B ultrasonic image enhancement processing method based on deep learning

Abstract

The invention provides a B ultrasonic image enhancement processing method based on deep learning, and belongs to the field of image enhancement. The method comprises a plurality of key modules: a feature preprocessing module extracts important features in a B ultrasonic image, enhances a key information area, and reduces redundancy and noise at the same time; the feature extraction module further extracts local and global features in the image by combining convolution operation and a self-attention mechanism, and improves details and a global structure of the image; the feature transformation module processes the feature map by using a convolution and activation function, optimizes the contrast and details of the image, and ensures that the detail part is accurately recovered; the method can recover key information in a low-resolution image, enhance image details and improve image quality, and is suitable for scenes with high requirements on B ultrasonic image quality and precision in medical image analysis.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and in particular relates to a B-ultrasound image enhancement processing method based on deep learning. Background Art

[0002] As an important part of artificial intelligence technology, deep learning has made breakthrough progress in computer vision, natural language processing, speech recognition and other fields. Deep learning uses multi-layer neural networks to automatically extract and learn data features, which can significantly improve the accuracy and efficiency of image processing tasks. Deep learning models such as convolutional neural networks (CNN) have shown powerful capabilities in applications such as image classification, segmentation and enhancement. Especially in the field of medical imaging, the application of deep learning technology has achieved remarkable results. For example, deep learning is widely used in the automatic diagnosis and enhancement of medical images such as X-rays, CT and MRI, which can effectively improve image quality, help doctors more accurately identify lesion areas, and provide auxiliary diagnosis support.

[0003] B-ultrasound (ultrasound) images are widely used in clinical medical diagnosis, especially in the fields of gynecology, obstetrics, and cardiovascular. Since the imaging process of B-ultrasound images is affected by factors such as signal noise, resolution limitation, and image blur, the image quality is usually poor, which makes it difficult for doctors to accurately determine the condition during diagnosis. Therefore, how to enhance the quality of B-ultrasound images and improve their clarity and recognizability of details has become an important research direction in the field of image processing. Traditional B-ultrasound image enhancement methods include image denoising, contrast enhancement, edge enhancement, etc. Although they can improve image quality to a certain extent, the effect is usually limited and requires manual adjustment of parameters, which is cumbersome to operate. With the development of deep learning technology, B-ultrasound image enhancement methods based on deep learning have gradually become an effective solution. By training deep neural networks and automatically optimizing the image processing process, efficient and accurate B-ultrasound image enhancement can be achieved, significantly improving the clarity and contrast of image details, and helping doctors to make better diagnoses. Summary of the invention

[0004] The present invention provides a B-ultrasound image enhancement processing method based on deep learning, which aims to improve the quality of B-ultrasound images and enhance the details and clarity of images through deep learning models and feature extraction technology, especially to accurately enhance key areas (such as lesion areas, tissue boundaries, etc.) in low-resolution images, so as to assist doctors in making more accurate judgments in medical diagnosis; this method can improve the contrast of images, sharpen edges, and reduce noise interference through innovative feature processing and feature transformation modules under the premise of ensuring the integrity of image information, and has significant practical application value.

[0005] The present invention aims to propose an enhancement model for B-ultrasound images and provide a B-ultrasound image enhancement processing method based on deep learning, which includes the following steps.

[0006] S1. Preparation of B-ultrasound image dataset: Use high-resolution B-ultrasound equipment to collect B-ultrasound images. Based on the high-resolution B-ultrasound images, a downsampling algorithm is used to convert them into low-resolution images, generating a pair of high-resolution B-ultrasound images and their corresponding low-resolution B-ultrasound images, forming a training set for B-ultrasound image enhancement.

[0007] S2. Construct a feature preprocessing module, including spatial attention and IR modules. The spatial attention module uses the Sigmod activation function to obtain the feature map weight and performs weighted calculation with the original feature map. Then, the IR module is used for feature extraction and dimensionality reduction. The IR module includes convolution and pooling operations.

[0008] S3. Construct a feature extraction module. The feature extraction module combines the advantages of convolution and self-attention to extract global and local features in B-ultrasound images.

[0009] S4. Construct a feature transformation module, including an upsampling module, a downsampling module, a 3×3 convolution and a ReLU activation function. First, downsample the B-ultrasound image features to reduce its resolution and extract the image macro features. Then, upsample it to obtain its detail information. By combining upsampling and downsampling, the image detail information is retained. Finally, it is processed by 3×3 convolution and ReLU activation function.

[0010] S5. Construct a B-ultrasound image enhancement model, including input, 3×3 convolution, feature preprocessing module, feature extraction module, feature transformation module, 3×3 convolution, fully connected layer and output.

[0011] S6, B-ultrasound image processing, obtaining the B-ultrasound image that needs to be enhanced, inputting it into the B-ultrasound image enhancement model, and outputting a high-resolution B-ultrasound image.

[0012] Preferably, in step S1, high-resolution B-ultrasound images are collected, and high-resolution B-ultrasound equipment such as high-end color ultrasound machines are used to collect B-ultrasound images in different clinical scenarios. Based on the collected high-resolution B-ultrasound images, a downsampling algorithm is used to convert them into low-resolution images. The B-ultrasound images collected by the low-resolution acquisition machine are simulated through downsampling technology to generate a pair of high-resolution B-ultrasound images and their corresponding low-resolution B-ultrasound images for training the deep learning model, thereby forming a training set for B-ultrasound image enhancement, which includes pairs of high-resolution B-ultrasound images and low-resolution B-ultrasound images for training the deep learning model.

[0013] Preferably, in step S2, for the feature preprocessing module, the low-resolution B-ultrasound image features are input, and are set to , , and Respectively represent The height, width and number of channels of Perform spatial attention operations, including: first process the input features through global average pooling and global maximum pooling , generate two feature maps, and obtain the spatial attention map after 1×1 convolution operation , , is the activation function, is a 1×1 convolution, is the global average pooling, is the global maximum pooling, and then the attention map With feature map Performing element-wise multiplication, we get , , represents an element-wise multiplication operation, Indicates input Using Attention Map The result after weighting; The IR module consists of a 3×3 convolutional layer and pooling operations for feature extraction and dimensionality reduction. , the feature map after being processed by the IR module has a reduced spatial resolution and an increased number of channels: , , and They are The height, width and number of channels of It is the B-ultrasound image feature output by the feature preprocessing module.

[0014] Preferably, in step S2, for the feature preprocessing module, the B-ultrasound image features are effectively extracted and optimized through the spatial attention mechanism and the inverted residual (IR) module; during the processing, two feature maps are first generated through global average pooling and global maximum pooling, and a spatial attention map is generated through a 1×1 convolution operation, so as to enhance the feature expression of important areas in the image; the spatial attention mechanism enables the model to adaptively focus on key areas in the image, increase the weight of image details and important information, and effectively remove irrelevant information; then, after processing by the inverted residual module, the spatial resolution of the feature map is reduced by half, while retaining key information and removing redundant information, further improving the accuracy and efficiency of feature expression; this module provides high-quality input features for subsequent image enhancement or disease detection tasks by reducing redundancy in the feature map and improving computational efficiency, ensuring that information transmission in the processing process is more efficient and accurate.

[0015] Preferably, in step S3, the feature extraction module inputs the preprocessed B-ultrasound image features obtained by the feature preprocessing module. , , , and Respectively The feature extraction module combines the advantages of convolution and self-attention to extract local and global features in B-ultrasound images. The steps include: S31, use two 1×1 convolutional layers to process separately , generating two projection feature maps and , , , Focus on low-level color and texture features, Focus on high-order edge and shape features. and Concatenate in the channel dimension to form the fusion feature F, , Represents a splicing operation, , F integrates the low-order features and high-order features of B-ultrasound image features; S32. Generate Query, Key and Value using 1×1 convolution from the fused feature F. Query, Key and Value are used for self-attention operation. Query, Key and Value are recorded as Q, K and V respectively. , , , , is a 1×1 convolution, and then generates the attention weight O, , is the activation function, is the transposed matrix of K, is the scaling factor; finally, the output B-ultrasound image features are generated through residual connection and 1×1 convolution: , is a 1×1 convolution, , , and They are The height, width and number of channels of It is the B-ultrasound image features output by the feature extraction module.

[0016] Preferably, in step S3, for the feature extraction module, the advantages of convolution and self-attention mechanisms are effectively combined through a hybrid attention module to simultaneously extract global and local features in the image; in this module, 1×1 convolution is first used to extract low-order features (color and texture) and high-order features (edge ​​and shape) respectively, and then these two features are concatenated (Concat) in the channel dimension to form fused features; the fused features are then input into the self-attention mechanism to effectively model global and local dependencies; finally, the features processed by attention are output through residual connections and 1×1 convolutions; this method combines the local feature extraction of convolution and the global modeling advantages of self-attention, which not only improves the accuracy of feature extraction, but also enhances the ability to restore image details and global structures.

[0017] Preferably, in step S4, the feature conversion module inputs the feature map after passing through the feature extraction module. , , , and Respectively The height, width and number of channels of the feature map are first converted into Downsampling is performed to capture the global structural information of the B-ultrasound image and obtain low-resolution B-ultrasound image features : , represents the downsampling operation, , , and They are The height, width and number of channels of the network are calculated; the low-resolution features are then restored to high resolution through transposed convolution upsampling to obtain more refined local information: , represents the upsampling operation, , , and Respectively The height, width and number of channels of the original feature map and multi-scale features Fusion, generating feature representations that contain both details and global structures: , Represents a splicing operation, ; Use 3×3 convolution on the feature map Perform transformations to learn complex relationships between features, and use activation functions to introduce nonlinearity after convolution operations: , is the activation function, is a 3×3 convolution, , , and They are The height, width and number of channels of It is the B-ultrasound image feature output by the feature transformation module.

[0018] Preferably, in step S4, for the feature transformation module, the feature transformation module enhances the details and global structure of the image; down-sampling is performed through convolution and pooling operations to convert the feature map into a low-resolution image to capture the global structural information of the image. This process helps to extract the macroscopic features of the image, avoid the redundancy of local details, and reduce the computational complexity; through up-sampling operations, the low-resolution features are restored to high resolution to ensure that the details are restored more finely; this strategy of combining up-sampling and down-sampling not only ensures that the global information of the image is effectively retained, but also can restore the details in the image at high resolution, avoid excessive amplification of noise or distortion, and ensure that the effect of image enhancement is more balanced, which can not only restore local details but also maintain the integrity of global information.

[0019] Compared with the prior art, the present invention has the following technical effects: The present invention provides a B-ultrasound image enhancement processing method based on deep learning, which aims to improve the quality of B-ultrasound images, especially enhance the details and clarity in the images, especially accurately enhance the key areas such as lesion areas and tissue boundaries in low-resolution images, so as to help doctors make more accurate judgments in medical diagnosis; the method first collects high-resolution B-ultrasound images and converts them into low-resolution images using a downsampling algorithm to generate high-resolution and low-resolution image pairs to form an image data set for training; a feature preprocessing module is constructed, which combines a spatial attention mechanism and an inverted residual module to adaptively enhance the key information areas in the image, remove irrelevant redundant information, and optimize feature representation; the feature extraction module combines convolution operations and self-attention mechanisms to simultaneously extract local features and global features of the image, and further improve the detail performance of the image; the feature transformation module processes the image features through convolution and activation functions to effectively improve the clarity and contrast of the image; finally, based on these processing modules, a complete B-ultrasound image enhancement model is constructed to output an enhanced image with high resolution and rich details. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a B-ultrasound image enhancement processing method based on deep learning provided by the present invention.

[0021] Figure 2 It is a structural diagram of the feature preprocessing module provided by the present invention.

[0022] Figure 3 It is a structural diagram of the feature extraction module provided by the present invention.

[0023] Figure 4 It is a structural diagram of the feature transformation module provided by the present invention.

[0024] Figure 5 It is a high-resolution B-ultrasound image effect diagram obtained after high-resolution processing of the low-resolution B-ultrasound image provided by the present invention. DETAILED DESCRIPTION

[0025] The present invention aims to propose a B-ultrasound image enhancement processing method based on deep learning, and realizes efficient image quality improvement by designing a B-ultrasound image enhancement model including multiple key modules; the model includes a feature preprocessing module, a feature extraction module, and a feature transformation module, and each module works together to improve the image processing effect; firstly, the feature preprocessing module is combined with a spatial attention mechanism and an inverted residual module to extract key information in the B-ultrasound image, and effectively remove noise and redundant features to ensure that important areas in the image are accurately enhanced; then, the feature extraction module is combined with a convolution operation and a self-attention mechanism to further enhance the expressiveness of image details, extract local and global features, and thus enhance the structural information and texture details of the image; the feature transformation module improves the contrast and sharpness of the image through a combination of convolution and activation functions, and accurately restores the details to ensure that the image quality is optimized at all levels; through the synergistic effect of the modules, the present invention can effectively improve the quality of B-ultrasound images, enhance image details, and improve key information in low-resolution images, and is suitable for scenes with high requirements for image quality in medical diagnosis.

[0026] Please refer to Figure 1, a B-ultrasound image enhancement processing method based on deep learning in an embodiment of the present application, the process is as follows.

[0027] S1. Ultrasound image dataset preparation: high-resolution B-ultrasound equipment is used to collect B-ultrasound images with a size of 640×640. Based on the high-resolution B-ultrasound images, a downsampling algorithm is used to convert them into low-resolution images, generating a pair of high-resolution B-ultrasound images and their corresponding low-resolution B-ultrasound images to form a training set for B-ultrasound image enhancement.

[0028] Furthermore, in step S1, high-resolution B-ultrasound images are collected. A high-resolution B-ultrasound device, such as a high-end color ultrasound machine, is used to collect B-ultrasound images in different clinical scenarios. The size of the B-ultrasound images is 640×640. Based on the collected high-resolution B-ultrasound images, a downsampling algorithm is used to convert them into low-resolution images. The downsampling technology is used to simulate the B-ultrasound images collected by the low-resolution acquisition machine to generate a pair of high-resolution B-ultrasound images and their corresponding low-resolution B-ultrasound images for training the deep learning model, thereby forming a training set for B-ultrasound image enhancement, including 100 pairs of high-resolution B-ultrasound images and low-resolution B-ultrasound images for training the deep learning model.

[0029] S2. Construct a feature preprocessing module, including spatial attention and IR modules. The spatial attention module uses the Sigmod activation function to obtain the feature map weight and performs weighted calculation with the original feature map. Then, the IR module is used for feature extraction and dimensionality reduction. The IR module includes convolution and pooling operations.

[0030] Further, in step S2, for the feature preprocessing module, see Figure 2 As shown, input low-resolution B-ultrasound image features, set , , and Respectively represent The height, width and number of channels of Perform spatial attention operations, including: first process the input features through global average pooling and global maximum pooling , generate two feature maps, and obtain the spatial attention map after 1×1 convolution operation , , is the activation function, is a 1×1 convolution, is the global average pooling, is the global maximum pooling, and then the attention map With feature map Performing element-wise multiplication, we get , , represents an element-wise multiplication operation, Indicates input Using Attention Map The result after weighting; The IR module consists of a 3×3 convolutional layer and pooling operations for feature extraction and dimensionality reduction. , the feature map after being processed by the IR module has a reduced spatial resolution and an increased number of channels: , , and They are The height, width and number of channels of It is the B-ultrasound image feature output by the feature preprocessing module.

[0031] S3. Construct a feature extraction module. The feature extraction module combines the advantages of convolution and self-attention to extract global and local features in B-ultrasound images.

[0032] Further, in step S3, for the feature extraction module, see Figure 3 As shown, the preprocessed B-ultrasound image features of the feature preprocessing module are input , , , and Respectively The feature extraction module combines the advantages of convolution and self-attention to extract local and global features in B-ultrasound images. The steps include: Use two 1×1 convolutional layers to process , generating two projection feature maps and , , , Focus on low-level color and texture features, Focus on high-order edge and shape features. and Concatenate in the channel dimension to form the fusion feature F, , Represents a splicing operation, , F integrates the low-order features and high-order features of B-ultrasound image features; The fused feature F is used to generate Query, Key and Value using 1×1 convolution. Query, Key and Value are used for self-attention operation. Query, Key and Value are denoted as Q, K and V respectively. , , , , is a 1×1 convolution, and then generates the attention weight O, , is the activation function, is the transposed matrix of K, is the scaling factor; finally, the output B-ultrasound image features are generated through residual connection and 1×1 convolution: , is a 1×1 convolution, , , and They are The height, width and number of channels of It is the B-ultrasound image features output by the feature extraction module.

[0033] S4. Construct a feature transformation module, including an upsampling module, a downsampling module, a 3×3 convolution and a ReLU activation function. First, downsample the B-ultrasound image features to reduce its resolution and extract the image macro features. Then, upsample it to obtain its detail information. By combining upsampling and downsampling, the image detail information is retained. Finally, it is processed by 3×3 convolution and ReLU activation function.

[0034] Further, in step S4, for the feature transformation module, see Figure 4 As shown, the feature map after the input passes through the feature extraction module , , , and Respectively The height, width and number of channels of the feature map are first converted into Downsampling is performed to capture the global structural information of the B-ultrasound image and obtain low-resolution B-ultrasound image features : , represents the downsampling operation, , , and They are The height, width and number of channels of the network are calculated; the low-resolution features are then restored to high resolution through transposed convolution upsampling to obtain more refined local information: , represents the upsampling operation, , , and Respectively The height, width and number of channels of the original feature map and multi-scale features Fusion, generating feature representations that contain both details and global structures: , Represents a splicing operation, ; Use 3×3 convolution on the feature map Perform transformations to learn complex relationships between features, and use activation functions to introduce nonlinearity after convolution operations: , , , and They are The height, width and number of channels of It is the B-ultrasound image feature output by the feature transformation module.

[0035] S5. Construct a B-ultrasound image enhancement model, including input, 3×3 convolution, feature preprocessing module, feature extraction module, feature transformation module, 3×3 convolution, fully connected layer and output.

[0036] Further, in step S5, for the B-ultrasound image enhancement model, a low-resolution B-ultrasound image is input , , , and Respectively The height, width and number of channels of the image are as follows. In this embodiment, H=640, W=640, C=3, that is, the input image is , the image First input into the 3×3 convolution to obtain low-resolution B-ultrasound image features ,Will Input to the feature preprocessing module and get ,Will Input into the feature extraction module and get ,Will Input to the feature transformation module and get ,Will Input to 3×3 convolution for processing, the processed B-ultrasound features are input to the fully connected layer, and finally output to obtain , .

[0037] Further, in step S5, the B-ultrasound image enhancement model is coded using the Pycharm application and Python language, using the Pytorch framework, and the model input is trained with a low-resolution B-ultrasound image with a resolution of 640×640×3, and the model is trained from scratch for 200 epochs.

[0038] S6, B-ultrasound image processing, obtaining the B-ultrasound image that needs to be enhanced, inputting it into the B-ultrasound image enhancement model, and outputting a high-resolution B-ultrasound image.

[0039] Further, if Figure 5 As shown, Figure 5 The left shows a low-resolution B-ultrasound image. Figure 5 The right side shows a high-resolution B-ultrasound image processed by the B-ultrasound image enhancement model. It can be seen that the image clarity and detail have been significantly improved, the details are clearer, and subsequent medical diagnosis is more accurate.

[0040] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A B-ultrasound image enhancement processing method based on deep learning, characterized in that: The following steps are involved: S1. Preparation of B-ultrasound image dataset: using high-resolution B-ultrasound equipment to collect B-ultrasound images, based on the high-resolution B-ultrasound images, using downsampling algorithms to convert them into low-resolution images, generating a pair of high-resolution B-ultrasound images and their corresponding low-resolution B-ultrasound images, forming a training set for B-ultrasound image enhancement; S2. Construct a feature preprocessing module, including a spatial attention module and an IR module. The spatial attention module uses the Sigmod activation function to obtain the feature map weight and performs weighted calculation with the original feature map. Then, the IR module is used to extract and reduce the dimension of the B-ultrasound image features. The IR module includes convolution and pooling operations. S3, construct a feature extraction module, which combines the advantages of convolution and self-attention to extract global and local features in B-ultrasound images; S4, construct a feature transformation module, including an upsampling module, a downsampling module, a 3×3 convolution and a ReLU activation function, first downsample the B-ultrasound image features to reduce its resolution, extract the image macro features, then obtain its detail information through upsampling, retain the B-ultrasound image detail information through a combination of upsampling and downsampling, and finally process it through a 3×3 convolution and a ReLU activation function; S5. Construct a B-ultrasound image enhancement model, including input, 3×3 convolution, feature preprocessing module, feature extraction module, feature transformation module, 3×3 convolution, fully connected layer and output; S6, B-ultrasound image processing, obtaining the B-ultrasound image that needs to be enhanced, inputting it into the B-ultrasound image enhancement model, and outputting a high-resolution B-ultrasound image.

2. The method for B-ultrasound image enhancement based on deep learning according to claim 1, characterized in that: In step S1, for the preparation of B-ultrasound image data set, high-resolution B-ultrasound images are collected, and high-resolution B-ultrasound equipment such as high-end color ultrasound machines are used to collect B-ultrasound images in different clinical scenarios. Based on the collected high-resolution B-ultrasound images, a downsampling algorithm is used to convert them into low-resolution B-ultrasound images. The B-ultrasound images collected by the low-resolution acquisition machine are simulated through downsampling technology to generate a pair of high-resolution B-ultrasound images and their corresponding low-resolution B-ultrasound images for training deep learning models, thereby forming a training set for B-ultrasound image enhancement. The training set includes pairs of high-resolution B-ultrasound images and low-resolution B-ultrasound images for training deep learning models.

3. The method for B-ultrasound image enhancement based on deep learning according to claim 1, characterized in that: In step S2, for the feature preprocessing module, the low-resolution B-ultrasound image features are input and set to , , and Respectively represent The height, width and number of channels of Perform spatial attention operations, including: first process the input features through global average pooling and global maximum pooling , generate two feature maps, and obtain the spatial attention map after 1×1 convolution operation , , is the activation function, is a 1×1 convolution, is the global average pooling, is the global maximum pooling, and then the attention map With feature map Performing element-wise multiplication, we get , , represents an element-wise multiplication operation, Indicates input Using Attention Map The result after weighting; The IR module consists of a 3×3 convolutional layer and pooling operations for feature extraction and dimensionality reduction. , the feature map after being processed by the IR module has a reduced spatial resolution and an increased number of channels: , , and They are The height, width and number of channels of It is the B-ultrasound image feature output by the feature preprocessing module.

4. The method for B-ultrasound image enhancement based on deep learning according to claim 1, characterized in that: In step S3, the feature extraction module inputs the preprocessed B-ultrasound image features obtained by the feature preprocessing module. , , , and Respectively The feature extraction module combines the advantages of convolution and self-attention to extract local and global features in B-ultrasound images. The steps include: S31, use two 1×1 convolutional layers to process separately , generating two projection feature maps and , , , Focus on low-level color and texture features, Focus on high-order edge and shape features. and Concatenate in the channel dimension to form the fusion feature F, , Represents a splicing operation, , F integrates the low-order features and high-order features of B-ultrasound image features; S32. Generate Query, Key and Value using 1×1 convolution from the fused feature F. Query, Key and Value are used for self-attention operation. Query, Key and Value are recorded as Q, K and V respectively. , , , , is a 1×1 convolution, and then generates the attention weight O, , is the activation function, is the transposed matrix of K, is the scaling factor; finally, the output B-ultrasound image features are generated through residual connection and 1×1 convolution: , is a 1×1 convolution, , , and They are The height, width and number of channels of It is the B-ultrasound image features output by the feature extraction module.

5. The method for B-ultrasound image enhancement based on deep learning according to claim 1, characterized in that: In step S4, the feature conversion module inputs the B-ultrasound image features after the feature extraction module. , , , and Respectively The height, width and number of channels of B-ultrasound image are firstly transformed through convolution and pooling operations. Downsampling is performed to capture the global structural information of the B-ultrasound image and obtain low-resolution B-ultrasound image features : , represents the downsampling operation, , , and They are The height, width and number of channels of the image are calculated; then the low-resolution features are restored to high resolution through transposed convolution upsampling to obtain more refined local information of B-ultrasound image features: , represents the upsampling operation, , , and Respectively The height, width and number of channels of the original feature map and multi-scale features Fusion, generating feature representations that contain both details and global structures: , Represents a splicing operation, ; Use 3×3 convolution on the feature map Perform transformations to learn complex relationships between features, and use the ReLU activation function to introduce nonlinearity after the convolution operation: , , , and They are The height, width and number of channels of It is the B-ultrasound image feature output by the feature transformation module.

Images