CN-121999008-A - Mammary gland ultrasonic focus segmentation method based on auxiliary branch contrast aggregation network

CN121999008ACN 121999008 ACN121999008 ACN 121999008ACN-121999008-A

Abstract

The invention belongs to the technical field of medical image processing, and particularly relates to a breast ultrasound focus segmentation method based on an auxiliary branch comparison aggregation network, which comprises the steps of inputting processed images into a segmentation network of an encoder-decoder framework, wherein an encoder in the segmentation network comprises feature extraction modules which are sequentially cascaded from a first stage to a fifth stage, and performing downsampling processing on features by using maximum pooling operation after each feature extraction, and then taking the processed images as output of the stage; the feature extraction module of the third stage of the encoder is connected with the double auxiliary branch module, and the module extracts foreground features and background features from the feature map output by the feature extraction module of the third stage. The invention can enhance the characteristic contrast between the focus and the surrounding tissues and improve the boundary segmentation precision.

Inventors

QIN DUI
CHEN KUN
RAN MENGJIA
WANG WEI

Assignees

重庆邮电大学

Dates

Publication Date: 20260508
Application Date: 20260210

Claims (8)

1. A breast ultrasound focus segmentation method based on an auxiliary branch contrast aggregation network is characterized by comprising the following steps: Preprocessing the mammary gland ultrasonic image, reducing image background noise interference and enhancing the image; Inputting the processed image into a segmentation network of an encoder-decoder architecture, wherein an encoder in the segmentation network comprises feature extraction modules which are sequentially cascaded from a first stage to a fifth stage, and performing downsampling processing on features by using a maximum pooling operation after each feature extraction is performed, and then taking the processed features as output of the stage; the method comprises the steps that a third-stage feature extraction module of an encoder is connected with a double auxiliary branch module, and foreground features and background features are extracted from a feature map output by the third-stage feature extraction module; The decoder in the segmentation network comprises feature extraction modules which are sequentially cascaded from a fourth stage to a first stage, and the output of the feature extraction module of the fifth stage of the encoder is spliced with the output of the feature extraction module of the fourth stage of the encoder to be used as the input of the feature extraction module of the fourth stage of the decoder; After up-sampling is carried out on the output of the fourth-stage feature extraction module of the decoder, the foreground and background contrast driving feature aggregation module utilizes foreground features and background features to screen the up-sampled feature map, and the screened feature map is spliced with the feature extraction module of the third stage of the encoder to serve as the input of the next stage; The output of the subsequent previous stage of the decoder is spliced with the output of the encoder corresponding to the current stage after up-sampling and then is used as the input of the current stage; And the first-stage feature extraction module of the decoder outputs a final segmentation result after the feature map is subjected to point convolution operation.
2. The breast ultrasound lesion segmentation method based on the auxiliary branch comparison aggregation network according to claim 1, wherein the feature extraction module of each stage of the encoder or the encoder comprises a cascaded multipath parallel feature extraction unit, an adaptive attention weight generation unit and a dynamic weighted fusion unit, wherein: The multi-channel parallel feature extraction unit is used for extracting features with different sizes from the input feature images by adopting convolution operations with convolution kernel sizes of 1 multiplied by 1, 3 multiplied by 3 and 5 multiplied by 5 respectively, and splicing the feature images with different sizes together as the output of the multi-channel parallel feature extraction unit after the convolution results undergo batch normalization operation and LeakyReLU activation operation; the self-adaptive attention weight generating unit firstly carries out channel dimension reduction on the spliced feature map through CBR operation, then laminates space dimension through global average pooling, finally generates self-adaptive weights corresponding to three paths by using a 1X 1 convolution layer and a Sigmoid activation function, and the CBR operation represents that convolution, batch normalization and LeakyReLU activation are sequentially carried out; And the dynamic weighting fusion unit is used for respectively weighting the feature graphs with different sizes obtained by the multipath parallel feature extraction unit by using the generated weights, and obtaining the output of the feature extraction module through adding and fusing the weighted feature graphs and CBR operation.
3. The breast ultrasound lesion segmentation method based on the auxiliary branch comparison aggregation network according to claim 2, wherein the process of obtaining the output of each level of feature extraction module by the dynamic weighted fusion unit comprises the following steps: ; ; ; Wherein, the Representing a splicing operation; representing the spliced characteristic diagram; The input feature graphs are processed through the 1 multiplied by 1, the 3 multiplied by 3 and the 5 multiplied by 5 convolution layers respectively and are all activated through batch normalization and LeakyReLU; representing the resulting adaptive weights, the weights being comprised of three sub-weight vectors 、、 Tensors of (a); Representing a global pooling operation; representing a point convolution up-dimensional transformation; Representing a Sigmoid function; An output characteristic diagram representing each level of coding layer; Representing the number of encoder stages.
4. The breast ultrasound lesion segmentation method based on the auxiliary branch comparison aggregation network according to claim 1, wherein the extracting of the foreground feature and the background feature from the feature map output by the feature extraction module of the third stage by the double auxiliary branch module comprises: ; ; Wherein, the Representing foreground branch output characteristics; representing a background branch output feature; Representing convolution feature transformation operation, namely performing feature extraction and dimension transformation on an input feature map through a CBR operation of 3×3 and a CBR operation of 1×1 in sequence; the weight generation operation for expressing the attention of the channels, namely generating attention weight for each channel of the input feature map through a cascade global average pooling layer, full connection layer sequence transformation and Sigmoid function; an output characteristic diagram of the third-level coding layer is shown.
5. The breast ultrasound focus segmentation method based on the auxiliary branch comparison aggregation network according to claim 1, wherein the foreground-background comparison driving feature aggregation module uses foreground features and background features to screen a feature map obtained by up-sampling the output of the fourth-stage feature extraction module of the decoder, specifically comprising: Processing foreground features and background features based on channel filtering, and inhibiting a noise channel; Generating a front Jing Zhuyi force weight and a background attention weight which are corresponding to each other and are based on the foreground characteristic and the background characteristic by utilizing a linear mapping mechanism; Extracting trunk features from a feature map output by a fourth-level feature extraction module of the decoder, converting the trunk features into value feature vectors through linear mapping, extracting a local value feature vector set through sliding window partitioning operation and remolding operation, wherein the window sliding operation is to divide a complete feature map into feature block sets through a window with a certain size, and the remolding operation is to change the dimension structure of the feature vectors; Weighting the local value feature vector set in a local window by using the attention weight based on the foreground feature, and generating a foreground enhanced intermediate feature map focused on a focus area through space dimension reconstruction; Performing sliding window blocking operation and remolding operation on the foreground enhanced intermediate feature image again to obtain a local value feature vector set of the foreground enhanced intermediate feature image, and secondarily weighting the set by using the attention weight based on the background feature to inhibit irrelevant background area interference in the feature image; And finally, performing space dimension reconstruction on the secondarily weighted features to obtain a contrast-enhanced refined feature map.
6. The method for breast ultrasound lesion segmentation based on the auxiliary branch contrast aggregation network according to claim 5, wherein the processing of the foreground features and the background features based on the channel filtering comprises: ; ; Wherein, the Representing foreground branch output characteristics; representing a background branch output feature; representing foreground branch output characteristics based on channel filtering; representing a background output characteristic based on channel filtering; Representing a global average pooling operation; 、 representing the weight matrix of the fully connected layer, Representing channel multiplication.
7. The method of breast ultrasound lesion segmentation based on an assisted branch contrast aggregation network according to claim 5, wherein generating corresponding local window-based front Jing Zhuyi and background attention weights based on foreground and background features comprises: ; ; Wherein, the Representing a front Jing Zhuyi force weight derived based on a linear mapping; Representing background attention weights based on linear mapping, softmax representing a weight normalization function, reshape representing a reshaping operation, i.e., changing the dimensional structure of the feature vector; representing foreground branch output characteristics based on channel filtering; representing a background output characteristic based on channel filtering; , the weight and bias of the foreground linear mapping layer are represented; , the weights and offsets of the background linear mapping layer are represented.
8. The breast ultrasound lesion segmentation method based on the auxiliary branch contrast aggregation network according to claim 1, wherein the loss function adopted in training the segmentation network is expressed as: ; ; ; Wherein, the Representing the total loss function of the device, Representing the main segmentation loss function, The weights of the auxiliary loss function are represented, Representing auxiliary branch loss; g represents a real segmentation result, and P represents a segmentation result output by a segmentation network; representing a binary cross entropy loss function; Representing the Dice loss function.

Description

Mammary gland ultrasonic focus segmentation method based on auxiliary branch contrast aggregation network Technical Field The invention belongs to the technical field of medical image processing, and particularly relates to a breast ultrasound focus segmentation method based on an auxiliary branch contrast aggregation network. Background Breast cancer is one of the most common malignant tumors in women, and early detection and accurate diagnosis are critical for improving survival rate of patients. With the development of artificial intelligence technology, computer-aided diagnosis systems based on convolutional neural networks have been widely used in lesion segmentation and classification tasks of breast medical images. While existing deep-learning models perform well in conventional natural image processing, their automatic segmentation faces many challenges when applied to breast medical ultrasound images. Breast ultrasound images typically exhibit low contrast, high speckle noise, and shading artifacts, resulting in unclear boundaries between different tissue organs. In addition, there are significant differences in the size, shape and texture of lesions in different cases in breast ultrasound images, making the segmentation method generally poorly performing. Breast lesions in different cases have significant morphological heterogeneity. The fixed receptive field adopted by the existing method cannot be adapted to the span of the focus scale, and a single convolution cannot be adapted to targets of all sizes at the same time. If the network is biased to extract local high-frequency details, the complete topological structure and global context of a large tumor cannot be captured, the large target is easily recognized incompletely or the internal cavity is easily caused, the malignant focus is poorly segmented, if the network acquires global semantics through deep downsampling, and the spatial position information and the edge details of the micro focus are greatly lost along with the reduction of the resolution of the feature map, so that the micro focus is missed. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a breast ultrasound focus segmentation method based on an auxiliary branch contrast aggregation network, which comprises the following steps: Preprocessing the mammary gland ultrasonic image, reducing image background noise interference and enhancing the image; inputting the processed image into a segmentation network of an encoder-decoder architecture, wherein the encoder in the segmentation network comprises feature extraction modules which are sequentially cascaded from a first stage to a fifth stage, and performing downsampling processing on features by using a maximum pooling operation after each feature extraction, and then taking the processed image as the output of the stage; the method comprises the steps that a third-stage feature extraction module of an encoder is connected with a double auxiliary branch module, and foreground features and background features are extracted from a feature map output by the third-stage feature extraction module; The decoder in the segmentation network comprises feature extraction modules which are sequentially cascaded from a fourth stage to a first stage, and the output of the feature extraction module of the fifth stage of the encoder is spliced with the output of the feature extraction module of the fourth stage of the encoder to be used as the input of the feature extraction module of the fourth stage of the decoder; After up-sampling is carried out on the output of the fourth-stage feature extraction module of the decoder, the foreground-background comparison driving feature aggregation module screens the feature map after up-sampling by utilizing foreground features and background features, and the screened feature map is spliced with the third-stage output features of the encoder to serve as the input of a next-stage decoding layer; The output of the subsequent previous stage of the decoder is spliced with the output of the encoder corresponding to the current stage after up-sampling and then is used as the input of the current stage; And the first-stage feature extraction module of the decoder outputs a final segmentation result after the feature map is subjected to point convolution operation. Compared with the prior art, the invention designs the feature extraction module with the multi-scale collaborative perception function, can accurately capture the feature information of the focuses of different scales by adapting the multi-path parallel convolution structures of the focuses of the breast of different sizes and combining the self-adaptive weight fusion mechanism, realizes the extraction of foreground and background features by the double auxiliary branch module, strengthens the focus and background related channel features through the channel attention, provides differentiated features for