A breast ultrasound image segmentation method based on diagnostic information perception

Abstract

This invention discloses a breast ultrasound image segmentation method based on diagnostic information perception, comprising: a diagnostic information perception-based image segmentation model; acquiring breast ultrasound image data and performing preprocessing operations; inputting the preprocessed image data into the encoder of the diagnostic information perception-based image segmentation model for multi-level image feature extraction, and finally outputting a bottleneck feature map; constructing an image-level diagnostic semantic vector based on the bottleneck feature map; dynamically modulating the decoding features of each decoding layer in the decoder of the diagnostic information perception-based image segmentation model using the image-level diagnostic semantic vector corresponding to the decoding features in each decoding layer; modeling the uncertainty of the segmentation result during model training to achieve uncertainty constraints; implementing differentiated segmentation constraint strategies for normal image data; and inputting the feature map output by the decoder into the segmentation head for segmentation processing to generate the final segmentation result of the breast ultrasound image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to a method for segmenting breast ultrasound images based on diagnostic information perception. Background Technology

[0002] Deep learning-based breast ultrasound image segmentation methods have been extensively studied. The most similar implementation to this invention typically employs a convolutional neural network encoder-decoder structure to perform end-to-end segmentation modeling of ultrasound images, implicitly assuming the presence of lesion regions in each input image, and constructing a segmentation loss function based on region overlap metrics. However, in real-world clinical scenarios, many breast ultrasound images are normal samples and do not contain lesion regions. Without effective modeling of the characteristics of normal samples, the aforementioned implementations are prone to generating false lesion segmentation results in normal images, leading to false positive predictions and affecting the reliability of the model in clinical applications. On the other hand, breast ultrasound examinations are usually accompanied by clear image-level diagnostic conclusions, such as normal, benign, or malignant. Some existing technologies optimize this diagnostic information in parallel with the segmentation task as an auxiliary task, but the diagnostic semantics are usually not explicitly introduced into the segmentation decoding process, making it difficult to effectively guide lesion localization at the pixel-level segmentation stage. Furthermore, the imaging uncertainties caused by speckle noise and echo attenuation in ultrasound images result in fluctuations in pixel-level prediction results. Existing research focuses more on the estimation or analysis of uncertainty, and rarely introduces uncertainty factors directly into the segmentation optimization objective to suppress noise interference and reduce false positive predictions in normal samples.

[0003] Existing breast ultrasound image segmentation techniques generally assume the presence of lesion areas in the input image, making it difficult to effectively process normal samples and prone to false positive predictions. Diagnostic semantic information is not explicitly introduced into the segmentation and decoding process, limiting its guiding role in lesion localization. At the same time, ultrasound imaging has high uncertainty, and existing methods utilize uncertainty in a relatively simple way, still falling short in suppressing noise and improving segmentation stability. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a breast ultrasound image segmentation method based on diagnostic information perception. The method introduces image-level diagnostic semantic information into the decoder to adjust the decoding features, and models the uncertainty of the output segmentation result in the model training, introducing the uncertainty information into the segmentation optimization process to improve the stability of the segmentation result in noisy and blurred boundary regions.

[0005] Technical Solution: To achieve the above objectives, the present invention provides a breast ultrasound image segmentation method based on diagnostic information perception, comprising: an image segmentation model based on diagnostic information perception; acquiring breast ultrasound image data to be processed and performing preprocessing operations on the image data; inputting the preprocessed image data into the encoder of the image segmentation model based on diagnostic information perception, extracting multi-level image features from the preprocessed image data through the encoder, and finally outputting a bottleneck feature map; constructing an image-level diagnostic semantic vector based on the bottleneck feature map output by the encoder; dynamically modulating the decoding features of the decoding layer by using the image-level diagnostic semantic vector corresponding to the decoding features in each decoding layer of the image segmentation model based on diagnostic information perception; modeling the uncertainty of the segmentation result during the training process of the image segmentation model based on diagnostic information perception, introducing the uncertainty information into the segmentation optimization process, and realizing uncertainty constraint on the segmentation result; and implementing a differentiated segmentation constraint strategy for normal image data during the model training process; inputting the feature map output by the decoder into the segmentation head for segmentation processing, generating the final segmentation result of the breast ultrasound image, and outputting a pixel-level segmentation map of the lesion region.

[0006] Furthermore, the bottleneck feature map output by the encoder is obtained, and a global average pooling operation is performed on the bottleneck feature map to obtain channel-level statistical features; the channel-level statistical features are then expanded to form an image-level diagnostic semantic vector; let the bottleneck feature map be F. b ∈R C×H×W Where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map, the global average pooling operation process is as follows:

[0007]

[0008] In the formula, F b (c, i, j) represents the feature value of the bottleneck feature map at the c-th channel and spatial location (i, j); z c This is represented as the channel-level statistical feature corresponding to the c-th channel.

[0009] Furthermore, image-level diagnostic semantic vector representations are introduced into the decoder as conditional information, and the decoding features in each decoding layer are dynamically modulated layer by layer, enabling the segmentation features to perceive the overall diagnostic context during the decoding process; this includes the following steps:

[0010] Step 1: Input the image-level diagnostic semantic vector z corresponding to the decoding features of the l-th decoding layer into the parameter mapping module, and perform feature mapping to generate the intermediate semantic features z1 required for modulation;

[0011] Step 2: Generate the channel-level scaling parameter γ based on the intermediate semantic features and the decoding features corresponding to the l-th decoding layer in the decoder.l and bias parameter β l ;

[0012] Step 3, Channel-level scaling parameter γ l and bias parameter β l Apply a nonlinear activation function and limit the channel-level scaling parameter γ l and bias parameter β l The range of values ​​for;

[0013] Step 4: Use the channel-level scaling parameter γ corresponding to the decoding feature of the l-th decoding layer. l and bias parameter β l For the decoding feature F corresponding to the l-th decoding layer in the decoder l Channel-by-channel dynamic modulation is performed to obtain the modulated decoded feature map.

[0014] Furthermore, in step 2, the channel-level scaling parameter γ l and bias parameter β l The calculation process is as follows:

[0015]

[0016] In the formula, , Both are represented as parameter mapping functions.

[0017] Furthermore, in step 4, the dynamic modulation decoding feature process is as follows:

[0018]

[0019] In the formula, This is represented as a channel-by-channel multiplication operation. This is represented as the decoded feature after modulation.

[0020] Furthermore, breast ultrasound image data to be processed is acquired, and preprocessing is performed on the image data; a training sample set is constructed based on several preprocessed image data sets, and the image segmentation model based on diagnostic information perception is trained using the training sample set; during the model training process, the uncertainty of the output segmentation result is modeled, including the following steps:

[0021] Step 1-1: The feature map output by the decoder is input into the segmentation head for segmentation processing to obtain pixel-level segmentation results. At the same time, the pixel-level uncertainty information corresponding to the pixel-level segmentation results is predicted simultaneously.

[0022] Steps 1-2: Based on the element-level uncertainty information, model the prediction uncertainty level of each pixel position to obtain a pixel-level prediction uncertainty representation;

[0023] Steps 1-3: During model training, pixel-level prediction uncertainty representation is introduced into the segmentation loss function to construct a weighted segmentation loss function. The weighted segmentation loss function is used to adaptively weight the loss term in the high uncertainty region of the image data during model training.

[0024] Furthermore, in steps 1-3, let the pixel-level binary classification cross-entropy loss be the pixel-level basic segmentation loss L. base The weighted segmentation loss function L, after introducing pixel-level prediction uncertainty representation, is defined as:

[0025]

[0026]

[0027] In the formula, w(i,j) represents the adaptive weight for each pixel position. This can be represented as the average of the weighted loss function over all pixels. This is represented as pixel-by-pixel multiplication. This represents the pixel-level variance at each pixel location. It is represented as a minimum value.

[0028] Beneficial Effects: This invention provides a breast ultrasound image segmentation method based on diagnostic information perception. It extracts image-level diagnostic semantic information from breast ultrasound images and uses it as a global prior to guide the segmentation network in adaptive adjustments during feature recovery, thereby improving the segmentation accuracy of lesion regions. By utilizing diagnostic semantic information to weight, scale, or bias-modulate intermediate features in the decoding stage, the segmentation features can perceive the overall diagnostic state, thus enhancing lesion-related features and suppressing irrelevant or abnormal responses. During segmentation prediction, a pixel-level uncertainty level is simultaneously modeled and introduced into the segmentation loss function. Unreliable prediction regions are adaptively weighted to improve the stability of the segmentation results in noisy and blurred boundary regions. Combining diagnostic category information and pixel-level uncertainty, a differentiated segmentation constraint strategy is applied to normal image data without lesions, effectively suppressing abnormal segmentation responses in normal regions and reducing false-positive segmentation results. The proposed method is independent of specific network structures and can be embedded in different encoder-decoder medical image segmentation networks, exhibiting good versatility and scalability. Attached Figure Description

[0029] Figure 1 This is a flowchart of a breast ultrasound image segmentation method based on diagnostic information perception.

[0030] Figure 2 A schematic diagram of the structure of an image segmentation model based on diagnostic information perception;

[0031] Figure 3 A qualitative comparison chart of segmentation results from different segmentation models. Detailed Implementation

[0032] The invention will now be further described with reference to the accompanying drawings.

[0033] like Figure 1 As shown, a breast ultrasound image segmentation method based on diagnostic information perception includes: a diagnostic information perception-based image segmentation model; acquiring breast ultrasound image data to be processed and performing preprocessing operations on the image data; inputting the preprocessed image data into the encoder of the diagnostic information perception-based image segmentation model, extracting multi-level image features from the preprocessed image data through the encoder, and finally outputting a bottleneck feature map; constructing an image-level diagnostic semantic vector based on the bottleneck feature map output by the encoder; dynamically modulating the decoding features of the decoding layer by using the image-level diagnostic semantic vector corresponding to the decoding features in each decoding layer of the diagnostic information perception-based image segmentation model; modeling the uncertainty of the segmentation result during the training process of the diagnostic information perception-based image segmentation model, introducing the uncertainty information into the segmentation optimization process, and realizing uncertainty constraint on the segmentation result; and implementing a differentiated segmentation constraint strategy for normal image data during the model training process; inputting the feature map output by the decoder into the segmentation head for segmentation processing, generating the final segmentation result of the breast ultrasound image, and outputting a pixel-level segmentation map of the lesion region.

[0034] like Figure 2 As shown, the image segmentation model based on diagnostic information perception includes an encoder, a diagnostic semantic module, a feature modulation module, a decoder, and an output head. The encoder takes the image data to be processed as input, passes through convolutional layers and undergoes four downsampling operations, and outputs five feature maps at different scales, representing multi-level image features. The feature map output by the last downsampling operation is the bottleneck feature map, which is the feature map with the lowest resolution and strongest semantics output from the last encoding layer of the encoder. The input of the diagnostic semantic module is connected to the last layer of the encoder. The bottleneck feature map input to the diagnostic semantic module is sequentially processed through global average pooling and flattening operations to obtain an image-level diagnostic semantic vector. The input of the feature modulation module is connected to the output of the diagnostic semantic module, and four sets of channel-level scaling parameters γ are generated based on the image-level diagnostic semantic vector. l and bias parameter β l These correspond to the four decoding layers in the decoder. The decoder comprises four decoding layers. Each decoding layer sequentially performs upsampling, and then performs skip connections and convolution operations with the feature maps output from the encoder's corresponding encoding layer. Finally, the decoded features output from each decoding layer are compared with their corresponding channel-level scaling parameter γ. l and bias parameter βl A modulation operation is performed, and the modulated decoded features are used as input to the next decoding layer; the output of the last decoding layer is used as the feature map output by the decoder. The output head includes a segmentation head, an uncertainty head, and a classification head; the post-decoding feature map is input to the segmentation head, uncertainty head, and classification head respectively, and the pixel-level segmentation map of the lesion region, pixel-level uncertainty information, and image-level diagnostic classification result are output respectively.

[0035] The process involves acquiring breast ultrasound image data to be processed and performing preprocessing operations on the image data. The preprocessing operations include adjusting the images to a fixed size and grayscale normalization. During the training phase, conventional data augmentation operations such as random horizontal flipping and rotation are performed to ensure that the preprocessed image data has a uniform image format and stable distribution, enabling the encoder to extract global features and bottleneck feature maps more stably, which is beneficial for the subsequent learning and generation of image-level diagnostic semantic vectors.

[0036] The preprocessed image data is input into the encoder of the diagnostic information-aware image segmentation model. The encoder extracts multi-level image features from the preprocessed image data. Specifically, the encoder extracts multi-scale feature representations from shallow to deep layer by layer to obtain multi-level image features. The shallow features are used to represent local texture information, and the deep features are used to represent the high-level semantic information of the image. The encoder in the diagnostic information-aware image segmentation model can adopt any form of feature extraction network, including but not limited to convolutional neural network structure, residual structure, or a combination thereof. The decoder in the diagnostic information-aware image segmentation model can also adopt the decoder corresponding to the encoder in the conventional segmentation model.

[0037] The feature map output by the decoder is input into the segmentation head. First, it passes through a 1×1 convolutional layer to map the high-dimensional features of the decoder output feature map into single-channel segmentation logits. During the inference phase, a Sigmoid activation function is applied to the single-channel segmentation logits, converting them into a probability map in the [0, 1] interval. Then, a threshold is applied for binarization, resulting in a final binary segmentation map, which generates a pixel-level segmentation mask for the lesion region, i.e., the pixel-level segmentation map of the lesion region. The threshold for binarization can be 0.5.

[0038] Obtain the bottleneck feature map output by the encoder, perform global average pooling on the bottleneck feature map to obtain channel-level statistical features; expand the channel-level statistical features to form an image-level diagnostic semantic vector; let the bottleneck feature map be F. b ∈R C×H×W Where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map, the global average pooling operation process is as follows:

[0039]

[0040] In the formula, F b (c, i, j) represents the feature value of the bottleneck feature map at the c-th channel and spatial location (i, j); z c Let z represent the channel-level statistical feature corresponding to the c-th channel; after global average pooling, all channel-level statistical features z are obtained. c The constituent dimension is R C×H×W A tensor of which spatial dimensions H and W are both compressed to 1; the unfolding operation specifically involves expanding the tensor of dimension R... C×H×W The tensor is flattened to preserve the channel dimension and merge the spatial dimensions, i.e., R C×1×1 The channel-level statistical features of the morphology are converted into R C A one-dimensional vector of shape, which is the image-level diagnostic semantic vector z∈R C .

[0041] Image-level diagnostic semantic vector representations are introduced into the decoder as conditional information. The decoding features in each decoding layer are dynamically modulated layer by layer, enabling the segmentation features to perceive the overall diagnostic context during the decoding process. The segmentation results can perceive the overall diagnostic state, improving the rationality of lesion localization. Through a diagnostic-perceived feature modulation mechanism, the response of lesion-related features is enhanced, reducing missegmentation under complex ultrasound texture conditions. The process includes the following steps:

[0042] Step 1: Input the image-level diagnostic semantic vector z corresponding to the decoded features of the l-th decoding layer into the parameter mapping module to perform feature mapping and generate the intermediate semantic features z1 required for modulation. The parameter mapping module performs feature mapping through at least two fully connected layers or 1×1 convolutional layers. The image-level diagnostic semantic vector z corresponding to the decoded features of each decoding layer is the same image-level diagnostic semantic vector z. The parameter mapping module uses two fully connected layers to implement the feature mapping operation. First, the image-level diagnostic semantic vector z is input into the first fully connected layer and processed by the ReLU nonlinear activation function to obtain an intermediate feature with a length twice the number of channels of the first decoding layer. Then, the intermediate feature is input into the second fully connected layer for processing. The second fully connected layer has no activation function and obtains the intermediate semantic feature z1 with a length twice the number of channels of the first decoding layer.

[0043] Step 2: Generate the channel-level scaling parameter γ based on the intermediate semantic features and the decoding features corresponding to the l-th decoding layer in the decoder. l and bias parameter β l The intermediate semantic feature z1 is split into a first half and a second half along the channel dimension. The first half is used as the channel-level scaling parameter γ of the first decoding layer. lThe latter half serves as the bias parameter β for the first decoding layer. l And the channel-level scaling parameter γ l and bias parameter β l The length of each feature channel is exactly the same as the number of feature channels in the first decoding layer.

[0044] Step 3, Channel-level scaling parameter γ l and bias parameter β l Apply a nonlinear activation function and limit the channel-level scaling parameter γ l and bias parameter β l The range of values ​​for the parameter; the nonlinear activation function is the tanh hyperbolic tangent function, which can map any real number input to the interval (−1, 1), and is the core means of limiting the range of parameter values;

[0045] The channel-level scaling parameter γ l The activation and value range limitations are first addressed by adjusting the original channel-level scaling parameter γ obtained from the splitting. l Perform tanh activation, constraining its value range to (−1, 1); then add 1.0 to the activated result, ultimately adjusting the channel-level scaling parameter γ. l The value range of is limited to (0, 2), and the calculation formula is:

[0046] γ l =1.0+tanh(γ l (raw)

[0047] In the formula, (γ) l (raw) represents the raw scaling parameter obtained from the intermediate semantic feature decomposition; this operation makes the channel-level scaling parameter γ l The value fluctuates around 1.0 to avoid excessive amplification or reduction of decoding features due to excessively large or small channel-level scaling parameters, thus ensuring the stability of the modulation process.

[0048] The channel-level bias parameter β l The activation and value range restrictions apply to the original channel-level bias parameter β obtained from the decomposition. l Directly performing tanh activation, without additional offset operations, limits its value range to (−1, 1), and the calculation formula is:

[0049] β l =tanh(β l (raw)

[0050] In the formula, (β) l (raw) represents the raw bias parameter obtained from the intermediate semantic feature decomposition; this operation avoids the channel-level bias parameter β. l The value should not be too large to prevent the overall offset of the decoded features from exceeding a reasonable range.

[0051] The technical significance of limiting the range of values, γ l ∈(0,2) ensures that the scaling of the decoded features is controllable, adjusting only within the range of 0 to 2 times the original features; β l ∈(−1,1) ensures that the bias adjustment amplitude is moderate. The combination of the two enables the dynamic modulation process of the decoding features to both perceive the global diagnostic semantics and avoid feature distortion due to abnormal parameters, ultimately improving the rationality and stability of the segmentation results.

[0052] Step 4: Use the channel-level scaling parameter γ corresponding to the decoding feature of the l-th decoding layer. l and bias parameter β l For the decoding feature F corresponding to the l-th decoding layer in the decoder l Channel-by-channel dynamic modulation is performed to obtain the modulated decoded feature map.

[0053] In step 2, the channel-level scaling parameter γ l and bias parameter β l The calculation process is as follows:

[0054]

[0055] In the formula, , Both are represented as parameter mapping functions.

[0056] In step 4, the dynamic modulation decoding feature process is as follows:

[0057]

[0058] In the formula, This is represented as a channel-by-channel multiplication operation. Represented as the decoded feature after modulation, F l The decoding feature corresponding to the l-th decoding layer is represented as the decoding feature output by the l-th decoding layer. By modulating the decoding features in the decoder, the image-level diagnostic semantic vector is introduced into the expression of the decoding features, thereby enhancing the discriminative features related to diagnosis and suppressing the feature response of irrelevant background regions.

[0059] Acquire breast ultrasound image data to be processed and preprocess the image data; construct a training sample set based on several preprocessed image data sets, and use the training sample set to train an image segmentation model based on diagnostic information perception; during the model training process, model the uncertainty of the output segmentation results, introduce pixel-level uncertainty modeling, and adaptively constrain noise and unreliable prediction regions to improve segmentation stability and robustness; including the following steps:

[0060] Step 1-1: The feature map output by the decoder is input into the segmentation head for segmentation processing to obtain pixel-level segmentation results. Simultaneously, pixel-level uncertainty information corresponding to the pixel-level segmentation results is predicted. The feature map output by the decoder is used as a shared feature base, and the feature map output by the decoder has dimensions [B, C, H, W]. The feature map output by the decoder is simultaneously input into the segmentation head and the uncertainty head in parallel. Both the segmentation head and the uncertainty head are independent 1×1 convolutional layers. The segmentation head outputs a single-channel pixel-level segmentation logits, corresponding to the original predicted value of the pixel-level segmentation result, and finally obtains the pixel-level segmentation map of the lesion region. The uncertainty head synchronously outputs a single-channel pixel-level log-variance log(σ²), which is the pixel-level uncertainty information. The two 1×1 convolutional layers synchronously complete the forward calculation to achieve a one-to-one correspondence between the segmentation prediction at each pixel position and the corresponding uncertainty information, and synchronous output.

[0061] Steps 1-2: Based on the element-level uncertainty information, model the prediction uncertainty level for each pixel location to obtain a pixel-level prediction uncertainty representation; firstly, perform range clipping on the pixel-level uncertainty information log(σ²), calculated as follows:

[0062]

[0063] In the formula, log (σ²) raw This is represented as the original pixel-level uncertainty information, log (σ²). clip This represents the pixel-level uncertainty information after clipping. `clamp` represents an operation that limits a value to a specified minimum and maximum value. Limiting the pixel-level uncertainty information to the range [-6.0, 6.0] avoids numerical explosion, and then applying this to the clipped log(σ²). clip Perform exponential operations to obtain pixel-level variance. The calculation formula is as follows:

[0064]

[0065] In the formula, σ² represents the pixel-level variance of each pixel position, and exp represents the exponential function. The pixel-level variance of each pixel position is calculated by the above formula, and the pixel-level variance σ² of each pixel position is the prediction uncertainty level of that pixel position. The larger σ² is, the less reliable the segmentation prediction is. The pixel-level variance σ² of all pixel positions constitutes a pixel-level prediction uncertainty representation with dimensions [B, 1, H, W].

[0066] Steps 1-3: During model training, pixel-level prediction uncertainty representation is introduced into the segmentation loss function to construct a weighted segmentation loss function. The weighted segmentation loss function is used to adaptively weight the loss term in the high uncertainty region of the image data during model training.

[0067] A new weighted segmentation loss function is constructed by introducing pixel-level uncertainty representation into the segmentation loss function. This weighted loss function is then used to adaptively weight the loss term in high-uncertainty regions during model training. Specifically, the Dice loss or cross-entropy loss is used as the pixel-level basic segmentation loss L. base Based on the obtained pixel-level uncertainty variance σ², adaptive weights are generated for each pixel location. The calculation process is as follows:

[0068]

[0069] In the formula, ε represents a minimum value on the order of 1e-6 to avoid infinite weights; σ²(i,j) represents the pixel-level variance at each pixel position; where σ² is large in high uncertainty regions, the corresponding weight w becomes smaller; and σ² is small in low uncertainty regions, the corresponding weight w becomes larger; then the pixel-level basic segmentation loss is multiplied pixel by pixel with the adaptive weight at each pixel position to obtain the weighted loss function for each pixel; the average of the weighted loss functions of all pixels is calculated to obtain the final weighted segmentation loss, realizing the adaptive weight reduction constraint on the loss term in high uncertainty regions.

[0070] In steps 1-3, let the pixel-level binary classification cross-entropy loss be the pixel-level basic segmentation loss L. base The weighted segmentation loss function L, after introducing pixel-level prediction uncertainty representation, is defined as:

[0071]

[0072] In the formula, w(i,j) represents the adaptive weight for each pixel position. This can be represented as the average of the weighted loss function over all pixels. This is represented as pixel-by-pixel multiplication.

[0073] Pixel-level segmentation labels are constructed based on the existing annotation information on the breast ultrasound images to be processed. The training sample set is then divided into normal and abnormal images based on the presence or absence of lesions. Normal images are those without lesions, while abnormal images are those with lesions. These are used for subsequent model training and differential constraints. A differential segmentation constraint strategy is implemented on normal images to suppress abnormal segmentation responses in lesion-free areas, thereby reducing the interference of normal samples on the model training process and minimizing false positive segmentation results.

[0074] The specific method for implementing differentiated segmentation constraint strategies for normal image data is as follows: During model training, normal image data is first identified based on the annotation information of the training samples, i.e., breast ultrasound images with no lesion annotations and all labels set to 0. For normal image data, during the calculation of the weighted segmentation loss function L, an adaptive weight for each pixel position is superimposed with a background enhancement coefficient α, amplifying the weight adjustment range, to obtain the final weighted segmentation loss function for normal image data; the calculation process is shown below:

[0075]

[0076] In the formula, α represents the background enhancement coefficient, where α>1, for example, a value of 2.0;

[0077] This differentiated segmentation constraint strategy further reduces the loss weight of high-uncertainty regions in normal image data, i.e. regions prone to false positives, by increasing the coefficient α in the differentiated weighting; at the same time, it increases the loss weight of low-uncertainty background regions, specifically suppressing abnormal segmentation responses in lesion-free regions, and ultimately reducing the interference of normal image data on model training and reducing false positive segmentation results.

[0078] Example

[0079] To verify the effectiveness and universality of the breast ultrasound image segmentation method based on diagnostic information perception proposed in this invention, publicly available breast ultrasound image datasets were selected for experimental verification. In the experiments, breast ultrasound image data was input into the image segmentation model after unified preprocessing. While segmenting the lesion region, diagnostic semantic information was introduced to modulate the intermediate features of the network, and pixel-level uncertainty representation was used to constrain and optimize the segmentation results. The method of this invention was embedded into segmentation networks with various encoder-decoder structures and compared with baseline models without diagnostic semantic information. Specifically, the UNet segmentation network incorporating diagnostic semantic information was compared with the UNet segmentation network incorporating the method of this invention; the AttUNet segmentation network incorporating diagnostic semantic information was compared with the AttUNet segmentation network incorporating the method of this invention. The experimental results are shown in Table 1 below.

[0080] Table 1 Comparison of Experimental Results

[0081]

[0082] As shown in Table 1, under different segmentation network structures, the breast ultrasound image segmentation method based on diagnostic information perception of the present invention has achieved varying degrees of improvement in segmentation accuracy and boundary consistency indicators. In particular, it shows significant advantages in boundary error and false positive control of normal samples, indicating that the method of the present invention can effectively improve the stability and reliability of segmentation results.

[0083] like Figure 3 The image shows a qualitative comparison of the segmentation results; from Figure 3 It can be observed that the segmentation model without introducing a breast ultrasound image segmentation method based on diagnostic information perception is prone to missegmentation in normal tissue and noisy areas, and there are discontinuities or blurring at the lesion boundary. In contrast, the segmentation results of the breast ultrasound image segmentation method based on diagnostic information perception of the present invention can more accurately locate the lesion area, predict the boundary more smoothly and continuously, and effectively suppress abnormal responses in normal areas, reducing false positive segmentation results.

[0084] The experimental results above demonstrate that the breast ultrasound image segmentation method based on diagnostic information perception of the present invention, by introducing diagnostic semantic information and pixel-level uncertainty information, can improve the accuracy, stability and robustness of breast ultrasound image segmentation without relying on a specific network structure, thus verifying the effectiveness and versatility of the method of the present invention.

[0085] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. A method for segmenting breast ultrasound images based on diagnostic information perception, characterized in that: The system includes an image segmentation model based on diagnostic information perception; acquiring breast ultrasound image data to be processed and performing preprocessing operations on the image data; inputting the preprocessed image data into the encoder of the image segmentation model based on diagnostic information perception, extracting multi-level image features from the preprocessed image data through the encoder, and finally outputting a bottleneck feature map; Image-level diagnostic semantic vectors are constructed based on the bottleneck feature map output by the encoder; the decoding features of the decoding layer are dynamically modulated using the image-level diagnostic semantic vectors corresponding to the decoding features in each decoding layer of the image segmentation model based on diagnostic information perception. During the training process of the image segmentation model based on diagnostic information perception, the uncertainty of the segmentation result is modeled, and the uncertainty information is introduced into the segmentation optimization process to achieve uncertainty constraint on the segmentation result; During model training, a differentiated segmentation constraint strategy is implemented for normal image data; the feature map output by the decoder is input into the segmentation head for segmentation processing to generate the final segmentation result of the breast ultrasound image and output the pixel-level segmentation map of the lesion area.

2. The breast ultrasound image segmentation method based on diagnostic information perception according to claim 1, characterized in that: Obtain the bottleneck feature map output by the encoder, perform global average pooling on the bottleneck feature map to obtain channel-level statistical features; expand the channel-level statistical features to form an image-level diagnostic semantic vector; let the bottleneck feature map be F. b ∈R C×H×W Where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map, the global average pooling operation process is as follows: In the formula, F b (c, i, j) represents the feature value of the bottleneck feature map at the c-th channel and spatial position (i, j); z c This is represented as the channel-level statistical feature corresponding to the c-th channel.

3. The breast ultrasound image segmentation method based on diagnostic information perception according to claim 1, characterized in that: Image-level diagnostic semantic vector representations are introduced into the decoder as conditional information, and the decoding features in each decoding layer are dynamically modulated layer by layer, enabling the segmentation features to perceive the overall diagnostic context during the decoding process; this includes the following steps: Step 1: Input the image-level diagnostic semantic vector z corresponding to the decoding features of the l-th decoding layer into the parameter mapping module, and perform feature mapping to generate the intermediate semantic features z1 required for modulation; Step 2: Generate the channel-level scaling parameter γ based on the intermediate semantic features and the decoding features corresponding to the l-th decoding layer in the decoder. l and bias parameter β l ; Step 3, Channel-level scaling parameter γ l and bias parameter β l Apply a nonlinear activation function and limit the channel-level scaling parameter γ l and bias parameter β l The range of values ​​for; Step 4: Use the channel-level scaling parameter γ corresponding to the decoding feature of the l-th decoding layer. l and bias parameter β l For the decoding feature F corresponding to the l-th decoding layer in the decoder l Channel-by-channel dynamic modulation is performed to obtain the modulated decoded feature map.

4. The breast ultrasound image segmentation method based on diagnostic information perception according to claim 3, characterized in that: In step 2, the channel-level scaling parameter γ l and bias parameter β l The calculation process is as follows: In the formula, , Both are represented as parameter mapping functions.

5. The breast ultrasound image segmentation method based on diagnostic information perception according to claim 3, characterized in that: In step 4, the dynamic modulation decoding feature process is as follows: In the formula, This is represented as a channel-by-channel multiplication operation. This is represented as the decoded feature after modulation.

6. The breast ultrasound image segmentation method based on diagnostic information perception according to claim 1, characterized in that: Acquire breast ultrasound image data to be processed and perform preprocessing operations on the image data; construct a training sample set based on several preprocessed image data sets, and use the training sample set to train an image segmentation model based on diagnostic information perception; during the model training process, model the uncertainty of the output segmentation results, including the following steps: Step 1-1: The feature map output by the decoder is input into the segmentation head for segmentation processing to obtain pixel-level segmentation results. At the same time, the pixel-level uncertainty information corresponding to the pixel-level segmentation results is predicted simultaneously. Steps 1-2: Based on the element-level uncertainty information, model the prediction uncertainty level of each pixel position to obtain a pixel-level prediction uncertainty representation; Steps 1-3: During model training, pixel-level prediction uncertainty representation is introduced into the segmentation loss function to construct a weighted segmentation loss function. The weighted segmentation loss function is used to adaptively weight the loss term in the high uncertainty region of the image data during model training.

7. The breast ultrasound image segmentation method based on diagnostic information perception according to claim 6, characterized in that: In steps 1-3, let the pixel-level binary classification cross-entropy loss be the pixel-level basic segmentation loss L. base The weighted segmentation loss function L, after introducing pixel-level prediction uncertainty representation, is defined as: In the formula, w(i,j) represents the adaptive weight for each pixel position. This can be represented as the average of the weighted loss function over all pixels. This is represented as pixel-by-pixel multiplication. This represents the pixel-level variance at each pixel location. It is represented as a minimum value.

Images