Ultrasonic image detection method and system based on artificial intelligence

Through the ultrasonic image detection method based on artificial intelligence, the self-encoding network and convolutional neural network extract features and combined with a random forest model for diagnosis, the problems of unstable ultrasonic image quality and low diagnostic accuracy are solved, and efficient and accurate Image analysis and auxiliary diagnosis.

CN118552504BActive Publication Date: 2024-11-15THE FIRST AFFILIATED HOSPITAL OF SHANTOU UNIV MEDICAL COLLEGE
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410685335.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-30
Publication Date
2024-11-15
Estimated Expiration
2044-05-30

AI Technical Summary

Technical Problem

Ultrasonic image quality is susceptible to equipment and operator factors, and there are unstable image quality, limited resolution, artifacts and noise interference, resulting in low diagnostic accuracy and efficiency, and the image feature generalization ability that relies on manual design is limited, making it difficult to adapt. Complex clinical scenarios.

Method used

Using an ultrasonic image detection method based on artificial intelligence, deep features are extracted through self-encoding networks, semantic segmentation and random forest models are combined with convolutional neural networks for integrated learning, imaging microscopy features are generated, the confidence of diagnostic results is calculated and three-dimensional reconstruction is carried out , draw a pseudo-color heat map and generate a structured diagnostic report.

Benefits of technology

It improves the quality and diagnostic accuracy of ultrasound images, reduces the subjectivity of manual operations, quantifies the credibility of diagnostic results, provides an important reference for clinical decision-making, and improves the efficiency and interpretability of diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118552504B_ABST
    Figure CN118552504B_ABST
Patent Text Reader

Abstract

The present invention provides an ultrasonic image detection method and system based on artificial intelligence, which relates to the technical field of ultrasonic image processing, and includes: acquiring an ultrasonic image sequence and preprocessing it in sequence to obtain a standard ultrasonic image sequence, extracting a deep feature vector and generating a first feature sequence, adjusting the length of the first feature sequence and fusing time series information to obtain ultrasonic image features; adding the ultrasonic image features to a convolutional neural network, performing semantic segmentation, generating contour boundaries, extracting a binary mask, multiplying the original ultrasonic image pixel by pixel to generate an ultrasonic image subset, performing feature extraction and weighted fusion to obtain an imaging genomics feature vector; adding the ultrasonic image features to a random forest model, diagnosing each region of interest and generating a diagnostic result, performing three-dimensional reconstruction of the lesion, calculating the confidence of the diagnostic result and mapping it to the surface of the three-dimensional model, drawing a pseudo-color heat map to obtain a three-dimensional visual model, and comprehensively analyzing to generate a structured diagnostic report.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

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

[0002] Ultrasound imaging is a non-invasive, real-time, low-cost medical imaging method that is widely used in clinical diagnosis and treatment. However, the quality of ultrasound images is easily affected by factors such as the imaging equipment, the operator's technical level and the patient's own conditions. There are problems such as unstable image quality, limited resolution, artifacts and noise interference, which bring challenges to accurate diagnosis. At the same time, the interpretation and analysis of ultrasound images are highly dependent on the doctor's experience and subjective judgment, and there is a risk of misdiagnosis and missed diagnosis, which affects the accuracy and efficiency of diagnosis.

[0003] In order to improve the quality and diagnostic accuracy of ultrasound images, researchers have proposed a variety of image processing and analysis techniques. Traditional ultrasound image processing methods include image denoising, enhancement, segmentation, etc., which improve the readability and interpretability of images by improving image quality and highlighting regions of interest. However, these methods mainly rely on manually designed image features and rules, have limited generalization capabilities, and are difficult to adapt to complex and changing clinical application scenarios.

[0004] In recent years, artificial intelligence technology, especially deep learning methods, has made significant progress in the field of medical image analysis. Therefore, there is an urgent need for a new ultrasound imaging detection technology that can perfectly apply artificial intelligence technology to the field of medical image analysis. Summary of the invention

[0005] The embodiments of the present invention provide an ultrasonic image detection method and system based on artificial intelligence, which can at least solve some of the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides an ultrasonic image detection method based on artificial intelligence, comprising:

[0007] Acquire an ultrasound image sequence of the tissue to be detected and preprocess each frame of the ultrasound image sequence in turn to obtain a standard ultrasound image sequence, extract deep feature vectors from the standard ultrasound image sequence through unsupervised learning in combination with a sequence autoencoder network, generate a first feature sequence according to a time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm and fuse the time sequence information to obtain ultrasound image features;

[0008] Adding the ultrasound image features to a pre-set convolutional neural network, performing semantic segmentation through the encoding structure, generating contour boundaries corresponding to different tissues and organs, extracting a binary mask in combination with connected domain analysis, multiplying the binary mask by the original ultrasound image pixel by pixel to generate an ultrasound image subset, performing feature extraction on the ultrasound image subset and performing weighted fusion in combination with an attention mechanism to obtain an imaging omics feature vector;

[0009] The imaging genomics feature vector is added to a preset random forest model, each region of interest is diagnosed and a diagnosis result is generated through an integrated learning algorithm, the lesion is three-dimensionally reconstructed in combination with the position coordinates of the region of interest, the confidence of the diagnosis result is calculated and mapped to the surface of the constructed three-dimensional model, a pseudo-color heat map is drawn based on the mapping result, a three-dimensional visual model is obtained, the diagnosis result and the three-dimensional visual model are comprehensively analyzed, and a structured diagnosis report is generated.

[0010] In an optional embodiment,

[0011] Acquire an ultrasound image sequence of the tissue to be detected and preprocess each frame of the ultrasound image sequence in turn to obtain a standard ultrasound image sequence, combine the sequence autoencoder network, extract the deep feature vector in the standard ultrasound image sequence through unsupervised learning, and generate a first feature sequence according to the time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm and fuse the time sequence information, and obtain the ultrasound image features including:

[0012] Acquiring an ultrasonic image sequence of the tissue to be detected by an ultrasonic imaging device, removing speckle noise in each frame of the ultrasonic image sequence by a mean filtering operation, enhancing the contrast of the image by a logarithmic transformation, and performing a standardization operation to obtain the standard ultrasonic image sequence;

[0013] Constructing a sequence autoencoder network comprising a plurality of alternating convolutional layers and pooling layers, adding the standard ultrasound image sequence to the sequence autoencoder network, mapping each frame image in the standard ultrasound image sequence to a low-dimensional feature space by layer-by-layer convolution and downsampling to obtain a compact feature representation, reconstructing the original image based on the compact feature representation, determining the reconstruction difference between the input image and the reconstructed image, constructing a reconstruction loss function with the goal of minimizing the reconstruction difference, performing feature extraction based on the reconstruction loss function to obtain the deep feature vector, combining the standard ultrasound image sequence with the corresponding time series, adding a timestamp to the deep feature vector to obtain the first feature sequence;

[0014] By calculating the similarity of the feature vectors in the first feature sequence and the healthy feature sequence corresponding to the healthy population at different time steps, determining the optimal path through the similarity matrix and maximizing the cumulative similarity of the two sequences, dynamically time warping the first feature sequence according to the optimal path, and adaptively adjusting the length of the first feature sequence to obtain a standard feature sequence;

[0015] The standard feature sequence is added to a pre-constructed recurrent neural network, the feature vector of the current frame is received at each time step and the hidden state of the current time step is updated in combination with the hidden state of the previous time step, the network parameters of the recurrent neural network are optimized in combination with the back propagation algorithm, the timing dependency information is extracted and fused, and the ultrasound image features are obtained.

[0016] In an optional embodiment,

[0017] The first feature sequence is dynamically time-warped according to the optimal path as shown in the following formula:

[0018]

[0019] Where D represents the alignment cost calculated by the optimal path W, W represents the optimal path, k represents the alignment point, represents the ikth frame in the first feature sequence X, represents the jth in the health feature sequence Y k frame, λ represents the first weight coefficient, which is used to balance the sequence similarity and alignment, μ represents the smoothness constraint weight coefficient, i k-1 represents the k-1th frame of the first feature sequence, j k-1 Represents the k-1th frame of the health feature sequence.

[0020] In an optional embodiment,

[0021] The ultrasonic image features are added to a pre-set convolutional neural network, semantic segmentation is performed through the coding structure, contour boundaries corresponding to different tissues and organs are generated, binary masks are extracted in combination with connected domain analysis, and the binary masks are multiplied pixel by pixel with the original ultrasonic image to generate ultrasonic image subsets including:

[0022] Adding the ultrasound image features to a pre-set convolutional neural network, wherein the convolutional neural network is composed of symmetrical upsampling and downsampling paths, extracting multi-scale features through multi-layer convolution and pooling operations for the ultrasound image features, adding a nonlinear activation function after each convolution layer to activate the multi-scale features to obtain high-level semantic features, and performing step-by-step fusion of features for the high-level semantic features through multi-layer deconvolution and jump connections to obtain semantic segmentation results;

[0023] Applying a morphological post-processing method to the semantic segmentation result, filling the breaks and holes through a closing operation, and extracting the boundary between the foreground area and the background area in the semantic segmentation result in combination with a gradient operation, to obtain a binary boundary map corresponding to the contours of different tissues and organs;

[0024] The connected domains of the semantic segmentation results are marked, and the spatially adjacent pixels with the same category labels are divided into the same connected region and the current connected region is assigned a region number. In combination with the target tissue and organ type to be analyzed, the connected region with the corresponding category label is selected as the region of interest. By analyzing the area, aspect ratio of the circumscribed rectangle and circularity of the region of interest, the regions that do not conform to the morphology of the target organ and the deviated regions are eliminated to obtain a binary mask of the region of interest. The binary mask is multiplied pixel by pixel with the original ultrasound image to obtain an ultrasound image subset containing the target tissue and organ.

[0025] In an optional embodiment,

[0026] Feature extraction is performed on the ultrasound image subset and weighted fusion is performed in combination with the attention mechanism to obtain the radiomics feature vector including:

[0027] Based on the ultrasound image subset, the morphological features are determined by calculating the area and perimeter of the region of interest, the first-order texture statistical features are determined by calculating the mean, variance and skewness of the pixel grayscale values ​​in the region of interest, the contrast, correlation and energy are extracted by calculating the grayscale co-occurrence matrix corresponding to the region of interest, the second-order texture statistical index is determined, and multi-scale frequency domain analysis is performed in combination with wavelet transform to obtain multi-dimensional image features;

[0028] For each dimension of the multi-dimensional image features, a corresponding attention network is added. By determining the correlation between the features of the current dimension and the target output, the weight corresponding to the current dimension is adaptively adjusted, and the weight of each dimension is normalized to obtain the importance score of the features in the current dimension. Weighted fusion is performed according to the importance score to obtain the imaging genomics feature vector.

[0029] In an optional embodiment,

[0030] The radiomics feature vector is added to a preset random forest model, each region of interest is diagnosed and a diagnosis result is generated by an integrated learning algorithm, the lesion is three-dimensionally reconstructed in combination with the position coordinates of the region of interest, the confidence of the diagnosis result is calculated and mapped to the surface of the constructed three-dimensional model, a pseudo-color heat map is drawn based on the mapping result to obtain a three-dimensional visual model, the diagnosis result and the three-dimensional visual model are comprehensively analyzed, and a structured diagnosis report is generated, including:

[0031] According to the clinical diagnostic criteria, the radiomics feature vector corresponding to each region of interest is labeled, the corresponding disease category and case score are determined, and multiple random subsets are generated. For each random subset, an independent decision tree classification model is trained, and the features in the random subset are divided into different leaf nodes by recursively selecting the best splitting feature and the splitting quality is determined according to the Gini coefficient. The optimal splitting point is selected, and the decision tree classification model is improved in combination with the pre-set constraint conditions, and the decision tree is repeatedly constructed to obtain the random forest model;

[0032] The radiomics feature is added to the random forest model, each decision tree makes independent predictions and outputs prediction results, and the diagnosis result corresponding to the current radiomics feature is determined through a voting mechanism;

[0033] Based on the diagnosis result and the position coordinates corresponding to the region of interest, the lesion area is marked in the three-dimensional space, the surface of the marked three-dimensional lesion area is reconstructed, the voxel marks of the three-dimensional lesion area are converted into a mesh representation of triangular facets through a triangulation algorithm to obtain a three-dimensional lesion surface model, the confidence of the diagnosis result corresponding to each region of interest is calculated and mapped to the three-dimensional lesion surface model, the three-dimensional lesion surface model is rendered in pseudo color based on the confidence, the confidence is mapped to different colors through the color coding method of the heat map, and the three-dimensional visual model is obtained;

[0034] Based on the diagnosis results and the three-dimensional visual model, the disease category and pathological score of each region of interest are extracted respectively, and the spatial distribution and morphological feature information of the lesions are extracted in combination with the confidence level and the disease incidence frequency, and a structured diagnosis report is comprehensively generated.

[0035] In an optional embodiment,

[0036] The confidence of the diagnosis result corresponding to each region of interest is calculated as shown in the following formula:

[0037]

[0038] Among them, H represents confidence, c represents disease category, x represents the radiomics characteristics of the region of interest, α represents the weight coefficient of the random forest model, RF represents the random forest model, t represents the decision tree, and w t represents the weight of decision tree t, I() represents the indicator function, β represents the radiomics feature weight, f represents the radiomics feature, F represents the radiomics feature set, r f represents the relevance weight of the radiomics feature f, Φ f(x) represents the diagnostic relevance function value corresponding to the radiomics feature f, γ represents the spatial consistency weight, and s(c|M) represents the spatial consistency function value based on the three-dimensional lesion model M.

[0039] A second aspect of this embodiment provides an ultrasonic imaging detection system based on artificial intelligence, including:

[0040] The first unit is used to obtain an ultrasound image sequence of a tissue to be detected and preprocess each frame of the ultrasound image sequence in sequence to obtain a standard ultrasound image sequence, extract a deep feature vector from the standard ultrasound image sequence through unsupervised learning in combination with a sequence autoencoder network, generate a first feature sequence according to a time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm, and fuse the time sequence information to obtain an ultrasound image feature;

[0041] The second unit is used to add the ultrasound image features to a pre-set convolutional neural network, perform semantic segmentation through the encoding structure, generate contour boundaries corresponding to different tissues and organs, extract binary masks in combination with connected domain analysis, multiply the binary masks by the original ultrasound image pixel by pixel to generate an ultrasound image subset, perform feature extraction on the ultrasound image subset and perform weighted fusion in combination with an attention mechanism to obtain an imaging genomics feature vector;

[0042] The third unit is used to add the imaging genomics feature vector to a preset random forest model, diagnose each region of interest and generate a diagnosis result through an integrated learning algorithm, perform three-dimensional reconstruction of the lesion in combination with the position coordinates of the region of interest, calculate the confidence of the diagnosis result and map it to the surface of the constructed three-dimensional model, draw a pseudo-color heat map based on the mapping result, obtain a three-dimensional visual model, comprehensively analyze the diagnosis result and the three-dimensional visual model, and generate a structured diagnosis report.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the embodiments of the present invention is:

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0050] In the present invention, by preprocessing the ultrasound image sequence, a standard ultrasound image sequence is obtained, the interference of image noise and artifacts is reduced, the quality and readability of the image are improved, and the foundation is laid for subsequent feature extraction and analysis. The automatically extracted imaging genomics features are more comprehensive and objective, and the diagnostic information contained in the image is fully utilized. Semantic segmentation is performed through the coding structure to generate contour boundaries corresponding to different tissues and organs. The binary mask is extracted in combination with the connected domain analysis to accurately segment the region of interest, reducing the subjectivity and variability of manual operation. The confidence of the diagnosis result is calculated to quantify the credibility of the diagnosis result, which provides an important reference for clinical decision-making. In summary, the present invention realizes intelligent analysis and auxiliary diagnosis of ultrasound images, improves the efficiency, accuracy and interpretability of diagnosis, provides strong support and reference for clinical diagnosis and treatment, and has significant technical advantages and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the process of an ultrasonic image detection method based on artificial intelligence according to an embodiment of the present invention;

[0052] Figure 2 It is a schematic diagram of the structure of an ultrasonic imaging detection system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0054] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0055] Figure 1 FIG. 1 is a flow chart of an ultrasonic imaging detection method based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S1. Obtain an ultrasound image sequence of the tissue to be detected and preprocess each frame of the ultrasound image sequence in turn to obtain a standard ultrasound image sequence, extract deep feature vectors from the standard ultrasound image sequence through unsupervised learning in combination with a sequence autoencoder network, generate a first feature sequence according to the time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm and fuse the time sequence information to obtain ultrasound image features;

[0057] The ultrasound image sequence is composed of a series of continuous ultrasound image frames, each frame is a two-dimensional ultrasound image at a specific time point. By combining these frames together, dynamic information about a certain organ or tissue can be obtained. The autoencoder network is a neural network used for efficient encoding of unsupervised learning data. The unsupervised learning is a machine learning method that does not rely on pre-labeled training data. The deep feature vector is a high-dimensional representation extracted by a deep neural network (such as an autoencoder network or a convolutional neural network). The dynamic time warping is an algorithm for measuring the similarity between two time series.

[0058] In an optional embodiment,

[0059] Acquire an ultrasound image sequence of the tissue to be detected and preprocess each frame of the ultrasound image sequence in turn to obtain a standard ultrasound image sequence, combine the sequence autoencoder network, extract the deep feature vector in the standard ultrasound image sequence through unsupervised learning, and generate a first feature sequence according to the time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm and fuse the time sequence information, and obtain the ultrasound image features including:

[0060] Acquiring an ultrasonic image sequence of the tissue to be detected by an ultrasonic imaging device, removing speckle noise in each frame of the ultrasonic image sequence by a mean filtering operation, enhancing the contrast of the image by a logarithmic transformation, and performing a standardization operation to obtain the standard ultrasonic image sequence;

[0061] Constructing a sequence autoencoder network comprising a plurality of alternating convolutional layers and pooling layers, adding the standard ultrasound image sequence to the sequence autoencoder network, mapping each frame image in the standard ultrasound image sequence to a low-dimensional feature space by layer-by-layer convolution and downsampling to obtain a compact feature representation, reconstructing the original image based on the compact feature representation, determining the reconstruction difference between the input image and the reconstructed image, constructing a reconstruction loss function with the goal of minimizing the reconstruction difference, performing feature extraction based on the reconstruction loss function to obtain the deep feature vector, combining the standard ultrasound image sequence with the corresponding time series, adding a timestamp to the deep feature vector to obtain the first feature sequence;

[0062] By calculating the similarity of the feature vectors in the first feature sequence and the healthy feature sequence corresponding to the healthy population at different time steps, determining the optimal path through the similarity matrix and maximizing the cumulative similarity of the two sequences, dynamically time warping the first feature sequence according to the optimal path, and adaptively adjusting the length of the first feature sequence to obtain a standard feature sequence;

[0063] The standard feature sequence is added to a pre-constructed recurrent neural network, the feature vector of the current frame is received at each time step and the hidden state of the current time step is updated in combination with the hidden state of the previous time step, the network parameters of the recurrent neural network are optimized in combination with the back propagation algorithm, the timing dependency information is extracted and fused, and the ultrasound image features are obtained.

[0064] The mean filter is a commonly used image smoothing technology used to remove noise in images. The speckle noise is a type of noise that is common in ultrasound and radar images and appears as granular interference. The compact feature representation refers to generating a low-dimensional and information-rich feature vector through an effective feature extraction method. The reconstruction loss function is used to measure the difference between the original input and the reconstructed output. The similarity matrix is ​​a matrix used to represent the similarity or dissimilarity between a group of objects. The optimal path is a path used to represent the best alignment between two time series in algorithms such as dynamic time warping (DTW). The time series dependency information refers to the correlation between time points in time series data.

[0065] Scanning the tissue to be detected by an ultrasonic imaging device to obtain a series of continuous ultrasonic images to form an ultrasonic image sequence, convolving the image with a mean filter to remove speckle noise in the image, performing logarithmic transformation on the grayscale value of the image to enhance the contrast of the image, normalizing the pixel value of the image to map the pixel value to the range of [0, 1] to achieve standardized representation between different images, and recording the preprocessed ultrasonic image sequence as a standard ultrasonic image sequence;

[0066] Constructing a sequence autoencoder network, including multiple alternating convolutional layers and pooling layers, using convolution kernels to perform convolution operations on the input image to extract local feature information, downsampling the output of the convolutional layer through the pooling layer to reduce the size of the feature map, and realizing feature compression and abstraction, gradually mapping the input image to a low-dimensional feature space through a combination of multiple layers of convolution and pooling, adding the standard ultrasound image sequence to the sequence autoencoder network, and obtaining a compact feature representation of each frame of the image through layer-by-layer convolution and downsampling, based on the compact feature representation, reconstructing an approximate representation of the original image through upsampling and deconvolution operations, calculating the reconstruction difference between the input image and the reconstructed image, constructing a reconstruction loss function, and optimizing the network with the goal of minimizing the reconstruction difference, extracting features from the standard ultrasound image sequence through the optimized sequence autoencoder network to obtain a deep feature vector, adding a timestamp to the deep feature vector in combination with the time series corresponding to the standard ultrasound image sequence, and forming a first feature sequence;

[0067] Construct an ultrasound image dataset of healthy people, extract features from ultrasound image sequences of healthy people to obtain a healthy feature sequence, calculate the similarity of feature vectors between the first feature sequence and the healthy feature sequence at different time steps, form a similarity matrix, find the optimal path in the similarity matrix to maximize the cumulative similarity of the two sequences, perform dynamic time warping on the first feature sequence according to the optimal path, adaptively adjust the sequence length, and obtain a standard feature sequence aligned with the healthy feature sequence;

[0068] A recurrent neural network is constructed, and the standard feature sequence is added to the recurrent neural network. The feature vector of the current frame is input at each time step. The hidden state of the current time step is updated by combining the current input and the hidden state of the previous time step at each time step. The parameters of the recurrent neural network are optimized through the back propagation algorithm, and the timing dependency information in the standard feature sequence is learned and extracted. At the last time step of the recurrent neural network, the final hidden state is used as the ultrasound image feature, which integrates the timing information of the entire sequence.

[0069] In this embodiment, the sequence autoencoder network maps the ultrasound image to a low-dimensional feature space through layer-by-layer convolution and downsampling, and obtains a compact and comprehensive feature representation that makes full use of the local and global information of the ultrasound image, can capture key diagnostic clues, and provides a rich feature basis for subsequent diagnostic analysis. By adaptively adjusting the length of the feature sequence, it is aligned with the healthy feature sequence and eliminates the difference in sequence length. The fusion of time series information helps to capture the dynamic changes in the development of the disease and improve the accuracy of diagnosis. In summary, this embodiment provides comprehensive, accurate, and efficient diagnostic support, helps to improve the quality and efficiency of ultrasound diagnosis, and provides a powerful auxiliary tool for clinical decision-making.

[0070] In an optional embodiment,

[0071] The first feature sequence is dynamically time-warped according to the optimal path as shown in the following formula:

[0072]

[0073] Where D represents the alignment cost calculated by the optimal path W, W represents the optimal path, k represents the alignment point, represents the ikth frame in the first feature sequence X, represents the jth in the health feature sequence Y k frame, λ represents the first weight coefficient, which is used to balance the sequence similarity and alignment, μ represents the smoothness constraint weight coefficient, i k-1 represents the k-1th frame of the first feature sequence, j k-1 Represents the k-1th frame of the health feature sequence.

[0074] In this embodiment, the most similar alignment method between the two sequences is determined by minimizing the cumulative similarity, so that the aligned sequences have the greatest similarity. By introducing an alignment constraint, excessive distortion of the sequence can be avoided, so that the aligned sequence can remain as consistent as possible in time. The first feature sequence is dynamically time-warped through the optimal path, and the sequence length can be adaptively adjusted to align it with the healthy feature sequence to obtain a standardized feature sequence representation. In summary, this embodiment effectively eliminates the difference in sequence length, improves the comparability and consistency of feature sequences, and provides a standardized feature basis for subsequent diagnostic analysis.

[0075] S2. adding the ultrasound image features to a pre-set convolutional neural network, performing semantic segmentation through the encoding structure, generating contour boundaries corresponding to different tissues and organs, extracting a binary mask in combination with the connected domain analysis, multiplying the binary mask by the original ultrasound image pixel by pixel to generate an ultrasound image subset, performing feature extraction on the ultrasound image subset and performing weighted fusion in combination with the attention mechanism to obtain an imaging genomics feature vector;

[0076] The semantic segmentation is a technology in computer vision that aims to classify each pixel in an image into a specific category. The contour boundary is the contour line of the boundary of an object in an image, indicating the dividing line between the object and the background or other objects. The connected domain analysis is a technology in image processing that is used to identify and mark all connected pixel regions in an image. The binary mask is an image with only two values, 0 and 1, and is commonly used for image segmentation and region extraction. The radiomics feature vector is a collection of quantitative features extracted from medical images, which is used to describe the shape, texture, intensity and statistical properties of the image.

[0077] In an optional embodiment,

[0078] The ultrasonic image features are added to a pre-set convolutional neural network, semantic segmentation is performed through the coding structure, contour boundaries corresponding to different tissues and organs are generated, binary masks are extracted in combination with connected domain analysis, and the binary masks are multiplied pixel by pixel with the original ultrasonic image to generate ultrasonic image subsets including:

[0079] Adding the ultrasound image features to a pre-set convolutional neural network, wherein the convolutional neural network is composed of symmetrical upsampling and downsampling paths, extracting multi-scale features through multi-layer convolution and pooling operations for the ultrasound image features, adding a nonlinear activation function after each convolution layer to activate the multi-scale features to obtain high-level semantic features, and performing step-by-step fusion of features for the high-level semantic features through multi-layer deconvolution and jump connections to obtain semantic segmentation results;

[0080] Applying a morphological post-processing method to the semantic segmentation result, filling the breaks and holes through a closing operation, and extracting the boundary between the foreground area and the background area in the semantic segmentation result in combination with a gradient operation, to obtain a binary boundary map corresponding to the contours of different tissues and organs;

[0081] The connected domains of the semantic segmentation results are marked, and the spatially adjacent pixels with the same category labels are divided into the same connected region and the current connected region is assigned a region number. In combination with the target tissue and organ type to be analyzed, the connected region with the corresponding category label is selected as the region of interest. By analyzing the area, aspect ratio of the circumscribed rectangle and circularity of the region of interest, the regions that do not conform to the morphology of the target organ and the deviated regions are eliminated to obtain a binary mask of the region of interest. The binary mask is multiplied pixel by pixel with the original ultrasound image to obtain an ultrasound image subset containing the target tissue and organ.

[0082] The high-level semantic features are high-level features extracted from images by deep learning models (such as convolutional neural networks). The multi-layer deconvolution is a technology for image generation and upsampling. By performing a deconvolution operation on the input feature map, its size is enlarged and the resolution of the original image is restored. The jump connection is a network structure used to connect feature maps of different levels in a deep neural network. The morphological post-processing method is a type of operation in image processing, which is used to perform morphological transformation on binary images to improve the structure and shape of the image. The closing operation is a type of morphological operation, including a dilation operation followed by an erosion operation. The binary boundary map is a boundary image extracted from the binary image, wherein the boundary pixel value is 1 and the non-boundary pixel value is 0. The ultrasound image subset refers to a specific subset selected from a group of ultrasound images for specific analysis and processing tasks.

[0083] Construct a convolutional neural network, including a downsampling path and an upsampling path. Through multi-layer convolution and pooling operations, the spatial size of the feature map is gradually reduced, the receptive field is increased, and multi-scale features are extracted. Through multi-layer deconvolution and jump connections, the spatial size of the feature map is gradually restored, and feature information of different scales is integrated. In the downsampling path, multi-layer convolution and pooling operations are performed on the input features to extract multi-scale features. After each convolution layer, a nonlinear activation function (such as ReLU) is added to activate the multi-scale features, introduce nonlinear expression capabilities, and obtain high-level semantic features. In the upsampling path, the high-level semantic features are upsampled through multi-layer deconvolution to restore the spatial size. The feature map of the corresponding layer in the downsampling path is spliced ​​with the feature map in the upsampling path through jump connections to achieve step-by-step fusion of features. In the output layer of the convolutional neural network, a semantic segmentation result of the same size as the input ultrasound image is obtained, and each pixel is assigned a corresponding tissue organ category label;

[0084] Define the structural element of morphological closing operation, such as circular or rectangular structural element, perform morphological closing operation on the semantic segmentation result, fill the small breaks and holes in the target area by dilation and then corrosion operation, make the target area more complete and continuous, calculate the gradient amplitude and direction of the semantic segmentation result, highlight the boundary between the foreground area and the background area, binarize the gradient amplitude, and obtain a binary boundary map to indicate the location of the contours of different tissues and organs;

[0085] Scan each pixel of the semantic segmentation result, divide the spatially adjacent pixels with the same category label into the same connected region, assign a unique region number to each connected region, distinguish different connected regions, select the connected region with the corresponding category label as the region of interest according to the target tissue and organ type to be analyzed, calculate the morphological features of each region of interest, such as the area, aspect ratio and circularity of the circumscribed rectangle, set the morphological feature thresholds of the target organ according to prior knowledge, such as the area range, aspect ratio range and circularity range, eliminate the regions that do not meet the morphological characteristics of the target organ, exclude the regions that are too small, too large, too narrow or too irregular, create an all-zero mask image of the same size as the original ultrasound image, set the pixel position corresponding to the screened region of interest to 1, and obtain the binary mask of the region of interest. For each pixel of the original ultrasound image, multiply it by the binary mask value of the corresponding position, keep the pixel with the mask value of 1 unchanged, and set the pixel with the mask value of 0 to 0, so as to extract the region of interest and obtain the ultrasound image subset containing the target tissue and organ.

[0086] In this embodiment, semantic segmentation of ultrasound images is achieved through a convolutional neural network, and each pixel is automatically classified into different tissue and organ categories to achieve pixel-level automatic segmentation, which greatly reduces the workload of manual operations, improves segmentation efficiency, and reduces the impact of subjective errors. Through deconvolution and jump connections, step-by-step fusion of features of different scales and restoration of spatial dimensions are achieved, which improves the refinement and consistency of the segmentation results. The gradient operation highlights the boundary between the foreground area and the background area, obtains clear tissue and organ contours, and provides accurate boundary information for subsequent morphological analysis. The region of interest extraction eliminates interference from the background area and other irrelevant areas, and provides an accurate target area for subsequent quantitative analysis. In summary, this embodiment improves segmentation efficiency, reduces subjective errors, and provides accurate and reliable regions of interest for subsequent quantitative analysis and diagnostic decisions, and has significant clinical application value.

[0087] In an optional embodiment,

[0088] Feature extraction is performed on the ultrasound image subset and weighted fusion is performed in combination with the attention mechanism to obtain the radiomics feature vector including:

[0089] Based on the ultrasound image subset, the morphological features are determined by calculating the area and perimeter of the region of interest, the first-order texture statistical features are determined by calculating the mean, variance and skewness of the pixel grayscale values ​​in the region of interest, the contrast, correlation and energy are extracted by calculating the grayscale co-occurrence matrix corresponding to the region of interest, the second-order texture statistical index is determined, and multi-scale frequency domain analysis is performed in combination with wavelet transform to obtain multi-dimensional image features;

[0090] For each dimension of the multi-dimensional image features, a corresponding attention network is added. By determining the correlation between the features of the current dimension and the target output, the weight corresponding to the current dimension is adaptively adjusted, and the weight of each dimension is normalized to obtain the importance score of the features in the current dimension. Weighted fusion is performed according to the importance score to obtain the imaging genomics feature vector.

[0091] The morphological features are quantitative features that describe the shape and structure of objects in the image. The skewness is a statistic that describes the degree of asymmetry of data distribution. The first-order texture statistical features are statistical features based on the grayscale value of the image and do not consider the spatial relationship between pixels. The second-order texture statistical features take into account the spatial relationship between pixels and are usually calculated through the co-occurrence matrix. The multi-scale frequency domain analysis is a method of analyzing image features at different scales and frequencies through transformation technology.

[0092] Sum the pixels of the binary mask image to obtain the total number of pixels in the region of interest, multiply the total number of pixels by the physical size of a single pixel (such as millimeters squared) to obtain the actual area of ​​the region of interest, perform edge extraction on the binary mask image to obtain the boundary pixels of the region of interest, calculate the total number of boundary pixels, multiply the total number by the physical size of a single pixel to obtain the perimeter of the region of interest, and calculate other morphological features such as circularity and compactness based on the area and perimeter to describe the shape characteristics of the region of interest;

[0093] Calculate the mean, variance and skewness of the pixel grayscale values ​​in the region of interest to obtain the first-order texture statistical features, select appropriate distance and direction parameters, such as distance of 1, directions of 0°, 45°, 90° and 135°, count the pixel pairs in the region of interest, obtain the frequency of occurrence of different grayscale combinations, normalize the frequency matrix, obtain the grayscale co-occurrence matrix, which reflects the spatial correlation of the pixel grayscale in the region, calculate the contrast, correlation and energy based on the grayscale co-occurrence matrix, obtain the second-order texture statistical features, select the wavelet basis function, perform wavelet decomposition on the image, obtain the wavelet coefficients at different scales and directions, extract the statistical features of the wavelet coefficients, such as mean, variance, energy, etc., to reflect the distribution characteristics of different frequency components, combine the wavelet domain features with the morphological features and texture statistical features of the spatial domain, and obtain a multi-dimensional image feature vector;

[0094] Design the architecture of the attention network, such as fully connected layers, activation functions, etc., use the features of the current dimension as the input of the attention network, and the target output as the supervision signal. By training the attention network, learn the correlation between the features of the current dimension and the target output, input the features of the current dimension into the corresponding attention network, obtain the importance score of the dimension, normalize the importance scores of all dimensions, obtain the weight coefficient of each dimension, multiply the features of each dimension by the corresponding weight coefficient to obtain the weighted features, concatenate or sum the weighted features to obtain the final radiomics feature vector.

[0095] In this embodiment, by extracting features of different dimensions, the morphology, grayscale distribution and texture characteristics of the region of interest are characterized from different angles, providing rich and complementary image information, and providing comprehensive quantitative indicators for subsequent disease diagnosis and prognosis prediction. By calculating the first-order statistical characteristics of the pixel grayscale values ​​in the region of interest and the second-order statistical characteristics of the grayscale co-occurrence matrix, a quantitative description of the texture characteristics of the target tissue and organ is achieved, providing an effective indicator for evaluating the texture abnormalities of tissues and organs. By introducing an attention network, the weight of each feature dimension is adaptively adjusted, highlighting the features with a high correlation with the target output, and improving the pertinence and effectiveness of the feature representation. In summary, this embodiment achieves a comprehensive and accurate quantitative characterization of the region of interest, and provides reliable data support and objective quantitative indicators for disease diagnosis, prognosis prediction and personalized medical decision-making.

[0096] S3. Add the radiomics feature vector to a pre-set random forest model, diagnose each region of interest and generate a diagnosis result through an integrated learning algorithm, perform three-dimensional reconstruction of the lesion based on the position coordinates of the region of interest, calculate the confidence of the diagnosis result and map it to the surface of the constructed three-dimensional model, draw a pseudo-color heat map based on the mapping result, obtain a three-dimensional visual model, comprehensively analyze the diagnosis result and the three-dimensional visual model, and generate a structured diagnosis report.

[0097] The random forest is an ensemble learning algorithm based on decision trees, which is used for classification and regression tasks. The ensemble learning algorithm is a technology that improves the overall model performance by combining the prediction results of multiple base learners (such as decision trees, neural networks, etc.). The pseudo-color heat map is a method of converting grayscale images or numerical data into color images, which is used to enhance the visualization of data.

[0098] In an optional embodiment,

[0099] The radiomics feature vector is added to a preset random forest model, each region of interest is diagnosed and a diagnosis result is generated by an integrated learning algorithm, the lesion is three-dimensionally reconstructed in combination with the position coordinates of the region of interest, the confidence of the diagnosis result is calculated and mapped to the surface of the constructed three-dimensional model, a pseudo-color heat map is drawn based on the mapping result to obtain a three-dimensional visual model, the diagnosis result and the three-dimensional visual model are comprehensively analyzed, and a structured diagnosis report is generated, including:

[0100] According to the clinical diagnostic criteria, the radiomics feature vector corresponding to each region of interest is labeled, the corresponding disease category and case score are determined, and multiple random subsets are generated. For each random subset, an independent decision tree classification model is trained, and the features in the random subset are divided into different leaf nodes by recursively selecting the best splitting feature and the splitting quality is determined according to the Gini coefficient. The optimal splitting point is selected, and the decision tree classification model is improved in combination with the pre-set constraint conditions, and the decision tree is repeatedly constructed to obtain the random forest model;

[0101] The radiomics feature is added to the random forest model, each decision tree makes independent predictions and outputs prediction results, and the diagnosis result corresponding to the current radiomics feature is determined through a voting mechanism;

[0102] Based on the diagnosis result and the position coordinates corresponding to the region of interest, the lesion area is marked in the three-dimensional space, the surface of the marked three-dimensional lesion area is reconstructed, the voxel marks of the three-dimensional lesion area are converted into a mesh representation of triangular facets through a triangulation algorithm to obtain a three-dimensional lesion surface model, the confidence of the diagnosis result corresponding to each region of interest is calculated and mapped to the three-dimensional lesion surface model, the three-dimensional lesion surface model is rendered in pseudo color based on the confidence, the confidence is mapped to different colors through the color coding method of the heat map, and the three-dimensional visual model is obtained;

[0103] Based on the diagnosis results and the three-dimensional visual model, the disease category and pathological score of each region of interest are extracted respectively, and the spatial distribution and morphological feature information of the lesions are extracted in combination with the confidence level and the disease incidence frequency, and a structured diagnosis report is comprehensively generated.

[0104] The random subset is a part of data randomly extracted from the data set, the optimal splitting feature refers to the feature selected for splitting the node during the decision tree construction process, so that the child node after the split is purer, the lesion area is the area showing abnormal lesions or pathological changes in the medical image, the surface reconstruction is the process of generating a three-dimensional surface model from two-dimensional image data or three-dimensional point cloud data, the triangulation algorithm is the process of dividing a point set in a plane or space into non-overlapping triangles or triangular facets, the voxel marker is the label or classification of each volume pixel in a three-dimensional image, the triangular facet is a geometric model composed of a group of triangles, and is often used to represent a three-dimensional surface, the pseudo-color rendering is the mapping of the pixel values ​​of a grayscale image to a color space to enhance the visualization of the image, and the color encoding is the process of assigning colors to data to represent different values ​​or categories.

[0105] Refer to the clinical diagnostic criteria to determine the disease category corresponding to each region of interest, and score each region of interest according to the severity of the disease and the clinical stage, and generate multiple random subsets, each of which contains a part of the annotated radiomics feature vectors;

[0106] For each random subset, train an independent decision tree classification model, recursively select the best split features, divide the features in the random subset into different leaf nodes, evaluate the split quality according to the Gini coefficient, select the optimal split point, make the child nodes after splitting purer, combine the pre-set constraints, such as the maximum depth of the tree, the minimum number of samples of leaf nodes, etc., improve the decision tree classification model, generate different training subsets and feature subsets through self-service sampling and random feature selection, train a decision tree for each subset, obtain multiple independent decision tree classifiers, and combine all decision trees into a random forest model;

[0107] For each decision tree, use the new radiomics feature to make an independent prediction and obtain the prediction result. Through a voting mechanism, such as majority vote decision or weighted average, the prediction results of all decision trees are integrated to determine the disease category and case score corresponding to the current radiomics feature as the final diagnosis result. The position coordinates of each region of interest are mapped into three-dimensional space to form a three-dimensional lesion area. According to the diagnosis result, the three-dimensional lesion area is marked to distinguish different disease categories and severity. Through the triangulation algorithm, the voxel mark of the three-dimensional lesion area is converted into a mesh representation of triangular facets. The mesh representation is optimized, such as smoothing, hole filling, etc., to obtain a three-dimensional lesion surface model.

[0108] Based on the prediction results of the random forest model, the confidence of the diagnosis results of each region of interest is calculated, and the confidence is mapped to the corresponding three-dimensional lesion surface model to form a confidence distribution map. The confidence is mapped to different colors through the color coding method of the heat map, and the three-dimensional lesion surface model is rendered in pseudo color to intuitively display the severity and distribution of the disease;

[0109] The disease category and pathological score of each region of interest are extracted as the key information of the diagnosis results. The reliability and importance of the diagnosis results of each region of interest are evaluated in combination with the confidence and disease frequency. The three-dimensional visual model is analyzed, and the spatial distribution characteristics of the lesions, such as position, size, shape, etc., are extracted. The morphological indicators of the lesions, such as volume, surface area, compactness, etc., are calculated to quantitatively describe the morphological characteristics of the lesions. The diagnosis results, confidence, spatial distribution and morphological characteristics information are integrated into a structured diagnosis report.

[0110] In this embodiment, the random forest model uses the association between imaging genomics features and disease categories to automatically diagnose new ultrasound images, predict disease categories and severity, provide objective and accurate diagnostic references, reduce the workload of doctors, and improve diagnostic efficiency and accuracy. Three-dimensional visualization enables doctors to understand the spatial characteristics of lesions more comprehensively and intuitively, providing an important reference for surgical planning and treatment plan formulation. Morphological feature quantification provides objective indicators such as lesion size and complexity, which facilitates doctors to assess the severity of the disease, monitor disease progression, and evaluate treatment effects. In summary, this embodiment realizes intelligent analysis and comprehensive presentation of ultrasound images through technical means such as intelligent diagnosis, three-dimensional visualization, morphological feature quantification, and structured diagnostic reports, providing strong support and innovative means for clinical diagnosis, disease research, precision medicine, and medical education.

[0111] In an optional embodiment,

[0112] The confidence of the diagnosis result corresponding to each region of interest is calculated as shown in the following formula:

[0113]

[0114] Among them, H represents confidence, c represents disease category, x represents the radiomics characteristics of the region of interest, α represents the weight coefficient of the random forest model, RF represents the random forest model, t represents the decision tree, and w t represents the weight of decision tree t, I() represents the indicator function, β represents the radiomics feature weight, f represents the radiomics feature, F represents the radiomics feature set, r f represents the relevance weight of radiomics feature f, φ f (x) represents the diagnostic relevance function value corresponding to the radiomics feature f, γ represents the spatial consistency weight, and s(c|M) represents the spatial consistency function value based on the three-dimensional lesion model M.

[0115] In this embodiment, the confidence of the disease category is obtained by weighted averaging the prediction results of the decision tree, which fully utilizes the integrated learning ability of the random forest model and improves the sensitivity and specificity of the confidence calculation for disease category prediction. By introducing the diagnostic relevance of imaging genomics features, the confidence calculation fully utilizes the association information between imaging genomics features and disease categories, improves the specificity and interpretability of the confidence assessment, and introduces spatial consistency information to evaluate the distribution of disease categories in the lesion area. For situations where there are multiple lesions or there is heterogeneity within the lesions, the spatial consistency information can provide additional judgment basis and improve the reliability of the confidence calculation. In summary, this embodiment provides a comprehensive and reliable confidence assessment indicator, which provides important reference information for clinical decision-making.

[0116] Figure 2 FIG. 1 is a schematic diagram of the structure of an ultrasonic imaging detection system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0117] The first unit is used to obtain an ultrasound image sequence of a tissue to be detected and preprocess each frame of the ultrasound image sequence in sequence to obtain a standard ultrasound image sequence, extract a deep feature vector from the standard ultrasound image sequence through unsupervised learning in combination with a sequence autoencoder network, generate a first feature sequence according to a time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm, and fuse the time sequence information to obtain an ultrasound image feature;

[0118] The second unit is used to add the ultrasound image features to a pre-set convolutional neural network, perform semantic segmentation through the encoding structure, generate contour boundaries corresponding to different tissues and organs, extract binary masks in combination with connected domain analysis, multiply the binary masks by the original ultrasound image pixel by pixel to generate an ultrasound image subset, perform feature extraction on the ultrasound image subset and perform weighted fusion in combination with an attention mechanism to obtain an imaging genomics feature vector;

[0119] The third unit is used to add the imaging genomics feature vector to a preset random forest model, diagnose each region of interest and generate a diagnosis result through an integrated learning algorithm, perform three-dimensional reconstruction of the lesion in combination with the position coordinates of the region of interest, calculate the confidence of the diagnosis result and map it to the surface of the constructed three-dimensional model, draw a pseudo-color heat map based on the mapping result, obtain a three-dimensional visual model, comprehensively analyze the diagnosis result and the three-dimensional visual model, and generate a structured diagnosis report.

[0120] According to a third aspect of the embodiments of the present invention,

[0121] An electronic device is provided, comprising:

[0122] processor;

[0123] a memory for storing processor-executable instructions;

[0124] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0125] A fourth aspect of the embodiments of the present invention is:

[0126] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0127] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An ultrasonic imaging detection method based on artificial intelligence, characterized in that: include: Acquire an ultrasound image sequence of the tissue to be detected and preprocess each frame of the ultrasound image sequence in turn to obtain a standard ultrasound image sequence, extract deep feature vectors from the standard ultrasound image sequence through unsupervised learning in combination with a sequence autoencoder network, generate a first feature sequence according to a time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm and fuse the time sequence information to obtain ultrasound image features; Adding the ultrasound image features to a pre-set convolutional neural network, performing semantic segmentation through the encoding structure, generating contour boundaries corresponding to different tissues and organs, extracting a binary mask in combination with connected domain analysis, multiplying the binary mask by the original ultrasound image pixel by pixel to generate an ultrasound image subset, performing feature extraction on the ultrasound image subset and performing weighted fusion in combination with an attention mechanism to obtain an imaging omics feature vector; Adding the radiomics feature vector to a pre-set random forest model, diagnosing each region of interest and generating a diagnosis result through an integrated learning algorithm, performing three-dimensional reconstruction of the lesion in combination with the position coordinates of the region of interest, calculating the confidence of the diagnosis result and mapping it to the surface of the constructed three-dimensional model, drawing a pseudo-color heat map based on the mapping result to obtain a three-dimensional visual model, comprehensively analyzing the diagnosis result and the three-dimensional visual model, and generating a structured diagnosis report; Acquire an ultrasound image sequence of the tissue to be detected and preprocess each frame of the ultrasound image sequence in turn to obtain a standard ultrasound image sequence, combine the sequence autoencoder network, extract the deep feature vector in the standard ultrasound image sequence through unsupervised learning, and generate a first feature sequence according to the time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm and fuse the time sequence information, and obtain the ultrasound image features including: Acquiring an ultrasonic image sequence of the tissue to be detected by an ultrasonic imaging device, removing speckle noise in each frame of the ultrasonic image sequence by a mean filtering operation, enhancing the contrast of the image by a logarithmic transformation, and performing a standardization operation to obtain the standard ultrasonic image sequence; Constructing a sequence autoencoder network comprising a plurality of alternating convolutional layers and pooling layers, adding the standard ultrasound image sequence to the sequence autoencoder network, mapping each frame image in the standard ultrasound image sequence to a low-dimensional feature space by layer-by-layer convolution and downsampling to obtain a compact feature representation, reconstructing the original image based on the compact feature representation, determining the reconstruction difference between the input image and the reconstructed image, constructing a reconstruction loss function with the goal of minimizing the reconstruction difference, performing feature extraction based on the reconstruction loss function to obtain the deep feature vector, combining the standard ultrasound image sequence with the corresponding time series, adding a timestamp to the deep feature vector to obtain the first feature sequence; By calculating the similarity of the feature vectors in the first feature sequence and the healthy feature sequence corresponding to the healthy population at different time steps, determining the optimal path through the similarity matrix and maximizing the cumulative similarity of the two sequences, dynamically time warping the first feature sequence according to the optimal path, and adaptively adjusting the length of the first feature sequence to obtain a standard feature sequence; Adding the standard feature sequence to a pre-built recurrent neural network, receiving the feature vector of the current frame at each time step and updating the hidden state of the current time step in combination with the hidden state of the previous time step, optimizing the network parameters of the recurrent neural network in combination with a back propagation algorithm, extracting and fusing the time-dependent information, and obtaining the ultrasonic image features; Feature extraction is performed on the ultrasound image subset and weighted fusion is performed in combination with the attention mechanism to obtain the radiomics feature vector including: Based on the ultrasound image subset, the morphological features are determined by calculating the area and perimeter of the region of interest, the first-order texture statistical features are determined by calculating the mean, variance and skewness of the pixel grayscale values ​​in the region of interest, the contrast, correlation and energy are extracted by calculating the grayscale co-occurrence matrix corresponding to the region of interest, the second-order texture statistical index is determined, and multi-scale frequency domain analysis is performed in combination with wavelet transform to obtain multi-dimensional image features; For each dimension of the multi-dimensional image features, a corresponding attention network is added, and the weight corresponding to the current dimension is adaptively adjusted by determining the correlation between the feature of the current dimension and the target output, and the weight of each dimension is normalized to obtain the importance score of the feature in the current dimension, and weighted fusion is performed according to the importance score to obtain the imaging omics feature vector; The radiomics feature vector is added to a preset random forest model, each region of interest is diagnosed and a diagnosis result is generated by an integrated learning algorithm, the lesion is three-dimensionally reconstructed in combination with the position coordinates of the region of interest, the confidence of the diagnosis result is calculated and mapped to the surface of the constructed three-dimensional model, a pseudo-color heat map is drawn based on the mapping result to obtain a three-dimensional visual model, the diagnosis result and the three-dimensional visual model are comprehensively analyzed, and a structured diagnosis report is generated, including: According to the clinical diagnostic criteria, the radiomics feature vector corresponding to each region of interest is labeled, the corresponding disease category and case score are determined, and multiple random subsets are generated. For each random subset, an independent decision tree classification model is trained, and the features in the random subset are divided into different leaf nodes by recursively selecting the best splitting feature and the splitting quality is determined according to the Gini coefficient. The optimal splitting point is selected, and the decision tree classification model is improved in combination with the pre-set constraint conditions, and the decision tree is repeatedly constructed to obtain the random forest model; The radiomics feature is added to the random forest model, each decision tree makes independent predictions and outputs prediction results, and the diagnosis result corresponding to the current radiomics feature is determined through a voting mechanism; Based on the diagnosis result and the position coordinates corresponding to the region of interest, the lesion area is marked in the three-dimensional space, the surface of the marked three-dimensional lesion area is reconstructed, the voxel marks of the three-dimensional lesion area are converted into a mesh representation of triangular facets through a triangulation algorithm to obtain a three-dimensional lesion surface model, the confidence of the diagnosis result corresponding to each region of interest is calculated and mapped to the three-dimensional lesion surface model, the three-dimensional lesion surface model is rendered in pseudo color based on the confidence, the confidence is mapped to different colors through the color coding method of the heat map, and the three-dimensional visual model is obtained; Based on the diagnosis results and the three-dimensional visual model, the disease category and pathological score of each region of interest are extracted respectively, and the spatial distribution and morphological feature information of the lesions are extracted in combination with the confidence level and the disease incidence frequency, and a structured diagnosis report is comprehensively generated.

2. The method according to claim 1, characterized in that The first feature sequence is dynamically time-warped according to the optimal path as shown in the following formula: Where D represents the alignment cost calculated by the optimal path W, W represents the optimal path, k represents the alignment point, represents the i-th k frame, represents the jth in the health feature sequence Y k frame, λ represents the first weight coefficient, which is used to balance the sequence similarity and alignment, μ represents the smoothness constraint weight coefficient, i k-1 represents the k-1th frame of the first feature sequence, j k-1 Represents the k-1th frame of the health feature sequence.

3. The method according to claim 1, characterized in that The ultrasonic image features are added to a pre-set convolutional neural network, semantic segmentation is performed through the coding structure, contour boundaries corresponding to different tissues and organs are generated, binary masks are extracted in combination with connected domain analysis, and the binary masks are multiplied pixel by pixel with the original ultrasonic image to generate ultrasonic image subsets including: Adding the ultrasound image features to a pre-set convolutional neural network, wherein the convolutional neural network is composed of symmetrical upsampling and downsampling paths, extracting multi-scale features through multi-layer convolution and pooling operations for the ultrasound image features, adding a nonlinear activation function after each convolution layer to activate the multi-scale features to obtain high-level semantic features, and performing step-by-step fusion of features for the high-level semantic features through multi-layer deconvolution and jump connections to obtain semantic segmentation results; Applying a morphological post-processing method to the semantic segmentation result, filling the breaks and holes through a closing operation, and extracting the boundary between the foreground area and the background area in the semantic segmentation result in combination with a gradient operation, to obtain a binary boundary map corresponding to the contours of different tissues and organs; The connected domains of the semantic segmentation results are marked, and the spatially adjacent pixels with the same category labels are divided into the same connected region and the current connected region is assigned a region number. In combination with the target tissue and organ type to be analyzed, the connected region with the corresponding category label is selected as the region of interest. By analyzing the area, aspect ratio of the circumscribed rectangle and circularity of the region of interest, the regions that do not conform to the morphology of the target organ and the deviated regions are eliminated to obtain a binary mask of the region of interest. The binary mask is multiplied pixel by pixel with the original ultrasound image to obtain an ultrasound image subset containing the target tissue and organ.

4. The method according to claim 1, characterized in that The confidence of the diagnosis result corresponding to each region of interest is calculated as shown in the following formula: Among them, H represents confidence, c represents disease category, x represents the radiomics feature of the region of interest, α represents the weight coefficient of the random forest model, RF represents the random forest model, t represents the decision tree, and w t represents the weight of decision tree t, I() represents the indicator function, β represents the radiomics feature weight, f represents the radiomics feature, F represents the radiomics feature set, r f represents the relevance weight of radiomics feature f, φ f (x) represents the diagnostic relevance function value corresponding to the radiomics feature f, γ represents the spatial consistency weight, and s(c|M) represents the spatial consistency function value based on the three-dimensional lesion model M.

5. An ultrasonic imaging detection system based on artificial intelligence, used to implement the method according to any one of claims 1 to 4, characterized in that: include: The first unit is used to obtain an ultrasound image sequence of a tissue to be detected and preprocess each frame of the ultrasound image sequence in sequence to obtain a standard ultrasound image sequence, extract a deep feature vector from the standard ultrasound image sequence through unsupervised learning in combination with a sequence autoencoder network, generate a first feature sequence according to a time sequence, adjust the length of the first feature sequence through a dynamic time warping algorithm, and fuse the time sequence information to obtain an ultrasound image feature; The second unit is used to add the ultrasound image features to a pre-set convolutional neural network, perform semantic segmentation through the encoding structure, generate contour boundaries corresponding to different tissues and organs, extract binary masks in combination with connected domain analysis, multiply the binary masks by the original ultrasound image pixel by pixel to generate an ultrasound image subset, perform feature extraction on the ultrasound image subset and perform weighted fusion in combination with an attention mechanism to obtain an imaging genomics feature vector; The third unit is used to add the imaging genomics feature vector to a preset random forest model, diagnose each region of interest and generate a diagnosis result through an integrated learning algorithm, perform three-dimensional reconstruction of the lesion in combination with the position coordinates of the region of interest, calculate the confidence of the diagnosis result and map it to the surface of the constructed three-dimensional model, draw a pseudo-color heat map based on the mapping result, obtain a three-dimensional visual model, comprehensively analyze the diagnosis result and the three-dimensional visual model, and generate a structured diagnosis report.

6. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Citation Information

Patent Citations

  • Semantic segmentation method and device for tree structure in three-dimensional tomography image

    CN113192069A

  • Portable three-dimensional carotid artery ultrasonic automatic diagnosis system and method

    CN115553816A

  • Ultrasonic organ structure point positioning method and device based on deep learning

    CN118037685A

  • Systems for medical image visualization

    US20230386153A1

  • 3D image compounding for ultrasound fetal imaging

    WO2017186518A1