An ultrasonic imaging diagnosis method based on artificial intelligence

Abstract

The present invention discloses an ultrasonic imaging diagnosis method based on artificial intelligence, comprising: step S1: collecting ultrasonic images of patients; step S2: extracting organ images and lesion images from ultrasonic images; step S3: generating morphological scores according to lesion images, and generating position scores according to organ images and lesion images; step S4: generating comprehensive scores according to morphological scores and position scores; the comprehensive score is used to characterize the probability of metastasis of lesions. The beneficial effects of the present invention are: comprehensive prediction of diseases based on morphological scores and position scores to improve the accuracy of prediction, and reasonable prediction of related lesions of various lesions or organs can be effectively made; unified transformation of large-scale statistical data is achieved based on image processing methods, so that large-scale collected data can be fused at the same scale. Effective analysis is carried out from a quantitative, objective and interpretable perspective based on image features, which has further analytical value and easy transformation of results.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence diagnosis, and in particular to an ultrasonic imaging diagnosis method based on artificial intelligence. Background Art

[0002] Medical diagnosis based on ultrasound images is painless and convenient, and is one of the important trends in clinical research. Ultrasound doctors make medical diagnoses based on characteristic signals such as the location, size, and morphological information of the lesion indicated in the ultrasound image. However, due to the severe shortage of ultrasound doctors and uneven experience, the interpretation of ultrasound images can cause large deviations, resulting in large errors in diagnostic results and extremely high rates of misdiagnosis and missed diagnosis. Clinical practice requires standardized auxiliary diagnostic methods with strong explanatory power.

[0003] In recent years, due to the continuous improvement of computing power and algorithms, artificial intelligence methods combined with medical imaging research have achieved many remarkable results. Artificial intelligence methods can effectively extract and analyze the features of ultrasound images, such as finding lesions (nodules or tumors), calculating the size of lesions, identifying morphological features, and distinguishing benign and malignant tumors. In many aspects, they can match or even exceed the accuracy of human doctors and have great potential. At present, the use of ultrasound images mainly stays in the statistical research of small samples for image morphological indicators, that is, the data of simple preprocessing of original images or segmentation images, and then directly using artificial intelligence methods such as machine learning and deep learning to predict image diagnosis based on morphological features. The results mainly reflect the characteristics of the lesions themselves and the prediction scores. At present, there are few predictions about related lesions, and there is no way to indicate the basis for scoring. Summary of the invention

[0004] In view of the above problems existing in the prior art, an ultrasonic imaging diagnosis method based on artificial intelligence is now provided.

[0005] The specific technical solutions are as follows:

[0006] An ultrasonic imaging diagnosis method based on artificial intelligence, comprising:

[0007] Step S1: Acquire ultrasound images of the patient;

[0008] Step S2: extracting organ images and lesion images from the ultrasound images;

[0009] Step S3: generating a morphological score according to the lesion image, and generating a position score according to the organ image and the lesion image;

[0010] Step S4: generating a comprehensive score according to the morphology score and the position score;

[0011] The comprehensive score is used to characterize the metastasis probability of the lesion.

[0012] Preferably, step S2 comprises:

[0013] Step S21 pre-processes the plurality of ultrasound images to remove invalid ultrasound images;

[0014] In step S22, a first image segmentation method is used to extract the organ image from the ultrasound image, and a second image segmentation method is used to extract the lesion image from the ultrasound image.

[0015] Preferably, the first image segmentation method in step S22 includes:

[0016] Step S221: preprocessing the ultrasound image to generate a preprocessed image;

[0017] Step S222: extracting the organ image from the pre-processed image using a network recognition model.

[0018] Preferably, the second image segmentation method in step S22 includes:

[0019] extracting the lesion image from the ultrasound image using an object segmentation model;

[0020] The object segmentation model includes:

[0021] a feature pyramid network, wherein the feature pyramid network receives the ultrasound image and generates an object segmentation image according to the ultrasound image;

[0022] a relationship module, the relationship module receiving the object segmentation image and generating a foreground enhanced image according to the object segmentation image;

[0023] A decoding module, the decoding module receives the foreground enhanced image and generates a semantically restored image;

[0024] An optimization module receives the semantically restored image and generates the lesion image.

[0025] Preferably, the step S3 comprises: using a pre-established binary classification model to generate a morphological score according to the lesion image;

[0026] The binary classification model includes:

[0027] A feature extraction layer, wherein the feature extraction layer performs a convolution operation on the ultrasound image, and then performs four residual block operations to output an extraction result;

[0028] A classification layer is used to generate the morphology score according to the extraction result.

[0029] Preferably, the step S3 further comprises: using a location recognition method to generate a location score according to the organ image and the lesion image;

[0030] The location identification method comprises:

[0031] Step S31: performing registration according to the ultrasound image and a preset organ template to generate a registered image;

[0032] Step S32: extracting position information from the registered image;

[0033] Step S33: using a graph algorithm model to generate the location score according to the location information.

[0034] Preferably, the graph algorithm model includes:

[0035] A bipartite graph submodule, wherein the bipartite graph submodule generates a bipartite graph according to the position information;

[0036] A graph convolution submodule, wherein a plurality of graph convolution layers and a classification pooling layer are provided in the graph convolution submodule, and are used to extract structural features of the bipartite graph;

[0037] A classification submodule is used to generate the position score according to the structural characteristics of the bipartite graph.

[0038] Preferably, the step S4 further comprises generating a probability distribution diagram using a visualization method.

[0039] The above technical solution has the following advantages or beneficial effects: Comprehensive prediction of diseases based on morphological scores and location scores to improve the accuracy of prediction, which can effectively and reasonably predict the relevant lesions of various lesions or organs; based on image processing methods, unified transformation of large-scale statistical data is achieved, so that large-scale collected data can be integrated at the same scale. Based on image features, effective analysis is carried out from a quantitative, objective and interpretable perspective, which has further analytical value and is easy to transform. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The embodiments of the present invention will be described more fully with reference to the attached drawings, which are provided for illustration and description only and are not intended to limit the scope of the present invention.

[0041] Figure 1 is an overall schematic diagram of an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of sub-steps of step S2 of an embodiment of the present invention;

[0043] Figure 3 is a schematic diagram of a first image segmentation method according to an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of an object segmentation model according to an embodiment of the present invention;

[0045] Figure 5 Schematic diagram of a binary classification model according to an embodiment of the present invention;

[0046] Figure 6 A schematic diagram of a location identification method according to an embodiment of the present invention;

[0047] Figure 7 Schematic diagram of a graph algorithm model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0050] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0051] At present, the technical route of combining artificial intelligence with ultrasound imaging mainly focuses on feature extraction and prediction model training based on the morphology of the lesion itself, while ignoring the important influence of some information such as the lesion growth area, the distance between lesions and the distance between lesions and various organs on the disease. This leads to a single research factor, a certain deviation between the prediction results and the medical clinical significance, and the loss of further analysis of more useful information.

[0052] At the same time, since the artificial intelligence algorithm itself is a "black box" method, that is, the prediction results cannot be analyzed for specific directional factors, it is also impossible to give a reasonable explanation for the prediction results when processing medical imaging data. The cause is more important in medical diagnosis, and the cause is sometimes the result of multiple factors working together, so the interpretability of the prediction results is forced and necessary.

[0053] Ultrasound scanning is the main diagnostic method for diseases such as thyroid, lung, breast, pancreas, and gastrointestinal tract. Due to the squeezing during the scanning process, even the organs of the same person have different sizes and resolutions. Artificial intelligence methods such as deep learning rely heavily on the prediction of the morphological characteristics of the lesions themselves, so this poses a considerable technical challenge for training and identifying data with distorted detection targets, different input sizes, and different resolutions.

[0054] On the other hand, the correlation analysis between lesions based on specific diseases is also neglected by current technology. Current research only predicts the diagnosis of specific diseases based on the lesions present in a given ultrasound image, and rarely conducts corresponding research on diseases associated with the lesions. This also results in the loss of relevant factors in clinical practice and the opportunity to use retrospective data for cutting-edge analysis.

[0055] In view of the above technical problems, the present invention provides an ultrasonic imaging diagnosis method based on artificial intelligence, such as Figure 1 As shown, including:

[0056] Step S1: Acquire ultrasound images of the patient;

[0057] Step S2: extracting organ images and lesion images from the ultrasound images;

[0058] Step S3: generating a morphological score according to the lesion image, and generating a position score according to the organ image and the lesion image;

[0059] Step S4: Generate a comprehensive score based on the morphology score and the location score.

[0060] Specifically, the ultrasound imaging diagnosis method proposed in this technical solution is mainly divided into three parts: morphological scoring, position scoring and comprehensive scoring. Among them, the morphological scoring is used to classify and study lesions and judge the current state of lesions; the position scoring is used to judge the development of the disease and predict the possible progression of the disease based on the relative position of the lesion and the organ; finally, a comprehensive score is generated based on the morphological scoring and position scoring.

[0061] Taking cervical lymph node metastasis (LNM) of papillary thyroid carcinoma as an example, the accuracy of the existing technology for preoperative prediction of cervical lymph node metastasis is often insufficient, and thus it is impossible to effectively avoid over-medicalization and missed diagnosis. Ultrasound imaging is usually used in the existing technology for LNM prediction, and its principle mainly includes: comparing ultrasound images with pathological section results, judging whether LNM exists based on the aspect ratio of thyroid nodules, nodule calcification lesions, edge clarity, and the distance of the nodule edge close to the thyroid capsule. However, the existing technology mainly relies on manual measurement and lacks the means to measure large-scale data, resulting in strong subjectivity of statistical results, only rough quantification, and lack of precise quantitative research results. In addition, there is a lack of systematic research on the observed characteristic results, and it is impossible to evaluate their usability and the interpretability of the results is poor.

[0062] In view of the above problems, in one embodiment, a deep learning algorithm can be used, including differential homeomorphic registration of complex images, semantic segmentation, optimization of foreground and background correlation binary classification network, graph convolutional neural network and other methods, respectively, for the thyroid nodule's own morphology, thyroid nodule position information on the thyroid gland (including minimum distance information), and comprehensive information of the first two to establish an overall prediction model, which includes two parts: morphological score and position score. And by calculating the area under the receiver operating characteristic curve (AUC), sensitivity and specificity, accuracy ACC, and comparing the results of human doctors predicting LNM based on Tirads as a baseline, the performance of the LNM prediction classification model is evaluated. And by calculating the model parameters to visualize the LNM risk distribution of nodules growing at different positions on the thyroid gland, an evaluation basis is provided for predicting LNM. Then the model with the best comprehensive effect is selected for verification on an independent external multi-center data set.

[0063] In a preferred embodiment, Figure 2 As shown, step S2 includes:

[0064] Step S21: pre-processing a plurality of ultrasonic images to remove invalid ultrasonic images;

[0065] Step S22: extracting organ images from the ultrasound image using a first image segmentation method, and extracting lesion images from the ultrasound image using a second image segmentation method.

[0066] Specifically, in step S21, the collected ultrasound image first needs to be cleaned according to preset conditions to retain valid data. The preset conditions include whether the ultrasound image is redundant, whether it is a low-quality ultrasound image, and whether the ultrasound image label is unclear.

[0067] Furthermore, by setting the first image segmentation method and the second image segmentation method, the organ or lesion can be effectively separated from the background image to obtain a better recognition effect.

[0068] In a preferred embodiment, Figure 3 As shown, the first image segmentation method in step S22 includes:

[0069] Step S221: preprocessing the ultrasound image to generate a preprocessed image;

[0070] Step S222: extracting organ images from the pre-processed image using a network recognition model.

[0071] As an optional implementation, the network identification model is a UNET model.

[0072] Specifically, the preprocessing method in step A1 includes: performing a transformation operation on the original ultrasound image, converting the image uniformly into a three-channel image, and cutting it to the same length and width. The left half of the UNET model in step A2 is provided with a feature extraction layer, which is preferably a 3X3 convolution layer, and the convolution layer also includes a Relu function module for processing the convolution result to remove negative values; the right half of the UNET model is an upsampling convolution layer, which is used to perform upsampling convolution operations. The UNET model halves the number of upsampling features each time, while the number of downsampling features doubles. The UNET model can combine the information of deep and shallow feature spaces, and use shallow features (cascaded with the same resolution) to improve the lack of upsampling information. The boundaries of ultrasound images are blurred and the gradients are complex. This method can provide segmentation effects at high resolution and multiple scales. It is also due to the unique model structure of UNET, and the number of parameters used is relatively small, which can well fit a model with excellent results on small sample data, which is very suitable for dealing with the problem of small natural data volume in the medical field.

[0073] In a preferred embodiment, the second image segmentation method in step S22 includes:

[0074] An object segmentation model is used to extract lesion images from ultrasound images.

[0075] like Figure 4 As shown, the object segmentation model includes:

[0076] A feature pyramid network 11, the feature pyramid network 11 receives the ultrasound image and generates an object segmentation image according to the ultrasound image;

[0077] A relationship module 12, the relationship module 12 receives the object segmentation image and generates a foreground enhancement image according to the object segmentation image;

[0078] A decoding module 13, the decoding module 13 receives the foreground enhanced image and generates a semantically restored image;

[0079] The optimization module 14 receives the semantic restoration image and generates a lesion image.

[0080] As an optional implementation, the object segmentation model may be a Farseg model.

[0081] Specifically, in the specific implementation process, the feature pyramid network 11 is used to perform multi-scale extraction of the lesion part from the ultrasound image, and then the relationship module 12 learns the symbiotic relationship between the scene and the foreground to generate a context associated with the foreground, that is, the lesion part, thereby enhancing the image characteristics of the lesion. The decoding module 13 is embodied as a lightweight decoder in the implementation process, which is used to restore the semantic features of the foreground enhanced image, thereby improving the spatial resolution of the semantic features. Finally, the optimization module 14 performs foreground perception optimization and suppresses other content in the background image, so that the recognition model is concentrated on the foreground part in the subsequent morphological scoring process. The Farseg model can be used to divide image pixels into two subsets of foreground objects and background areas based on the foreground-background relationship, which solves the problem that the foreground targets in ultrasound images are small and few, and the background targets are large and complex, and the accuracy of nodule segmentation is greatly improved.

[0082] In a preferred embodiment, Figure 5 As shown, step S3 includes: using a pre-established binary classification model to generate a morphological score according to the lesion image;

[0083] The binary classification model consists of the following settings:

[0084] Feature extraction layer 21, which performs convolution operation on the ultrasound image and then performs four residual block operations to output the extraction result;

[0085] The classification layer 22 generates a morphological score based on the extraction results.

[0086] Specifically, the binary classification model is a binary classification model for predicting the probability of metastasis of a specific lesion by extracting the relationship between lesion features and lesion metastasis using the TLR deep learning method. Among them, the feature extraction layer 21 can use Resnet50 as the feature extraction layer during the implementation process, and complete the operations of convolution and residual blocks.

[0087] As an optional implementation, the classification layer 22 includes a global average pooling layer, three fully connected layers, and a dropout layer connected in series. By setting the above classification layer 22, a better morphology prediction effect is achieved, and then a morphology score is generated.

[0088] In a preferred embodiment, step S3 further comprises using a location recognition method to generate a location score according to the organ image and the lesion image;

[0089] If Figure 6 As shown, the location identification method includes:

[0090] Step S31: performing registration according to the ultrasound image and a preset organ template to generate a registered image;

[0091] Step S32: extracting position information from the self-registered image;

[0092] Step S33: using a graph algorithm model to generate a location score based on the location information.

[0093] Specifically, the registration operation in step B1 mainly relies on a preset organ template. By registering the segmented organ image to the organ template, the lesion is also mapped to the corresponding position in the template according to the deformation field obtained by the registration, thereby obtaining the absolute position information in the current registered image.

[0094] As an optional implementation, the registration operation is implemented using the ANTs software package, based on the principle of symmetric diffeomorphism. The registration method is 'SyN', the interpolation method is linear interpolation, the optimization metirc in the registration process is mutual information, and the final result evaluation index is MSE.

[0095] Taking thyroid nodules as an example, the use of ultrasound equipment to detect thyroid nodules causes different imaging due to compression of the neck, and the target tissue can be arbitrarily positioned to zoom in and out. Therefore, thyroid nodule ultrasound images cannot be compared in a large-scale standardized manner across multiple populations. Most of the images obtained have different thyroid edge shapes and resolutions. Some images can only show the nodule part and cannot determine the position of the entire thyroid gland. By setting the registration operation, the large-scale collected image data can be effectively fused to the same scale, which is convenient for the processing of artificial intelligence models.

[0096] In a preferred embodiment, the graph algorithm model includes:

[0097] The bipartite graph submodule 31, the graph structure submodule 31 generates a bipartite graph according to the position information;

[0098] A graph convolution submodule 32, wherein a plurality of graph convolution layers and a classification pooling layer are provided in the graph convolution submodule 32, and is used to extract structural features of a bipartite graph;

[0099] The classification submodule 33 generates a position score according to the structural features of the bipartite graph.

[0100] Specifically, the graph structure submodule 31 is embodied as an input layer during the implementation process, which converts the position information obtained by the registration into a bipartite graph, in which there are two types of nodes, which are used to characterize the position coordinates of the organ and the lesion respectively. There is a directed edge between the lesion node and each organ node, with the organ node pointing to the lesion node. Subsequently, the node features of the input bipartite graph are extracted by multiple graph convolution layers set in the graph convolution submodule 32, and the most representative first K node features are extracted by the classification pooling layer (sort-pooling) layer as the overall features of the input bipartite graph. The classification submodule 33 is a multi-layer perceptron model (MLP), which directly predicts the transfer probability and generates a position score based on the output of the above-mentioned graph convolution module. Based on the above-mentioned graph algorithm model, the positional relationship between the lesion and the organ can be converted into a graph structure, and the positional relationship can be constructed in a higher dimension, so as to more effectively explore the corresponding relationship implicit in each position information.

[0101] In a preferred embodiment, step S4 further includes generating a probability distribution graph using a visualization method.

[0102] As an optional implementation, the visualization method includes using SaliencyMaps to perform a visualization operation.

[0103] Specifically, SaliencyMaps is a gradient-based visualization method that quantifies the importance of nodes in the graph through gradients. The derivative of each organ node and lesion node in the bipartite graph is calculated. The influence of the change in the node on the final result is represented by the quantification of the derivative, where xn is the input graph and yk is the output of the k-th model. The model then rearranges the derivatives to obtain the Maps graph. This method directly extracts the parameters in the trained CNN layer, so no additional annotation is required and the solution is fast. After obtaining the importance of a node, the node is mapped back from the graph network to the pixel in the image to obtain the importance of the pixel in the image for classification. Finally, the importance value is used to calculate the probability of metastasis and obtain a visualization graph. Taking the preoperative prediction of cervical lymph node metastasis (LNM) of papillary thyroid carcinoma as an example, the above method can generate a LNM probability distribution graph, which intuitively reflects the probability of LNM occurring at each node on the thyroid gland.

[0104] The ultrasound imaging diagnosis method provided by this technical solution integrates the morphology and location information of thyroid nodules, and the AUC (95% CI) on the training and external test sets are 94.38%, 94.11%, and 88.89%, respectively, compared with the AUC (95% CI) of 49.43%, 49.91%, and 52.42% of TIRADS evaluated by ultrasound doctors. In the AI ​​model, the AUC of the AI ​​model that only uses morphological information is 57.74%, 57.86%, and 58.07%, and the AUC of the AI ​​model that only uses location information is 94.24%, 93.67%, and 88.79%. It can be seen that the ultrasound imaging diagnosis method used in this technical solution has a better accuracy rate than the existing technology.

[0105] The beneficial effects of the present invention are:

[0106] 1. High accuracy in predicting multi-factorial diseases:

[0107] The automatic detection method based on artificial intelligence, combined with morphological information and location information, greatly improves the accuracy of predicting diseases. In the prediction of specific diseases (see Example 1 for details), compared with the accuracy of about 50% of the image analysis results based on doctor experience and the accuracy of about 58% of the prediction model established based on morphological information alone, the accuracy of this solution is about 94%. It provides a comprehensive solution for the shortage of ultrasound doctors and the difficulty of seeing a doctor in some backward areas.

[0108] 2. Provide explainable causes for the diagnosis results:

[0109] This solution overcomes the technical barrier of poor interpretability of artificial intelligence methods. Based on the multiple features of the image, it conducts effective analysis from a quantitative, objective, and interpretable perspective, which can provide analysts with strong reference value.

[0110] 3. Technical difficulties in solving large-scale statistical data and comparing different scales:

[0111] When using ultrasound equipment to detect thyroid nodules, the imaging varies due to the compression of the neck, and the target tissue can be arbitrarily positioned to zoom in and out. Therefore, it is impossible to achieve large-scale standardized comparison of thyroid nodule ultrasound images for multiple populations. Most of the images obtained have different thyroid edge shapes and resolutions. Some images can only show the nodule part and cannot determine the position of the entire thyroid gland. In order to count and compare the positional relationship between the thyroid gland and nodules on a large scale, we need to fuse all the images at the same scale. Using mathematical methods such as differential homeomorphism and affine transformation, we achieved a good result with an MSE of only 0.07.

[0112] 4. Use existing data to make reasonable predictions about related diseases:

[0113] This solution can use ultrasound imaging data to make reasonable predictions about related lesions of other lesions or organs, providing new ideas and innovations for scientific research.

[0114] 5. The transformation of results is relatively easy:

[0115] The prediction model trained by this scheme can be widely used in various ultrasound equipment or Internet cloud platforms such as 5G to realize remote consultation in areas with inconvenient conditions, and also provide auxiliary tools for ultrasound diagnosis.

[0116] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. An ultrasonic imaging diagnosis method based on artificial intelligence, characterized in that: include: Step S1: Acquire ultrasound images of the patient; Step S2: extracting organ images and lesion images from the ultrasound images; Step S3: generating a morphological score according to the lesion image, and generating a position score according to the organ image and the lesion image; Step S4: generating a comprehensive score according to the morphology score and the position score; The comprehensive score is used to characterize the metastasis probability of the lesion; The step S3 comprises: using a pre-established binary classification model to generate a morphological score according to the lesion image; The binary classification model includes: A feature extraction layer, wherein the feature extraction layer performs a convolution operation on the ultrasound image, and then performs four residual block operations to output an extraction result; a classification layer, the classification layer generating the morphological score according to the extraction result; The step S3 further comprises: using a location recognition method to generate a location score according to the organ image and the lesion image; The location identification method comprises: Step S31: performing registration according to the ultrasound image and a preset organ template to generate a registered image; Step S32: extracting position information from the registered image; Step S33: using a graph algorithm model to generate the location score according to the location information; The graph algorithm model includes: A bipartite graph submodule, wherein the bipartite graph submodule generates a bipartite graph according to the position information; A graph convolution submodule, wherein a plurality of graph convolution layers and a classification pooling layer are provided in the graph convolution submodule, and are used to extract structural features of the bipartite graph; A classification submodule is used to generate the position score according to the structural characteristics of the bipartite graph.

2. The ultrasonic imaging diagnostic method according to claim 1, characterized in that: The step S2 comprises: Step S21 pre-processes the plurality of ultrasound images to remove invalid ultrasound images; In step S22, a first image segmentation method is used to extract the organ image from the ultrasound image, and a second image segmentation method is used to extract the lesion image from the ultrasound image.

3. The ultrasonic imaging diagnostic method according to claim 2, characterized in that: The first image segmentation method in step S22 includes: Step S221: preprocessing the ultrasound image to generate a preprocessed image; Step S222: extracting the organ image from the pre-processed image using a network recognition model.

4. The ultrasonic imaging diagnostic method according to claim 2, characterized in that: The second image segmentation method in step S22 includes: extracting the lesion image from the ultrasound image using an object segmentation model; The object segmentation model includes: a feature pyramid network, wherein the feature pyramid network receives the ultrasound image and generates an object segmentation image according to the ultrasound image; a relationship module, the relationship module receiving the object segmentation image and generating a foreground enhanced image according to the object segmentation image; A decoding module, the decoding module receives the foreground enhanced image and generates a semantically restored image; An optimization module receives the semantically restored image and generates the lesion image.

5. The ultrasonic imaging diagnostic method according to claim 1, characterized in that: The step S4 also includes generating a probability distribution diagram using a visualization method.

Images