Ultrasonic image processing method based on image analysis

Abstract

The invention discloses an ultrasonic image processing method based on image analysis, and the method comprises the following steps: S1, carrying out the noise suppression, enhancement and normalization of an original ultrasonic image, and improving the image quality; s2, performing initial boundary positioning on the image by adopting a reverse firefly algorithm, and determining a target area; s3, extracting an accurate organization structure based on the deep segmentation network in combination with boundary guidance information; s4, extracting form, texture, gray scale and depth features of the segmented region, and constructing a high-dimensional medical feature vector; s5, selecting an optimal feature subset by using a grey wolf optimization algorithm, and removing redundancy; s6, training the diagnosis model and optimizing hyper-parameters of the diagnosis model through a grey wolf algorithm; and S7, performing identification and risk grading on the new image by using the model, and completing auxiliary diagnosis. According to the invention, accurate recognition and intelligent diagnosis of the ultrasonic image are realized, and the processing efficiency and the diagnosis accuracy are remarkably improved.

Description

Technical Field

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

[0002] In the field of medical imaging, ultrasound images are widely used in clinical diagnosis, screening, and intraoperative navigation due to their safety, low cost, and real-time performance. However, because the ultrasound imaging process is affected by multiple factors, including equipment accuracy, operator experience, and tissue interface reflections, the acquired images often suffer from strong noise interference, low contrast, and blurred boundaries. These issues significantly limit the accuracy and automation of subsequent image analysis, especially in the segmentation of target tissue structures and lesion identification.

[0003] Traditional ultrasound image processing methods rely primarily on manual experience and basic image processing techniques for denoising, enhancement, and target recognition. These methods often use fixed-parameter filters, threshold segmentation, edge detection operators, and other methods to process image data. These methods suffer from poor processing accuracy and adaptability, making it difficult to cope with the complex differences between different lesion morphologies and organ tissues, and are also sensitive to noise. In target region recognition, existing rule-based algorithms are mostly only applicable to images with clear contours and stable tissue structures. They are ineffective for processing images with blurred boundaries or non-structural interference, and are prone to missed or false detections. Furthermore, some methods attempt to introduce machine learning models to classify or identify images, but often fail to effectively integrate the logical connections between image segmentation, feature extraction, and diagnostic decision-making, resulting in insufficient systematicity and a lack of a closed loop.

[0004] In recent years, with the development of deep learning and swarm intelligence optimization technologies, some studies have begun to attempt to introduce methods such as convolutional neural networks (CNNs), U-Net, attention mechanisms, and genetic algorithms into the field of ultrasound image processing, and have made some progress. However, these methods still face multiple challenges in practical applications. On the one hand, most existing studies focus on model structure design, while the adaptive modeling and processing capabilities of image noise are limited, and there are still problems such as rough pre-processing and insufficient edge restoration. On the other hand, model training often relies on a large amount of manually annotated data and lacks automatic selection and compression strategies for "potential diagnostic features" in images, resulting in redundant feature dimensions and affecting model generalization performance and operational efficiency. At the same time, existing methods lack effective global strategies for model parameter adjustment and performance optimization. The stability of classification results relies on manual experience settings, resulting in insufficient adaptability across different organs and different lesion types.

[0005] Based on the above problems, the existing technology urgently needs an ultrasonic image intelligent processing method that forms an efficient closed loop between image preprocessing, target segmentation, feature extraction, and diagnostic classification. It can accurately identify and segment target tissue structures in high-noise, low-contrast ultrasonic images, and at the same time has strong adaptive feature selection and optimization modeling capabilities. To this end, the present invention proposes an image analysis method that integrates the reverse firefly algorithm and the gray wolf optimization algorithm, combines deep learning segmentation networks with multi-source feature extraction technology, and constructs an end-to-end ultrasonic image intelligent diagnosis process. While improving the diagnostic accuracy, it optimizes the model structure and operating efficiency, overcoming the shortcomings of the existing technology in terms of robustness, automation and system integration.

[0006] Therefore, how to provide an ultrasound image processing method based on image analysis is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0007] One objective of the present invention is to propose an ultrasound image processing method based on image analysis. This method fully integrates the inverse firefly algorithm, the gray wolf optimization algorithm, and a deep image segmentation network to construct a complete intelligent processing flow encompassing image preprocessing, target tissue structure segmentation, medical feature extraction, feature selection, and classification diagnosis. By introducing a swarm intelligence algorithm to optimize initial image boundary positioning and model hyperparameters, and utilizing a deep learning network to achieve high-precision segmentation and risk assessment, the present invention is capable of performing accurate analysis even in ultrasound images with high noise and blurred boundaries, demonstrating the advantages of high robustness, high automation, and good clinical adaptability.

[0008] An ultrasound image processing method based on image analysis according to an embodiment of the present invention includes the following steps:

[0009] S1. Acquire original ultrasound image data, preprocess the image, and generate preprocessed image data;

[0010] S2. Based on the pre-processed image data, using the reverse firefly algorithm to perform initial region positioning and output candidate region boundaries of the target tissue structure;

[0011] S3. Input the candidate region boundary into a deep image segmentation network based on an attention mechanism, perform image segmentation operations, and extract the precise target tissue structure region;

[0012] S4, obtaining high-dimensional medical image features from the target tissue structure region;

[0013] S5. Based on the high-dimensional medical image feature selection optimization model, the gray wolf optimization algorithm is used to filter redundant features and output the medical feature set with optimized dimension;

[0014] S6. Inputting the dimensionally optimized medical feature set into the auxiliary diagnosis model, using the Gray Wolf Optimization Algorithm to tune and train the structural parameters and hyperparameters of the auxiliary diagnosis model to generate an optimized medical image classification model;

[0015] S7. Apply the medical image classification model to the ultrasound image data to be diagnosed to achieve automatic recognition, classification judgment and risk level output of the target tissue structure.

[0016] Optionally, S1 specifically includes: acquiring original ultrasound image data, suppressing speckle noise in the image using median filtering, improving the contrast and edge details of the image through an image enhancement algorithm, performing grayscale normalization processing, distributing the grayscale values within a uniform range, and generating a preprocessed image data set.

[0017] Optionally, the S2 specifically includes:

[0018] S21. Preprocessed image dataset Each image in Perform feature extraction and construct a multidimensional feature vector set F suitable for target area recognition in a noisy environment i ;

[0019] S22, set the brightness function and search space constraints of the reverse firefly algorithm, and set the multidimensional feature vector set F i In the input optimization model, the population position set X is initialized as {x1,x2,…,x m}, where x j represents the coordinates of the center point of the candidate target area corresponding to the j-th firefly in the image, and m is the population size;

[0020] S23. Execute the iterative update formula based on brightness attraction in the reverse firefly algorithm:

[0021]

[0022] in, represents the position information of the i-th firefly at the t-th iteration, represents the position information of the j-th firefly at the t-th iteration, β0 is the initial attraction, γ is the light intensity attenuation coefficient, r ij is the Euclidean distance between the i-th and j-th fireflies, and α·∈ is a disturbance factor that follows a Gaussian distribution;

[0023] S24. In each round of iteration, a reverse learning strategy is introduced to calculate the reverse individual position:

[0024] x′ i =lb+ub-x i ;

[0025] Among them, lb and ub are the lower and upper bounds of the search space respectively, and x′ i is the reverse position of the i-th individual, which is used to improve the global search capability;

[0026] S25, after iteration until convergence or meeting the termination condition, output the optimal target area boundary point set B i ={b1,b2,…,b k}, where b k Representing an image The kth target organizational structure boundary point located in .

[0027] Optionally, the S3 specifically includes:

[0028] S31, the optimal target area boundary point set B i ={b1,b2,…,b k} with image Combined as input, construct the image segmentation input tensor T i ;

[0029] S32, dividing the image into input tensors T i Input to a deep image segmentation network with an attention mechanism. The segmentation network adopts a variant structure based on the U-Net architecture, including a downsampling encoder module, skip connections, an upsampling decoder module and an attention gating unit;

[0030] S33, introduce the boundary guidance module into the segmentation network, and perform the boundary point set B in the target area. i Perform weighted superposition processing on the feature maps to form the boundary attention weight map A i , used to guide the network to focus on the target boundary area;

[0031] S34, perform the forward propagation process, based on the input tensor T i and boundary attention map A i , generate the segmentation output map S i , the S i For images A pixel-level classification map of the same size, where each pixel value represents the probability of belonging to the target tissue structure;

[0032] S35, segmentation output image S i Perform threshold processing to obtain the final binary segmentation mask image M i ,in:

[0033]

[0034] Among them, τ is the set segmentation probability threshold, (x, y) represents the image pixel coordinates;

[0035] S36, the segmentation mask image M i With image Perform registration and superposition to form a target tissue structure image with clear structure.

[0036] Optionally, the S4 specifically includes:

[0037] S41, segmentation mask image M i Applied to the corresponding image Extract the target tissue structure area image R i , the R i is the area in the image where the pixel value is 1;

[0038] S42, in area R i Calculate the morphological characteristics, including the area A i , perimeter P i , boundary compactness C i , where the boundary tightness is defined as:

[0039]

[0040] S43, based on region R i Grayscale distribution calculation grayscale statistical characteristics, including mean μ i ,variance Skewness γ i and kurtosis κ i , used to describe the statistical characteristics of the pixel grayscale values in the region;

[0041] S44, in area R i Construct gray-level co-occurrence matrix G i , calculate texture features, including energy E i , contrast D i , entropy value H i and correlation

[0042] S45, for area R i Perform grayscale histogram normalization operation to obtain the normalized grayscale frequency distribution vector Among them, h j represents the normalized frequency of gray level j in the region, and m is the number of gray levels of the image;

[0043] S46, the area R i Input into the convolutional neural network extraction module with the same structure as the image segmentation network to obtain the deep feature vector Where d is the feature dimension of the output of a specified layer in the network;

[0044] S47, combine morphological features, grayscale statistical features, texture features, histogram features and depth features into a complete set of high-dimensional medical image feature vectors V i .

[0045] Optionally, the S5 specifically includes:

[0046] S51, high-dimensional medical image feature vector V i Constructed as feature dataset V={V1,V2,…,V n}, where V i represents the feature vector of the i-th sample, n is the number of samples;

[0047] S52, set the feature selection objective function J(V ′ ), taking the classification accuracy and feature subset dimension compression rate as the joint optimization goal, Perform optimization;

[0048] S53, initializing the population individuals of the gray wolf optimization algorithm, each individual represents a binary feature selection mask vector, which is used to control whether the corresponding feature is retained;

[0049] S54, based on the search strategy of the gray wolf optimization algorithm, comprehensively referencing the feature selection performance of the current optimal individual and other candidate individuals, dynamically updating the position of each individual, and iteratively executing the feature subset search process;

[0050] S55. When the optimization process meets the preset termination condition, the optimal feature subset V is output. * , and use this feature subset as the input feature of the classification model.

[0051] Optionally, the feature selection objective function set in step S52 is a multi-objective optimization function with an adaptive weighting factor in the form of:

[0052] J(V ′ )=α(t)·Acc(V ′ )-β(t)·R(V ′ );

[0053] Among them, Acc(V ′ ) indicates the use of feature subset V ′ The classification accuracy of the constructed diagnostic model on the validation set, represents the feature compression rate, α(t) and β(t) are adaptive weight factors that change dynamically with the number of iterations t, satisfying:

[0054]

[0055] Wherein, T is the maximum number of iterations, and the mechanism focuses on the compression rate.

[0056] Optionally, the S6 specifically includes:

[0057] S61, the optimal feature subset V * As input features, construct training sets and validation sets;

[0058] S62. Constructing a medical image-assisted diagnosis model based on a convolutional neural network, wherein the model includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and includes an adjustable hyperparameter set;

[0059] S63, using the Gray Wolf Optimization Algorithm to perform a global search and optimization on the hyperparameters of the auxiliary diagnosis model, wherein the hyperparameters include but are not limited to a learning rate, a dropout rate, a batch size, and a maximum number of training rounds;

[0060] S64. During the optimization process, the performance of different hyperparameter combinations on the validation set is comprehensively evaluated through iterative updates of multiple individuals in the gray wolf population to determine the optimal hyperparameter combination.

[0061] S65. Based on the optimal hyperparameter combination, the auxiliary diagnosis model is trained on the training set, and its classification accuracy and stability are evaluated on the validation set to obtain a trained optimized diagnosis model;

[0062] S66. Use the trained optimized diagnosis model as the final classification judgment model, and output the classification results and risk level information.

[0063] Optionally, the S7 specifically includes:

[0064] S71, normalizing the image to be diagnosed Input into the image segmentation network constructed in claim 4 to obtain the corresponding segmentation mask image M new ;

[0065] S72, using the segmentation mask image M new Extract the target tissue structure region R in the image new , perform feature extraction and generate medical image feature vector V new ;

[0066] S73, the feature vector V new and feature subset V * The corresponding features are used as input and passed into the trained optimization diagnosis model;

[0067] S74: The trained optimized diagnostic model performs inference processing on the input features and outputs a classification result label and a corresponding risk level score, wherein the classification result includes category information of the target tissue structure;

[0068] S75. Visually fuse the classification results output by the model with the original image to generate an ultrasound image display result with classification labels and risk warnings.

[0069] The beneficial effects of the present invention are:

[0070] (1) This invention introduces a reverse firefly algorithm during the ultrasound image preprocessing stage. By optimizing the search for image texture, grayscale, and edge features, it effectively improves the processing capability for noisy, low-contrast images. Compared to traditional filtering and enhancement methods, this invention can adaptively adjust processing parameters to improve image clarity and edge preservation, laying a foundation for higher-quality images for subsequent segmentation and recognition.

[0071] (2) The present invention constructs a complete processing chain from preprocessing, segmentation, feature extraction, feature selection to classification judgment. It is driven by a deep image segmentation network and a gray wolf optimization algorithm, and organically integrates image structure information with high-dimensional features, avoiding the problems of step separation and low information transmission efficiency in traditional methods, and effectively improving the recognition accuracy of target tissue structure and the reliability of medical decision-making.

[0072] (3) The present invention dynamically adjusts feature selection and classification model parameters through the Gray Wolf Optimization Algorithm, improving its adaptability to different organ types, lesion areas, and image quality, and possessing strong versatility and transplantability. Furthermore, the present invention can complete parameter optimization and region identification without extensive manual intervention, reducing reliance on professional operators, and possessing a high level of intelligence and practical clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 This is an overall flow chart of an ultrasonic image processing method based on image analysis proposed by the present invention;

[0075] Figure 2 This is a flowchart of the reverse firefly algorithm positioning of the ultrasonic image processing method based on image analysis proposed by the present invention;

[0076] Figure 3 This is a flowchart of feature selection and model parameter adjustment for the gray wolf optimization of the ultrasonic image processing method based on image analysis proposed by the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0078] refer to Figure 1-3 , an ultrasonic image processing method based on image analysis, comprising the following steps:

[0079] S1. Acquire original ultrasound image data, preprocess the image, and generate preprocessed image data;

[0080] S2. Based on the pre-processed image data, using the reverse firefly algorithm to perform initial region positioning and output candidate region boundaries of the target tissue structure;

[0081] S3. Input the candidate region boundary into a deep image segmentation network based on an attention mechanism, perform image segmentation operations, and extract the precise target tissue structure region;

[0082] S4, obtaining high-dimensional medical image features from the target tissue structure region;

[0083] S5. Based on the high-dimensional medical image feature selection optimization model, the gray wolf optimization algorithm is used to filter redundant features and output the medical feature set with optimized dimension;

[0084] S6. Inputting the dimensionally optimized medical feature set into the auxiliary diagnosis model, using the Gray Wolf Optimization Algorithm to tune and train the structural parameters and hyperparameters of the auxiliary diagnosis model to generate an optimized medical image classification model;

[0085] S7. Apply the medical image classification model to the ultrasound image data to be diagnosed to achieve automatic recognition, classification judgment and risk level output of the target tissue structure.

[0086] This invention establishes a complete ultrasound image processing process based on image analysis, achieving an intelligent closed-loop process from image preprocessing to auxiliary diagnosis output. Compared with traditional image processing methods, it has stronger system integration and data processing consistency. Especially in low-quality ultrasound image scenarios, it can significantly improve image resolution and diagnostic accuracy, providing doctors with more reliable auxiliary diagnosis basis.

[0087] In this embodiment, S1 specifically includes: obtaining original ultrasound image data, suppressing speckle noise in the image using median filtering, improving the contrast and edge details of the image through image enhancement algorithm, performing grayscale normalization processing, distributing the grayscale values within a uniform range, and generating a preprocessed image data set.

[0088] In this embodiment, S2 specifically includes:

[0089] S21. Preprocessed image dataset Each image in Perform feature extraction and construct a multidimensional feature vector set F suitable for target area recognition in a noisy environment i ;

[0090] S22, set the brightness function and search space constraints of the reverse firefly algorithm, and set the multidimensional feature vector set F i In the input optimization model, the population position set X is initialized as {x1,x2,…,x m}, where x j represents the coordinates of the center point of the candidate target area corresponding to the j-th firefly in the image, and m is the population size;

[0091] S23. Execute the iterative update formula based on brightness attraction in the reverse firefly algorithm:

[0092]

[0093] in, represents the position information of the i-th firefly at the t-th iteration, represents the position information of the j-th firefly at the t-th iteration, β0 is the initial attraction, γ is the light intensity attenuation coefficient, r ij is the Euclidean distance between the i-th and j-th fireflies, and α·∈ is a disturbance factor that follows a Gaussian distribution;

[0094] S24. In each round of iteration, a reverse learning strategy is introduced to calculate the reverse individual position:

[0095] x′ i =lb+ub-x i ;

[0096] Among them, lb and ub are the lower and upper bounds of the search space respectively, and x′ i is the reverse position of the i-th individual, which is used to improve the global search capability;

[0097] S25, after iteration until convergence or meeting the termination condition, output the optimal target area boundary point set B i ={b1,b2,…,b k}, where b k Representing an image The kth target organizational structure boundary point located in .

[0098] This paper introduces a reverse firefly algorithm to intelligently locate the initial boundary of target regions in ultrasound images, addressing the unstable boundaries and noise sensitivity of traditional thresholding and region growing methods. This method leverages swarm intelligence to achieve relatively accurate ROI extraction even in noisy images, providing a more optimal initial position for the subsequent segmentation network and improving overall recognition accuracy and efficiency.

[0099] In this embodiment, S3 specifically includes:

[0100] S31, the optimal target area boundary point set B i ={b1,b2,…,b k} with image Combined as input, construct the image segmentation input tensor T i ;

[0101] S32, dividing the image into input tensors T i Input to a deep image segmentation network with an attention mechanism. The segmentation network adopts a variant structure based on the U-Net architecture, including a downsampling encoder module, skip connections, an upsampling decoder module and an attention gating unit;

[0102] S33, introduce the boundary guidance module into the segmentation network, and perform the boundary point set B in the target area. i Perform weighted superposition processing on the feature maps to form the boundary attention weight map A i , used to guide the network to focus on the target boundary area;

[0103] S34, perform the forward propagation process, based on the input tensor T i and boundary attention map A i , generate the segmentation output map S i , the S i For images A pixel-level classification map of the same size, where each pixel value represents the probability of belonging to the target tissue structure;

[0104] S35, segmentation output image S i Perform threshold processing to obtain the final binary segmentation mask image M i ,in:

[0105]

[0106] Among them, τ is the set segmentation probability threshold, (x, y) represents the image pixel coordinates;

[0107] S36, the segmentation mask image M i With image Perform registration and superposition to form a target tissue structure image with clear structure.

[0108] This paper combines an attention mechanism with a deep image segmentation network to fully utilize initial boundary information and enhance the model's sensitivity to edge regions. Compared to traditional U-Net or unguided segmentation models, this paper demonstrates stronger structural restoration capabilities in ultrasound images with blurred boundaries and unclear structures, significantly improving the segmentation accuracy and visualization of tissue regions.

[0109] In this embodiment, the S4 specifically includes:

[0110] S41, segmentation mask image M i Applied to the corresponding image Extract the target tissue structure area image R i , the R i is the area in the image where the pixel value is 1;

[0111] S42, in area R i Calculate the morphological characteristics, including the area A i , perimeter P i , boundary compactness C i , where the boundary tightness is defined as:

[0112]

[0113] S43, based on region R i Grayscale distribution calculation grayscale statistical characteristics, including mean μ i ,variance Skewness γ i and kurtosis κ i , used to describe the statistical characteristics of the pixel grayscale values in the region;

[0114] S44, in area R i Construct gray-level co-occurrence matrix G i , calculate texture features, including energy E i , contrast D i , entropy value H i and correlation

[0115] S45, for area R i Perform grayscale histogram normalization operation to obtain the normalized grayscale frequency distribution vector Among them, h j represents the normalized frequency of gray level j in the region, and m is the number of gray levels of the image;

[0116] S46, the area R i Input into the convolutional neural network extraction module with the same structure as the image segmentation network to obtain the deep feature vector Where d is the feature dimension of the output of a specified layer in the network;

[0117] S47, combine morphological features, grayscale statistical features, texture features, histogram features and depth features into a complete set of high-dimensional medical image feature vectors V i

[0118] This paper combines traditional image processing with deep learning extraction methods to establish a multidimensional medical image feature set encompassing morphological, texture, grayscale, and depth semantic information. Compared to single-texture or grayscale analysis techniques, this multi-feature fusion method can more comprehensively reflect tissue structural differences, improve the ability to characterize lesions, and provide a solid data foundation for subsequent classification.

[0119] In this embodiment, the S5 specifically includes:

[0120] S51, high-dimensional medical image feature vector V i Constructed as feature dataset V={V1,V2,…,V n}, where V i represents the feature vector of the i-th sample, n is the number of samples;

[0121] S52, set the feature selection objective function J(V ′ ), taking the classification accuracy and feature subset dimension compression rate as the joint optimization goal, Perform optimization;

[0122] S53, initializing the population individuals of the gray wolf optimization algorithm, each individual represents a binary feature selection mask vector, which is used to control whether the corresponding feature is retained;

[0123] S54, based on the search strategy of the gray wolf optimization algorithm, comprehensively referencing the feature selection performance of the current optimal individual and other candidate individuals, dynamically updating the position of each individual, and iteratively executing the feature subset search process;

[0124] S55. When the optimization process meets the preset termination condition, the optimal feature subset V is output. * , and use this feature subset as the input feature of the classification model.

[0125] This paper uses the Gray Wolf Optimization algorithm to filter and compress feature subsets from high-dimensional feature sets, addressing the issues of feature redundancy, low computational efficiency, and overfitting caused by excessive dimensionality in existing models. Compared to manual or filter-based feature selection methods, this paper can search for the globally optimal feature combination, improving the performance and generalization of diagnostic models under complex data.

[0126] In this embodiment, the feature selection objective function set in step S52 is a multi-objective optimization function with an adaptive weighting factor in the form of:

[0127] J(V ′ )=α(t)·Acc(V ′ )-β(t)·R(V ′ );

[0128] Among them, Acc(V ′ ) indicates the use of feature subset V ′ The classification accuracy of the constructed diagnostic model on the validation set, represents the feature compression rate, α(t) and β(t) are adaptive weight factors that change dynamically with the number of iterations t, satisfying:

[0129]

[0130] Wherein, T is the maximum number of iterations, and the mechanism focuses on the compression rate.

[0131] This method introduces a dynamically adaptive weighted objective function to automatically balance compression rate and classification accuracy during feature selection. Compared to fixed weighting factors or single-objective optimization methods, this method improves the intelligence of the feature compression process, enhances the algorithm's adaptability in different task scenarios, and effectively improves the stability and practicality of the final diagnostic model.

[0132] In this embodiment, S6 specifically includes:

[0133] S61, the optimal feature subset V * As input features, construct training sets and validation sets;

[0134] S62. Constructing a medical image-assisted diagnosis model based on a convolutional neural network, wherein the model includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and includes an adjustable hyperparameter set;

[0135] S63, using the Gray Wolf Optimization Algorithm to perform a global search and optimization on the hyperparameters of the auxiliary diagnosis model, wherein the hyperparameters include but are not limited to a learning rate, a dropout rate, a batch size, and a maximum number of training rounds;

[0136] S64. During the optimization process, the performance of different hyperparameter combinations on the validation set is comprehensively evaluated through iterative updates of multiple individuals in the gray wolf population to determine the optimal hyperparameter combination.

[0137] S65. Based on the optimal hyperparameter combination, the auxiliary diagnosis model is trained on the training set, and its classification accuracy and stability are evaluated on the validation set to obtain a trained optimized diagnosis model;

[0138] S66. Use the trained optimized diagnosis model as the final classification judgment model, and output the classification results and risk level information.

[0139] This method uses the Gray Wolf Optimization algorithm to automatically tune deep model hyperparameters during diagnostic model construction, avoiding the inefficient tuning methods of traditional methods that rely on experience or grid search. This method significantly improves the training efficiency and ultimate accuracy of classification models for medical imaging tasks. It offers advantages such as high automation, a stable training process, and highly consistent results, and is applicable to different disease types and ultrasound imaging modalities.

[0140] In this embodiment, the S7 specifically includes:

[0141] S71, normalizing the image to be diagnosed Input into the image segmentation network constructed in claim 4 to obtain the corresponding segmentation mask image M new ;

[0142] S72, using the segmentation mask image M new Extract the target tissue structure region R in the image new , perform feature extraction and generate medical image feature vector V new ;

[0143] S73, the feature vector V new In the feature subset V * The corresponding features are used as input and passed into the trained optimization diagnosis model;

[0144] S74: The trained optimized diagnostic model performs inference processing on the input features and outputs a classification result label and a corresponding risk level score, wherein the classification result includes category information of the target tissue structure;

[0145] S75. Visually fuse the classification results output by the model with the original image to generate an ultrasound image display result with classification labels and risk warnings.

[0146] This invention achieves a visual presentation of structured diagnostic results by fusing model outputs with the original image during the inference phase. Compared to traditional models that only output labels, this invention improves the interpretability and clinical readability of the results, facilitating quick understanding and judgment by doctors, and enhancing the practical application value and acceptability of auxiliary diagnosis systems in clinical practice.

[0147] Example 1:

[0148] To verify the feasibility of the present invention, it was applied to the ultrasound imaging-assisted diagnosis of breast lumps at the imaging center of a tertiary hospital. The hospital has long conducted breast disease screening, accumulating a large number of ultrasound images annually. The hospital routinely uses manual image reading with the assistance of semi-automatic software to determine benign or malignant tumors. However, due to the high noise, low contrast, and blurred tumor boundaries of breast ultrasound images, traditional methods suffer from low recognition accuracy, high misdiagnosis rates, and strong reliance on the physician's subjective judgment. Manual interpretation is particularly prone to errors when the boundaries are blurred and the grayscale of the lesion and background tissue are close.

[0149] In this scenario, the method of the present invention is deployed as a complete image processing and intelligent diagnosis system. The system first preprocesses the original breast ultrasound image uploaded by the doctor, including speckle noise removal, image enhancement and normalization, to make the image structure clearer. Then, the inverse firefly algorithm is used to initially locate the potential lesion area, and high-precision lesion segmentation is achieved through a deep segmentation network integrated with an attention mechanism. Subsequently, morphological features, texture features, statistical grayscale features and semantic features of the intermediate layer of the deep neural network are extracted from the segmented area to construct a high-dimensional feature set. The gray wolf optimization algorithm is used to screen the features, eliminate redundant information, retain the feature dimensions most relevant to the classification task, and automatically optimize the model's hyperparameters to train the diagnostic model. The model ultimately outputs a benign or malignant judgment and risk grading of breast nodules, and visually integrates the diagnostic results with the image for display.

[0150] During the project, the hospital selected 1,280 breast ultrasound images collected between August and October 2023 as the evaluation dataset. Of these, 720 were benign nodules with a confirmed pathological diagnosis, and 560 were malignant lesions. The images had a resolution of 1024×768, and were graded as high, medium, and low quality. All samples were approved for research use by the hospital's ethics committee, and patient identification information was removed.

[0151] The results of the method presented in this paper were compared with those of traditional diagnostic methods and manual interpretation by doctors. The traditional method involved manual threshold segmentation combined with a support vector machine classifier. The results of the doctor's interpretation were independently interpreted by three chief physicians, with the majority opinion serving as the reference standard. Test results showed that, under the same image conditions, the method presented in this paper outperformed the comparative methods in terms of segmentation accuracy, diagnostic accuracy, diagnostic speed, and error rate control.

[0152] Table 1: Performance comparison of the present invention and existing methods in breast ultrasound image diagnosis

[0153]

[0154] According to the data results in the above "Table 1: Performance comparison table of the present invention and existing methods in breast ultrasound image diagnosis", it can be fully seen that the present invention has obvious advantages over traditional methods and manual interpretation in multiple key indicators.

[0155] First, in the image preprocessing stage, the method of the present invention takes an average of only 0.82 seconds per image, and the processing efficiency is higher than the 0.59 seconds of the traditional algorithm. Although it is slightly slower than the traditional method, the processing of the present invention includes more complex adaptive enhancement and noise optimization steps, and this step is a one-time offline operation, which has little effect on the overall response speed. In terms of initial target area positioning, the present invention adopts the reverse firefly algorithm, and its IOU index reaches 87.4%, which is nearly 19% higher than the traditional algorithm, indicating that the recognition of fuzzy boundary areas is more accurate. In terms of image segmentation accuracy, the Dice coefficient is 91.2%, which is significantly better than the doctor's manual interpretation (85.6%) and the traditional algorithm (74.3%), indicating that the deep segmentation network of the present invention combined with the initial boundary guidance strategy has higher stability and accuracy in region extraction. In terms of final diagnostic performance, the average diagnostic accuracy of the present invention reached 94.5%, the benign recognition rate was 95.3%, and the malignant recognition rate was 93.1%, which were significantly better than other comparison methods, showing good practicality and classification and differentiation capabilities.

[0156] The misdiagnosis rate for this method is 3.5%, significantly lower than the 7.1% for doctors and 11.4% for traditional algorithms. This demonstrates that the intelligent diagnostic model significantly reduces the risk of misdiagnosis while improving accuracy. Furthermore, the average image diagnosis time for this method is only 1.97 seconds, significantly lower than the 18.5 seconds for doctors and better than the 4.2 seconds for traditional algorithms, demonstrating significant efficiency advantages in large-scale image processing scenarios.

[0157] In terms of feature dimensionality compression, this paper utilizes the Gray Wolf Optimization algorithm to achieve a 72.3% feature dimensionality reduction rate, which not only reduces the computational burden but also enhances model generalization performance. Traditional methods achieve a compression rate of only 38.9%, and still contain a large amount of redundant information that affects training efficiency. This paper automatically completes 120 hyperparameter adjustments during the model construction phase, resulting in a high-performance and stable structure. Traditional methods, however, lack a parameter adjustment mechanism and require manual trial and error, resulting in low efficiency.

[0158] Finally, in the subjective evaluation of doctor satisfaction, the present invention scored 9.3 points, which was significantly higher than the 6.7 points of the traditional algorithm, proving that it has also been highly recognized by clinicians in terms of visual result presentation, rationality of diagnostic recommendations and interface interaction experience.

[0159] In summary, the present invention has achieved breakthrough optimization in multiple dimensions, including image quality processing, structure extraction, classification accuracy, response speed, and model intelligence. It is an efficient, reliable, and highly automated intelligent processing method for clinical breast ultrasound image-assisted diagnosis.

[0160] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An ultrasonic image processing method based on image analysis, characterized in that: The steps include: S1. Acquire original ultrasound image data, preprocess the image, and generate preprocessed image data; S2. Based on the pre-processed image data, using the reverse firefly algorithm to perform initial region positioning and output candidate region boundaries of the target tissue structure; S3. Input the candidate region boundary into a deep image segmentation network based on an attention mechanism, perform image segmentation operations, and extract the precise target tissue structure region; S4, obtaining high-dimensional medical image features from the target tissue structure region; S5. Based on the high-dimensional medical image feature selection optimization model, the gray wolf optimization algorithm is used to filter redundant features and output the medical feature set with optimized dimension; S6. Inputting the dimensionally optimized medical feature set into the auxiliary diagnosis model, using the Gray Wolf Optimization Algorithm to tune and train the structural parameters and hyperparameters of the auxiliary diagnosis model to generate an optimized medical image classification model; S7. Apply the medical image classification model to the ultrasound image data to be diagnosed to achieve automatic recognition, classification judgment and risk level output of the target tissue structure.

2. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: Said S1 specifically includes: acquiring original ultrasound image data, suppressing speckle noise in the image by using median filtering, improving the contrast and edge details of the image by using an image enhancement algorithm, performing grayscale normalization processing, distributing the grayscale values within a uniform range, and generating a preprocessed image data set.

3. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: The S2 specifically includes: S21. Preprocessed image dataset Each image in Perform feature extraction and construct a multidimensional feature vector set F suitable for target area recognition in a noisy environment i ; S22, set the brightness function and search space constraints of the reverse firefly algorithm, and set the multidimensional feature vector set F i In the input optimization model, the population position set X is initialized as {x1,x2,…,x m }, where x j represents the coordinates of the center point of the candidate target area corresponding to the j-th firefly in the image, and m is the population size; S23. Execute the iterative update formula based on brightness attraction in the reverse firefly algorithm: in, represents the position information of the i-th firefly at the t-th iteration, represents the position information of the j-th firefly at the t-th iteration, β0 is the initial attraction, γ is the light intensity attenuation coefficient, r ij is the Euclidean distance between the i-th and j-th fireflies, and α·∈ is a disturbance factor that follows a Gaussian distribution; S24. In each round of iteration, a reverse learning strategy is introduced to calculate the reverse individual position: x′ i =lb+ub-x i ; Among them, lb and ub are the lower and upper bounds of the search space respectively, and x′ i is the reverse position of the i-th individual, which is used to improve the global search capability; S25, after iteration until convergence or meeting the termination condition, output the optimal target area boundary point set B i ={b1,b2,…,b k }, where b k Representing an image The kth target organizational structure boundary point located in .

4. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: The S3 specifically includes: S31, the optimal target area boundary point set B i ={b1,b2,…,b k } with image Combined as input, construct the image segmentation input tensor T i ; S32, dividing the image into input tensors T i Input to a deep image segmentation network with an attention mechanism. The segmentation network adopts a variant structure based on the U-Net architecture, including a downsampling encoder module, skip connections, an upsampling decoder module and an attention gating unit; S33, introduce the boundary guidance module into the segmentation network, and perform the boundary point set B in the target area. i Perform weighted superposition processing on the feature maps to form the boundary attention weight map A i , used to guide the network to focus on the target boundary area; S34, perform the forward propagation process, based on the input tensor T i and boundary attention map A i , generate the segmentation output map S i , the S i For images A pixel-level classification map of the same size, where each pixel value represents the probability of belonging to the target tissue structure; S35, segmentation output image S i Perform threshold processing to obtain the final binary segmentation mask image M i ,in: Among them, τ is the set segmentation probability threshold, (x, y) represents the image pixel coordinates; S36, the segmentation mask image M i With image Perform registration and superposition to form a target tissue structure image with clear structure.

5. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: The S4 specifically includes: S41, segmentation mask image M i Applied to the corresponding image Extract the target tissue structure area image R i , the R i is the area in the image where the pixel value is 1; S42, in area R i Calculate the morphological characteristics, including the area A i , perimeter P i , boundary compactness C i , where the boundary tightness is defined as: S43, based on region R i Grayscale distribution calculation grayscale statistical characteristics, including mean μ i ,variance Skewness γ i and kurtosis κ i , used to describe the statistical characteristics of the pixel grayscale values in the region; S44, in area R i Construct gray-level co-occurrence matrix G i , calculate texture features, including energy E i , contrast D i , entropy value H i and correlation S45, for area R i Perform grayscale histogram normalization operation to obtain the normalized grayscale frequency distribution vector Among them, h j represents the normalized frequency of gray level j in the region, and m is the number of gray levels of the image; S46, the area R i Input into the convolutional neural network extraction module with the same structure as the image segmentation network to obtain the deep feature vector Where d is the feature dimension of the output of a specified layer in the network; S47, combine morphological features, grayscale statistical features, texture features, histogram features and depth features into a complete set of high-dimensional medical image feature vectors V i .

6. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: The S5 specifically includes: S51, high-dimensional medical image feature vector V i Constructed as feature dataset V={V1,V2,…,V n }, where V i represents the feature vector of the i-th sample, n is the number of samples; S52, set the feature selection objective function J(V ′ ), taking the classification accuracy and feature subset dimension compression rate as the joint optimization goal, Perform optimization; S53, initializing the population individuals of the gray wolf optimization algorithm, each individual represents a binary feature selection mask vector, which is used to control whether the corresponding feature is retained; S54, based on the search strategy of the gray wolf optimization algorithm, comprehensively referencing the feature selection performance of the current optimal individual and other candidate individuals, dynamically updating the position of each individual, and iteratively executing the feature subset search process; S55. When the optimization process meets the preset termination condition, the optimal feature subset V is output. * , and use this feature subset as the input feature of the classification model.

7. The ultrasonic image processing method based on image analysis according to claim 6, characterized in that: The feature selection objective function set in step S52 is a multi-objective optimization function with an adaptive weighting factor in the form of: J(V ′ )Nα(t)·Acc(V ′ )-β(t)·R(V). ′ )4 Among them, Acc(V ′ ) indicates the use of feature subset V ′ The classification accuracy of the constructed diagnostic model on the validation set, represents the feature compression rate, α(t) and β(t) are adaptive weight factors that change dynamically with the number of iterations t, satisfying: Wherein, T is the maximum number of iterations, and the mechanism focuses on the compression rate.

8. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: The S6 specifically includes: S61, the optimal feature subset V * As input features, construct training sets and validation sets; S62. Constructing a medical image-assisted diagnosis model based on a convolutional neural network, wherein the model includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and includes an adjustable hyperparameter set; S63, using the Gray Wolf Optimization Algorithm to perform a global search and optimization on the hyperparameters of the auxiliary diagnosis model, wherein the hyperparameters include but are not limited to a learning rate, a dropout rate, a batch size, and a maximum number of training rounds; S64. During the optimization process, the performance of different hyperparameter combinations on the validation set is comprehensively evaluated through iterative updates of multiple individuals in the gray wolf population to determine the optimal hyperparameter combination. S65. Based on the optimal hyperparameter combination, the auxiliary diagnosis model is trained on the training set, and its classification accuracy and stability are evaluated on the validation set to obtain a trained optimized diagnosis model; S66. Use the trained optimized diagnosis model as the final classification judgment model, and output the classification results and risk level information.

9. The ultrasonic image processing method based on image analysis according to claim 1, characterized in that: The S7 specifically includes: S71, normalizing the image to be diagnosed Input into the image segmentation network constructed in claim 4 to obtain the corresponding segmentation mask image M new ; S72, using the segmentation mask image M new Extract the target tissue structure region R in the image new , perform feature extraction and generate medical image feature vector V new ; S73, the feature vector V new and feature subset V * The corresponding features are used as input and passed into the trained optimization diagnosis model; S74: The trained optimized diagnostic model performs inference processing on the input features and outputs a classification result label and a corresponding risk level score, wherein the classification result includes category information of the target tissue structure; S75. Visually fuse the classification results output by the model with the original image to generate an ultrasound image display result with classification labels and risk warnings.

Images