Ultrasound Imaging Disease Identification Methods and Systems

Abstract

This invention discloses a method and system for ultrasound imaging disease identification. The method includes: semantically segmenting an image to be detected based on an image segmentation network model to obtain a target contour image; using a preliminary disease screening model to detect lesions in the target contour image to obtain an initial disease diagnosis result; and correcting the initial disease diagnosis result to obtain a final disease diagnosis result. This identification method, by semantically segmenting the image to be detected and extracting the target contour region, reduces information interference from other regions and improves the accuracy of subsequent model predictions. Detection of the target contour region quickly provides an initial diagnosis result, reducing the scope of subsequent processing. Finally, by correcting the initial diagnosis result, deviations in the preliminary disease screening model can be corrected, reducing the influence of physician subjectivity on the diagnosis result and improving the accuracy of the final diagnosis.

Description

Technical Field

[0001] This invention relates to the field of deep learning in medical image processing, and more particularly to a method and system for identifying diseases in ultrasound images. Background Technology

[0002] During medical ultrasound scans, different tissues exhibit different acoustic properties, resulting in varying appearances on ultrasound images. Doctors often analyze and diagnose diseases based on these ultrasound images, specifically by observing information such as the brightness, shape, boundaries, and texture of the ultrasound image.

[0003] Existing ultrasound imaging methods for disease identification mainly rely on doctors' experience and manual analysis. However, the analysis and diagnosis process is easily affected and limited by the operator's or doctor's skill level or experience, making it difficult to standardize the observation and diagnosis process and subject to subjectivity. As a result, problems such as long diagnosis time, low diagnostic accuracy, and poor universality may occur, which may lead to incorrect treatment plans and affect the patient's health. Summary of the Invention

[0004] One of the objectives of this invention is to provide a method for identifying diseases using ultrasound imaging, in order to solve the problem that the existing technology cannot take into account both human experience and algorithm analysis in collaborative diagnosis, which leads to medical staff being unable to obtain accurate diagnostic results based on the images to be tested.

[0005] One of the objectives of this invention is to provide an ultrasound imaging disease recognition system.

[0006] To achieve one of the above-mentioned objectives, the present invention provides a method for identifying diseases in ultrasound imaging, comprising: performing semantic segmentation on an image to be detected based on an image segmentation network model to obtain a target contour image; using a preliminary disease screening model to detect lesions on the target contour image to obtain an initial disease diagnosis result; and correcting the initial disease diagnosis result to obtain a final disease diagnosis result.

[0007] As a further improvement of one embodiment of the present invention, the image segmentation network model is configured as a network model built on the BisenetV2 network framework.

[0008] As a further improvement of one embodiment of the present invention, before the step of "semantically segmenting the image to be detected based on the image segmentation network model to obtain the target contour image", the method further includes: constructing an initial image segmentation network model using the BisenetV2 network framework; replacing the context embedding module of the initial image segmentation network model with a pyramid pooling module; wherein, the feature extraction backbone network of the initial image segmentation network model includes the context embedding module and several encoders; during the training of the initial image segmentation network model, several auxiliary segmentation heads are set on several encoders to obtain an improved image segmentation network model; wherein, the several encoders correspond to several encoding stages and are used to extract semantic information of different stages of the original image to be detected; obtaining first labeled sample data and inputting the first labeled sample data into the improved image segmentation network model for training to obtain the image segmentation network model; wherein, the first labeled sample data includes a first category label, and the first category label carries the name of the tissue or organ corresponding to the sample image.

[0009] As a further improvement of one embodiment of the present invention, the "semantic segmentation of the image to be detected based on the image segmentation network model to obtain the target contour image" specifically includes: removing several auxiliary segmentation heads configured in the image segmentation network model during the prediction process of the image segmentation network model to obtain an image segmentation network model without auxiliary segmentation heads; and performing semantic segmentation of the image to be detected based on the image segmentation network model without auxiliary segmentation heads to obtain the target contour image.

[0010] As a further improvement of one embodiment of the present invention, the "semantic segmentation of the image to be detected based on the image segmentation network model to obtain the target contour image" specifically includes: extracting spatial feature maps corresponding to spatial information and semantic feature maps corresponding to semantic information from the image to be detected using the image segmentation network model; and performing an aggregation operation based on the spatial feature maps and the semantic feature maps to obtain the target contour image.

[0011] As a further improvement of one embodiment of the present invention, the image segmentation network model includes a detail branch, a semantic branch, and a bilateral guided aggregation layer; the step of "extracting spatial feature maps corresponding to spatial information and semantic feature maps corresponding to semantic information using the image segmentation network model according to the image to be detected" specifically includes: extracting spatial information corresponding to different stages of the image to be detected using the detail branch to obtain several spatial feature maps; extracting semantic information corresponding to different stages of the image to be detected using the semantic branch to obtain several semantic feature maps; the step of "calculating the target contour image using an aggregation operation method based on the spatial feature maps and the semantic feature maps" specifically includes: fusing the several spatial feature maps and the several semantic feature maps based on a pyramid pooling module to obtain several fused feature maps; wherein, the pyramid pooling module is set in the last layer of the semantic branch; and performing multi-scale information encoding aggregation operation on the several fused feature maps using the bilateral guided aggregation layer to obtain the target contour image.

[0012] As a further improvement of one embodiment of the present invention, before the step of “semantically segmenting the image to be detected based on the image segmentation network model to obtain the target contour image”, the method further includes: constructing an optimized image classification model; acquiring the original image to be detected, and inputting the original image to be detected into the optimized image classification model for classification and recognition to obtain the image to be detected.

[0013] As a further improvement to one embodiment of the present invention, the "construction of an optimized image classification model" specifically includes: constructing an initial image classification model using the GhostNet network framework; halving the number of input and output channels of all convolutional and fully connected layers in the initial image classification model, replacing the single-dimensional attention module with a multi-dimensional attention module, and replacing the ReLU activation function with a Hard-swish activation function to obtain an improved image classification model; acquiring second labeled sample data, and inputting the second labeled sample data into the improved image classification model for training to obtain the optimized image classification model; wherein, the second labeled sample data includes a first judgment label, which carries a judgment result on whether the tissue part in the corresponding sample image is a first tissue part; wherein, the single-dimensional attention module includes an SENet attention module; and the multi-dimensional attention module includes at least one of an SKNet attention module, a CBAM attention module, and an ECA attention module.

[0014] As a further improvement of one embodiment of the present invention, the preliminary disease screening model includes an optimized focal disease detection model; the initial disease diagnosis result includes an initial focal disease diagnosis result; and the final disease diagnosis result includes a final focal disease diagnosis result.

[0015] As a further improvement of one embodiment of the present invention, before the step of "using a preliminary disease screening model to detect lesions in the target contour image at least to obtain an initial disease diagnosis result", the method further includes: constructing an initial focal disease detection model using a YOLOv5s network framework; replacing the convolutional layer modules in the feature extraction backbone network and detection head component of the initial focal disease detection model with Ghost module components, and setting the small target detection head in the second layer feature map output by the initial focal disease detection model to obtain an improved focal disease detection model; wherein, the Ghost module component includes a Ghost module ontology and a multi-dimensional attention module; acquiring third labeled sample data, and inputting the third labeled sample data into the improved focal disease detection model for training to obtain the optimized focal disease detection model; wherein, the third labeled sample data includes a second category label, and the second category label carries the focal disease diagnosis result of the corresponding sample image; the step of "using a preliminary disease screening model to detect lesions in the target contour image at least to obtain an initial disease diagnosis result" specifically includes: inputting the target contour image into the optimized focal disease detection model for prediction to obtain the initial focal disease diagnosis result.

[0016] As a further improvement of one embodiment of the present invention, the preliminary disease screening model includes an optimized diffuse disease detection model; the initial disease diagnosis result includes an initial diffuse disease diagnosis result; and the final disease diagnosis result includes a final diffuse disease diagnosis result.

[0017] As a further improvement of one embodiment of the present invention, before the step of "using a preliminary disease screening model to detect lesions in the target contour image at least to obtain an initial disease diagnosis result", the method further includes: constructing an initial diffuse disease detection model using the GhostNet network framework; halving the number of input and output channels of all convolutional and fully connected layers in the initial diffuse disease detection model, replacing the single-dimensional attention module with a multi-dimensional attention module, and replacing the ReLU activation function with a Hard-swish activation function to obtain an improved diffuse disease detection model; acquiring fourth labeled sample data, and inputting the fourth labeled sample data into the improved diffuse disease detection model for training, and executing a label smoothing algorithm during training to obtain the optimized diffuse disease detection model; wherein, the fourth labeled sample data includes a third category label, and the third category label carries the diffuse disease diagnosis result of the corresponding sample image; the step of "using a preliminary disease screening model to detect lesions in the target contour image at least to obtain an initial disease diagnosis result" specifically includes: inputting the image to be detected and the target contour image into the optimized diffuse disease detection model for prediction to obtain the initial diffuse disease diagnosis result.

[0018] As a further improvement of one embodiment of the present invention, the step of "correcting the initial disease diagnosis result to obtain the final disease diagnosis result" specifically includes: acquiring and determining whether a plurality of original images to be detected are consecutive frames; wherein, the plurality of original images to be detected include the current image to be corrected and a plurality of historical images; based on the consecutive frame determination result, determining and correcting the initial disease diagnosis result according to the correction strategy to obtain the final disease diagnosis result.

[0019] As a further improvement of one embodiment of the present invention, the step of "acquiring and determining whether several original images to be detected are consecutive frames" specifically includes: calculating the similarity between two adjacent frames of original images to be detected based on the several original images to be detected, and obtaining several image similarity values; determining whether any of the image similarity values ​​is less than or equal to a first similarity preset threshold; if yes, then determining that the several original images to be detected are not consecutive frames; if no, then determining that the several original images to be detected are consecutive frames.

[0020] As a further improvement of one embodiment of the present invention, the step of "determining and correcting the initial disease diagnosis result according to the continuous frame judgment result and obtaining the final disease diagnosis result" specifically includes: when the plurality of original images to be detected are not continuous frames, the initial disease diagnosis result corresponding to the current image to be corrected is taken as its final disease diagnosis result; when the plurality of original images to be detected are continuous frames, the initial disease diagnosis result of the current image to be corrected is corrected according to the image similarity value to obtain the final disease diagnosis result.

[0021] As a further improvement of one embodiment of the present invention, the step of "correcting the initial disease diagnosis result of the current detection image to be corrected according to the image similarity value to obtain the final disease diagnosis result" specifically includes: determining whether the image similarity value between the current detection image to be corrected and the previous historical detection image is greater than or equal to a second similarity preset threshold; wherein, the previous historical detection image and the current detection image to be corrected are two adjacent frames of detection images taken in chronological order; if so, the initial disease diagnosis result of the previous historical detection image is taken as the final disease diagnosis result of the current detection image to be corrected.

[0022] As a further improvement of one embodiment of the present invention, the initial disease diagnosis result includes the initial diffuse disease diagnosis result; the final disease diagnosis result includes the final diffuse disease diagnosis result; after the step of "determining whether the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to the second similarity preset threshold", the method further includes: if not, then statistically analyzing the initial diffuse disease diagnosis results corresponding to the plurality of original images to be detected, and taking the initial diffuse disease diagnosis result that appears most frequently in the statistical results as the final diffuse disease diagnosis result of the current image to be corrected.

[0023] As a further improvement of one embodiment of the present invention, the initial disease diagnosis result includes an initial focal disease diagnosis result; the final disease diagnosis result includes a final focal disease diagnosis result; after "determining whether the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to a second similarity preset threshold", the method further includes: if not, calculating and correcting the initial focal disease diagnosis result of the current image to be corrected based on the focal diagnostic indicators of the plurality of original images to be detected, to obtain the final focal disease diagnosis result; wherein, the focal diagnostic indicators include at least one of crossover ratio and lesion distance.

[0024] As a further improvement of one embodiment of the present invention, the step of "calculating and correcting the focal disease diagnosis result of the current detection image to be corrected based on the focal diagnostic indicators of the plurality of original images to be detected, and obtaining the final focal disease diagnosis result" specifically includes: determining a plurality of focal disease types and a plurality of focal disease locations corresponding to the initial focal disease diagnosis result of the current detection image to be corrected; wherein, the focal disease type includes a first focal disease type; the focal disease location includes a first focal disease location; calculating the focal diagnostic indicators of the current detection image to be corrected and the plurality of historical detection images based on the first focal disease location to obtain a first focal diagnosis result; correcting the first focal disease type corresponding to the current detection image to be corrected based on the first focal diagnosis result to obtain a corrected first focal disease type; iterating until all focal disease types have been corrected to obtain the final focal disease diagnosis result of the current detection image to be corrected.

[0025] As a further improvement of one embodiment of the present invention, the step of "correcting the first focal disease type corresponding to the current detection image to be corrected based on the first focal diagnosis result, and obtaining the corrected first focal disease type" specifically includes: determining whether the first focal diagnosis result meets the focal diagnosis conditions; wherein, the focal diagnosis conditions characterize the positional distribution of the first focal disease type in the plurality of original detection images; if yes, then statistically analyzing the focal disease types corresponding to the position of the first focal disease in the plurality of original detection images, and using the focal disease type that appears most frequently in the statistical results to correct the first focal disease type of the current detection image to be corrected, to obtain the final first focal disease type; if no, then deleting the first focal disease type corresponding to the current detection image to be corrected.

[0026] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides an ultrasound imaging disease recognition system, comprising: an acquisition unit for acquiring an original image to be detected; a processing unit for processing the original image to be detected to obtain an image to be detected, and inputting the image to be detected into an image segmentation network model to obtain a target contour image; a disease diagnosis unit for detecting lesions on the target contour image using a preliminary disease screening model to obtain an initial disease diagnosis result; a disease diagnosis correction unit for correcting the initial disease diagnosis result to obtain a final disease diagnosis result; and a display unit for outputting and displaying the final disease diagnosis result.

[0027] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0028] This invention employs an ultrasound imaging disease identification method. By using an image segmentation network model to perform semantic segmentation on the image to be detected, the target contour region can be extracted, reducing interference from other irrelevant information such as non-target regions and improving the accuracy of subsequent model predictions. Then, the target contour region is detected according to a preliminary disease screening model, which can quickly provide an initial diagnostic result, reducing the scope of subsequent processing. Finally, by correcting the initial diagnostic result, deviations in the preliminary disease screening model can be corrected, reducing the influence of physician subjectivity on the diagnostic result and improving the accuracy of the final diagnosis. The entire disease identification process integrates image processing and machine learning, achieving automated and intelligent analysis of ultrasound images, reducing reliance on operator or physician experience, shortening diagnostic time, and improving work efficiency. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the steps of an ultrasound imaging disease identification method according to an embodiment of the present invention.

[0030] Figure 2 This is a detailed schematic diagram of step S1 of the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of some steps in the ultrasound imaging disease identification method according to the first embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of some steps in the ultrasound imaging disease identification method according to the second embodiment of the present invention.

[0033] Figure 5(a) shows a liver ultrasound image before the liver organ segmentation model in an embodiment of the ultrasound imaging disease identification method of the present invention.

[0034] Figure 5(b) is a liver target contour image based on a liver organ segmentation model in an embodiment of the ultrasound imaging disease identification method of the present invention.

[0035] Figure 6 This is a schematic diagram of some steps in the ultrasound imaging disease identification method according to the third embodiment of the present invention.

[0036] Figure 7 This is a detailed step diagram of step P11 of the ultrasound imaging disease identification method in one embodiment of the present invention.

[0037] Figure 8(a) is a schematic diagram of the detailed steps of the optimized focal disease detection model of the ultrasound imaging disease identification method in one embodiment of the present invention.

[0038] Figure 8(b) is a schematic diagram of the detailed steps of the optimized diffuse disease detection model of the ultrasound imaging disease identification method in one embodiment of the present invention.

[0039] Figure 9 This is a schematic diagram of some steps in the ultrasound imaging disease identification method according to the fourth embodiment of the present invention.

[0040] Figure 10 This is a detailed schematic diagram of step S31 of the ultrasound imaging disease identification method in one embodiment of the present invention.

[0041] Figure 11 This is a detailed schematic diagram of step S32 of the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0042] Figure 12(a) is a detailed schematic diagram of the steps of the initial focal disease diagnosis result correction step S322 in the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0043] Figure 12(b) is a detailed schematic diagram of the steps in step S322 of the initial diffuse disease diagnosis result correction step in the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0044] Figure 13 This is a detailed schematic diagram of step S3222B' of the ultrasound imaging disease identification method in one embodiment of the present invention.

[0045] Figure 14 This is a detailed schematic diagram of step S3222B3' of the ultrasound imaging disease identification method in one embodiment of the present invention.

[0046] Figure 15(a) is a flowchart illustrating a preferred embodiment of the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0047] Figure 15(b) is a schematic diagram of the initial diagnostic result correction process of a preferred embodiment of the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0048] Figure 16(a) is an ultrasound image after correction of the initial focal disease diagnosis result of a preferred embodiment of the ultrasound imaging disease identification method according to an embodiment of the present invention.

[0049] Figure 16(b) is an ultrasound image after correction of the initial diagnostic result of a diffuse disease in a preferred embodiment of the ultrasound imaging disease identification method according to an embodiment of the present invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0051] It should be noted that the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the description of specific embodiments of the present invention, the terms "first," "second," "third," "fourth," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0052] In the field of ultrasound medical imaging, ultrasound imaging disease diagnosis methods based on deep learning and image processing technologies have broad application scenarios. They can help primary care physicians solve the problem of real-time diagnosis, effectively improve doctors' diagnostic efficiency and accuracy, and shorten patients' waiting time.

[0053] Based on this, the present invention provides a method for identifying diseases using ultrasound imaging, such as... Figure 1 As shown, the ultrasound imaging disease identification method specifically includes the following steps:

[0054] Step S1: Perform semantic segmentation on the image to be detected based on the image segmentation network model to obtain the target contour image;

[0055] Step S2: Use a preliminary disease screening model to detect lesions in the target contour image at least once to obtain an initial disease diagnosis result;

[0056] Step S3: Correct the initial disease diagnosis result to obtain the final disease diagnosis result.

[0057] Thus, by employing an image segmentation network model to perform semantic segmentation on the image to be detected, the target contour region can be extracted, reducing interference from other irrelevant information such as non-target regions and improving the accuracy of subsequent model predictions. Then, by detecting the target contour region based on the preliminary disease screening model, an initial diagnostic result can be quickly provided, reducing the scope of subsequent processing. Finally, correcting the initial diagnostic result can correct the deviation in the model's initial recognition, reduce the influence of the doctor's subjective factors on the diagnostic result, and improve the accuracy of the final diagnosis.

[0058] The image segmentation network model is configured as a network model built on the BisenetV2 network framework. Thus, by building a model based on the BisenetV2 network framework, images can be segmented quickly while maintaining segmentation accuracy.

[0059] Based on this, such as Figure 2 As shown, before performing step S1, the image segmentation network model needs to be constructed and trained. That is, the ultrasound image disease recognition method may specifically include the following steps:

[0060] Step S01: Construct an initial image segmentation network model using the BisenetV2 network framework;

[0061] Step S02: Replace the context embedding module of the initial image segmentation network model with the pyramid pooling module;

[0062] Step S03: During the training of the initial image segmentation network model, several auxiliary segmentation heads are set on several encoders to obtain an improved image segmentation network model.

[0063] Step S04: Obtain the first labeled sample data and input the first labeled sample data into the improved image segmentation network model for training to obtain the image segmentation network model.

[0064] Thus, by replacing the context embedding module with a pyramid pooling module, contextual information at different scales can be captured, which is beneficial to improving image segmentation accuracy. In addition, by adding segmentation heads to different aggregation dilation layers, the training speed can be accelerated, multi-scale semantic information fusion can be achieved, and the segmentation results can be more accurate.

[0065] The initial image segmentation network model's feature extraction backbone includes the context embedding module and several encoders; these encoders correspond to several encoding stages, used to extract semantic information from different stages of the original image to be detected; the first labeled sample data includes a first category label, which carries the name of the corresponding organ in the sample image. Furthermore, steps S02 and S03 are not sequentially dependent and their execution order can be interchanged.

[0066] It should be noted that after the image segmentation network model is trained, a refinement step can be provided for step S1, which may specifically include:

[0067] Step P1: During the prediction process of the image segmentation network model, several auxiliary segmentation heads configured in the image segmentation network model are removed to obtain an image segmentation network model without auxiliary segmentation heads.

[0068] Step S1': Based on the image segmentation network model without auxiliary segmentation head, perform semantic segmentation on the image to be detected to obtain the target contour image.

[0069] Thus, by removing the auxiliary segmentation head during the model prediction process, a lighter feature representation can be obtained, reducing computational and storage costs and shortening the model prediction time.

[0070] Semantic segmentation, in this context, is a method for classifying each pixel in an image to be detected and determining the category of each pixel. In semantic segmentation tasks, both low-level details and high-level semantics are essential. To achieve high accuracy and efficiency in semantic segmentation, a BisenetV2 network architecture that processes spatial details and classification semantics separately is proposed.

[0071] The BisenetV2 (Bilateral Segmentation Network V2) framework is an image semantic segmentation network model, an improvement upon the Bisenet network. It includes two parallel network paths: a shallow network extracts image detail information, and a deep network extracts semantic information. A pyramid pooling module replaces the original context embedding module, fusing the information extracted from these two network paths. This allows for the transmission of semantic information at multiple scales and from multiple perspectives, enhancing the expressive power of detailed features.

[0072] The context embedding module, a component of the BisenetV2 network framework, is used to directly transmit contextual information in the detail path and semantic path, enhancing the ability to express detailed features. The pyramid pooling module uses pooling kernels of different sizes to perform spatial pyramid pooling on the feature map, capturing contextual information at different scales, which enhances the model's ability to perceive global context.

[0073] The segmentation head can be understood as an upsampling and convolution operation used to progressively restore the resolution of the feature map and finally output the semantic segmentation result. During the network model training process, the encoders corresponding to different encoding stages contain feature information of different granularities. By adding segmentation heads at different stages, multi-scale semantic information fusion can be achieved. In the model prediction process, removing the segmentation head can reduce the amount of computation and memory consumption, and speed up the model's prediction speed.

[0074] Furthermore, to accelerate training, this invention employs a supervised training method, meaning the sampled data has a first category label carrying the name of the corresponding organ in the sample image. In a specific example, the first category label of the sample image may include at least one of the following: liver, gallbladder, pancreas, spleen, kidney, etc. Step S1' can be understood as a derivative step of step S1, and P1 can be understood as a step to be performed before using the trained image segmentation network model for model prediction.

[0075] Based on this, combined Figure 1 and Figure 3 As shown, in the first embodiment of the present invention, the ultrasound imaging disease identification method specifically includes the following steps:

[0076] Step S11: Based on the image to be detected, the image segmentation network model is used to extract its spatial feature map corresponding to spatial information and its semantic feature map corresponding to semantic information;

[0077] Step S12: Based on the spatial feature map and the semantic feature map, perform an aggregation operation to obtain the target contour image.

[0078] Thus, by extracting spatial feature maps, fine contour details can be provided, and by extracting semantic feature maps, semantic information can be provided; combining the two can improve the accuracy of target contour image extraction.

[0079] The image segmentation network model includes a detail branch, a semantic branch, and a bilateral guided aggregation layer. Specifically, the detail branch is responsible for spatial details and represents low-level information. This branch requires sufficient channel capacity to encode rich spatial detail information. Because it only focuses on low-level details, it can be designed as a shallow structure. The semantic branch runs in parallel with the detail branch and is used to capture high-level semantic information. This branch has a lower channel capacity, while spatial details are provided by the detail branch. The feature representations of the detail branch and the semantic branch are complementary. The outputs of the two branches have different levels of feature representation. Using simple addition or concatenation fusion methods easily ignores the diversity of the two types of information, leading to performance degradation. Therefore, the bilateral guided aggregation layer can effectively merge the two types of feature representations.

[0080] Combination Figure 1 and Figure 4 As shown, in the second embodiment of the present invention, the ultrasound imaging disease identification method specifically includes the following steps:

[0081] Step S111: The detail branch is used to extract spatial information corresponding to different stages of the image to be detected, and several spatial feature maps are obtained.

[0082] Step S112: Extract semantic information corresponding to different stages of the image to be detected using the semantic branch to obtain several semantic feature maps;

[0083] Based on this, step S12 can specifically include:

[0084] Step S121: Based on the pyramid pooling module, fuse the several spatial feature maps and the several semantic feature maps to obtain several fused feature maps;

[0085] Step S122: Based on the several fused feature maps, the target contour image is obtained by performing multi-scale information encoding and aggregation operations on them using the bilateral guided aggregation layer.

[0086] Thus, the detail branch captures low-level details (i.e., spatial details) and generates high-resolution feature representations; the semantic branch and pyramid pooling module obtain high-level semantic context to obtain feature maps with different receptive fields and capture rich contextual information; finally, the bilateral guided aggregation layer enhances the interconnection between the two branches and fuses the two types of feature representations, improving the accuracy of prediction.

[0087] The pyramid pooling module is located at the last layer of the semantic branch; that is, the last stage of the semantic branch is the output of the pyramid pooling module. In a specific example, the image segmentation network model may include a liver organ segmentation model; the target contour image may include a liver contour image. Figures 5(a) and 5(b) show the liver ultrasound image before image segmentation of the liver organ based on the liver organ segmentation model, and the liver target contour image after image segmentation based on the liver organ segmentation model. Of course, this invention does not exclude other image segmentation models and contour images, and does not impose specific limitations on them.

[0088] Optionally, in order to narrow the prediction range of the image segmentation model for the image to be detected, the present invention also provides a preliminary screening operation on the original detection image before step S1. For example... Figure 6 As shown, in the third embodiment of the present invention, the ultrasound imaging disease identification method may specifically include:

[0089] Step P11: Construct and optimize the image classification model;

[0090] Step P12: Obtain the original image to be detected, and input the original image to be detected into the optimized image classification model for classification and recognition to obtain the image to be detected.

[0091] Thus, by constructing an optimized image classification model and classifying and recognizing the original image to be detected, the processing range of the subsequent image segmentation model can be narrowed, thereby accelerating the speed of disease recognition.

[0092] In order to further accelerate the prediction speed and accuracy of the image classification model, this invention uses the lightweight GhostNet network framework to construct the initial image classification model, and then improves and trains it to obtain the optimized image classification model.

[0093] Specifically, such as Figure 7 As shown, step P11 can specifically include the following steps:

[0094] Step P111: Construct an initial image classification model using the GhostNet network framework;

[0095] Step P112: Halve the number of input and output channels of all convolutional and fully connected layers in the initial image classification model, replace the single-dimensional attention module with a multi-dimensional attention module, and replace the ReLU activation function with the Hard-swish activation function to obtain the improved image classification model.

[0096] Step P113: Obtain the second labeled sample data and input the second labeled sample data into the improved image classification model for training to obtain the optimized image classification model.

[0097] Thus, by adopting and improving the lightweight GhostNet network framework, the number of acceptable parameter channels and computational cost can be reduced, thereby lowering model complexity. Replacing the single-dimensional attention module with a multi-dimensional attention module allows for learning feature representations from multiple perspectives, enhancing the model's expressive power. Furthermore, replacing the ReLU activation function with the Hard-swish activation function can give the model better non-linear expressive power.

[0098] The second labeled sample data includes a first judgment label, which carries a determination result of whether the corresponding tissue part in the sample image is a first tissue part. In a specific example, the optimized image classification model includes an abdominal classification model; based on image features and physiological knowledge, sample images containing organs such as the liver, gallbladder, spleen, and kidneys are defined as abdominal images, and others are defined as non-abdominal images, to obtain the second sample labeled data. The first judgment label can indicate whether the current sample image is an abdominal image. Of course, the present invention is not limited to the abdominal classification model and the classification label for whether an image is abdominal.

[0099] Furthermore, the single-dimensional attention module includes the SENet attention module, which is a single-dimensional channel attention mechanism that learns the dependencies between channels through global average pooling and reweights channel features to enhance the representation of the target.

[0100] The multi-dimensional attention module includes at least one of the SKNet (Selective Kernel Networks) attention module, CBAM (Convolutional Block Attention Module), and ECA (Efficient Channel Attention). Specifically, the SKNet attention module performs two transformations on the feature map to capture spatial attention and channel attention respectively, fusing the two to obtain enhanced feature representation. The CBAM attention module captures spatial attention of the feature map through max pooling and average pooling, and then obtains channel attention through the same operations, sequentially stacking the two to form a hybrid attention mechanism. The ECA attention module obtains channel attention through a global average pooling, a learnable 1D convolutional kernel, and a learnable activation function, achieving low computational cost and high efficiency. Therefore, all multi-dimensional attention modules can simultaneously consider both spatial and channel dimensions, obtaining hybrid attention of features through different operations, thus enhancing the model's expressive power.

[0101] The GhostNet network framework is a lightweight neural network architecture. Based on Ghost modules, this framework reduces the overall acceptable number of channels and computational cost of the model without altering the size and channel size of the output feature maps from the convolutions. Optionally, the GhostNet framework can be replaced by lightweight network frameworks such as MobileNet or ShuffleNet.

[0102] The activation function can be understood as introducing nonlinear factors. Because linear network models lack sufficient expressive power, introducing a nonlinear function as an activation function enhances the expressive power of the network model. Specifically, the ReLU (Rectified Linear Unit) activation function is a piecewise linear function that changes all negative values ​​to 0 while leaving positive values ​​unchanged. The Hard-swish activation function is smooth and continuous, capable of learning more complex nonlinear mapping relationships. Therefore, the Hard-swish activation function can learn richer feature representations, resulting in a model with stronger fitting ability, lower training error, and higher test accuracy.

[0103] It should be noted that, in one specific embodiment, the preliminary disease screening model in steps S2 and S3 includes an optimized focal disease detection model; the initial disease diagnosis result includes an initial focal disease diagnosis result; and the final disease diagnosis result includes a final focal disease diagnosis result.

[0104] In another specific embodiment, the preliminary disease screening model includes an optimized diffuse disease detection model; the initial disease diagnosis result includes an initial diffuse disease diagnosis result; and the final disease diagnosis result includes a final diffuse disease diagnosis result.

[0105] In a preferred embodiment, the two methods described above can be used in combination. That is, the initial focal disease diagnosis result and the initial diffuse disease diagnosis result are detected simultaneously by the optimized focal disease detection model and the optimized diffuse disease detection model, and then corrected respectively to obtain the final focal disease diagnosis result and the final diffuse disease diagnosis result.

[0106] Thus, by combining the advantages of the two models, they can complement each other, improving the accuracy and comprehensiveness of disease detection. In addition, obtaining detection results from both aspects simultaneously can accelerate the disease detection process and save multi-stage prediction time.

[0107] On the one hand, as shown in Figure 8(a), for the detection of focal lesions, before step S2, an optimized focal lesion detection model needs to be constructed and trained, which may include the following steps:

[0108] Step P211: Construct an initial focal lesion detection model using the YOLOV5s network framework;

[0109] Step P212: Replace the convolutional layer module in the feature extraction backbone network and detection head component of the initial focal lesion detection model with the Ghost module component, and set the small target detection head in the second layer feature map output by the initial focal lesion detection model to obtain the improved focal lesion detection model.

[0110] Step P213: Obtain the third labeled sample data and input the third labeled sample data into the improved focal disease detection model for training to obtain the optimized focal disease detection model;

[0111] Step S2': Input the target contour image into the optimized focal lesion detection model for prediction to obtain the initial focal lesion diagnosis result.

[0112] Thus, by constructing an initial focal lesion detection model based on the YOLOv5s network framework, the relationship between the model's prediction accuracy and prediction speed can be balanced. In addition, replacing the key cross-stage modules with Ghost module components that include multi-dimensional attention mechanisms can enhance the feature extraction capabilities of the feature extraction backbone network and focus on important features. Moreover, by adding a small target detection head, the detection capability for small lesions can be enhanced.

[0113] The Ghost module component includes a Ghost module ontology and a multi-dimensional attention module; the third labeled sample data includes a second category label, which carries the focal disease diagnosis result of the corresponding sample image. In a specific example, the optimized focal disease detection model includes a liver focal detection model, where the second category label corresponding to the third labeled sample data may include intrahepatic hyperechoic lesions, liver cysts, liver hemangiomas, liver cancer, etc.

[0114] Furthermore, the described YOLOv5s network framework is an object detection algorithm. The network structure employs a lightweight framework to improve detection speed and accuracy. The network framework can be divided into two parts: a feature extraction backbone network and a detection head component. The detection head component includes multiple convolutional layers and detection layers, with the detection layers primarily implementing the object detection process. Of course, this invention does not exclude other object detection algorithm network frameworks.

[0115] It should be emphasized that when performing focal disease detection, small targets (such as small lesions) are prone to missed detection or poor detection results. The second layer of features in the initial focal disease detection model introduces a small target detection head, which enables the second layer detection head to detect small targets quickly and effectively.

[0116] On the other hand, as shown in Figure 8(b), for the detection of diffuse diseases, before step S2, the optimized diffuse disease detection model needs to be constructed and trained, which may specifically include the following steps:

[0117] Step P221: Construct an initial diffuse disease detection model using the GhostNet network framework;

[0118] Step P222: Halve the number of input and output parameters of all convolutional and fully connected layers in the initial diffuse disease detection model, replace the single-dimensional attention module with a multi-dimensional attention module, and replace the ReLU activation function with a Hard-swish activation function to obtain the improved diffuse disease detection model.

[0119] Step P223: Obtain the fourth labeled sample data and input the fourth labeled sample data into the improved diffuse disease detection model for training. During the training process, execute the label smoothing algorithm to obtain the optimized diffuse disease detection model.

[0120] Step S2”: Input the image to be detected and the target contour image into the optimized diffuse disease detection model for prediction to obtain the initial diffuse disease diagnosis result.

[0121] Thus, by basing the model on and improving the GhostNet network framework, the optimized diffuse disease detection model can be obtained, which can reduce the amount of computation and the number of acceptable parameters, thereby reducing the model complexity. In addition, using the image to be detected and the target contour image as the four-channel input parameters of the model can provide richer feature information, thereby improving the accuracy of the diffuse disease detection model prediction.

[0122] The fourth labeled sample data includes a third category label, which carries the diagnostic result of diffuse disease for the corresponding sample image. In a specific example, the optimized diffuse disease detection model includes a liver diffuse classification model; based on image features and physiological knowledge, the third category label corresponding to the fourth labeled sample data may include sonographic changes of fatty liver, sonographic changes of diffuse liver, sonographic changes of cirrhosis, sonographic changes of polycystic liver, and no obvious abnormalities in the liver (no diffuse lesions in the liver), etc.

[0123] It should be noted that, considering that the diagnosis of diffuse liver diseases mainly relies on the analysis and judgment of the internal texture of the liver, as well as the overall echo of the liver and the echo of other abdominal organs, the first to third channels of the optimized diffuse disease detection model are used to input the RGB information of the image to be detected, and the fourth channel is used to input the information of the target contour detection image. This can highlight the texture and echo information of the liver itself, which is very important for the diagnosis of diffuse liver diseases.

[0124] Furthermore, since the differences between categories of diffuse diseases are not significant, this invention employs the label smoothing algorithm to suppress overfitting of the optimized diffuse disease detection model. Specifically, the label smoothing algorithm adds a small amount of uniformly distributed random noise to the labels of the training samples, transforming the target output by the network model from the original absolute one-hot encoding to a smoothed target. For example, if the original sample data is divided into 5 categories (0, 1, 2, 3, and 4), and the target category is category 1, the original one-hot encoding of the output target (each value ranges from 0 to 1, where 0 represents a probability of 0 for the current category and 1 represents a probability of 1 for the current category) is [0, 1, 0, 0, 0]. After executing the label smoothing algorithm, the output target becomes [0.025, 0.9, 0.025, 0.025, 0.025]. It is evident that after the label smoothing algorithm, the output target of the training network is no longer either 0 or 1, thus making the network training target less absolute and mitigating overfitting.

[0125] Here, steps S2' and S2" can be understood as derived steps of step S2; the execution order of steps P211 to S2' and steps P221 to S2" can be interchanged, or steps P211 to S2' can be interspersed between steps P221 to S2" . In other words, the initial focal disease diagnosis result and the initial diffuse disease diagnosis result are two parallel methods, and there is no sequential dependency between them. Figure 1 and Figure 9 As shown, in the fourth embodiment of the present invention, the ultrasound imaging disease identification method may specifically include the following steps:

[0126] Step S31: Obtain and determine whether several original images to be detected are consecutive frames;

[0127] Step S32: Based on the judgment results of the consecutive frames, determine and correct the initial disease diagnosis result according to the correction strategy to obtain the final disease diagnosis result.

[0128] The aforementioned original images to be detected include the current image to be corrected and several historical images. Thus, correcting the initial disease diagnosis based on continuous detection images integrates results from multiple time points, avoids temporary noise, and ensures a continuous and comprehensive diagnosis.

[0129] like Figure 10 As shown, for step S31, the present invention also provides a refined step, which may specifically include:

[0130] Step S311: Based on the plurality of original images to be detected, calculate the similarity between two adjacent frames of original images to be detected to obtain a plurality of image similarity values;

[0131] Step S312: Determine whether the similarity value of any of the images is less than or equal to the first similarity preset threshold;

[0132] If so, proceed to step S313A to determine that the plurality of original images to be detected are not consecutive frames;

[0133] If not, proceed to step S313B to determine that the plurality of original images to be detected are consecutive frames.

[0134] Thus, by calculating image similarity values, the continuity and consistency among several original images to be detected can be effectively determined. The calculation method is simple and easy to implement, providing a basis for the correction of subsequent diagnostic results. The first similarity preset threshold can be dynamically adjusted according to the actual application scenario; preferably, it can be set to 0.3.

[0135] Furthermore, such as Figure 11 As shown, for step S32, the present invention also provides detailed steps, which may specifically include:

[0136] Step S321: When the plurality of original images to be detected are not consecutive frames, the initial disease diagnosis result corresponding to the current image to be corrected is taken as its final disease diagnosis result.

[0137] Step S322: When the plurality of original images to be detected are consecutive frames, the initial disease diagnosis result of the current image to be corrected is corrected according to the image similarity value to obtain the final disease diagnosis result.

[0138] In this way, the original images to be detected that are not consecutive frames are not corrected, while the original images to be detected that are consecutive frames are corrected based on their image similarity. This approach is highly adaptive, meets practical needs, and is highly accurate.

[0139] The image similarity value can be understood as calculating the similarity between any two adjacent frames of the original, ordered (e.g., in chronological order of capture time) original images to be detected. Specifically, algorithms such as mean squared error or structural similarity can be used to map the calculation results to a range of 0 to 1 to obtain the image similarity value. Structural similarity can be understood as an indicator of the similarity between two images, mainly used to detect the similarity between two detection images of the same size, or the degree of distortion in the detection images.

[0140] As shown in Figure 12(a), on the one hand, when correcting the initial diffuse disease diagnosis result, the part of step S322 that "corrects the initial disease diagnosis result of the current detection image to be corrected according to the image similarity value to obtain the final disease diagnosis result" can specifically include:

[0141] Step S3221: Determine whether the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to the second similarity preset threshold.

[0142] If so, proceed to step S3222A and use the initial disease diagnosis result of the previous historical detection image as the final disease diagnosis result of the current detection image to be corrected.

[0143] If not, proceed to step S3222B, statistically analyze the initial diffuse disease diagnosis results corresponding to the plurality of original images to be detected, and take the initial diffuse disease diagnosis result that appears most frequently in the statistical results as the final diffuse disease diagnosis result of the current image to be corrected.

[0144] Thus, by determining the correlation between the current image to be corrected and historical images through image similarity values, the inherent correlation of image data can be fully utilized to improve the accuracy and reliability of diagnosis. Furthermore, this method combines a single current image to be corrected with a comprehensive assessment of multiple historical images, resulting in a more complete disease diagnosis.

[0145] The previous historical detection image and the current detection image to be corrected are two adjacent frames of detection images taken in chronological order. The second similarity preset threshold can also be dynamically adjusted according to the actual application scenario; preferably, it can be set to 0.95.

[0146] As shown in Figure 12(b), on the other hand, when correcting the initial focal disease diagnosis result, since the same image to be detected may contain multiple lesions, it is necessary to further determine the positional distribution of the same lesion in several frames of detection images. Based on this, the part of step S322 that "corrects the initial disease diagnosis result of the current detection image to be corrected according to the image similarity value to obtain its final disease diagnosis result" can specifically include:

[0147] Step S3221: Determine whether the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to the second similarity preset threshold.

[0148] If so, proceed to step S3222A and use the initial disease diagnosis result of the previous historical detection image as the final disease diagnosis result of the current detection image to be corrected.

[0149] If not, proceed to step S3222B', calculate and correct the initial focal disease diagnosis result of the current detection image to be corrected based on the focal diagnostic indicators of the several original images to be detected, and obtain the final focal disease diagnosis result.

[0150] Thus, by calculating the focal diagnostic indicators of the original images to be detected, the spatial relationship between different lesions is reflected, which is helpful in determining whether the lesions belong to the same disease type and the development trend of focal diseases. In addition, combining continuous frames (temporal features) and intersection-over-union ratio (spatial features) can improve the accuracy of prediction results.

[0151] The focal diagnostic indicators include at least one of the overlap ratio and the distance between lesions. Specifically, the overlap ratio can be understood as the degree of overlap between adjacent lesions, which can be quantified by the ratio of the overlapping area to the total lesion area. A higher overlap ratio indicates a larger overlapping area between lesions, making it more likely that they belong to the same pathological process. The distance between lesions can be understood as the closest distance between the boundaries of different lesions; the closer the distance, the more likely the lesions are to have spatial diffusion and continuity.

[0152] In a preferred embodiment, the crossover ratio can be used to calculate the area of ​​lesions with larger areas, while the distance between lesions can be used to calculate the area of ​​lesions with smaller areas. In addition, steps S3221 to S3222B and steps S3221 to S3222B' can be understood as derivative steps of step S322.

[0153] Furthermore, such as Figure 13 As shown, for step S3222B', the present invention provides a refined step, which may specifically include:

[0154] S3222B1', based on the initial focal disease diagnosis result of the current image to be corrected, determine several focal disease types and several focal disease locations.

[0155] S3222B2', based on the location of the first focal lesion, calculate the focal diagnostic index of the current detection image to be corrected and the plurality of historical detection images to obtain the first focal diagnostic result;

[0156] S3222B3', Based on the first focal diagnosis result, the first focal disease type corresponding to the current detection image to be corrected is corrected to obtain the corrected first focal disease type;

[0157] S3222B4', iterate until all focal lesion types have been corrected, and obtain the final focal lesion diagnosis result of the current image to be corrected.

[0158] Thus, by calculating and optimizing the preliminary diagnostic results of each lesion based on focal diagnostic indicators, the accuracy of diagnosis can be gradually improved. In addition, by optimizing at the global level, the errors of a single lesion can be avoided from affecting the overall diagnostic results.

[0159] The focal lesion type includes a first focal lesion type; the focal lesion location includes a first focal lesion location. The first focal lesion type can be understood as any lesion in the initial focal lesion diagnosis result corresponding to the current image to be corrected, and its corresponding location distribution is the first focal lesion location.

[0160] Furthermore, such as Figure 14 As shown, step S3222B3' can specifically include the following steps:

[0161] Step S3222B31': Determine whether the first focal diagnosis result meets the focal diagnosis criteria;

[0162] If so, proceed to step S3222B32A', count the focal disease types corresponding to the location of the first focal disease in the plurality of original images to be detected, and use the focal disease type that appears most frequently in the statistical results to correct the first focal disease type of the current image to be corrected, so as to obtain the final first focal disease type.

[0163] If not, proceed to step S3222B32B' and delete the first focal lesion type corresponding to the current image to be corrected.

[0164] The focal diagnostic criteria characterize the locational distribution of the first focal lesion type in the plurality of original images to be detected. Thus, by determining the locational distribution pattern of lesions in the plurality of original images to be detected, a correction strategy is determined, exhibiting strong adaptability and high accuracy.

[0165] For example, Figures 15(a) and 15(b) show a flowchart of an ultrasound imaging disease identification method and a flowchart of correcting the initial disease diagnosis result in a preferred embodiment. The ultrasound imaging disease identification process of this preferred embodiment is summarized below with reference to Figure 15(a), taking the liver as an example.

[0166] A labeled sample dataset (including first, second, third, and fourth labeled sample data) is constructed. Based on this dataset, an abdominal classification model (i.e., an optimized image classification model), a liver organ segmentation model (i.e., an image segmentation network model), a diffuse liver classification model (i.e., an optimized focal disease detection model), and a focal liver detection model (i.e., an optimized diffuse disease detection model) are trained sequentially. The abdominal classification model is used to classify and identify the original image to be detected, obtaining the image to be detected corresponding to the abdomen. Then, the target contour image is obtained through the liver organ segmentation model. An initial diagnosis result is obtained based on the diffuse liver classification model and the focal liver detection model. Finally, the initial diagnosis result is corrected to obtain the final disease diagnosis result.

[0167] The correction process of this preferred embodiment is summarized with reference to Figure 15(b). Input the current image to be corrected and the corresponding initial disease diagnosis result; calculate the similarity between the current image to be corrected and the previous historical image to obtain an image similarity value; cache a certain number of image similarity values ​​and corresponding initial disease diagnosis results (i.e., obtain several original images to be detected, as well as several image similarity values ​​and several initial disease diagnosis results). When the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to 0.95 (a second similarity preset threshold), the two images are considered very similar, and the diagnosis result of the previous historical image is directly taken as the final disease diagnosis result of the current image to be corrected (direct correction);

[0168] When the image similarity value between the current image to be corrected and the previous historical image is less than or equal to 0.3 (the first similarity preset threshold), these two images can be considered completely different. Therefore, the initial disease diagnosis result of the current image to be corrected will not be corrected. That is, the initial disease diagnosis result of the current image to be corrected will be directly used as the final disease diagnosis result.

[0169] When the image similarity value between the current image to be corrected and the previous historical image is greater than 0.3 and less than 0.95 (between the first and second similarity preset thresholds), it can be considered that the two images have a certain similarity and can be preliminarily considered to be in the same video or the same cross-section real-time scanning. The above corresponding schemes can be used for correction according to the type of initial disease diagnosis result.

[0170] Specifically, when the initial disease diagnosis result is diffuse, the initial diffuse disease diagnosis result that appears most frequently among the initial diffuse disease diagnosis results corresponding to several original images to be detected is taken as the final diffuse disease diagnosis result of the current image to be corrected and output. When the initial disease diagnosis result is focal, the cross-union ratio or the distance between lesions in several original images is calculated. If the cross-union ratio or the distance between lesions meets the preset conditions, the focal disease diagnosis result that appears most frequently in several consecutive images is taken as the final focal disease diagnosis result and output. If the cross-union ratio or the distance between lesions does not meet the preset conditions, the lesion is deleted from the current image to be corrected. The process is iterated until all lesions in the current image to be corrected are corrected.

[0171] Figure 16(a) shows the ultrasound image after correction processing of the initial focal disease diagnosis result in this preferred embodiment; Figure 16(b) shows the ultrasound image after correction processing of the initial diffuse disease diagnosis result in this preferred embodiment.

[0172] The present invention also provides an ultrasound imaging disease recognition system, comprising: an acquisition unit for acquiring an original image to be detected; a processing unit for processing the original image to be detected to obtain an image to be detected, and inputting the image to be detected into an image segmentation network model to obtain a target contour image; a disease diagnosis unit for performing lesion detection on the target contour image using a preliminary disease screening model to obtain an initial disease diagnosis result; a disease diagnosis correction unit for correcting the initial disease diagnosis result to obtain a final disease diagnosis result; and a display unit for outputting and displaying the final disease diagnosis result.

[0173] Thus, by sequentially employing an image segmentation network model and a preliminary disease screening model to analyze and diagnose the images to be tested, automated processing and analysis of the images can be achieved, reducing the burden on doctors, decreasing diagnostic errors, and improving work efficiency. Finally, by using a correction algorithm to revise the initial diagnostic results, the accuracy and reliability of the diagnosis can be improved, and misjudgments and subjective biases can be reduced.

[0174] In summary, the ultrasound imaging disease identification method provided by this invention employs an image segmentation network model for semantic segmentation of the image to be detected, extracting the target contour region and reducing interference from other irrelevant information such as non-target regions, thereby improving the accuracy of subsequent model predictions. Based on the detected image and the target contour model, preliminary focal disease detection models and preliminary diffuse disease detection models are used for detection, respectively, to quickly provide an initial diagnostic result, reducing the scope of subsequent processing. Finally, by correcting the initial diagnostic result, deviations in the preliminary diagnostic result can be corrected, reducing the influence of physician subjectivity on the diagnostic result and improving the accuracy of the final diagnosis.

[0175] Furthermore, the entire disease identification process integrates the advantages of deep learning and traditional image processing, achieving automated and intelligent analysis of ultrasound images, reducing reliance on operator or physician experience, shortening diagnostic time, and improving work efficiency. Simultaneously, by independently or jointly detecting diffuse and focal diseases, it helps provide accurate disease localization and type determination, facilitating better treatment plans for physicians and improving patient outcomes.

[0176] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0177] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying diseases using ultrasound imaging, characterized in that, Including the following steps: The target contour image is obtained by semantic segmentation of the image to be detected based on the image segmentation network model. A preliminary disease screening model is used to detect lesions in the target contour image to obtain an initial disease diagnosis result. The initial disease diagnosis result is corrected to obtain the final disease diagnosis result; The step of revising the initial disease diagnosis result to obtain the final disease diagnosis result includes: Acquire and determine whether several original images to be detected are consecutive frames; wherein, the several original images to be detected include the current image to be corrected and several historical images; When the original images to be detected are not consecutive frames, the initial disease diagnosis result corresponding to the current image to be corrected is taken as the final disease diagnosis result. When the original images to be detected are consecutive frames, the initial disease diagnosis result of the current image to be detected is corrected based on the image similarity value to obtain the final disease diagnosis result.

2. The ultrasound imaging disease identification method according to claim 1, characterized in that, The image segmentation network model is configured as a network model built on the BisenetV2 network framework.

3. The ultrasound imaging disease identification method according to claim 2, characterized in that, Before the step of "semantically segmenting the image to be detected based on an image segmentation network model to obtain a target contour image", the method further includes: An initial image segmentation network model was constructed using the BisenetV2 network framework; The context embedding module of the initial image segmentation network model is replaced with a pyramid pooling module; wherein, the feature extraction backbone network of the initial image segmentation network model includes the context embedding module and several encoders; During the training of the initial image segmentation network model, several auxiliary segmentation heads are set on several encoders to obtain an improved image segmentation network model; wherein, the several encoders correspond to several encoding stages and are used to extract semantic information of the original image to be detected at different stages. First labeled sample data is obtained and input into the improved image segmentation network model for training to obtain the image segmentation network model; wherein, the first labeled sample data includes a first category label, and the first category label carries the name of the tissue or organ corresponding to the sample image.

4. The ultrasound imaging disease identification method according to claim 1, characterized in that, The phrase "semantically segmenting the image to be detected based on an image segmentation network model to obtain a target contour image" specifically includes: During the prediction process of the image segmentation network model, several auxiliary segmentation heads configured in the image segmentation network model are removed to obtain an image segmentation network model without auxiliary segmentation heads. The image segmentation network model without auxiliary segmentation head is used to perform semantic segmentation on the image to be detected to obtain the target contour image.

5. The ultrasound imaging disease identification method according to claim 1, characterized in that, The phrase "semantically segmenting the image to be detected based on an image segmentation network model to obtain a target contour image" specifically includes: Based on the image to be detected, the image segmentation network model is used to extract its spatial feature map corresponding to spatial information and its semantic feature map corresponding to semantic information; The target contour image is obtained by performing an aggregation operation based on the spatial feature map and the semantic feature map.

6. The ultrasound imaging disease identification method according to claim 5, characterized in that, The image segmentation network model includes a detail branch, a semantic branch, and a bilateral guided aggregation layer; the step of "extracting spatial feature maps corresponding to spatial information and semantic feature maps corresponding to semantic information from the image to be detected using the image segmentation network model" specifically includes: The detailed branch is used to extract spatial information corresponding to different stages of the image to be detected, resulting in several spatial feature maps; The semantic branches are used to extract semantic information corresponding to different stages of the image to be detected, resulting in several semantic feature maps; The phrase "calculating the target contour image using an aggregation operation method based on the spatial feature map and the semantic feature map" specifically includes: Based on the pyramid pooling module, the several spatial feature maps and the several semantic feature maps are fused to obtain several fused feature maps; wherein, the pyramid pooling module is set in the last layer of the semantic branch; Based on the aforementioned fused feature maps, the target contour image is obtained by performing multi-scale information encoding and aggregation operations on them using the bilateral guided aggregation layer.

7. The ultrasound imaging disease identification method according to claim 1, characterized in that, Before the step of "semantically segmenting the image to be detected based on an image segmentation network model to obtain a target contour image", the method further includes: Build and optimize image classification models; The original image to be detected is obtained, and the original image to be detected is input into the optimized image classification model for classification and recognition to obtain the image to be detected.

8. The ultrasound imaging disease identification method according to claim 7, characterized in that, The "construction of an optimized image classification model" specifically includes: An initial image classification model was built using the GhostNet network framework; The number of input and output channels of all convolutional and fully connected layers in the initial image classification model is halved, and the single-dimensional attention module in the model is replaced with a multi-dimensional attention module, and the ReLU activation function is replaced with the Hard-swish activation function to obtain an improved image classification model. A second labeled sample data is obtained and input into the improved image classification model for training to obtain the optimized image classification model; wherein, the second labeled sample data includes a first judgment label, which carries a judgment result on whether the tissue part in the corresponding sample image is a first tissue part; The single-dimensional attention module includes the SENet attention module; the multi-dimensional attention module includes at least one of the SKNet attention module, the CBAM attention module, and the ECA attention module.

9. The ultrasound imaging disease identification method according to claim 1, characterized in that, The preliminary disease screening model includes an optimized focal disease detection model; the initial disease diagnosis result includes the initial focal disease diagnosis result; and the final disease diagnosis result includes the final focal disease diagnosis result.

10. The ultrasound imaging disease identification method according to claim 9, characterized in that, Before the step of "using a preliminary disease screening model to detect lesions in at least the target contour image to obtain an initial disease diagnosis result", the method further includes: An initial focal lesion detection model was constructed using the YOLOv5s network framework; The convolutional layer modules in the feature extraction backbone network and detection head component of the initial focal lesion detection model are replaced with Ghost module components, and the small target detection head is set in the second layer feature map output by the initial focal lesion detection model to obtain an improved focal lesion detection model; wherein, the Ghost module component includes a Ghost module ontology and a multi-dimensional attention module; A third labeled sample data is obtained and input into the improved focal disease detection model for training to obtain the optimized focal disease detection model; wherein, the third labeled sample data includes a second category label, and the second category label carries the focal disease diagnosis result of the corresponding sample image; The phrase "using a preliminary disease screening model to detect lesions in at least the target contour image to obtain an initial disease diagnosis result" specifically includes: The target contour image is input into the optimized focal lesion detection model for prediction to obtain the initial focal lesion diagnosis result.

11. The ultrasound imaging disease identification method according to claim 1, characterized in that, The preliminary disease screening model includes an optimized diffuse disease detection model; the initial disease diagnosis result includes the initial diffuse disease diagnosis result; and the final disease diagnosis result includes the final diffuse disease diagnosis result.

12. The ultrasound imaging disease identification method according to claim 11, characterized in that, Before the step of "using a preliminary disease screening model to detect lesions in at least the target contour image to obtain an initial disease diagnosis result", the method further includes: An initial diffuse disease detection model was built using the GhostNet network framework; The number of input and output channels of all convolutional and fully connected layers in the initial diffuse disease detection model is halved, and the single-dimensional attention module in the model is replaced with a multi-dimensional attention module, and the ReLU activation function is replaced with a Hard-swish activation function, to obtain an improved diffuse disease detection model. A fourth labeled sample data is obtained and input into the improved diffuse disease detection model for training. During the training process, a label smoothing algorithm is executed to obtain the optimized diffuse disease detection model. The fourth labeled sample data includes a third category label, which carries the diffuse disease diagnosis result of the corresponding sample image. The phrase "using a preliminary disease screening model to detect lesions in at least the target contour image to obtain an initial disease diagnosis result" specifically includes: The image to be detected and the target contour image are input into the optimized diffuse disease detection model for prediction to obtain the initial diffuse disease diagnosis result.

13. The ultrasound imaging disease identification method according to claim 1, characterized in that, The phrase "acquiring and determining whether several original images to be detected are consecutive frames" specifically includes: Based on the aforementioned original images to be detected, the similarity between two adjacent frames of original images to be detected is calculated to obtain several image similarity values. Determine whether any of the image similarity values ​​is less than or equal to a first similarity preset threshold; If so, then the plurality of original images to be detected are determined to be non-continuous frames; If not, then the plurality of original images to be detected are determined to be consecutive frames.

14. The ultrasound imaging disease identification method according to claim 1, characterized in that, The phrase "correcting the initial disease diagnosis result of the current detection image to be corrected based on the image similarity value to obtain its final disease diagnosis result" specifically includes: Determine whether the image similarity value between the current detection image to be corrected and the previous historical detection image is greater than or equal to a second similarity preset threshold; wherein, the previous historical detection image and the current detection image to be corrected are two adjacent frames of detection images taken in chronological order; If so, the initial disease diagnosis result of the previous historical detection image shall be used as the final disease diagnosis result of the current detection image to be corrected.

15. The ultrasound imaging disease identification method according to claim 14, characterized in that, The initial disease diagnosis result includes the initial diffuse disease diagnosis result; the final disease diagnosis result includes the final diffuse disease diagnosis result; after "determining whether the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to the second similarity preset threshold", the method further includes: If not, the initial diffuse disease diagnosis results corresponding to the original images to be detected are statistically analyzed, and the initial diffuse disease diagnosis result that appears most frequently in the statistical results is taken as the final diffuse disease diagnosis result of the current image to be corrected.

16. The ultrasound imaging disease identification method according to claim 14, characterized in that, The initial disease diagnosis result includes the initial focal disease diagnosis result; the final disease diagnosis result includes the final focal disease diagnosis result; after "determining whether the image similarity value between the current image to be corrected and the previous historical image is greater than or equal to the second similarity preset threshold", the method further includes: If not, the initial focal disease diagnosis result of the current detection image to be corrected is calculated and corrected based on the focal diagnostic indicators of the plurality of original images to be detected, so as to obtain the final focal disease diagnosis result; wherein, the focal diagnostic indicators include at least one of crossover ratio and lesion distance.

17. The ultrasound imaging disease identification method according to claim 16, characterized in that, The phrase "calculating and correcting the focal disease diagnosis result of the current detection image to be corrected based on the focal diagnostic indicators of the plurality of original images to be detected, to obtain the final focal disease diagnosis result" specifically includes: Based on the initial focal disease diagnosis result of the current image to be corrected, several focal disease types and several focal disease locations are determined; wherein, the focal disease type includes a first focal disease type; and the focal disease location includes a first focal disease location. Based on the location of the first focal lesion, the focal diagnostic index of the current detection image to be corrected and the plurality of historical detection images are calculated to obtain the first focal diagnostic result. Based on the first focal diagnosis result, the first focal disease type corresponding to the current detection image to be corrected is corrected to obtain the corrected first focal disease type; The process is iterated until all focal lesion types have been corrected, resulting in the final focal lesion diagnosis of the current image to be corrected.

18. The ultrasound imaging disease identification method according to claim 17, characterized in that, The phrase "based on the first focal diagnosis result, correcting the first focal disease type corresponding to the current detection image to be corrected, to obtain the corrected first focal disease type" specifically includes: Determine whether the first focal diagnosis result meets the focal diagnosis criteria; wherein, the focal diagnosis criteria characterize the location distribution of the first focal disease type in the plurality of original images to be detected; If so, the focal disease types corresponding to the location of the first focal disease in the plurality of original images to be detected are statistically analyzed, and the first focal disease type of the current image to be corrected is corrected by the focal disease type that appears most frequently in the statistical results, so as to obtain the final first focal disease type. If not, delete the first focal lesion type corresponding to the current image to be corrected.

19. An ultrasound imaging disease recognition system, characterized in that, include: The acquisition unit is used to acquire the original image to be detected; The processing unit is used to process the original image to be detected to obtain the image to be detected, and input the image to be detected into the image segmentation network model to obtain the target contour image; The disease diagnosis unit is used to detect lesions on the target contour image using a preliminary disease screening model to obtain an initial disease diagnosis result. A disease diagnosis correction unit is used to correct the initial disease diagnosis result to obtain a final disease diagnosis result; it is used to acquire and determine whether a plurality of original images to be detected are consecutive frames; wherein, the plurality of original images to be detected include the current image to be corrected and a plurality of historical images; when the plurality of original images to be detected are not consecutive frames, the initial disease diagnosis result corresponding to the current image to be corrected is taken as its final disease diagnosis result; when the plurality of original images to be detected are consecutive frames, the initial disease diagnosis result of the current image to be corrected is corrected according to the image similarity value to obtain the final disease diagnosis result. The display unit is used to output and display the final disease diagnosis result.

Images