A breast ABUS image gain grading and correction method

Abstract

This invention discloses a method for gain grading and correction of breast ABUS images, belonging to the field of breast ABUS image technology. The method includes: S1: data acquisition and physician evaluation; S2: data preprocessing; S3: constructing a gain classification model; S4: feature enhancement processing; S5: model training and parameter adjustment; S6: model inference and performance evaluation; S7: generating a reference image with appropriate gain. This invention solves the problem of significant differences in breast ultrasound image quality caused by variations in physician scanning operations, affecting examination efficiency and accuracy. This invention not only automatically and in real-time evaluates image gain quality during the examination but also corrects images with abnormal gain to qualified gain reference images through gamma correction, assisting operators in quickly adjusting equipment gain parameters. This effectively improves image quality consistency and examination efficiency, providing more comprehensive technical support for standardized breast screening operations.

Description

Technical Field

[0001] This invention relates to the field of breast ABUS imaging technology, specifically to a method for gain grading and correction of breast ABUS images. Background Technology

[0002] Breast cancer has become the most common malignant tumor worldwide, seriously threatening women's health. Statistics show that in 2024, there were as many as 2.3 million new cases of breast cancer globally, accounting for 11.7% of all cancer cases, ranking first among malignant tumors in women. Automated Breast Ultrasound (ABUS), as a high-resolution three-dimensional ultrasound technology, has advantages such as standardized operation, good repeatability, and full breast coverage, and has become an important auxiliary means for breast cancer screening and diagnosis.

[0003] During ABUS image acquisition, the gain parameter is preset by the operator. Gain is crucial to ABUS image quality and diagnostic accuracy, controlling the amplification of the ultrasound echo signal. Too low a gain setting results in an overall dark image, unclear tissue layers, and may mask small, hypoechoic lesions; while too high a gain setting makes the image too bright, significantly increases noise, reduces tissue contrast, and affects lesion identification. Abnormal gain images do not meet quality control standards, necessitating patient recall for re-examination, which not only increases screening costs but also reduces clinical efficiency.

[0004] Currently, gain adjustment in ABUS images relies heavily on physicians' subjective judgment, lacking a unified and objective evaluation standard. Existing ultrasound image enhancement techniques, such as filtering and denoising, contrast enhancement, and histogram equalization, fall under the category of traditional image processing methods, primarily used for post-processing of acquired images, and cannot address the issue of improper gain settings at the source of acquisition. Some studies have attempted to utilize machine learning methods for image quality assessment, but most are limited to predicting single quality indicators, exhibiting limited generalization ability, and do not address the adjustment of gain parameters. Although some research has explored handheld ultrasound image quality assessment, a grading system specifically for ABUS image gain quality remains lacking. Summary of the Invention

[0005] The purpose of this invention is to provide a gain grading and correction method for breast ultrasound images, which can realize automatic quality assessment of breast ultrasound images and provide operational suggestions for gain parameters, thereby effectively overcoming the limitations of relying on subjective experience to adjust gain parameters in clinical practice and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for gain grading and correction of breast ABUS images, comprising:

[0008] S1. Data Acquisition and Physician Evaluation: Collect three-dimensional DICOM data of breast ultrasound and evaluate it based on image quality assessment standards by ultrasound physicians;

[0009] S2. Data preprocessing: Obtain three-dimensional breast DICOM volume data, perform cross-sectional slicing, and set gain labels according to the actual gain values;

[0010] S3. Constructing a gain classification model: A multi-level feature extraction architecture is adopted. In the preprocessing network, a 7×7 convolutional kernel with a stride of 2 is used for initial feature extraction to capture the overall structure and contextual information of the image. A progressive channel expansion strategy is adopted, with the number of channels increasing from 32 to 256 through four layers to balance computational efficiency and feature representation capability. Each gain-aware residual block integrates convolutional operations and attention mechanisms to achieve synergy between feature extraction and filtering.

[0011] S4. Feature Enhancement Processing: A frequency domain enhancement module is introduced into the gain classification model. The spatial domain features are converted to the frequency domain for analysis and processing through Fourier transform, which is used to enhance the model's multi-scale feature extraction capability.

[0012] S5. Model Training and Parameter Tuning: Model training uses a loss function with integrated class weights, introduces label smoothing technology to transform hard labels into soft labels, and combines cosine annealing learning rate scheduling with gradient accumulation technology to maintain training stability under limited hardware resources. The performance of the validation set is monitored throughout the training process, and an early stopping mechanism is enabled to ensure the generalization performance of the final model.

[0013] S6. Model Inference and Performance Evaluation: The model performance is evaluated using an independent test set. The contribution of each module to the final performance is evaluated through ablation experiments. The trained model is integrated into the ultrasound workstation to provide real-time gain level identification.

[0014] S7. Generate a moderate gain reference map: The gain quality of ABUS images is automatically graded using a gain classification model. The boundary of the correction parameters is dynamically determined based on the statistical features of the images. An incremental mapping relationship is established from the standardized gain to the gamma correction value. For images with different gain levels, the corresponding γ value is automatically calculated and processed to output a moderate gain reference map with consistent visual quality.

[0015] Preferably, in step S1, the following operations are performed: Three-dimensional DICOM data of breast ultrasound is collected and evaluated by an ultrasound physician based on image quality assessment standards.

[0016] We collected three-dimensional DICOM data of breast ultrasound from different equipment models, patient body sizes, and lesion types. For each patient, in addition to routine clinical examinations, we acquired a series of images with different gain settings at 10dB intervals. Ultrasound physicians labeled each image to the corresponding five gain levels based on image quality assessment standards. We established a labeling quality control system to ensure labeling consistency and accuracy, and formed training, validation, and test datasets.

[0017] Preferably, in step S2, the three-dimensional breast DICOM volume data is acquired and processed into cross-sectional slices, and the following operations are performed:

[0018] Three-dimensional breast DICOM volume data were acquired from clinical ABUS equipment, and cross-sectional two-dimensional images consistent with the standard reading perspective were extracted. The metadata and pixel array of the DICOM file were parsed using Python, and cross-sectional slices were reconstructed and extracted along the Z-axis and uniformly cropped to a fixed size of 1100*350. To simulate the impact of different gain settings on image performance in clinical practice, data augmentation methods including random horizontal flipping, center cropping, brightness adjustment, and contrast adjustment were used to expand the dataset and improve the model's ability to discriminate gain. The preprocessed and standardized cross-sectional images were used as model input.

[0019] Preferably, in step S2, the gain label is set according to the actual gain value, and the following operations are performed:

[0020] After the gain quality label parameters are systematically collected, the actual data are labeled according to the true gain value, and the insufficient gain label is set. The gain value in the range of [10dB, 30dB] belongs to class 0.

[0021] Set the low gain label; gain values ​​in the range of [31dB, 40dB] belong to Class 1.

[0022] Set the optimal gain label; gain values ​​in the range of [42dB, 50dB] belong to category 2.

[0023] Set the high gain label; gain values ​​in the range of [55dB, 78dB] belong to category 3.

[0024] Set the excessive gain label; gain values ​​in the range of [80dB, 95dB] belong to category 4.

[0025] Preferably, in step S3, a gain classification model is constructed, and the following operations are performed:

[0026] The convolution operation of the preprocessed network follows the formula below:

[0027] (1)

[0028] in, , To output the spatial coordinates of the feature map ( =0,1,..., ; =0,1,..., ), For output channel index ( =0,1,...,15), Input feature map (i.e., single-channel medical imaging) Step size ( =2), The weights of the convolution kernel ( , For output ( , For bias ( ;

[0029] A spatial context attention pooling module is introduced at the top layer of the backbone network. By performing hierarchical pooling on the feature map, it adaptively aggregates context information of different granularities. Its three parallel links process features at different levels respectively: the global link captures the overall statistical features of the image, the medium link identifies regional anatomical structures, and the local link focuses on fine textures.

[0030] Among them, for the input feature map The multi-level pooling operation is as follows:

[0031] The pooling calculation formula for global links is as follows:

[0032] (2)

[0033] The pooling calculation formula for medium-sized links is as follows:

[0034] (3)

[0035] The pooling calculation formula for local links is as follows:

[0036] (4)

[0037] After performing multi-level pooling operations on the above three links, all are upsampled and recovered. For each level Calculate attention weights ( =1,2,4), the calculation formula is as follows:

[0038] (5)

[0039] in, Representing hierarchy Pooling operations, and for Convolution weights, where σ is the activation function. This represents the convolution operation. Indicates batch normalization;

[0040] Multi-level attention fusion employs a learnable weighting mechanism, calculated using the following formula:

[0041] = * (6)

[0042] in, 1 1 convolution weight, For attention fusion output;

[0043] The final feature weighting calculation formula is as follows:

[0044] = ⊙ (7)

[0045] in, Weighted output for features , and These represent the height and width of the feature map, respectively. For the number of channels, For the input feature map, ⊙ denotes element-wise multiplication;

[0046] Compress spatial features into channel-dimensional feature vectors ( The number of output channels is The calculation formula is as follows:

[0047] (8)

[0048] Where, ":" indicates Take all channels. It is a three-dimensional feature map (tensor), whose shape is usually represented as ( ).

[0049] Preferably, in step S4, a frequency domain enhancement module is introduced into the gain classification model. This module uses Fourier transform to convert spatial domain features to the frequency domain for analysis and processing, thereby enhancing the model's multi-scale feature extraction capability. The following operations are performed:

[0050] Spatial features of the input Each channel undergoes a two-dimensional discrete Fourier transform, transforming the spatial domain... Convert to frequency domain , =0,1,..., -1, =0,1,..., -1 represents the frequency coordinate. The channel index is calculated using the following formula:

[0051] (9)

[0052] Amplitude spectrum Phase spectrum The calculation formula is as follows:

[0053] (10)

[0054] (11)

[0055] Gaussian weighted enhancement, calculating the frequency center distance The calculation formula is as follows:

[0056] (12)

[0057] Maximum distance The calculation formula is as follows:

[0058] (13)

[0059] Based on the fact that low-frequency components in medical images contain anatomical structural information, a low-frequency-first enhancement strategy is adopted, while high-frequency components contain noise and details. σ= Control the range of enhancement. The Gaussian weighted enhancement function is calculated using the following formula, where the distance from the frequency point to the center is taken as the reference point:

[0060] (14)

[0061] The formula for calculating the enhanced amplitude spectrum is as follows:

[0062] (15)

[0063] The formula for calculating the inverse Fourier transform is as follows:

[0064] (16)

[0065] The phase spectrum preserves the positional information of image edges and structures. The enhanced amplitude spectrum is combined with the original phase spectrum, and then features are reconstructed through inverse transform. The resulting frequency domain enhanced features are then weighted and fused with the original spatial features. The weights are... and The parameters are learnable and satisfy the following conditions: + =1, used to balance the contributions of both, the fusion formula is as follows:

[0066] (17)

[0067] The gain-aware residual block in the model employs a design that includes cross-layer connections. These connections implement identity mappings to optimize gradient propagation in deep networks and support the construction of even deeper network architectures. For the first Layer input, Represents the residual function. For the first The layer output is calculated using the following formula:

[0068] (18)

[0069] During forward propagation, for The activation function, after introducing cross-layer connections, the first layer... The calculation of the +1 layer activation value becomes:

[0070] )+ (19)

[0071] Introducing cross-layer connections adds an identity term to gradient calculation. Let be the identity matrix, used to ensure that the gradient does not vanish. The gradient propagation calculation formula is as follows:

[0072] (20).

[0073] Preferably, in step S5, model training and parameter tuning involve the following operations:

[0074] Design a regularized gain-aware loss function that integrates three components: label smoothing, modulated cross-entropy, and class weights.

[0075] Among them, label smoothing technology uses hard labels Convert to soft tags Introducing uncertainty into the training objective, hard labels Convert to soft tags The calculation formula is as follows:

[0076] (twenty one)

[0077] in, For smoothing parameters, For the number of categories, Index for real categories;

[0078] The formula for calculating the reduction in weight of easily classified samples using modulated cross-entropy loss is as follows:

[0079] (twenty two)

[0080] in, This represents the model's predicted probability for the true class. For focusing parameters;

[0081] right The gradient can be obtained by taking the derivative. The formula for calculating the gradient is as follows:

[0082] (twenty three)

[0083] when When the gradient approaches 1, it is used to reduce the weight of easily classified samples.

[0084] For class imbalance, inverse frequency weighting is used, and the formula for calculating inverse frequency weighting is as follows:

[0085] (twenty four)

[0086] in, The total number of samples, K represents the number of samples in category c, and K represents the number of categories.

[0087] The regularized gain-aware loss function is designed, and the specific calculation formula is as follows:

[0088] (25)

[0089] This loss function is used to smooth labels and make gradient updates more stable. Modulated cross-entropy loss automatically focuses on hard samples and alleviates imbalance through weight adjustment.

[0090] Preferably, in step S7, a reference map with moderate gain is generated, and the following operations are performed:

[0091] Image statistical feature extraction: Let the input image be... The depth of the position The size is The pixel value range is Then the mean of the image and standard deviation The calculation formula is as follows:

[0092] = (26)

[0093] = (27)

[0094] Determining the boundary of the correction value: Let the maximum pixel value of the image be... The maximum correction value is The minimum correction value is , position deep The calculation formula is as follows:

[0095] (28)

[0096] =1.1+1.4×(1 (29)

[0097] =0.4+0.6× (30)

[0098] Establish an exponential mapping from gain to correction value: Data acquisition gain The physical gain of ultrasonic equipment Adjustment range ( Gain The corresponding correction value is γ( ), =0 corresponds to , =100 corresponds to λ is the mapping coefficient, and the formula for calculating the incremental mapping between the two is as follows:

[0099] λ= (31)

[0100] γ( )= (32)

[0101] Applying exponential mapping to obtain a correction value with appropriate gain: Appropriate gain ,Depend on The formula is as follows:

[0102] = × (33)

[0103] Based on moderate gain Anomalous image correction: For images with evaluated gain anomalies, adaptive gain correction is performed using a correction algorithm to convert them into reference images with appropriate gain. The correction principle is as follows:

[0104] ≥0,0≤ ≤1) (34)

[0105] in, To correct the parameters, The scaling factor (usually taken as...) =1), Given an input image, output The result after correction. This is the maximum pixel value.

[0106] Compared with the prior art, the beneficial effects of the present invention are:

[0107] This invention can not only evaluate image gain quality in real time and automatically during the examination, but also correct images with abnormal gain to qualified gain reference images through gamma correction, assisting operators to quickly adjust equipment gain parameters, thereby effectively improving the consistency of image quality and examination efficiency, and providing more comprehensive technical support for the standardized operation of breast cancer screening. Attached Figure Description

[0108] Figure 1 This is a schematic diagram of the actual ultrasound images acquired according to the present invention;

[0109] Figure 2 A schematic diagram of a cross-section illustrating five gain categories of the present invention;

[0110] Figure 3 This is a schematic diagram illustrating that the true gain of the image in this invention is Class 1 and the predicted gain is Class 1.

[0111] Figure 4 This is a schematic diagram illustrating the generation of a moderately gain image from a low-gain image according to the present invention.

[0112] Figure 5 This is a flowchart of the gain grading and correction of breast ABUS images according to the present invention. Detailed Implementation

[0113] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0114] To address the issue of significant variations in breast ultrasound image quality due to differences in physician scanning techniques, which affects examination efficiency and accuracy, please refer to [link to relevant documentation]. Figures 1-5 This embodiment provides the following technical solution:

[0115] A method for gain grading and correction of breast ABUS images, comprising:

[0116] S1: Data collection and physician evaluation;

[0117] We collected three-dimensional DICOM data of breast ultrasound from different equipment models, patient body sizes, and lesion types. For each patient, in addition to routine clinical examinations, we acquired a series of images with different gain settings at 10dB intervals. Based on image quality assessment standards, ultrasound physicians labeled each image to the corresponding five gain levels, established a labeling quality control system to ensure labeling consistency and accuracy, and formed training, validation, and test datasets.

[0118] S2: Data preprocessing;

[0119] 1) Cross-sectional slicing treatment:

[0120] Three-dimensional breast DICOM volume data were acquired from clinical ABUS equipment, and two-dimensional cross-sectional (Axial Plane) images consistent with the standard reading perspective were extracted. The metadata and pixel array of the DICOM file were parsed using Python, and cross-sectional slices were reconstructed and extracted along the Z-axis, uniformly cropped to a fixed size of 1100*350. To simulate the impact of different gain settings on image performance in clinical settings, data augmentation methods such as random horizontal flipping, center cropping, brightness adjustment, and contrast adjustment were employed to expand the dataset and improve the model's ability to discriminate gain. The preprocessed and standardized cross-sectional images were used as model input. Figure 1 As shown.

[0121] 2) Gain label setting:

[0122] After systematically collecting gain quality label parameters, the actual data are labeled according to the true gain values. The following labels are set: insufficient gain (10dB, 30dB), low gain (31dB, 40dB), optimal gain (42dB, 50dB), high gain (55dB, 78dB), and excessive gain (80dB, 95dB). Figure 2 As shown.

[0123] It should be noted that in breast ultrasound imaging, gain refers to the overall brightness of the image. The goal is to maintain the background fat lobules in a medium grayscale. ABUS images are defined according to the gain category as follows:

[0124] Category 0: Insufficient gain ultrasound images. Fat lobules will appear dark gray or black. Because the preset overall gain value before scanning is too low, it cannot fully amplify the weak echo signals returning from deep breast tissue. A large amount of true tissue information is lost, and lesions in deep areas may not be displayed at all, easily leading to missed diagnoses. Visualization is as follows: Figure 2 As shown.

[0125] Class 1: Low-gain ultrasound images. Fat lobules appear light gray. Due to the preset gain being low but not extreme, the image can display the main structures, but the contrast and layering are poor. Hypoechoic lesions may not contrast clearly with the background tissue, increasing the difficulty of diagnosis. Visualization is as follows: Figure 2 As shown.

[0126] Category 2: Ultrasound images with optimal gain. Fat lobules appear as medium grayscale. Because the gain is set to the ideal level, the true echo characteristics of different tissues are obtained. The image has rich detail, and both superficial and deep areas are well displayed, minimizing missed diagnoses and misdiagnoses. Visualization is as follows: Figure 2 As shown.

[0127] Category 3: High-gain ultrasound images. Fat lobules appear as brighter gray. Due to the excessively high preset gain, the overall image becomes brighter, and speckle noise is amplified simultaneously. The contrast between lesions and normal tissue in the image decreases, resulting in poor visualization. Figure 2 As shown.

[0128] Category 4: Excessive gain ultrasound images. Fat lobules will appear as bright gray or white. This is due to a severe error in the gain setting; the gain is adjusted to near its maximum value, causing the signal amplification system to saturate. The image is severely overexposed, with a significant loss of detail, resulting in poor visualization. Figure 2 As shown.

[0129] S3: Construct a gain classification model;

[0130] The high-precision model for gain level classification adopts a multi-level feature extraction architecture with one input channel. In the preprocessing network, a 7×7 convolutional kernel with a stride of 2 is used for initial feature extraction to expand the shallow receptive field and capture the overall structure and contextual information of the image. A progressive channel expansion strategy is adopted, with the number of channels increasing from 32 to 256 through four layers to balance computational efficiency and feature representation capability. Each gain-aware residual block integrates convolutional operations and attention mechanisms to achieve synergy between feature extraction and selection.

[0131] The convolution operation of the preprocessed network follows the formula below:

[0132] (1)

[0133] in, , To output the spatial coordinates of the feature map ( =0,1,..., ; =0,1,..., ), For output channel index ( =0,1,...,15), Input feature map ), namely single-channel medical imaging, Step size ( =2), The weights of the convolution kernel ( , For output ( , For bias ( .

[0134] To effectively identify the hierarchical characteristics of anatomical structures in ABUS images, a spatial contextual attention pooling module is introduced at the top layer of the backbone network. This module adaptively aggregates contextual information of different granularities by performing hierarchical pooling on the feature map. Its three parallel links process features at different levels respectively: the global link captures the overall statistical features of the image, the medium link identifies regional anatomical structures, and the local link focuses on fine textures.

[0135] Specifically, for the input feature map The multi-level pooling operation is as follows:

[0136] The pooling calculation formula for global links is as follows:

[0137] (2)

[0138] The pooling calculation formula for medium-sized links is as follows:

[0139] (3)

[0140] The pooling calculation formula for local links is as follows:

[0141] (4)

[0142] After performing multi-level pooling operations on the above three links, all are upsampled and recovered. For each level Calculate attention weights ( =1,2,4), the calculation formula is as follows:

[0143] (5)

[0144] in, Representing hierarchy Pooling operations, and for Convolution weights, where σ is the activation function. This represents the convolution operation. Indicates batch normalization;

[0145] Multi-level attention fusion employs a learnable weighting mechanism, calculated using the following formula:

[0146] = * (6)

[0147] in, 1 1 convolution weight, For attention fusion output;

[0148] The final feature weighting calculation formula is as follows:

[0149] = ⊙ (7)

[0150] in, Weighted output for features , and These represent the height and width of the feature map, respectively. For the number of channels, For the input feature map, ⊙ denotes element-wise multiplication;

[0151] Compress spatial features into channel-dimensional feature vectors ( The number of output channels is The calculation formula is as follows:

[0152] (8)

[0153] Where, ":" indicates Take all channels. It is a three-dimensional feature map (tensor), whose shape is usually represented as ( ).

[0154] It should be noted that the image classification network introduces basic building blocks with cross-layer connections, which effectively alleviates the gradient vanishing and degradation problems of deep networks. In each unit, the input information can be directly propagated forward through the cross-layer path, so that the network only needs to learn the adjustment between the input and output, rather than the complete complex mapping. This not only ensures the effective backpropagation of gradients in extremely deep networks, but also significantly reduces the training difficulty. As the network depth increases, the model can extract more complex and abstract feature information from images. This powerful feature learning ability makes it perform well in tasks such as image classification, and is especially suitable for application scenarios that require fine-grained image classification.

[0155] It should be noted that feature extraction refers to automatically learning and selecting the most effective feature representations from the original ultrasound image. It is a hierarchical process that progresses from shallow to deep. In the initial convolutional layers of the network, shallow feature extraction uses a set of learnable convolutional kernels to perform convolution operations on the input image. Basic visual features, such as edges, textures, brightness, and contrast, are extracted by calculating the dot product between the convolutional kernel and the local region of the image. As the network depth increases, deep feature extraction uses multiple convolutional layers and nonlinear activation functions to stack. The network can learn more complex and abstract feature representations, such as high-level semantic features like tissue structure and anatomical morphology. Each layer of feature extraction generates a new feature map. These feature maps reflect the feature response intensity of the input image at different locations, providing multi-level feature support for the final classification decision and enabling more accurate data analysis and category prediction in gain-level classification tasks.

[0156] It should be noted that cross-layer connections, by introducing shortcut connections in the forward path, effectively improve two core problems in the training process of deep networks. In terms of gradient propagation, by establishing a direct path from input to output, the training signal can be effectively propagated in the network, alleviating the phenomenon of signal weakening layer by layer. This allows the original input features to be directly passed to subsequent layers, so that the network only needs to learn the adjustment amount based on the original features, rather than the complete feature transformation, thereby significantly improving the efficiency of feature utilization.

[0157] It should be noted that global feature compression is a feature compression technique that calculates the average value of all spatial locations on each feature map, summing the two-dimensional feature grid into a scalar representing the global response level of this channel. Compared with traditional fully connected layers, this method can significantly reduce the number of parameters that the model needs to learn, thereby effectively reducing the risk of the model over-memorizing the training data. At the same time, while achieving feature dimensionality reduction, it preserves the global context information of each channel and enables the model to process images of arbitrary input sizes.

[0158] S4: Feature enhancement processing;

[0159] A frequency domain enhancement module is introduced into the gain classification model. The spatial domain features are transformed to the frequency domain for analysis and processing through Fourier transform, thereby enhancing the model's multi-scale feature extraction capability.

[0160] Specifically, the spatial features of the input Each channel undergoes a two-dimensional discrete Fourier transform, transforming the spatial domain... Convert to frequency domain , =0,1,..., -1, =0,1,..., -1 represents the frequency coordinate. The channel index is calculated using the following formula:

[0161] (9)

[0162] Amplitude spectrum Phase spectrum The calculation formula is as follows:

[0163] (10)

[0164] (11)

[0165] Gaussian weighted enhancement, calculating the frequency center distance The calculation formula is as follows:

[0166] (12)

[0167] Maximum distance The calculation formula is as follows:

[0168] (13)

[0169] Since low-frequency components in medical images primarily contain anatomical structural information, a low-frequency enhancement strategy is employed. High-frequency components, on the other hand, mainly contain noise and detail characteristics. σ = Control the range of enhancement. The Gaussian weighted enhancement function is calculated using the following formula, where the distance from the frequency point to the center is taken as the reference point:

[0170] (14)

[0171] The formula for calculating the enhanced amplitude spectrum is as follows:

[0172] (15)

[0173] The formula for calculating the inverse Fourier transform is as follows:

[0174] (16)

[0175] The phase spectrum preserves the positional information of image edges and structures. To maintain structural integrity, the enhanced amplitude spectrum is combined with the original phase spectrum, and then features are reconstructed through inverse transform. The resulting frequency domain enhanced features are then weighted and fused with the original spatial features, where the weights are... and The parameters are learnable and satisfy the following conditions: + =1, to balance the contributions of both, the fusion formula is as follows:

[0176] (17)

[0177] The gain-aware residual block in the model employs a design that includes cross-layer connections. These connections implement identity mappings, optimize gradient propagation in deep networks, effectively alleviate the vanishing gradient problem, and support the construction of deeper network architectures. For the first Layer input, Represents the residual function. For the first The layer output is calculated using the following formula:

[0178] (18)

[0179] During forward propagation, for The activation function, after introducing cross-layer connections, the first layer... The calculation of the +1 layer activation value becomes:

[0180] )+ (19)

[0181] Introducing cross-layer connections adds an identity term to gradient calculation. Assuming the identity matrix is ​​used, and to ensure the gradient does not vanish, the gradient propagation calculation formula is as follows:

[0182] (20)

[0183] It should be noted that the loss function is a function used to optimize deep learning models. It is used to quantify the degree of difference between the model's prediction results and the true labels, and to provide optimization directions during model training. The model parameters are adjusted by minimizing the loss function. Binary cross-entropy loss is commonly used for binary classification tasks, while classification cross-entropy loss is commonly used for multi-class classification tasks. Therefore, for the five-class classification task of gain images, the classification cross-entropy loss function is adopted.

[0184] It should be noted that the activation function is a nonlinear transformation component in a neural network, responsible for introducing nonlinear expressive power, converting the weighted input signal of a neuron into an output signal, and passing it to the next layer of neurons, enabling the network to learn and simulate complex nonlinear input-output mapping relationships. In the gain level classification task, the last layer of the model normalizes the calculation through the activation function, converting the original class score into the probability percentage corresponding to each class, and the final output is the probability distribution of all gain levels.

[0185] S5: Model training and parameter tuning;

[0186] To address class imbalance and overfitting issues in medical image classification, the model training employs a loss function that integrates class weights to enhance learning on minority class samples. Label smoothing techniques are introduced to transform hard labels into soft labels, improving the model's generalization ability and prediction accuracy. Cosine annealing learning rate scheduling (with an initial learning rate set to 0.001) is combined with gradient accumulation to maintain training stability under limited hardware resources. Validation set performance is monitored throughout training, and an early stopping mechanism is enabled to effectively prevent overfitting and ensure the final model's generalization performance.

[0187] The regularized gain-aware loss function designed in this invention integrates three components: label smoothing, modulated cross-entropy, and class weights, which are used to alleviate the problems of model overconfidence, learning from difficult samples, and class imbalance, respectively.

[0188] Among them, label smoothing technology uses hard labels Convert to soft tags Introducing uncertainty into the training objective, hard labels Convert to soft tags The calculation formula is as follows:

[0189] (twenty one)

[0190] in For smoothing parameters, For the number of categories, Index for real categories;

[0191] The formula for calculating the reduction in weight of easily classified samples using modulated cross-entropy loss is as follows:

[0192] (twenty two)

[0193] in This represents the model's predicted probability for the true class. For focusing parameters;

[0194] right The gradient can be obtained by taking the derivative. The formula for calculating the gradient is as follows:

[0195] (twenty three)

[0196] when When the gradient approaches 1, the gradient approaches 0, thus reducing the weight of easily classified samples.

[0197] To address the class imbalance problem, this invention employs inverse frequency weighting, the calculation formula of which is as follows:

[0198] (twenty four)

[0199] in, The total number of samples, K is the number of samples in category c, and K is the number of categories.

[0200] Finally, the regularized gain-aware loss function is designed, and the specific calculation formula is as follows:

[0201] (25)

[0202] This loss function can smooth labels and make gradient updates more stable. The modulated cross-entropy loss automatically focuses on difficult samples and alleviates the imbalance problem through weight adjustment.

[0203] S6: Model Inference and Performance Evaluation;

[0204] Model performance was evaluated using an independent test set. The performance evaluation metrics primarily included accuracy, precision, recall, F1 score, and confusion matrix. Ablation experiments were conducted to assess the contribution of each module to the final performance. The trained model was then integrated into an ultrasound workstation to provide real-time gain level identification, such as... Figure 3 As shown.

[0205] S7: Generate a reference chart with moderate gain;

[0206] The gain quality of ABUS images is automatically graded using a gain classification model. To achieve adaptive correction, the boundaries of the correction parameters are first dynamically determined based on the image statistical features. Then, an incremental mapping relationship is established from the standardized gain to the gamma correction value. For images with different gain levels, the system automatically calculates the corresponding γ value for processing, and finally outputs a reference image with moderate gain and consistent visual quality. The specific steps are as follows:

[0207] 1) Image statistical feature extraction:

[0208] Let the input image be The depth of the position The size is The pixel value range is Then the mean of the image and standard deviation The calculation formula is as follows:

[0209] = (26)

[0210] = (27)

[0211] 2) Determining the boundary of the correction value:

[0212] Let the maximum pixel value of the image be The maximum correction value is The minimum correction value is , position deep The calculation formula is as follows:

[0213] (28)

[0214] =1.1+1.4×(1 (29)

[0215] =0.4+0.6× (30)

[0216] 3) Establish an exponential mapping from gain to correction value:

[0217] Data gain The physical gain of ultrasonic equipment Adjustment range ( Gain The corresponding correction value is γ( ), =0 corresponds to , =100 corresponds to λ is the mapping coefficient, and the formula for calculating the incremental mapping between the two is as follows:

[0218] λ= (31)

[0219] γ( )= (32)

[0220] 4) Apply exponential mapping to obtain a correction value with appropriate gain:

[0221] Moderate gain ,Depend on The formula is as follows:

[0222] = × (33)

[0223] 5) Based on appropriate gain Achieve abnormal image correction:

[0224] For the image with abnormal gain identified during evaluation, an adaptive gain correction algorithm is used to convert it into a reference image with appropriate gain, such as... Figure 4 As shown. The correction principle is as follows:

[0225] ≥0,0≤ ≤1) (34)

[0226] in To correct the parameters, The scaling factor (usually taken as...) =1), Given an input image, output For the corrected result, This is the maximum pixel value.

[0227] It should be noted that the specific implementation logic for correcting to a reference image with moderate gain when an abnormal gain is detected is as follows:

[0228] When the image gain prediction is insufficient gain (class 0), the predicted gain is... [10, 30] ,Pick = The correction value is set to γ ​​by formula (32). <1, applying this γ value through formula (34) can improve the contrast of the dark areas of the image and effectively improve the visibility of details.

[0229] When the image gain prediction is low gain (Class 1), the prediction gain is... [31, 40] ,Pick = The correction value is set to γ ​​by formula (32). <1, The correction of formula (34) can brighten the image and make the overall contrast more natural and smooth.

[0230] When the image gain is predicted as optimal gain (class 2), the predicted gain is... [42, 50d] ,Pick = The correction value is set to γ ​​by formula (32). The correction transformation of formula (34) is approximately the identity transformation, which preserves the original image features to the greatest extent.

[0231] When the image gain prediction is high gain (class 3), the prediction gain is... [55, 78] ,Pick = The correction value is set to γ ​​by formula (32). >1. The correction of formula (34) prevents image overexposure and loss of detail, and adjusts the overall brightness distribution to a moderate range.

[0232] When the graph gain prediction is an excessive gain (class 4), the predicted gain is... [80, 95] ,Pick = The correction value is set to γ ​​by formula (32). >1. The correction of formula (34) can effectively suppress the overall overexposure phenomenon and moderately enhance the details in the shadows.

[0233] It should be noted that recognition tasks are one of the core applications in the field of artificial intelligence. They involve the automatic classification and categorization of input data. Using deep learning algorithms and techniques, input data samples are automatically identified and classified. Based on the number of categories, they can be divided into binary classification tasks (such as two classes) and multi-class classification tasks (such as five classes). The goal is to train an accurate and robust artificial intelligence model that can reliably classify input data into a predefined set of categories. Gain recognition is a task that automatically analyzes acquired ultrasound image data and determines its gain quality level.

[0234] It should be noted that gain correction is a nonlinear image processing method based on power-law transform. Its core principle is to optimize image quality by adjusting the mapping relationship between input and output signals. Specifically, gain classification and correction tasks typically involve the following aspects:

[0235] 1) Quality level definition: In gain identification, it is necessary to clearly define the level standard of gain quality. According to clinical needs, gain quality can be divided into five categories: too low gain, low gain, moderate gain, high gain and too high gain.

[0236] 2) Data input: The input for this task is the collected ABUS ultrasound image data. These data need to be standardized and preprocessed before being input into the model, including grayscale normalization, image size unification and data augmentation, to ensure the consistency and reliability of model processing.

[0237] 3) Model Training: The core of this task is to train a deep learning model that can accurately map image data to the corresponding gain quality level. Supervised learning methods are adopted, and the model is trained using a gain quality classification dataset annotated by experts. Through backpropagation and optimization algorithms, the model automatically learns deep features related to gain quality and continuously adjusts the network parameters to minimize the difference between the prediction results and the true labels.

[0238] 4) Model Evaluation: After the model training is completed, its performance needs to be comprehensively evaluated on an independent test set. The evaluation metrics used include confusion matrix, accuracy, precision, recall and F1 score. These metrics can objectively reflect the classification accuracy and stability of the model in the gain quality recognition task.

[0239] 5) Applying gain correction: This refers to converting images with abnormal gain into reference images with appropriate gain through correction algorithms. Specifically, the gain classification model obtains the ABUS image gain quality classification prediction results, and the four types of abnormal gain, namely, too low, low, high and too high, are processed by gain correction to obtain a reference image with appropriate gain.

[0240] 6) Model Application: The model can be integrated into the ABUS scanning workstation to achieve real-time grading and gain correction of ABUS image gain quality. The correction algorithm generates a reference image with appropriate gain, providing gain adjustment suggestions and references for the operating physician. This effectively solves image quality problems caused by differences in operator experience and ensures the consistency of image quality and diagnostic reliability in breast ultrasound screening.

[0241] In summary, this invention enables automatic grading of ABUS image gain quality. Abnormal gain conditions include excessively low gain, low gain, high gain, and excessively high gain, while acceptable gain is considered moderate. The image classification model, through its unique cross-layer connection structure, can effectively learn visual features related to image quality, including key indicators such as contrast, noise performance, and brightness distribution. When the model identifies abnormal gain conditions, it automatically adjusts the image to a moderately gain reference image using gamma correction technology. This not only allows for real-time and automatic evaluation of image gain quality during the examination process but also enables the conversion of abnormal gain images into acceptable gain reference images through gamma correction. This assists operators in quickly adjusting equipment gain parameters, thereby effectively improving image quality consistency and examination efficiency, and providing more comprehensive technical support for the standardized operation of breast cancer screening.

[0242] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0243] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for gain grading and correction of breast ABUS images, characterized in that, include: S1. Data Acquisition and Physician Evaluation: Collect three-dimensional DICOM data of breast ultrasound and evaluate it based on image quality assessment standards by ultrasound physicians; S2. Data preprocessing: Obtain three-dimensional breast DICOM volume data, perform cross-sectional slicing, and set gain labels according to the actual gain values; S3. Constructing a gain classification model: A multi-level feature extraction architecture is adopted. In the preprocessing network, a 7×7 convolutional kernel with a stride of 2 is used for initial feature extraction to capture the overall structure and contextual information of the image. A progressive channel expansion strategy is adopted, with the number of channels increasing from 32 to 256 through four layers to balance computational efficiency and feature representation capability. Each gain-aware residual block integrates convolutional operations and attention mechanisms to achieve synergy between feature extraction and filtering. S4. Feature Enhancement Processing: A frequency domain enhancement module is introduced into the gain classification model. The spatial domain features are converted to the frequency domain for analysis and processing through Fourier transform, which is used to enhance the model's multi-scale feature extraction capability. S5. Model Training and Parameter Tuning: Model training uses a loss function with integrated class weights, introduces label smoothing technology to transform hard labels into soft labels, and combines cosine annealing learning rate scheduling with gradient accumulation technology to maintain training stability under limited hardware resources. The performance of the validation set is monitored throughout the training process, and an early stopping mechanism is enabled to ensure the generalization performance of the final model. S6. Model Inference and Performance Evaluation: The model performance is evaluated using an independent test set. The contribution of each module to the final performance is evaluated through ablation experiments. The trained model is integrated into the ultrasound workstation to provide real-time gain level identification. S7. Generate a moderate gain reference map: The gain quality of ABUS images is automatically graded using a gain classification model. The boundary of the correction parameters is dynamically determined based on the statistical features of the images. An incremental mapping relationship is established from the standardized gain to the gamma correction value. For images with different gain levels, the corresponding γ value is automatically calculated and processed to output a moderate gain reference map with consistent visual quality.

2. The method for gain grading and correction of breast ABUS images according to claim 1, characterized in that, In step S1, three-dimensional DICOM data of breast ultrasound are collected and evaluated by an ultrasound physician based on image quality assessment standards. The following operations are performed: We collected three-dimensional DICOM data of breast ultrasound from different equipment models, patient body sizes, and lesion types. For each patient, in addition to routine clinical examinations, we acquired a series of images with different gain settings at 10dB intervals. Ultrasound physicians labeled each image to the corresponding five gain levels based on image quality assessment standards. We established a labeling quality control system to ensure labeling consistency and accuracy, and formed training, validation, and test datasets.

3. The method for gain grading and correction of breast ABUS images according to claim 2, characterized in that, In step S2, the three-dimensional breast DICOM volume data is acquired and processed into cross-sectional slices. The following operations are performed: Three-dimensional breast DICOM volume data were acquired from clinical ABUS equipment, and cross-sectional two-dimensional images consistent with the standard reading perspective were extracted. The metadata and pixel array of the DICOM file were parsed using Python, and cross-sectional slices were reconstructed and extracted along the Z-axis and uniformly cropped to a fixed size of 1100*350. To simulate the impact of different gain settings on image performance in clinical practice, data augmentation methods including random horizontal flipping, center cropping, brightness adjustment, and contrast adjustment were used to expand the dataset and improve the model's ability to discriminate gain. The preprocessed and standardized cross-sectional images were used as model input.

4. The method for gain grading and correction of breast ABUS images according to claim 3, characterized in that, In step S2, the gain label is set according to the actual gain value, and the following operations are performed: After the gain quality label parameters are systematically collected, the actual data are labeled according to the true gain value, and the insufficient gain label is set. The gain value in the range of [10dB, 30dB] belongs to class 0. Set the low gain label; gain values ​​in the range of [31dB, 40dB] belong to Class 1. Set the optimal gain label; gain values ​​in the range of [42dB, 50dB] belong to category 2. Set the high gain label; gain values ​​in the range of [55dB, 78dB] belong to category 3. Set the excessive gain label; gain values ​​in the range of [80dB, 95dB] belong to category 4.

5. The method for gain grading and correction of breast ABUS images according to claim 4, characterized in that, In step S3, a gain classification model is constructed, and the following operations are performed: The convolution operation of the preprocessed network follows the formula below: （1）； in, , To output the spatial coordinates of the feature map ( =0,1,..., ; =0,1,..., ), For output channel index ( =0,1,...,15), Input feature map (i.e., single-channel medical imaging) Step size ( =2), The weights of the convolution kernel ( , For output ( , For bias ( ; A spatial context attention pooling module is introduced at the top layer of the backbone network. By performing hierarchical pooling on the feature map, it adaptively aggregates context information of different granularities. Its three parallel links process features at different levels respectively: the global link captures the overall statistical features of the image, the medium link identifies regional anatomical structures, and the local link focuses on fine textures. Among them, for the input feature map The multi-level pooling operation is as follows: The pooling calculation formula for global links is as follows: （2）； The pooling calculation formula for medium-sized links is as follows: （3）； The pooling calculation formula for local links is as follows: （4）； After performing multi-level pooling operations on the above three links, all are upsampled and recovered. For each level Calculate attention weights ( =1,2,4), the calculation formula is as follows: （5）； in, Representing hierarchy Pooling operations, and for Convolution weights, where σ is the activation function. This represents the convolution operation. Indicates batch normalization; Multi-level attention fusion employs a learnable weighting mechanism, calculated using the following formula: = * ) （6）； in, 1 1 convolution weight, For attention fusion output; The final feature weighting calculation formula is as follows: = ⊙ （7）； in, Weighted output for features , and These represent the height and width of the feature map, respectively. For the number of channels, For the input feature map, ⊙ denotes element-wise multiplication; Compress spatial features into channel-dimensional feature vectors ( The number of output channels is The calculation formula is as follows: （8）； Where, ":" indicates Take all channels. It is a three-dimensional feature map (tensor), whose shape is usually represented as ( ).

6. The method for gain grading and correction of breast ABUS images according to claim 5, characterized in that, In step S4, a frequency domain enhancement module is introduced into the gain classification model. This module uses Fourier transform to convert spatial domain features to the frequency domain for analysis and processing, thereby enhancing the model's multi-scale feature extraction capability. The following operations are performed: Spatial features of the input Each channel undergoes a two-dimensional discrete Fourier transform, transforming the spatial domain... Convert to frequency domain , =0,1,..., -1, =0,1,..., -1 represents the frequency coordinate. The channel index is calculated using the following formula: （9）； Amplitude spectrum Phase spectrum The calculation formula is as follows: （10）； （11）； Gaussian weighted enhancement, calculating the frequency center distance The calculation formula is as follows: （12）； Maximum distance The calculation formula is as follows: （13）； Based on the fact that low-frequency components in medical images contain anatomical structural information, a low-frequency-first enhancement strategy is adopted, while high-frequency components contain both noise and detail. σ= Control the range of enhancement. The Gaussian weighted enhancement function is calculated using the following formula, where the distance from the frequency point to the center is taken as the reference point: （14）； The formula for calculating the enhanced amplitude spectrum is as follows: （15）； The formula for calculating the inverse Fourier transform is as follows: （16）； The phase spectrum preserves the positional information of image edges and structures. The enhanced amplitude spectrum is combined with the original phase spectrum, and then features are reconstructed through inverse transform. The resulting frequency domain enhanced features are then weighted and fused with the original spatial features. The weights are... and The parameters are learnable and satisfy the following conditions: + =1, used to balance the contributions of both, the fusion formula is as follows: （17）； The gain-aware residual block in the model employs a design that includes cross-layer connections. These connections implement identity mappings to optimize gradient propagation in deep networks and support the construction of even deeper network architectures. For the first Layer input, Represents the residual function. For the first The layer output is calculated using the following formula: （18）； During forward propagation, for The activation function, after introducing cross-layer connections, the first layer... The calculation of the +1 layer activation value becomes: )+ （19）； Introducing cross-layer connections adds an identity term to gradient calculation. Let be the identity matrix, used to ensure that the gradient does not vanish. The gradient propagation calculation formula is as follows: （20）。 7. The method for gain grading and correction of breast ABUS images according to claim 6, characterized in that, In S5, model training and parameter tuning involve the following operations: Design a regularized gain-aware loss function that integrates three components: label smoothing, modulated cross-entropy, and class weights. Among them, label smoothing technology uses hard labels Convert to soft tags Introducing uncertainty into the training objective, hard labels Convert to soft tags The calculation formula is as follows: （21）； in, For smoothing parameters, For the number of categories, Index for real categories; The formula for calculating the reduction in weight of easily classified samples using modulated cross-entropy loss is as follows: （22）； in, This represents the model's predicted probability of the true class. For focusing parameters; right The gradient can be obtained by taking the derivative. The formula for calculating the gradient is as follows: （23）； when When the gradient approaches 1, it is used to reduce the weight of easily classified samples. For class imbalance, inverse frequency weighting is used, and the formula for calculating inverse frequency weighting is as follows: （24）； in, The total number of samples, K represents the number of samples in category c, and K represents the number of categories. The regularized gain-aware loss function is designed, and the specific calculation formula is as follows: （25）； This loss function is used to smooth labels and make gradient updates more stable. Modulated cross-entropy loss automatically focuses on hard samples and alleviates imbalance through weight adjustment.

8. The method for gain grading and correction of breast ABUS images according to claim 7, characterized in that, In step S7, a reference map with moderate gain is generated, and the following operations are performed: Image statistical feature extraction: Let the input image be... The depth of the position The size is The pixel value range is Then the mean of the image and standard deviation The calculation formula is as follows: = （26）； = （27）； Determining the boundary of the correction value: Let the maximum pixel value of the image be... The maximum correction value is The minimum correction value is , position deep The calculation formula is as follows: （28）； =1.1+1.4×(1 ) （29）； =0.4+0.6× （30）； Establish an exponential mapping from gain to correction value: Data acquisition gain The physical gain of ultrasonic equipment Adjustment range ( Gain The corresponding correction value is γ( ), =0 corresponds to , =100 corresponds to λ is the mapping coefficient, and the formula for calculating the incremental mapping between the two is as follows: λ= (31); c( )= (32); Applying exponential mapping to obtain a correction value with appropriate gain: Appropriate gain ,Depend on The formula is as follows: = × （33）； Based on moderate gain Anomalous image correction: For images with evaluated gain anomalies, adaptive gain correction is performed using a correction algorithm to convert them into reference images with appropriate gain. The correction principle is as follows: ≥0,0≤ ≤1) （34）； in, To correct the parameters, The scaling factor (usually taken as...) =1), Given an input image, output The result after correction. This is the maximum pixel value.

Images