in the first region, the second region and the third region from the target boundary to obtain a first channel distance map corresponding to the first region, a second channel distance map corresponding to the second region and a third channel distance map corresponding to the third region;

[0020] obtaining an initial distance map corresponding to the label image according to the first channel distance map, the second channel distance map and the third channel distance map.

[0021] Optionally, in an embodiment of the present application, the number of prediction probability maps is a plurality, and each prediction probability map corresponds to a symbolic Euclidean distance map. Figure One Optionally, in an embodiment of the present application, the boundary perception information comprises error product data and a direction guidance factor.

[0022] The obtaining of boundary perception information according to the prediction probability map and the boundary supervision information specifically comprises:

[0023] calculating the product of each prediction probability map and the corresponding symbolic Euclidean distance map according to the plurality of prediction probability maps and the corresponding plurality of symbolic Euclidean distance maps to obtain error product data;

[0024] processing the error product data to obtain a direction guidance factor.

[0025] Optionally, in an embodiment of the present application, the processing of the error product data to obtain the direction guidance factor specifically comprises:

[0026] nonlinearly compressing the error product data to obtain real number data in a positive number interval;

[0027] adjusting the real number data to a positive and negative number range to obtain the direction guidance factor.

[0028] Optionally, in an embodiment of the present application, the obtaining of the target model according to the boundary awareness information, the training set and the initial model specifically comprises:

[0029] obtaining a detail loss according to the error product data and a direction guidance factor loss according to the direction guidance factor;

[0030] updating the initial model according to the detail loss and the direction guidance factor loss to obtain a target model with parameters to be confirmed;

[0031] training the target model with parameters to be confirmed according to the training set for a preset number of training rounds to obtain a target model with parameters confirmed.

[0032] Optionally, in an embodiment of the present application, the target model with parameters to be confirmed is represented as:

[0033]

[0034] wherein, is a total loss function of the target model, is a Dice loss, and a is a weight coefficient of the Dice loss; is a cross-entropy loss, and β is a weight coefficient of the cross-entropy loss; is a detail loss, and γ is a weight coefficient of the detail loss; is a direction guidance factor loss, and δ is a weight coefficient of the direction guidance factor loss.

[0035] The second aspect of the embodiments of the present application further provides an edge segmentation system of an ultrasound image, wherein the edge segmentation system of the ultrasound image is applied to the edge segmentation method of the ultrasound image in any of the above schemes; and the edge segmentation system of the ultrasound image comprises:

[0036] a probability prediction module, configured to construct an initial model, acquire a training set, input the training set into the initial model, and obtain a predicted probability map;

[0037] a data preprocessing module, configured to perform data preprocessing on the training set to obtain boundary supervision information;

[0038] A boundary awareness module is used to obtain boundary awareness information based on the predicted probability map and the boundary supervision information;

[0039] The model training module is used to obtain the target model based on the boundary awareness information, the training set, and the initial model;

[0040] The image segmentation module is used to acquire the user's echocardiogram and input the echocardiogram into the target model to obtain a segmentation probability map.

[0041] A third aspect of this application also provides a terminal, wherein the terminal includes: a memory, a processor, and an edge segmentation program for ultrasound images stored in the memory and executable on the processor, wherein when the edge segmentation program for ultrasound images is executed by the processor, it implements the steps of the edge segmentation method for ultrasound images as described above.

[0042] A fourth aspect of this application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an edge segmentation program for an ultrasound image, and the edge segmentation program for the ultrasound image, when executed by a processor, implements the steps of the edge segmentation method for the ultrasound image as described above.

[0043] Beneficial effects: This application provides an edge segmentation method, system, terminal, and medium for ultrasound images. The edge segmentation method for ultrasound images in this application is a U-shaped network optimization method for atrial edge segmentation in 2D ultrasound images. By preprocessing the training set to provide boundary supervision information, the boundary ambiguity problem is solved. The boundary supervision information enhances the model's ability to perceive boundaries (detail loss and direction guidance factor loss), enabling interpretable control of boundary optimization. Furthermore, by using multi-loss joint optimization to balance overall and local accuracy, the target model predicts the output segmentation probability map, thereby improving the boundary segmentation accuracy of ultrasound images. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a preferred embodiment of the edge segmentation method for ultrasound images in this application;

[0046] Figure 2 This is a schematic diagram of the boundary perception module in a preferred embodiment of the edge segmentation method for ultrasound images in this application;

[0047] Figure 3 is a structural diagram of a preferred embodiment of an edge segmentation system of an ultrasound image of the present application;

[0048] Figure 4 is a structural diagram of a preferred embodiment of a terminal of the present application.

[0049] Explanation of reference signs:

[0050] 100, probability prediction module; 200, data preprocessing module; 300, boundary perception module; 400, model training module; 500, image segmentation module. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and effects of the present application clearer and more explicit, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. The described embodiments are only possible technical implementations of the present application, and are not all possible implementations. Based on the embodiments in the present application, those skilled in the art can certainly combine the embodiments of the present application to obtain other embodiments without creative labor, and these embodiments are also within the protection scope of the present application.

[0052] Firstly, the terms involved in the embodiments of the present application are introduced:

[0053] CHD: Congenital Heart Disease, congenital heart disease;

[0054] 4CH: Apical Four-Chamber View, apical four-chamber view (one of the standard sections of echocardiography);

[0055] TAPSE: Tricuspid Annular Plane Systolic Excursion, tricuspid annular plane systolic excursion (an index for evaluating right ventricular function);

[0056] HLHS: Hypoplastic Left Heart Syndrome, left heart hypoplasia syndrome;

[0057] AVSD: Atrioventricular Septal Defect, atrioventricular septal defect;

[0058] LV: Left Ventricle, left ventricle; LA: Left Atrium, left atrium; RA: Right Atrium, right atrium; RV: Right Ventricle, right ventricle; TV: Tricuspid Valve, tricuspid valve;

[0059] Dice: Dice Similarity Coefficient, evaluation index of semantic segmentation model, Dice coefficient;

[0060] CE Loss: Cross Entropy Loss, cross entropy loss, CE loss;

[0061] HD95: Hausdorff Distance 95%, Hausdorff distance, HD95 coefficient.

[0062] In the related art, from the perspective of data segmentation of cardiac ultrasound images, the current cardiac structure segmentation task mainly relies on echocardiogram images, especially the apical four-chamber view (4CH) as one of the standard sections for screening of neonatal congenital heart disease (CHD). This section can completely display the left atrium (LA), left ventricle (LV), right atrium (RA), right ventricle (RV), atrial septum, interventricular septum, mitral valve and tricuspid valve (TV) and other key anatomical structures in one scan, and can be used for rapid assessment of atrioventricular valve regurgitation, heart chamber ratio and myocardial contraction synchrony. A large number of evidence-based medical studies have shown that the 4CH section has a diagnostic sensitivity of more than 90% and a specificity of more than 85% for high-risk CHD such as hypoplastic left heart syndrome (HLHS), atrioventricular septal defect (AVSD) and Ebstein's anomaly. However, to achieve a closed loop from "image" to "diagnosis", accurate anatomical structure segmentation of the 4CH section is still needed, and key clinical parameters (such as LV volume, RA / LA area ratio, tricuspid annular displacement TAPSE) are extracted to assist clinical decision-making. However, the acquisition of high-quality echocardiogram annotation data faces many challenges. First, the heart structure of newborns is small, the heart rate is fast, and the chest wall is thin, so the image is prone to lateral loss, artifacts and other problems, resulting in uneven image quality. Second, due to factors such as blurred endocardial boundary, papillary muscle obstruction and sound beam angle change, the difference in delineation of heart chamber boundary and valve attachment point by doctors with different years of experience can reach 15%-30%, which seriously affects the annotation consistency. In addition, the manual annotation process is time-consuming, usually taking 5-8 minutes for a single image, and requires long-term clinical experience accumulation. These factors together limit the size of high-quality annotation data that can be used to train deep learning models, so that a large number of unannotated or low-quality annotated images cannot effectively participate in model training, thereby affecting the generalization ability of the segmentation model.

[0063] It is particularly worth noting that in the segmentation of the boundary between the atrium and the ventricle, the existing method still has significant deficiencies. The ventricular structure is relatively stable, the boundary is clear, and it occupies a large area in the image, so the existing segmentation method has good recognition effect. However, the atrial structure has significant dynamic variability in shape, size, and location, especially in the neonatal population, due to the immature development of the heart, faster heart rate, and limited image resolution, there is significant deformation and displacement between different frames of the atrial region, making it difficult to accurately capture its boundary. In addition, the atrial wall is thin, the gray difference between it and the surrounding tissue is small, and it is often disturbed by structures such as papillary muscle and tendon, further exacerbating the boundary blur problem. In some images, the atrial boundary even visually "fuses" with the pericardium or ventricular structure, making it difficult for traditional segmentation methods based on gray threshold, edge detection or active contour to accurately identify its true contour. Even though deep learning-based models have improved overall segmentation performance, the segmentation accuracy of the atrial region is still significantly lower than that of the ventricle, severely affecting the clinical usability of the overall segmentation result. More complex is that the performance of the atrial structure under different pathological conditions also varies significantly. For example, the atrial shape is abnormal and the boundary is irregular, making it difficult for traditional segmentation methods to adapt to such complex changes, and prone to misclassification or omission. In addition, due to the large amplitude of the atrial wall movement, especially at the end of systole and the beginning of diastole, its boundary changes dramatically in the time dimension, which poses a higher challenge to segmentation methods based on single-frame images or simple temporal modeling.

[0064] That is, existing echocardiogram segmentation methods still have deficiencies in identifying the atrial and ventricular boundaries, especially when faced with dynamic changes in the atrial region, blurred boundaries, and complex structures. Traditional algorithms and existing deep learning models are difficult to achieve high-precision and robust automatic segmentation.

[0065] It can be understood that in the field of medical image segmentation, especially in the automatic segmentation task of key anatomical structures such as atrium and ventricle in echocardiogram, the U-shaped network has shown significant advantages in global semantic understanding and spatial information recovery due to its encoder-decoder structure and skip connection mechanism, and has become one of the mainstream segmentation frameworks. However, the existing U-shaped network-based methods still face a common technical problem in practical applications: due to the inevitable loss of high-frequency detail information in the multiple downsampling process of the network, combined with the low contrast, speckle noise and boundary blur of the ultrasound image itself, the model often has problems such as edge breakage, contour discontinuity or geometric deformation in the segmentation of tissue boundaries, which seriously affects the clinical usability of the segmentation result.

[0066] To solve the problem of insufficient boundary segmentation accuracy in the prior art, the application concept is that: by embedding a learnable boundary perception module in the encoder path of the U-shaped network, explicit modeling and feature enhancement of the edge region of the target structure are realized. The module can capture and strengthen key clues such as gradient change, texture difference, and contour continuity related to the boundary in the early stage of feature extraction, so that the network can more accurately restore the fine boundary structure in the subsequent decoding and upsampling process.

[0067] Further, to guide the network to pay more attention to the segmentation quality of the boundary region during training, a composite loss function mechanism that fuses boundary perception is adopted. Based on the traditional Dice loss and cross-entropy loss used for overall segmentation performance optimization, a detail loss term (DetailLoss) between the boundary prediction map generated by the boundary perception module and the real boundary label is introduced, forming a multi-objective joint optimization goal. The boundary loss term can apply stronger gradient feedback to the edge missegmentation region, so that the model not only optimizes the overall segmentation accuracy during backpropagation, but also synchronously improves the restoration ability of the boundary region.

[0068] That is, the application introduces a boundary perception module into the network structure and cooperates with a boundary-oriented loss optimization mechanism to construct a boundary optimization scheme that cooperatively drives the structure and loss. The scheme does not need to significantly change the original U-shaped network architecture, has strong compatibility, is easy to integrate, and can significantly improve the boundary segmentation accuracy, continuity, and anatomical rationality of key structures such as the heart in complex medical images without significantly increasing the computational overhead, which is of great significance for improving the clinical applicability of automatic segmentation models.

[0069] The new segmentation method of the application can dynamically adapt to changes in atrial morphology, enhance edge modeling capability, and improve boundary recognition accuracy to meet the clinical needs of accurate identification and quantitative analysis of atrial structure in neonatal congenital heart disease screening.

[0070] The technical solutions of the application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] The edge segmentation method of the ultrasonic image described in the preferred embodiment of the application, as shown in Figure 1 The edge segmentation method of the ultrasonic image includes the following steps:

[0072] In step S101, an initial model is constructed, and a training set is obtained. The training set is input to the initial model to obtain a prediction probability map.

[0073] Specifically, the initial model is a deep learning model for segmenting an ultrasound image (echocardiogram), and the initial model includes a Dice loss and a cross-entropy loss. The training set includes samples (label images) and corresponding true results, and the initial model is input by the training set to output a prediction probability map used for subsequent detail loss construction.

[0074] In step S102, data preprocessing is performed on the training set to obtain boundary supervision information.

[0075] In a possible implementation, the training set includes a plurality of label images, and the boundary supervision information includes a plurality of signed Euclidean distance maps. A plurality of initial distance maps are obtained according to the plurality of label images, wherein the distance values in the initial distance maps include positive numbers and negative numbers; the plurality of initial distance maps are resampled to a uniform resolution to obtain a plurality of intermediate distance maps; and the plurality of intermediate distance maps are respectively linearly mapped to obtain a plurality of signed Euclidean distance maps.

[0076] Specifically, the data preprocessing is performed to construct a boundary supervision signal (boundary supervision information) with geometric semantics. First, the original annotation data is finely preprocessed to generate a signed Euclidean distance map (Signed Euclidean Distance Map), which is used as core guiding information for a subsequent boundary perception module.

[0077] In a possible implementation, the label image includes a first region, a second region, and a third region. For each label image, the Euclidean distances of pixels in the first region, the second region, and the third region from a target boundary are calculated to obtain a first channel distance map corresponding to the first region, a second channel distance map corresponding to the second region, and a third channel distance map corresponding to the third region; and the initial distance map corresponding to the label image is obtained according to the first channel distance map, the second channel distance map, and the third channel distance map.

[0078] It can be understood that each label image includes a plurality of regions to be segmented, such as a background (first region), a left atrium (second region), and a left ventricle (third region). A distance map is generated for each region to be segmented, and then the three channels are stacked together to form a distance map.

[0079] Specifically, in the distance map generation process, for each labeled image in the training set, its corresponding Euclidean distance transform is calculated. Further, for each pixel point (i, j), the shortest Euclidean distance to the nearest target boundary is calculated, and a sign attribute is assigned: if the pixel is located inside the target region (such as inside the left atrium or left ventricle), the distance value is negative; if it is located outside the target region, the distance value is positive; if it is exactly located on the boundary, the distance value is 0. This signed distance not only reflects the geometric distance (absolute value) of the pixel from the real boundary, but also clearly indicates its relative position (inside / outside) with respect to the boundary through the sign, thereby providing directional prior knowledge for the model. After the distance maps (such as left ventricle LV, left atrium LA) of all categories to be evaluated are independently calculated, they are stacked into a multi-channel tensor along the channel dimension, and saved in npy format. Each distance map file is consistent with the corresponding input image in spatial resolution (HxW) and file naming, ensuring accurate matching during training. It should be noted that the distance map of the present application as a "soft label" form of supervision signal can guide the model to perceive the deviation direction and degree of the current prediction result from the real boundary in backpropagation, and is the basis for realizing the fine optimization of the boundary.

[0080] It can be understood that the default generated distance map is in the format of (C, H, W) (channel first), while most deep learning frameworks use the memory layout of (H, W, C) in image processing. In order to keep the data format uniform, facilitate batch loading and tensor operation, the channel order transposition operation is performed on the distance map to adjust it to the structure of (H, W, C).

[0081] Since there are differences in image sizes in the original data set (due to different patients or device acquisition conditions), in order to ensure the consistency of model input and improve the stability of training, size resampling is performed, that is, all distance maps are uniformly resampled to the target resolution (such as 256x256) using the bilinear interpolation method. This operation enhances the robustness of the model to scale changes while preserving the geometric relationship of the boundary.

[0082] Then numerical normalization is performed to linearly map the numerical range of the resampled distance map to the interval [-1, 1]. This normalization strategy helps maintain the stability of the gradient during backpropagation, avoids oscillation in the optimization process due to the large dimension of the distance value, and improves the training convergence efficiency and generalization ability.

[0083] The present application provides a basis for subsequent boundary optimization by generating high-quality supervision signals (distance maps) through original labels-signed distance maps-format transposition / resampling / normalization.

[0084] The present application aims to solve the problems of poor stability and weak adaptability to irregular boundaries in the actual training process. By introducing a preprocessing mechanism and loss function structure reconstruction, the model's perception ability of the target boundary and the robustness of the training process are significantly improved. In the resampling and normalization-based distance map preprocessing mechanism, unlike traditional methods that directly use the original Euclidean distance map as the boundary supervision signal, the distance map preprocessing strategy of the present application is used as the front-end input processing link of the boundary optimization module. Specifically, first, the Euclidean distance transform map of the original binary segmentation label is calculated, and then the distance map is resampled to match the resolution of the intermediate feature map in the network, avoiding information misplacement or interpolation errors caused by scale mismatch. More importantly, by introducing a dynamic normalization mechanism, the distance values are mapped to a unified and compact numerical interval (such as [0, 1]), effectively suppressing the large distance value range caused by the complexity difference of the boundary between different samples. After this preprocessing process, the resulting distance map not only has higher spatial consistency and smoothness, but also enhances the comparability across samples and scales, providing a high-quality supervision signal basis for stable calculation of subsequent boundary loss. This preprocessing module can be seamlessly integrated into the skip connection path of the U-shaped network or the front end of loss calculation as a plug-in component, significantly improving the model's sensitivity to fine-grained boundaries.

[0085] In step S103, boundary perception information is obtained according to the prediction probability map and the boundary supervision information.

[0086] In one possible implementation, the number of prediction probability maps is multiple, and each prediction probability map corresponds to a symbolic Euclidean distance map. The boundary perception information includes error product data and direction guidance factors. According to the multiple prediction probability maps and the corresponding multiple symbolic Euclidean distance maps, the product of each prediction probability map and the corresponding symbolic Euclidean distance map is calculated to obtain error product data; and the error product data is processed to obtain direction guidance factors. Figure One

[0087] It should be noted that, as shown in Figure 2 The present application constructs a boundary perception module based on symbolic distance supervision, and realizes fine-grained and interpretable regulation of segmentation boundaries by introducing detail loss and direction guidance factors.

[0088] ​Specifically, an input loading and caching mechanism is used, in the training phase, the system automatically matches the distance map file with the same name according to the input image path, ensuring that each training sample can obtain the corresponding spatial distance supervision signal. In order to improve the data reading efficiency, all distance maps are cached in the memory after the first loading, reducing the I / O overhead and accelerating the training speed.

[0089] In the detail loss construction process, let the prediction probability map output by the model be p c ∈[0,1] H×W , and the corresponding real symbolic distance map be Define the error product term of each pixel point as: e i,j = p c (i,j) x d c (i,j). The product term integrates the prediction confidence and geometric position information, and its sign contains the direction requirement of boundary adjustment: in the model training process, when d c >0 (external), if e i,j >0, it means that the prediction area is over-expanded and needs to be contracted inward; when d c <0 (internal), if e i,j <0, it means that the prediction area is insufficient and needs to be expanded outward. Further, the absolute value of the error product is taken as the basis of the loss term to form the detail loss: Where N represents the number of error products, and this term encourages the model to minimize the overall deviation of the predicted boundary and the real boundary.

[0090] In one possible implementation, the error product data is nonlinearly compressed to obtain real data in the positive number interval; the real data is offset adjusted to the positive and negative number range to obtain a direction guide factor.

[0091] Specifically, in order to provide a stable and learnable directional guidance signal in backpropagation, the present application designs a direction guide factor g i,j , whose calculation process is as follows: first, the original error product e i,j is nonlinearly compressed by applying a Sigmoid function: This step maps the infinite range of real numbers to the [0,1] interval, making the signal smoother and the gradient more stable. Then subtract 0.5, that is, g i,j = s i,j -0.5 adjusts the output range to [-0.5, 0.5].

[0092] It is worth noting that in the present application, the positive and negative of the direction guide factor determines the adjustment direction, and the sign of the direction guide factor g i,j directly indicates how the model should adjust the prediction boundary: if g i,j >0: indicates that the current prediction is too small (under-segmentation), and the prediction area needs to be expanded; if gi,k <0: indicates that the current prediction is too large (over-segmentation), and the prediction region needs to be contracted. This symbol-based feedback mechanism enables the model to dynamically perceive "which way to adjust" according to the local geometric relationship. The absolute value reflects the adjustment urgency, |g i,j The size of | quantifies the adjustment intensity required at this position: the larger the absolute value, the more significant the gap between the current prediction and the true boundary, and the more urgent the adjustment demand; the closer the absolute value is to 0, the more the prediction fits the boundary, and the smaller the adjustment amplitude. For example, in regions with ambiguous boundaries, noise interference, or complex structures, |g i,j | significantly increases, prompting the model to focus attention on correction. The advantage of normalization, using the Sigmoid function and subtracting 0.5, not only ensures the continuity and differentiability of the direction signal, but also avoids the problem of gradient explosion caused by extreme values, improving the stability of the training process.

[0093] The direction guidance factor achieves fine and interpretable control of the boundary region by separating the "loss amount" (used for optimization target) from the "direction signal" (used for guiding optimization path), significantly enhancing the model's ability to capture weak boundaries and ambiguous edges.

[0094] The present application explicitly enhances the boundary perception ability by combining the interaction-detail loss of the prediction probability map and the distance map (global deviation optimization) with the direction guidance factor (local adjustment guidance).

[0095] In the present application, an innovative boundary perception module is introduced into the UNet network architecture, which skillfully connects the encoder and the decoder. This design significantly enhances the model's sensitivity and expression ability to the edge details of the target object, ensuring that the smoothness and geometric rationality of the segmentation boundary are improved without sacrificing the overall accuracy and continuity of the segmentation result. Especially in complex medical image segmentation tasks, this module shows stronger robustness and adaptability. In addition, due to its lightweight structure, it can be seamlessly integrated into existing U-shaped network architecture, and high-quality segmentation results with well-preserved edges can be generated without additional post-processing steps, greatly improving its practical value and promotion potential.

[0096] In step S104, a target model is obtained according to the boundary perception information, the training set, and the initial model.

[0097] In one possible implementation, a detail loss is obtained according to the error product data, and a direction guidance factor loss is obtained according to the direction guidance factor; the initial model is updated according to the detail loss and the direction guidance factor loss to obtain a target model with parameters to be confirmed; the target model with parameters to be confirmed is trained for a preset number of training rounds according to the training set to obtain a target model with confirmed parameters.

[0098] It is worth noting that the traditional method directly multiplies the distance map dc and the prediction probability map pc to obtain the loss product, which is easily affected by the irregularity of the boundary geometry and the absolute size of the distance value, leading to an imbalance in the local loss contribution, and further causing gradient explosion or training shock. Therefore, the present application first normalizes the distance map in the preprocessing stage to preliminarily smooth its dynamic range; further, the original product form is decoupled into two independent and explainable components: a detail loss term and a direction guiding factor. Among them, the detail loss term retains the basic interaction of dc and pc, while the direction guiding factor applies a Sigmoid function to the loss product of the normalized distance map and the probability map to achieve soft differentiation and secondary smoothing of the inside and outside regions of the boundary. More importantly, these two components are given independent learnable or configurable weight coefficients, allowing the influence strength of the two to be flexibly adjusted according to task requirements during training. This decoupled structure not only makes the overall change of the loss function more stable, improving the convergence and stability of the training process, but also realizes the fine-tuning of the model's boundary learning behavior. The refined boundary loss function ultimately formed shows stronger robustness and higher segmentation accuracy in various complex boundary scenarios, especially in the recovery of small structures and high-curvature boundaries.

[0099] In a possible implementation, the target model to be confirmed is represented as:

[0100]

[0101] wherein, is the total loss function of the target model, is the Dice loss, and a is the weight coefficient of the Dice loss; is the cross-entropy loss, and β is the weight coefficient of the cross-entropy loss; is the detail loss, and γ is the weight coefficient of the detail loss; is the direction guiding factor loss, and δ is the weight coefficient of the direction guiding factor loss.

[0102] Specifically, the total loss function of the target model is composed of four weighted terms, among which the direction guiding factor loss is newly added. The Dice loss focuses on the region overlap rate, improving the overall segmentation consistency; the cross-entropy loss enhances the class discrimination ability; the detail loss (absolute value of the product) focuses on boundary optimization; the direction guiding factor loss is used to provide a stable and learnable directional guiding signal; the weight coefficients are determined by grid search on the validation set to determine the optimal combination, ensuring balanced contribution of each loss.

[0103] In the training process, the training is cycled in epochs, and the maximum number of training rounds is set to 1000; after the end of each epoch, the performance is evaluated on the validation set, and the model weight with the best performance is saved; Gradient Clipping is introduced to prevent gradient explosion; Early Stopping is set, which terminates the training when the validation loss does not decrease for a number of rounds, preventing overfitting and ensuring that the boundary perception module does not damage the stability of the backbone network. It can be understood that the validation set is similar to the training set, the data amount of the validation set is less than that of the training set, and the validation set is used to screen the target model whose parameters have been determined.

[0104] On the basis of the traditional Dice loss and Cross-Entropy (CE) loss, the present application faces the Detail Loss for boundary fine reconstruction. The Dice loss and the CE loss as the core optimization indicators play an important role in the overall regional segmentation of the left atrium and the left ventricle of the cardiac ultrasound image. However, in order to further improve the segmentation accuracy and smoothness of the key anatomical structure, especially the edge region of the left atrium, the present application combines the detail error term based on the local structure difference calculation and the Directional Guidance Factor, which can guide the model prediction to tend to the geometric trend of the real anatomical structure. This composite loss strategy not only ensures the accurate segmentation of the atrium and ventricle overall region, but also effectively reduces the boundary sawtooth and over-smoothing phenomenon, and the experimental results prove its effectiveness.

[0105] In step S105, an echocardiogram of a user is acquired, and the echocardiogram is input to the target model to obtain a segmentation probability map.

[0106] Specifically, through model prediction and result output, in the inference stage, the best model weight saved in the training process is used to forward infer the test set image, generate a segmentation probability map, and the prediction result is saved in PNG format, the color mode adopts P palette mode, and the annotation category is consistent with the original label.

[0107] The present application implements the Gradient Clipping mechanism for the possible gradient abnormality problem in the model training process, and continuously monitors the Dice value on the validation set through the Early Stopping mechanism, and terminates the training as soon as the Dice value is found to be significantly decreased, thereby improving the stability and reliability of the training process. It is worth noting that through a large number of experimental verification, under the condition of reasonable setting of loss weight and training parameters, these protection measures are usually not triggered, which indicates that the proposed phased loss weighting strategy itself has strong stability and anti-overfitting ability. Therefore, the training strategy of the present application is not only reasonable, but also efficient, and provides strong support for the medical image segmentation field.

[0108] The application introduces a designed boundary perception module on the basis of the nnUNetv2 model architecture, and combines a boundary-oriented loss optimization mechanism to construct a boundary segmentation optimization scheme driven by structure and loss in cooperation, aiming to improve the description ability of the U-shaped network to the target boundary in the medical image segmentation task, especially for the fine segmentation of key structures such as left atrium (LA) and left ventricle (LV) in echocardiogram.

[0109] In order to verify the technical effect of the application, a clinical data set is used for experimental evaluation, which includes 4CH view ultrasound video sequences of 50 patients (35 healthy and 15 sick), and 1888 static images are formed after frame extraction. All images are stored in PNG format, wherein the input image is in RGB three-channel format, and the annotation image adopts P palette mode, and the annotation categories include: 0-background (Background), 1-left ventricle (LV), and 2-left atrium (LA). All annotation work is independently completed by 3 to 4 experienced ultrasonographers and reviewed by experts to ensure the accuracy and clinical reliability of the annotation results.

[0110] The model performance evaluation adopts two core indicators: the Dice similarity coefficient is used to measure the overlap accuracy of the overall segmentation, and the 95% Hausdorff distance (HD95) is used to evaluate the geometric consistency of the boundary segmentation. Selecting HD95 instead of the traditional HD can effectively exclude the influence of extreme outliers, making the boundary performance evaluation more robust and representative.

[0111] The experimental results show that, without changing the main structure of nnUNetv2, after introducing the boundary perception module proposed in the application, the segmentation performance of the model on the test set is as follows:

[0112] Baseline model (without boundary perception module): LV and LA overall Dice: 0.94, LV single indicator: Dice=0.94, HD95=162, LA single indicator: Dice=0.89, HD95=812.

[0113] Target model of the application (including boundary perception module and boundary optimization loss): LV and LA overall Dice: 0.94 (unchanged), LV single indicator: Dice=0.94, HD95=160 (slightly improved), LA single indicator: Dice=0.89, HD95=636 (significantly decreased, optimization rate up to 21.3%).

[0114] The above results show that the application significantly improves the recognition and segmentation ability of the model to the target boundary without sacrificing the overall segmentation accuracy, especially in the complex structure and fuzzy boundary of the left atrium region, which shows better edge fitting. This shows that the proposed boundary perception mechanism can effectively enhance the sensitivity of the network to edge features, and alleviate the edge missing or misclassification phenomenon caused by ultrasound image noise, low contrast and boundary discontinuity.

[0115] It can be understood that the application is based on the U-shaped network architecture, by constructing the soft distance graph of the training sample and jointly calculating the boundary perception loss function with the probability graph output by the network, and introducing a multi-round smoothing processing strategy to optimize the boundary gradient, thereby significantly improving the positioning accuracy of the model to the key anatomical boundaries such as endocardium and epicardium. The core of the boundary perception mechanism is to integrate geometric distance information into the loss function design, and strengthen the learning ability of the model to the edge region. However, this design idea has good universality and scalability, and is not limited to the U-shaped network structure itself. In fact, this boundary perception strategy can be flexibly migrated to other mainstream segmentation network frameworks, for example, it can be embedded into Swin-Unet, TransUNet and other Transformer-based models, and the distance perception loss can be introduced at the decoder end to make up for the shortcomings of the pure attention mechanism in local boundary preservation, and it can also be integrated into SAM (Segment Anything Model) as post-processing loss or fine-tuning stage supervision signal to enhance its perception ability to fine boundaries in medical images. In addition, this boundary optimization idea can also be extended to DeepLab series and other adaptive segmentation frameworks, as an auxiliary loss term to improve the robustness of the model in low-contrast and severe motion artifact ultrasound images. From the perspective of technology evolution and competition avoidance, possible future deformation schemes include: using different distance transformation methods to construct soft labels (such as weighted distance graph, direction-sensitive distance field), replacing the smoothing strategy (such as using Gaussian filtering, bilateral filtering or a learnable smoothing network layer), or combining edge detection networks (such as HED) to generate more accurate boundary supervision signals.

[0116] The loss function used in the application is proposed based on the specific task characteristics of atrial and ventricular segmentation, which can effectively optimize the segmentation ability of the model to the classification sub-label. The loss function can be defined according to the specific task, and the specific indicators of the loss function can be improved, such as using HD95 loss to supplement or replace the detail loss, etc. As long as the core idea is to use geometric prior information to enhance boundary learning, there are many variants of the loss function, and using different loss functions in different tasks will have different effects, but it has not been proven which loss is the best effect.

[0117] The application effectively enhances the sensitivity of the model to the target boundary by calculating the boundary loss between the soft distance map and the probability map of the training image and combining a multi-round smoothing processing strategy, thereby improving the fineness of the segmentation result. In the current implementation, the boundary loss is weighted and fused into the total loss function by a static weight, and the weight value is determined on the validation set by grid search, which can achieve better performance under a specific data distribution. Although the current static weighting strategy is simple to implement and stable in training, from the optimization point of view, introducing a dynamic weight adjustment mechanism can theoretically adaptively adjust the boundary loss during the training process according to the model learning state (such as loss change rate, gradient amplitude or boundary region uncertainty), thereby achieving more fine-grained training control. The learning rate scheduling idea, gating mechanism or reinforcement learning strategy can also be used to achieve this goal.

[0118] The application is not only applicable to the segmentation task of 2D medical images such as echocardiograms, but also the core design concept of the application, i.e. the boundary perception module, can be extended to 3D tasks by calculating the boundary loss between the soft distance map and the probability map of the training picture and through multi-round smoothing processing. For example, when dealing with complex heart structures or blood vessel systems, the boundary perception module can be directly applied to 3D ultrasound images, CT scans or multi-dimensional MR images without significant modification, thereby further improving the segmentation accuracy by accurately capturing the boundary information in three-dimensional space. In addition, considering the application requirements in different medical scenarios, the application is not limited to specific types of medical images or analysis of specific anatomical sites. By appropriately adjusting the parameters or optimizing the algorithm model, especially the calculation of the soft distance map and the probability map and the optimization strategy of the boundary loss, the design idea of the application can be extended to other medical image processing fields such as brain tumor segmentation and abdominal multi-organ segmentation. Therefore, whether for more complex image data formats or diverse clinical application scenarios, the application provides a flexible and widely applicable basic framework.

[0119] Secondly, the edge segmentation system of the ultrasound image according to the embodiment of the application is described with reference to the accompanying drawings, which is applied to the edge segmentation method of the ultrasound image in any of the above-mentioned schemes.

[0120] Figure 3 is a structural diagram of the edge segmentation system of the ultrasound image according to the embodiment of the application.

[0121] As Figure 3 shown, the edge segmentation system of the ultrasound image includes a probability prediction module 100, a data preprocessing module 200, a boundary perception module 300, a model training module 400 and an image segmentation module 500.

[0122] Specifically, the probability prediction module 100 is configured to construct an initial model, obtain a training set, input the training set into the initial model, and obtain a prediction probability map;

[0123] The data preprocessing module 200 is configured to perform data preprocessing on the training set to obtain boundary supervision information.

[0124] The boundary perception module 300 is configured to obtain boundary perception information according to the prediction probability map and the boundary supervision information.

[0125] The model training module 400 is configured to obtain a target model according to the boundary perception information, the training set and the initial model.

[0126] The image segmentation module 500 is configured to obtain an echocardiogram of a user, input the echocardiogram into the target model, and obtain a segmentation probability map.

[0127] Figure 4 The terminal provided by the embodiment of the present application is shown in a structural diagram. The terminal can include:

[0128] The memory 501, the processor 502 and the computer program stored in the memory 501 and executable on the processor 502.

[0129] The processor 502 implements the edge segmentation method of the ultrasonic image provided in the above embodiment when executing the program.

[0130] Further, the terminal further includes:

[0131] The communication interface 503 is configured to communicate between the memory 501 and the processor 502.

[0132] The memory 501 is configured to store the computer program executable on the processor 502.

[0133] The memory 501 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0134] If the memory 501, the processor 502 and the communication interface 503 are implemented independently, the communication interface 503, the memory 501 and the processor 502 can be connected with each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 4 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0135] Optionally, in a specific implementation, if the memory 501, the processor 502 and the communication interface 503 are integrated on a chip, the memory 501, the processor 502 and the communication interface 503 can complete communication between each other through an internal interface.

[0136] The processor 502 can be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0137] The embodiment also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the edge segmentation method of the ultrasonic image.

[0138] An embodiment of the present application provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the edge segmentation method of the ultrasonic image. Figure 1 corresponding to any of the embodiments of the edge segmentation method of the ultrasonic image.

[0139] In the description of the application, reference to "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that a particular feature, structure, material, or characteristic being described is included in at least one embodiment or example of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment or example. Furthermore, the described specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples. In addition, the usage of "N" means at least two, for example, two, three or the like, unless explicitly stated otherwise.

[0140] Furthermore, the terms "first", "second", or the like, are used merely as a designation of certain elements or features of the application, and do not imply or connote relative importance or a specific order of precedence. Thus, features defined with "first", "second", etc. can include at least one of the features, either explicitly or implicitly.

[0141] Any process or method descriptions or blocks in flow charts or otherwise described herein represent embodiments of modules, segments, or portions of code which include one or more executable instructions for implementing specific logic functions or steps, and alternate implementations are possible. In some embodiments, the processes or methods described in flow charts or otherwise described herein are not necessarily performed in the order shown or discussed, including, for example, performing or depending from other operations or stages, in parallel, in reverse order, or in other orders.

[0142] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of instructions to implement logic functions, and can be embodied in any computer-readable storage medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, processor- containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable storage medium" can be any means that can contain, store, communicate, propagate or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable storage medium can be a computer- readable storage medium that can be any media that can be used to store the desired program instructions in a form readable by a computer. More specific examples (a non-exhaustive list) of the computer-readable storage medium include the following: an electrical connection having one or more wires (electrical, optical, and / or other communications medium), a portable computer diskette, a hard disk, a RAM, a ROM, an erasable programmable ROM (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable storage medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via, for example, optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0143] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. As such, if implemented in hardware, and in another embodiment, any of the following technologies, known in the art, or their combinations can be used: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0144] Those skilled in the art can understand that all or part of the steps carried out by the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer-readable storage medium. When the programs are executed, one or a combination of the steps of the method embodiments is included.

[0145] In addition, each of the function units in each of the embodiments of the present application can be integrated in one processing module, or each unit can be physically present separately, or two or more units can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software function module. When the integrated module is realized in the form of a software function module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0146] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

[0147] It should be understood that the application of the present application is not limited to the above examples, and those skilled in the art can improve or change the above examples according to the above description, and all these improvements and changes should belong to the protection scope of the claims attached to the present application.

[0148] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An edge segmentation method of an ultrasound image, characterized by, The edge segmentation method of the ultrasound image comprises: An initial model is constructed, and a training set is obtained, the training set is input into the initial model to obtain a prediction probability map; Data preprocessing is performed on the training set to obtain boundary supervision information; Boundary perception information is obtained according to the prediction probability map and the boundary supervision information; A target model is obtained according to the boundary perception information, the training set and the initial model; An echocardiogram of a user is obtained, and the echocardiogram is input into the target model to obtain a segmentation probability map.

2. The edge segmentation method of an ultrasound image according to claim 1, characterized by, The training set comprises a plurality of label images, and the boundary supervision information comprises a plurality of symbolic Euclidean distance maps; The data preprocessing on the training set to obtain the boundary supervision information specifically comprises: A plurality of initial distance maps are obtained according to the plurality of label images, wherein the distance values in the initial distance maps comprise positive numbers and negative numbers; The plurality of initial distance maps are resampled to a uniform resolution to obtain a plurality of intermediate distance maps; The plurality of intermediate distance maps are respectively linearly mapped to obtain a plurality of symbolic Euclidean distance maps.

3. The edge segmentation method of an ultrasound image according to claim 2, characterized in that, The label images comprise a first region, a second region and a third region; The plurality of initial distance maps obtained according to the plurality of label images specifically comprise: For each label image, the Euclidean distances of the pixels in the first region, the second region and the third region from a target boundary are respectively calculated to obtain a first channel distance map corresponding to the first region, a second channel distance map corresponding to the second region and a third channel distance map corresponding to the third region; The initial distance map corresponding to the label image is obtained according to the first channel distance map, the second channel distance map and the third channel distance map.

4. The edge segmentation method of an ultrasound image according to claim 2, characterized by, The number of prediction probability maps is a plurality, and the plurality of prediction probability maps correspond one-to-one to a plurality of symbolic Euclidean distance maps, and the boundary perception information comprises an error product data direction guidance factor; The boundary perception information obtained according to the prediction probability map and the boundary supervision information specifically comprises: According to the plurality of prediction probability maps and the corresponding plurality of symbolic Euclidean distance maps, the product of each prediction probability map and the corresponding symbolic Euclidean distance map is calculated to obtain error product data; The error product data is processed to obtain a direction guidance factor.

5. The edge segmentation method of an ultrasound image according to claim 4, characterized in that, The error product data is processed to obtain a direction guidance factor, specifically comprising: The error product data is nonlinearly compressed to obtain real data in a positive number interval; The real data is offset adjusted to a positive and negative number range to obtain a direction guidance factor.

6. The edge segmentation method of an ultrasound image according to claim 4, characterized by, The target model obtained according to the boundary perception information, the training set and the initial model specifically comprises: A detail loss is obtained according to the error product data, and a direction guidance factor loss is obtained according to the direction guidance factor; The initial model is updated according to the detail loss and the direction guidance factor loss to obtain a target model with parameters to be confirmed; The target model with parameters to be confirmed is trained according to the training set for a preset number of training rounds to obtain a target model with confirmed parameters.

7. The edge segmentation method of an ultrasound image according to claim 6, characterized in that, The target model to be confirmed is represented as: wherein, is a total loss function of the target model, is a Dice loss, and a is a weight coefficient of the Dice loss; is a cross-entropy loss, and β is a weight coefficient of the cross-entropy loss; is a detail loss, and γ is a weight coefficient of the detail loss; is a direction guidance factor loss, and δ is a weight coefficient of the direction guidance factor loss.

8. An edge segmentation system of an ultrasound image, characterized by, The edge segmentation system of the ultrasonic image is applied to the edge segmentation method of the ultrasonic image according to any one of claims 1-7; the edge segmentation system of the ultrasonic image comprises: A probability prediction module is configured to construct an initial model, obtain a training set, input the training set into the initial model, and obtain a prediction probability map; A data preprocessing module is configured to perform data preprocessing on the training set to obtain boundary supervision information; A boundary perception module is configured to obtain boundary perception information according to the prediction probability map and the boundary supervision information; A model training module is configured to obtain a target model according to the boundary perception information, the training set and the initial model; An image segmentation module is configured to obtain an echocardiogram of a user, input the echocardiogram into the target model, and obtain a segmentation probability map.

9. A terminal, characterized by comprising: The terminal comprises a memory, a processor, and an ultrasonic image edge segmentation program stored on the memory and executable on the processor, and the ultrasonic image edge segmentation program implements the steps of the edge segmentation method of the ultrasonic image according to any one of claims 1-7 when executed by the processor.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores an ultrasonic image edge segmentation program, and the ultrasonic image edge segmentation program implements the steps of the edge segmentation method of the ultrasonic image according to any one of claims 1-7 when executed by the processor.

Images