Self-adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification

Abstract

The invention relates to a self-adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification, and belongs to the field of deep learning image classification. According to the method, an adaptive residual gated feature fusion (AR-GFF) module and a deep learning architecture of an adaptive lightweight enhanced feature block (ALEFB) are combined. According to the method, the classification accuracy of the fetal ultrasonic image is improved, and the method is more suitable for being deployed in a real-time diagnosis scene, especially in a region with limited medical resources.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning image classification, and specifically relates to an adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification. Background Art

[0002] Fetal ultrasound imaging has become an indispensable tool for prenatal screening due to its safety, non-invasiveness, convenience, and lack of radiation. Through standardized image acquisition, it is possible to obtain views of various fetal anatomical structures, assist in prenatal diagnosis, and minimize the risk of missed diagnosis and misdiagnosis. Common screening biomarkers include fetal biometry, fetal weight estimation, and Doppler blood flow analysis.

[0003] However, fetal ultrasound image classification faces a series of challenges. First, fetal ultrasound images usually contain noise, blurred boundaries, and similarities between different structures, especially in the structures of different anatomical regions, where the anatomical structures of fetal ultrasound images are often highly similar. For example, the structures of the lateral ventricles and thalamus are very similar, which makes it difficult for traditional convolutional neural networks (CNNs) and Transformer-based models to accurately distinguish them. Second, the boundaries of ultrasound images are often blurred, making it difficult to capture subtle structural differences. Therefore, traditional models often show low classification accuracy when dealing with these complex anatomical structures. In addition, the limitation of computing resources is also a key challenge in clinical applications. While ensuring high classification accuracy, the model also needs to reduce computational complexity to ensure the feasibility of the model in real-time applications. Therefore, how to design a deep learning model that can both improve classification accuracy and maintain efficient computing has become an important topic in fetal ultrasound image analysis. Summary of the invention

[0004] The purpose of the present invention is to solve the anatomical similarity problem and computational efficiency bottleneck in fetal ultrasound image classification, and to provide an adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification. Combining the deep learning architecture of the adaptive residual gated feature fusion (AR-GFF) module and the adaptive lightweight enhanced feature block (ALEFB), the method of the present invention improves the accuracy of fetal ultrasound image classification and is more suitable for deployment in real-time diagnosis scenarios, especially in areas with limited medical resources.

[0005] To achieve the above object, the technical solution of the present invention is: an adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification, comprising:

[0006] An adaptive residual gated feature fusion module AR-GFF is proposed, and a bidirectional adaptive gated fusion layer BAGF and a residual feature enhancement module are introduced;

[0007] Design an adaptive lightweight enhanced feature block ALEFB, including optimizing the shape of the convolution kernel, optimizing the activation function, and adding adaptive parameters at the end of the basic block;

[0008] The DropBlock regularization method is introduced to reduce the model's dependence on local features, and a composite loss function is designed to solve the problem of overlapping category distributions.

[0009] In one embodiment of the present invention, optimizing the shape of the convolution kernel includes replacing the square convolution kernel with a rectangular convolution kernel.

[0010] In one embodiment of the present invention, the composite loss function includes Bhattacharyya Loss.

[0011] In one embodiment of the present invention, the activation function is optimized, that is, the Swish activation function is replaced by the Mish activation function.

[0012] In one embodiment of the present invention, the method is implemented in the following steps:

[0013] Step S0, collecting a standard mid-pregnancy fetal ultrasound image dataset from level II prenatal screening that meets the inclusion criteria;

[0014] Step S1, performing data preprocessing on the standard mid-pregnancy fetal ultrasound image data set collected in step S0 to obtain a preprocessed data set;

[0015] Step S2: Based on the data set preprocessed in step S1, an adaptive lightweight feature fusion model for enhancing fetal ultrasound image classification is constructed to achieve fetal ultrasound image classification.

[0016] In one embodiment of the present invention, in step S0, the standard level II prenatal screening includes 10 key 2D standard slices: (a) abdominal circumference cross section, (b) bilateral eye cross section, (c) bilateral kidney sagittal section, (d) bilateral kidney cross section, (e) cerebellum cross section, (f) lateral ventricle cross section, (g) facial sagittal section, (h) nasolabial coronal section, (i) spinal cord longitudinal section and (j) thalamus cross section.

[0017] In one embodiment of the present invention, step S1 is specifically implemented as follows:

[0018] Step S11, the size of the original ultrasound images in the standard mid-pregnancy fetal ultrasound image dataset is uniformly adjusted to 320×320 pixels, and then the images are converted into RGB format, and data enhancement operations are performed, including random cropping and scaling, horizontal flipping, rotation, and adjustment of hue, saturation, and brightness;

[0019] Step S12: Using the AutoAugment strategy with ImageNetPolicy to increase the diversity of the standard mid-pregnancy fetal ultrasound image dataset.

[0020] In one embodiment of the present invention, step S2 is specifically implemented as follows:

[0021] Step S21, designing an adaptive residual gated feature fusion module AR-GFF based on the EfficientNet network, in which a bidirectional adaptive gated fusion layer BAGF and a residual feature enhancement module are introduced to better capture and integrate subtle features in fetal ultrasound images;

[0022] Step S22, designing an adaptive lightweight enhanced feature block ALEFB, including optimizing the shape of the convolution kernel, replacing the Swish activation function with the Mish activation function, and adding adaptive parameters at the end of the basic block; wherein, the shape of the convolution kernel is optimized, that is, the standard 1x1 convolution kernel is replaced with a rectangular convolution kernel (k h , 1) and (1, k w ), (k h , 1) The kernel extracts the vertical features, (1, k h ) The kernel extracts features in the horizontal direction, and the size of the rectangular convolution kernel in the horizontal and vertical directions can be adjusted independently;

[0023] Step S23: Introduce DropBlock regularization and design a composite loss function in the entire model to reduce the model's dependence on local features and solve the problem of overlapping category distributions.

[0024] In one embodiment of the present invention, the AR-GFF is specifically implemented as follows:

[0025] The algorithm of BAGF is implemented as follows:

[0026] First, the input feature map F input Perform average pooling and concatenation operations:

[0027] F xy =Concat(X.AvgPool(F input ),Y.AvgPool(F input ))

[0028] Among them, X.AvgPool represents the average pooling along the horizontal direction, Y.AvgPool represents the average pooling along the vertical direction, and the Concat operation combines the pooling features along the channel dimension to form F xy ;

[0029] Next, a convolution operation followed by a Sigmoid activation function is applied to the concatenated features Fxy , to calculate the gating signals along the X-axis and Y-axis directions:

[0030] G x =σ(Conv2d(F xy )),G y =σ(Conv2d(F xy ))

[0031] Among them, Conv2d(F xy ) represents the cascade feature map F xy 2D convolution operation, which will optimize F xy , and prepares it for gating. σ(·) is the Sigmoid activation function, which outputs a value between 0 and 1, allowing the gating signal G x and G y Acts as a weight or filter to control the flow of feature information, where G x and G y is the generated gating signal, corresponding to the X-axis and Y-axis directions respectively; the adaptive parameters α and β are introduced to adjust the ratio of elements in the two spatial directions, respectively controlling the influence of the gating signal in each direction on the features in the X-axis and Y-axis directions:

[0032] F′ x =α·G x ⊙F x ,F′ y =β·G y ⊙F y

[0033] Among them, F x and F y denote the original feature maps along the X-axis and Y-axis respectively, ⊙ denotes element-wise multiplication, which applies the gating effect to each spatial feature individually;

[0034] Finally, the adjusted feature F' x and F' y concatenate along the channel dimension to produce the final output feature of the BAGF module, denoted as F BAGF Output:

[0035] F BAGF_output =Concat(F′ x ,F′ y )

[0036] The algorithm implementation of the residual feature enhancement module is:

[0037] First, perform batch normalization operation BatchNorm and nonlinear activation operation Nonlinear on BAGF output;

[0038] Factivated =Nonlinear(BatchNorm(F BAGF_output ))

[0039] Next, convolution and sigmoid activation operations are applied to F activated , to generate an updated gating signal and

[0040]

[0041] After this, use the gate signal to control F activated Perform the re-weighting operation Re-weight to generate the output feature F output :

[0042] F output =Re-weight(F activated )

[0043] Finally, by transforming the original input feature map F input Add to F output The residual connection is introduced in order to obtain the final module output F AR-GFF Output:

[0044] F AR-GFF_output =F output +F input .

[0045] In one embodiment of the present invention, the composite loss function includes cross entropy loss and Bhattacharyya Loss, and the composite loss function is expressed as:

[0046] L=α·L CE +(1-α)·L BC

[0047] Among them, α is a weight to ensure that the cross entropy loss guides the training, while Bhattacharyya Loss enhances the model's attention to the distribution similarity between classes, and L CE is the cross entropy loss, L BC is Bhattacharyya Loss, which is calculated as follows:

[0048]

[0049] Among them, L BC represents the Bhattacharyya loss value, which is used to measure the distribution similarity between the predicted class and the target class, C is the total number of classes, and p i represents the predicted probability of class i, represents the confidence of the model in class prediction, q irepresents the actual probability distribution of the target for class i, providing the actual likelihood.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The adaptive residual gated feature fusion (AR-GFF) module proposed in this paper significantly improves the model's ability to capture tiny features in complex fetal ultrasound images by introducing a bidirectional adaptive gated fusion (BAGF) layer and a residual feature enhancement module. This design effectively solves the problem of insufficient classification accuracy of traditional methods when dealing with fuzzy boundaries and areas with similar anatomical structures.

[0052] 2. Based on the AR-GFF module, the present invention further designs an adaptive lightweight enhanced feature block (ALEFB). By optimizing the shape of the convolution kernel (such as replacing the square convolution kernel with a rectangular convolution kernel), replacing the Swish activation function with the Mish activation function, and adding adaptive parameters at the end of the basic block, the number of model parameters and computational complexity are significantly reduced, while maintaining or even improving the classification accuracy.

[0053] 3. To enhance the robustness of the model in different clinical scenarios, the present invention introduces the DropBlock regularization method to reduce the model's dependence on local features, and designs a composite loss function (including Bhattacharyya Loss) to solve the problem of overlapping category distributions. These improvements enable the model to perform well in resource-constrained environments.

[0054] 4. The model of the present invention not only surpasses the traditional method in the accuracy of fetal ultrasound image classification, but also makes it more suitable for deployment in real-time diagnosis scenarios through lightweight design, especially in areas with limited medical resources. In addition, by combining with doctor-assisted diagnosis, the model of the present invention significantly improves the doctor's diagnostic efficiency and accuracy, helps to reduce the misdiagnosis rate and optimize the clinical workflow. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is the data distribution diagram obtained in step S0.

[0056] Figure 2 This is the structural diagram of AR-GFF.

[0057] Figure 3 The figure shows the structural comparison of the model evolution; (a) is the BAGF layer, (b) is the R-BAGF layer, (c) is the enhanced MBConv with AR-GFF, and (d) is the adaptive lightweight enhancement functional block. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0059] The present invention provides an adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification, comprising:

[0060] An adaptive residual gated feature fusion module AR-GFF is proposed, and a bidirectional adaptive gated fusion layer BAGF and a residual feature enhancement module are introduced;

[0061] Design an adaptive lightweight enhanced feature block ALEFB, including optimizing the shape of the convolution kernel, optimizing the activation function, and adding adaptive parameters at the end of the basic block;

[0062] The DropBlock regularization method is introduced to reduce the model's dependence on local features, and a composite loss function is designed to solve the problem of overlapping category distributions.

[0063] The following is the specific implementation process of the present invention.

[0064] The present invention provides an adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification, comprising the following steps:

[0065] Step S0. According to the Level II prenatal screening guidelines provided by the Ultrasound Branch of the Chinese Medical Physician Association, 10 key 2D standard slices are retained as needed: (a) abdominal circumference cross section, (b) bilateral eye cross section, (c) bilateral kidney sagittal section, (d) bilateral kidney cross section, (e) cerebellum cross section, (f) lateral ventricle cross section, (g) facial sagittal section, (h) nasolabial coronal section, (i) spinal cord longitudinal section and (j) thalamus cross section. The standard cross-sectional images of each research subject are annotated and archived by the participating physicians to ensure consistency and accuracy. The annotation process involves multiple physicians, adheres to standardized protocols, and all annotations are strictly reviewed to ensure compliance with medical standards. The present invention collected 1,869 standard mid-pregnancy fetal ultrasound images from Level II screening that met the inclusion criteria. As Figure 1 shown.

[0066] Step S1: special data processing, specifically including:

[0067] Step S11: During data preprocessing, all original ultrasound images are uniformly resized to 320×320 pixels to ensure consistent input to the model. The images are then converted to RGB format and various data augmentation techniques are applied, including random cropping and scaling, horizontal flipping, rotation (from -15° to 15°), and adjustment of hue, saturation, and brightness. These enhancements are used to enhance the robustness and generalization ability of the model.

[0068] Step S12, in addition, the AutoAugment strategy with ImageNetPolicy is used to further increase the diversity of training data. When the scaled image does not meet the expected size, gray padding will maintain a consistent input size without changing the aspect ratio of the original image.

[0069] Step S3: Model establishment, specifically including:

[0070] Step S21, design an adaptive residual gated feature fusion (AR-GFF) mechanism based on the EfficientNet network, specifically for the optimization of feature extraction and fusion capabilities to improve accuracy. In AR-GFF, a bidirectional adaptive gated fusion (BAGF) layer and a residual feature enhancement module are introduced to better capture and integrate subtle features in fetal ultrasound images. The AR-GFF model structure is shown in the figure below. Figure 2 The BAGF module structure diagram and the improved R-BAGF diagram are shown in Figure 3 (a) and 3(b).

[0071] The algorithm of the bidirectional adaptive gating fusion layer is implemented as:

[0072] First, average pooling and concatenation operations are performed on the input feature map:

[0073] F xy =Concat(X.AvgPool(F input ),Y.AvgPool9F input ))

[0074] Where X.AvgPool represents the average pooling along the horizontal direction, and Y.AvgPool represents the average pooling along the vertical direction. The Concat operation combines these pooled features along the channel dimension to form Fxy.

[0075] Next, a convolution operation followed by a Sigmoid activation function is applied to the concatenated features Fxy to calculate the gating signals along the X-axis and Y-axis directions:

[0076] G x =σ(Conv2d(F xy )),G y =σ(Conv2d(F xy ))

[0077] Where Conv2d(F xy ) represents the cascade feature map F xy 2D convolution operation. This operation will optimize F xyσ(·) is the Sigmoid activation function, which outputs a value between 0 and 1, allowing the gating signal G x and G y Acts as a weight or filter to control the flow of feature information. x and G y are the generated gating signals, corresponding to the X-axis and Y-axis directions respectively.

[0078] Adaptive parameters α and β are introduced to adjust the ratio of features in the two spatial directions. These parameters control the influence of each directional gating signal on the features in the X-axis and Y-axis directions respectively:

[0079] F′ x =α·G x ⊙F x ,F′ y =β·G y ⊙F y

[0080] Among them, F x and F y Represents the original feature map along the X-axis and Y-axis respectively. x and G y are the gating signals calculated in the previous step for the X-axis and Y-axis directions. The symbols α and β represent adaptive parameters used to scale the contribution of the gating feature in each direction, allowing flexible use of feature representation. In addition, ⊙ represents element-wise multiplication, which applies the gating effect to each spatial feature individually.

[0081] Finally, the adjusted feature F' x and F' y concatenate along the channel dimension to produce the final output feature of the BAGF module, denoted as F BAGF Output:

[0082] F BAGF_output =Concat(F′ x ,F′ y )

[0083] The algorithm implementation of the residual feature enhancement module is:

[0084] First, batch normalization and nonlinear activation operations are performed on the FBAGF output:

[0085] F activated =Nonlinear(BatchNorm(F BAGF_output ))

[0086] Next, convolution and sigmoid activation operations are applied to Factivated to generate an updated gating signal and

[0087]

[0088] After that, the gating signal is used to perform a reweighting operation on Factivated to produce the output feature F output :

[0089] F output =Re-weight(F activated )

[0090] Finally, by transforming the original input feature map F input Add to F output The residual connection is introduced in order to obtain the final module output F AR-GFF Output:

[0091] F AR-GFF_output =F output +F input .

[0092] Step S22: In order to further reduce the computational complexity while maintaining or even improving the classification accuracy, the present invention develops an adaptive lightweight enhancement function block (ALEFB). This module achieves a lightweight design through several key improvements, including replacing the Swish activation function with Mish, optimizing the convolution kernel shape, and adding adaptive parameters at the end of the basic block. The architecture diagram of ALEFB is shown in Figure 3 Specifically, the present invention replaces the standard 1x1 convolution kernel with a rectangular convolution kernel (k h , 1) and (1, k w ). (k h , 1) kernel mainly extracts features in the vertical direction, while (1, k h ) kernel mainly extracts features in the horizontal direction. This design not only preserves the lightweight nature of the model, but also allows for more detailed capture of spatial information in the input feature map. The sizes of the rectangular convolution kernels in the horizontal and vertical directions can be adjusted independently, allowing the network to more effectively process anisotropic features in the image. This flexibility improves the efficiency of feature extraction and reduces computational complexity, thereby improving model performance while achieving a lightweight design.

[0093] Step S23: In order to further improve the robustness and classification performance of the model, this invention introduces DropBlock regularization and designs a composite loss function in the entire model. Bhattacharyya Loss is calculated as follows:

[0094]

[0095] The final composite loss function is expressed as: L = α·L CE +(1-α)·L BC

[0096] In the design of the present invention, α is set to 0.9 to ensure that the cross entropy loss mainly guides the training, while Bhattacharyya Loss enhances the model's attention to the distribution similarity between classes. This design effectively enhances the model's discriminative ability, making it more robust and accurate in distinguishing subtle class differences. Especially in the classification of fetal ultrasound images, the introduction of Bhattacharyya Loss enables the model to better handle the challenges brought by high inter-class similarity and distribution overlap, thereby achieving excellent performance in practical applications.

[0097] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. An adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification, characterized in that: include: An adaptive residual gated feature fusion module AR-GFF is proposed, and a bidirectional adaptive gated fusion layer BAGF and a residual feature enhancement module are introduced; Design an adaptive lightweight enhanced feature block ALEFB, including optimizing the shape of the convolution kernel, optimizing the activation function, and adding adaptive parameters at the end of the basic block; The DropBlock regularization method is introduced to reduce the model's dependence on local features, and a composite loss function is designed to solve the problem of overlapping category distributions.

2. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 1, characterized in that: Optimizing the convolution kernel shape involves replacing the square convolution kernel with a rectangular convolution kernel.

3. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 1, characterized in that: Composite loss functions include Bhattacharyya Loss.

4. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 1, characterized in that: Optimize the activation function, that is, replace the Swish activation function with the Mish activation function.

5. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 1, characterized in that: The method is implemented in the following steps: Step S0, collecting a standard mid-pregnancy fetal ultrasound image dataset from level II prenatal screening that meets the inclusion criteria; Step S1, performing data preprocessing on the standard mid-pregnancy fetal ultrasound image data set collected in step S0 to obtain a preprocessed data set; Step S2: Based on the data set preprocessed in step S1, an adaptive lightweight feature fusion model for enhancing fetal ultrasound image classification is constructed to achieve fetal ultrasound image classification.

6. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 5, characterized in that: In step S0, the standard level II prenatal screening includes 10 key 2D standard sections: (a) abdominal circumference transverse section, (b) bilateral eye transverse section, (c) bilateral kidney sagittal section, (d) bilateral kidney transverse section, (e) cerebellum transverse section, (f) lateral ventricle transverse section, (g) facial sagittal section, (h) nasolabial coronal section, (i) spinal cord longitudinal section and (j) thalamus transverse section.

7. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 5, characterized in that: Step S1 is specifically implemented as follows: Step S11, the size of the original ultrasound images in the standard mid-pregnancy fetal ultrasound image dataset is uniformly adjusted to 320×320 pixels, and then the images are converted into RGB format, and data enhancement operations are performed, including random cropping and scaling, horizontal flipping, rotation, and adjustment of hue, saturation, and brightness; Step S12: Using the AutoAugment strategy with ImageNetPolicy to increase the diversity of the standard mid-pregnancy fetal ultrasound image dataset.

8. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 5, characterized in that: Step S2 is specifically implemented as follows: Step S21, designing an adaptive residual gated feature fusion module AR-GFF based on the EfficientNet network, in which a bidirectional adaptive gated fusion layer BAGF and a residual feature enhancement module are introduced to better capture and integrate subtle features in fetal ultrasound images; Step S22, designing an adaptive lightweight enhanced feature block ALEFB, including optimizing the shape of the convolution kernel, replacing the Swish activation function with the Mish activation function, and adding adaptive parameters at the end of the basic block; wherein, the shape of the convolution kernel is optimized, that is, the standard 1x1 convolution kernel is replaced with a rectangular convolution kernel (k h , 1) and (1, k w ), (k h , 1) The kernel extracts the vertical features, (1, k h ) The kernel extracts features in the horizontal direction, and the size of the rectangular convolution kernel in the horizontal and vertical directions can be adjusted independently; Step S23: Introduce DropBlock regularization and design a composite loss function in the entire model to reduce the model's dependence on local features and solve the problem of overlapping category distributions.

9. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 8, characterized in that: The AR-GFF is specifically implemented as follows: The algorithm of BAGF is implemented as follows: First, the input feature map F input Perform average pooling and concatenation operations: F xy =Concat(X.AvgPool(F input ),Y.AvgPool(F input )) Among them, X.AvgPool represents the average pooling along the horizontal direction, Y.AvgPool represents the average pooling along the vertical direction, and the Concat operation combines the pooling features along the channel dimension to form F xy ; Next, a convolution operation followed by a Sigmoid activation function is applied to the concatenated features F xy , to calculate the gating signals along the X-axis and Y-axis directions: G x =σ(Conv2d(F xy )),G y =σ(Conv2d(F xy )) Among them, Conv2d(F xy ) represents the cascade feature map F xy 2D convolution operation, which will optimize F xy , and prepares it for gating. σ(·) is the Sigmoid activation function, which outputs a value between 0 and 1, allowing the gating signal G x and G y Acts as a weight or filter to control the flow of feature information, where G x and G y is the generated gating signal, corresponding to the X-axis and Y-axis directions respectively; the adaptive parameters α and β are introduced to adjust the ratio of elements in the two spatial directions, respectively controlling the influence of the gating signal in each direction on the features in the X-axis and Y-axis directions: F′ x =α·G x ⊙F x ,F y =β·G y ⊙F y Among them, F x and F y denote the original feature maps along the X-axis and Y-axis respectively, ⊙ denotes element-wise multiplication, which applies the gating effect to each spatial feature individually; Finally, the adjusted feature F' x and F' y concatenate along the channel dimension to produce the final output feature of the BAGF module, denoted as F BAGF Output: F BAGF_output =Concat(F′ x ,F′ y ) The algorithm implementation of the residual feature enhancement module is: First, perform batch normalization operation BatchNorm and nonlinear activation operation Nonlinear on BAGF output; F activated =Nonlinear(BatchNorm(F BAGF_output )) Next, convolution and sigmoid activation operations are applied to F activated , to generate an updated gating signal and After this, use the gate signal to control F activated Perform the re-weighting operation Re-weight to generate the output feature F output : F output =Re-weight(F activated ) Finally, by transforming the original input feature map F input Add to F output The residual connection is introduced in order to obtain the final module output F AR-GFF Output: F AR-GFF_output =F output +F input 。 10. The adaptive lightweight feature fusion method for enhancing fetal ultrasound image classification according to claim 8, characterized in that: The composite loss function includes cross entropy loss and Bhattacharyya Loss. The composite loss function is expressed as: L=α·L CE +(1-a)·L BC Among them, α is a weight to ensure that the cross entropy loss guides the training, while Bhattacharyya Loss enhances the model's attention to the distribution similarity between classes, and L CE is the cross entropy loss, L BC It is Bhattacharyya Loss, which is calculated as follows: Among them, L BC represents the Bhattacharyya loss value, which is used to measure the distribution similarity between the predicted class and the target class, C is the total number of classes, and p i represents the predicted probability of class i, represents the confidence of the model in class prediction, q i represents the actual probability distribution of the target for class i, providing the actual likelihood.

Images