Focus segmentation lightweight method applied to mammary gland medical detection image

Abstract

The invention discloses a lesion segmentation lightweight method applied to a mammary gland medical detection image, and relates to the technical field of medical image analysis. The method comprises the following specific steps: (1) acquiring a mammary gland medical detection image data set, performing preprocessing operations such as size standardization and data enhancement on an image, and dividing the image into a training set and a verification set; (2) a lightweight medical image segmentation model of a U-shaped coding and decoding architecture is constructed based on deep learning, an encoder of the model adopts an axial depth separable convolution block, and a decoder integrates a hierarchical scale perception fusion block; and (3) inputting the preprocessed training set image into the model, and training the constructed lightweight segmentation model. And (4) inputting the verification set into the trained model, evaluating segmentation precision through indexes, and adjusting and optimizing hyper-parameters according to a result to obtain a verified model. And (5) carrying out preprocessing such as size normalization and noise suppression on the to-be-segmented breast medical detection image. And (6) inputting the preprocessed image into the verified lightweight model, and outputting a pixel-level focus segmentation result to assist clinical diagnosis. According to the method, the model parameter quantity and computing resource requirements are remarkably reduced through lightweight architecture design, the reasoning speed is increased while the segmentation precision is optimized, the method is suitable for application scenes with limited resources, and efficient technical support is provided for rapid and accurate diagnosis of breast cancer lesions.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing and computer vision, and particularly relates to a lightweight method for lesion segmentation of breast medical detection images. Background Art

[0002] Early and accurate diagnosis is crucial for the treatment and prognosis of breast cancer. Breast ultrasound examination has become one of the common means for breast disease screening and diagnosis due to its advantages such as radiation-free, simple operation, low cost, and real-time imaging. Ultrasonic medical detection images can clearly show the hierarchical structure of breast tissue and the morphology of lesions, providing important diagnostic basis for doctors.

[0003] In the field of lesion segmentation of breast medical detection images, the existing technologies mainly rely on deep learning algorithms. UNet based on conventional convolution and its variant networks are commonly used segmentation models. By stacking multiple layers of small convolution kernels, these networks can extract local features of images, but the network structure is complex, the number of parameters is large, and the computational efficiency is low, making it difficult to meet the requirements of real-time segmentation. Models based on the attention mechanism aim to enhance the network's ability to capture key regions and improve the segmentation accuracy. However, the large number of computational operations introduced make it difficult for the model to run on devices with limited computational resources, such as being difficult to deploy on portable ultrasound devices or mobile auxiliary diagnosis platforms.

[0004] In addition, although a variety of lightweight neural network structures have been proposed, which can effectively reduce the computational amount and the number of parameters of the model, when directly applied to the lesion segmentation of breast medical detection images, due to the characteristics of uneven imaging quality, blurred lesion boundaries, and large noise interference in breast medical detection images, lightweight models often have difficulty accurately capturing the subtle features and complex structures of lesions, resulting in a significant decrease in segmentation accuracy.

[0005] Therefore, in actual clinical applications, especially in scenarios with limited resources, the existing lesion segmentation technologies for breast medical detection images face the problem of difficult balance between performance and efficiency. How to achieve the lightweight of the lesion segmentation model for breast medical detection images on the premise of ensuring the segmentation accuracy, so that it can run quickly and accurately on various terminal devices has become a key problem to be solved urgently. Summary of the Invention

[0006] In view of this, the present invention provides a lightweight method for lesion segmentation of breast medical detection images.

[0007] The purpose of the present invention is to reduce the computational amount and the number of parameters of the model on the premise of ensuring the segmentation accuracy by improving the feature extraction and multi-scale fusion mechanism, making it suitable for resource-limited devices, assisting doctors to quickly and accurately identify lesions, and improving the efficiency and accuracy of early breast cancer diagnosis.

[0008] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0009] 1. Obtain a breast medical detection image dataset, perform preprocessing operations such as size standardization and data augmentation on the images, and divide them into a training set and a validation set; the medical images used include breast medical detection images of the BUSI dataset, and the dataset is divided into a training set (for model parameter learning) and a validation set (for evaluating the model's generalization ability) according to an 8:2 ratio, ensuring balanced data distribution and covering different lesion sizes, shapes, and ultrasound imaging artifact scenarios.

[0010] 2. Build a lightweight medical image segmentation model based on deep learning with a U-shaped encoder-decoder architecture. The model consists of an encoder, a decoder, and skip connections with skip connections. The encoder and decoder are each divided into five levels. The encoder of this model uses an axial depthwise separable convolution module, and the decoder integrates a hierarchical scale-aware fusion unit.

[0011] (1) The encoder uses an axial depthwise separable convolution module, decomposes the standard 2D convolution into depthwise separable convolutions in the horizontal and vertical axes, and combines pointwise convolutions to achieve long-distance feature extraction.

[0012] (2) The decoder integrates a hierarchical scale-aware fusion unit, extracts local and global features through 3×3 and 5×5 multi-scale convolutions, and uses a dual-channel attention mechanism to dynamically calculate feature weights to achieve adaptive fusion of multi-scale information.

[0013] (3) Set up a cross-layer connection structure to maintain the coherence of the feature channels at the corresponding levels of the encoder and decoder, thereby improving the feature loss situation that is prone to occur in deep networks and enhancing the segmentation accuracy.

[0014] 3. Input the preprocessed training set images into the model to train the constructed lightweight segmentation model.

[0015] (1) First, in the encoding stage, by virtue of the characteristics of the axial depthwise separable convolution module, integrate the functions of the hierarchical scale-aware fusion module in the decoding stage, expand the image feature perception range, and effectively obtain the overall semantic information of breast lesions and cross-region spatial association features.

[0016] (2) Then, introduce a cross-entropy loss calculation mechanism, and by comparing the segmentation output generated by the model with the pre-annotated mask data, accurately measure the difference degree between the two, and intuitively present the segmentation error in numerical form.

[0017] (3) Finally, configure the SGD optimization algorithm, set the initial learning rate to 0.01, and combine it with the polynomial decay rule to cyclically adjust the weight parameters during model training. Continuously perform 300 rounds of iterative training until the value of the loss function stabilizes, so that the model can accurately extract breast lesion features and ensure good performance in terms of accuracy and stability of the segmentation results.

[0018] 4. Input the validation set into the trained model, evaluate the segmentation accuracy through metrics, and optimize the iterative hyperparameters according to the results to obtain a validated model.

[0019] (1) First, send the images in the validation dataset into the trained model and perform forward operations along the encoder-decoder architecture. Use the axial depthwise separable convolution module to obtain multi-scale global features, and through the hierarchical scale-aware fusion component, realize the dynamic integration of local details and overall information of the image, and then output a pixel-level segmentation result map.

[0020] (2) Then, adopt evaluation criteria such as intersection over union (IoU) and F1 score to comprehensively quantify and analyze the model performance from dimensions such as the accuracy of breast lesion boundary recognition, the efficiency of capturing small targets, and the overall matching degree of the segmentation results.

[0021] (3) Finally, according to the evaluation data obtained during the validation process, make targeted adjustments to the model hyperparameters, continuously optimize the model performance, and ensure its stable and efficient processing ability in the task of breast medical detection image lesion segmentation.

[0022] 5. Preprocess the breast medical detection images to be segmented.

[0023] (1) First, perform normalization processing on the breast medical detection images. To meet the input requirements of the model network architecture, use cropping or scaling means to uniformly standardize the image size to 256×256 pixels to ensure that the image size matches the model. Then carry out image denoising work. Considering the characteristic of speckle noise in ultrasound images, select the median filtering algorithm or normalization processing method to effectively suppress noise interference and improve the image quality.

[0024] (2) Finally, perform adaptive conversion of the image channels. According to the input standard of the model, convert the image into a single-channel grayscale mode or a three-channel form to eliminate the possible influence of channel dimension differences on model inference and provide standardized image data input for the subsequent lesion segmentation task.

[0025] 6. Input the preprocessed images into the validated lightweight model to output pixel-level lesion segmentation results to assist clinical diagnosis.

[0026] As can be seen from the above technical solutions, the present invention provides a lightweight method for lesion segmentation in breast medical detection images, which has the following advantages compared with the prior art:

[0027] (1) Lightweight architecture design: By using the axial depthwise separable convolution module to decompose the standard convolution into horizontal, vertical, and axial operations, combined with pointwise convolution, it significantly reduces the model parameters and computational complexity while expanding the feature perception range, adapts to resource-constrained scenarios such as mobile devices and edge devices, solves the problem of difficult deployment of traditional convolutional networks on devices with limited computing resources, and improves the engineering practicability of the model. It can assist doctors in quickly locating lesions, especially suitable for real-time ultrasound diagnosis scenarios in early breast cancer screening.

[0028] (2) Multi-scale feature dynamic fusion: The decoder uses a hierarchical scale-aware fusion module to extract fine-grained details and global context through multi-scale convolution, and uses a two-channel attention mechanism to dynamically calculate adaptive weights, effectively balancing the importance of features at different scales. It realizes refined feature extraction for targets of different sizes in the image, effectively balances the fusion weights of local details and global context information, enhances the model's segmentation ability for fuzzy boundaries and small lesions in medical images, and significantly improves the segmentation accuracy in complex backgrounds.

[0029] (3) Efficient global feature modeling: Without relying on high-complexity self-attention mechanisms, it breaks through the local receptive field limitation of traditional convolutions through an improved convolution block design, and realizes the modeling of long-range spatial dependence relationships at a lower computational cost, solves the problem of insufficient global feature capture in traditional convolutional networks, and enables the model to more accurately understand the spatial association between lesions and surrounding tissues. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flowchart of a lightweight method for lesion segmentation in breast medical detection images provided by the present invention;

[0031] Figure 2 It is a structural diagram of a lightweight medical image segmentation model based on a U-shaped encoder-decoder architecture constructed by deep learning provided by the present invention;

[0032] Figure 3 It is a detailed diagram of the encoder of a lightweight medical image segmentation model based on a U-shaped encoder-decoder architecture constructed by deep learning provided by the present invention;

[0033] Figure 4 It is a detailed diagram of the decoder of a lightweight medical image segmentation model based on a U-shaped encoder-decoder architecture constructed by deep learning provided by the present invention;

[0034] Figure 5 It is a detailed diagram of the module of a lightweight medical image segmentation model based on a U-shaped encoder-decoder architecture constructed by deep learning provided by the present invention. Specific Embodiment

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0036] The process of the present invention is as Figure 1 shown. The present invention discloses a lightweight method for lesion segmentation of breast medical detection images. The specific implementation steps are as follows:

[0037] 1. Obtain a breast medical detection image dataset, perform preprocessing operations such as image size normalization and data augmentation on the images, and divide them into a training set and a validation set; the medical images used include breast medical detection images of the BUSI dataset, and the dataset is divided into a training set (for model parameter learning) and a validation set (for evaluating the model generalization ability) according to the ratio of 8:2 to ensure balanced data distribution and cover different lesion sizes, shapes, and ultrasound imaging artifact scenarios.

[0038] 2. Build a lightweight medical image segmentation model based on deep learning with a U-shaped encoder-decoder architecture. Referring to Figure 2 shown, the model consists of an encoder, a decoder, and skip connections with skip connections. The encoder and decoder are each divided into five levels. The encoder of this model uses an axial depthwise separable convolution module, and the decoder integrates a hierarchical scale-aware fusion unit.

[0039] (1) Stable feature representation provides a good start for subsequent complex operations, ensures that the network can effectively capture long-range dependencies and perform multi-scale feature fusion, thereby improving the accuracy of medical image segmentation. Therefore, at the top of the encoder, a 3×3 convolution block (stride of 1, padding of 1) is applied, followed by GELU activation and batch normalization to extract initial features and maintain consistency with skip connections, laying the foundation for subsequent operations.

[0040] (2) Referring to Figure 3As shown, the encoder uses an Axial Separable Spatial Convolution (ASSC) block. Traditional convolution-based networks have difficulty capturing long-range dependencies due to their limited receptive fields, while Transformer-based methods introduce quadratic complexity. To address this issue, the ASSC module decomposes the standard 2D convolution into two orthogonal depth convolutions along the horizontal and vertical axes, effectively expanding the receptive field while maintaining computational efficiency. Different from traditional large-kernel convolutions that significantly increase the computational overhead, the ASSC module uses depthwise separable convolutions, retaining the ability to model long-range interactions while reducing the number of parameters. In addition, the separable convolution structure enhances the anisotropic feature extraction ability, enabling it to more effectively capture structural details in medical images, such as slender anatomical structures.

[0041] Each axial depth separable convolution block consists of two depthwise separable convolutions (applied along each spatial axis respectively) and a pointwise convolution for feature integration. Given the input feature map x, the axial depth separable convolution block processes it as follows:

[0042] X h,w = σ(BN(X * W h + X * W w ))

[0043] X ASSC = X h,w * W p

[0044] where W h and W w are depth convolution kernels applied along the horizontal and vertical axes respectively, and W p is the pointwise convolution kernel. σ represents the GELU activation function, and BN represents batch normalization. The residual connection in each ASSC module ensures stable gradient propagation, preventing feature degradation when the network is deepened. In addition, the combination of separate horizontal and vertical convolutions enhances the feature extraction ability for fine-grained segmentation tasks, which is particularly beneficial in medical imaging applications that require capturing structural continuity.

[0045] (3) In the encoding stage, the downsampling operation is completed. The encoder consists of five hierarchical levels, and each level achieves downsampling through a Max pooling operation. Specifically, after feature extraction, a max pooling layer with a stride of 2 is used, gradually reducing the spatial resolution (e.g., from the input image size to 1 / 2, 1 / 4, 1 / 8, 1 / 16, 1 / 32 of the original size in sequence). Each level contains multiple axial separable spatial convolution (ASSC) modules, residual connections, and channel expansion and contraction mechanisms.

[0046] (4) Refer toFigure 4 As shown, the decoder integrates a Hierarchical Scale-Aware Fusion (HSAF) module. The traditional skip connections in the U-Net architecture concatenate encoder and decoder features of the same resolution, which usually leads to a semantic gap between high-level and low-level features. Additionally, single-scale fusion is difficult to capture context dependencies. To address these challenges, the HSAF module uses an adaptive weighting mechanism to selectively integrate multi-scale features. Different from simple concatenation, the HSAF module ensures better alignment between global and local representations, thereby improving segmentation accuracy, especially for small structures.

[0047] The HSAF module uses depthwise separable convolutions with kernel sizes of 3×3 and 5×5 to extract multi-scale features. The extracted features F 3×3 and F 5×5 are adaptively fused as follows:

[0048] F HSA F = BN(α·F 3×3 +(1 - α)·F 5×5 )

[0049] where BN(·) represents batch normalization for feature stabilization, and α is the adaptive weight that controls the importance of the scale.

[0050] (5) To optimize feature selection across channels, the HSAF module adopts a lightweight attention mechanism through Dual-Pooling Channel Attention (DPCA). As shown in Figure 5 , the DPCA module enhances feature fusion by recalibrating channel importance.

[0051] First, adaptive average pooling (Adaptive Avg Pool) and adaptive max pooling (Adaptive Max Pool) operations are respectively performed on the input feature map to obtain two different pooling results. Then, these two pooling results are concatenated to form a combined feature representation. Subsequently, two pointwise convolution (Pointwise Conv) operations are sequentially performed on the concatenated feature representation. The GELU activation function is applied after the first pointwise convolution, and the Sigmoid activation function is applied after the second pointwise convolution to obtain the weight values for adjusting the features. Finally, element-wise scaling is performed on the original input feature map using these weight values to complete the recalibration of the features, and the processed feature map is output. The processing is as follows:

[0052] Concat = [AAP(F), AMP(F)]

[0053] X1 = GELU(Conv pointwise (Concat))

[0054] α = σ(Conv pointwise (X1))

[0055] F DPCA = α · F

[0056] Among them, F represents the feature map input to the DPCA module; AAP(F) is the result of performing adaptive average pooling on F; AMP(F) is the result of performing adaptive max pooling on F; Concat is the feature after concatenating the two pooling results; Conv pointwise represents a pointwise convolution operation; X1 is the feature after the first pointwise convolution and activation by GELU; α is the weight value obtained after the second pointwise convolution and activation by the Sigmoid activation function (σ); F DPCA is the output feature map after element-wise scaling, and (·) represents element-wise multiplication.

[0057] (6) The skip connection preserves the feature channels of the same level between the encoder and the decoder, alleviating the problem of feature degradation in deep networks.

[0058] (7) To restore the spatial resolution, an upsampling operation is performed in the decoding stage. First, bilinear interpolation is applied to the input feature map (the output F HSAF ) from the HSAF module to expand the spatial resolution to the target size (such as restoring from H / 16×W / 16 to H / 8×W / 8). Then, the interpolated feature map is adjusted in the number of channels and optimized in features through a 1×1 convolution. Finally, the output feature map F upsample of the upsampling operation is obtained, and the specific formula is as follows:

[0059] F upsample = σ(BN(Interp(F HSAF , mode = bilinear) * W 1×1 ))

[0060] Among them, Interp(·, mode = bilinear) is the bilinear interpolation operation. W 1×1 is the weight matrix of the 1×1 convolution, which is used to adjust the number of channels and optimize the features. At the same time, batch normalization (BN) and the GELU activation function (σ) are introduced to ensure the stability and non-linear expression of the features.

[0061] 3. Input the preprocessed training set images into the model to train the constructed lightweight segmentation model.

[0062] (1) First, the image is downsampled step by step through the axial depthwise separable convolution blocks in the encoder. The axial depthwise separable convolution is used to expand the receptive field, capture the global context information of breast lesions, and long-range spatial dependence relationships.

[0063] (2) Subsequently, the cross-entropy loss function is adopted to calculate the difference between the segmentation result output by the model and the annotated mask, and quantify the segmentation error.

[0064] (3) Finally, the SGD optimizer is used, with the initial learning rate set to 0.01 and combined with the polynomial decay strategy. During the training process, the model weight parameters are iteratively updated. After 300 iteration cycles until the loss function converges, the model can accurately represent the features of breast lesions, ensuring the accuracy and robustness of the segmentation result.

[0065] 4. Input the validation set into the trained model, evaluate the segmentation accuracy through metrics, and optimize the iterative hyperparameters according to the results to obtain the validated model.

[0066] (1) First, forward inference is performed on the image through the encoder-decoder path. After extracting multi-scale global features through the axial depthwise separable convolution blocks and dynamically fusing local and global information through the hierarchical scale-aware fusion blocks, a pixel-level segmentation probability map is generated.

[0067] (2) Subsequently, based on quantization metrics such as intersection over union (IoU) and F1 score, evaluate the accuracy of the model in localizing the boundaries of breast lesions, the detection ability of small targets, and the overall segmentation consistency; finally, adjust the hyperparameters according to the evaluation results of the validation set to optimize the model's capabilities and ensure that the model maintains stable segmentation performance.

[0068] 5. Preprocess the breast medical detection images to be segmented.

[0069] (1) First, perform size normalization. Adjust the image to the preset input size of 256×256 pixels for the model through cropping or scaling operations to ensure that the image size matches the model network structure.

[0070] (2) Secondly, perform noise suppression. For the specific speckle noise in ultrasound images, median filtering or normalization operations are used to reduce noise interference and improve the image quality; finally, perform channel adaptation. According to the model input requirements, convert the image into a single-channel grayscale image or a three-channel format to eliminate the influence of channel dimension differences on model inference and provide standardized input data for subsequent segmentation tasks.

[0071] 6. Input the preprocessed image into the validated lightweight model to output the pixel-level lesion segmentation result to assist clinical diagnosis.

[0072] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A lightweight method for lesion segmentation in breast medical detection images, characterized in that, It includes the following specific operation processes: (1) Collect the breast medical detection image data set, perform preliminary processing work such as size regularization and data augmentation on the obtained images, and then divide these image data into a training data group and a validation data group. (2) Construct a lightweight U-shaped encoder-decoder deep learning model suitable for medical image segmentation, where the encoder applies an axial depthwise separable convolution module, and the decoder integrates a hierarchical scale-aware fusion unit. (3) In the model training session, sequentially import the images in the pre-processed training data set into the constructed lightweight segmentation model for training. (4) After the model training is completed, send the validation data set into the model, detect the segmentation accuracy of the model with various evaluation metrics, and optimize and iterate the hyperparameters according to the evaluation results to obtain a validated available model. (5) For breast medical detection images that need to perform lesion segmentation, perform preprocessing operations such as size regularization and noise elimination. (6) Input the preprocessed images into the validated lightweight model, and the model outputs a lesion segmentation result accurate to the pixel level, providing strong data support for clinical diagnosis work.

2. The lightweight method for lesion segmentation of mammary medical detection images according to claim 1, characterized in that It is characterized in that the medical image data used comes from the BUSI breast medical detection image data set. The data set is divided into a training data subset and a validation data subset according to a ratio of 8:2, where the training subset is used for learning and optimizing model parameters, and the validation subset is used for evaluating the generalization performance of the model.

3. The method for lightweight lesion segmentation of breast medical detection images according to claim 1, characterized in that In step (2), a U-shaped encoder-decoder network architecture based on deep learning is constructed to achieve lightweight medical image segmentation. The model adopts a symmetric five-level encoder-decoder structure and realizes the fusion of low-level features and high-level features through a skip connection mechanism.

4. The lightweight method for lesion segmentation of mammary medical detection images according to claim 1, characterized in that In step (2), the encoder uses an axial depthwise separable convolution module to disassemble the conventional two-dimensional convolution into a depthwise separable convolution form along the horizontal and vertical directions, and then cooperates with a pointwise convolution operation to effectively extract long-distance features of the image. The decoder integrates a hierarchical scale-aware fusion unit, and uses convolution kernels of different sizes of 3×3 and 5×5 to simultaneously extract local details and overall features of the image. And a dual-path attention architecture is introduced to dynamically allocate weights according to the feature characteristics, promoting the integration of multi-scale information as needed. In addition, a cross-layer connection structure is set up to maintain the coherence of the feature channels of the corresponding levels of the encoder and decoder, thereby improving the feature loss situation that is prone to occur in the deep network and enhancing the segmentation accuracy.

5. According to claim 1, a lightweight method for lesion segmentation of breast medical detection images, characterized in that, In the implementation process of step (3) above, the image data in the pre-divided training data group are sequentially imported into the constructed model to carry out model training work. First, in the encoding stage, the characteristics of the axial depth-separable convolution module are used, and in the decoding stage, the functions of the hierarchical scale-aware fusion module are integrated to expand the image feature perception range and effectively obtain the overall semantic information and cross-regional spatial correlation features of breast lesions. Then, the cross-entropy loss calculation mechanism is introduced to accurately measure the difference between the segmentation output generated by the model and the pre-labeled mask data, and the segmentation error is intuitively presented in numerical form. Finally, the SGD optimization algorithm is configured, the initial learning rate is set to 0.01, and the weight parameters are cyclically adjusted during model training with the polynomial attenuation rule. Continuous 300 rounds of iterative training are carried out until the loss function value tends to be stable, so that the model can accurately extract the characteristics of breast lesions and ensure the good performance of segmentation results in accuracy and stability.

6. According to claim 1, a lightweight method for lesion segmentation of breast medical detection images, characterized in that In step (4), the specific implementation method for the verification phase operation is as follows: First, the images in the validation dataset are fed into the trained model and forward operations are performed along the encoder-decoder architecture. The axial depth-separable convolution module is used to obtain global features at multiple scales, and the hierarchical scale-aware fusion component is used to dynamically integrate local details and overall information of the image, and then output the pixel-level segmentation result map. Afterwards, the model performance is comprehensively quantitatively analyzed from the dimensions of breast lesion boundary recognition accuracy, small target capture efficiency, and overall matching of segmentation results using evaluation criteria such as intersection over union (IoU) and F1 score. Finally, based on the evaluation data obtained during the validation process, the model hyperparameters are adjusted in a targeted manner to continuously optimize the model performance and ensure its stable and efficient processing capabilities in the task of lesion segmentation in breast medical detection images.

7. According to claim 1, a lightweight method for lesion segmentation of breast medical detection images, characterized in that In the step (5), standardized preliminary processing is carried out for the breast medical examination images that need to be processed for lesion segmentation. First, the breast medical examination images are normalized. In order to meet the input requirements of the model network architecture, the image size is standardized to 256×256 pixels by cropping or scaling to ensure that the image size is compatible with the model. Then, image denoising is carried out. In view of the characteristics of speckle noise in ultrasound images, the median filtering algorithm or normalization processing method is used to effectively suppress noise interference and improve image quality. Finally, the image channel is adaptively converted. According to the input standard of the model, the image is converted into a single-channel grayscale mode or a three-channel form to eliminate the possible impact of channel dimension differences on model reasoning, and provide standardized image data input for subsequent lesion segmentation tasks.

Images