CN-121999218-A - Breast cancer ultrasonic image intelligent segmentation method and system based on weak supervision learning

CN121999218ACN 121999218 ACN121999218 ACN 121999218ACN-121999218-A

Abstract

The application provides an intelligent breast cancer ultrasonic image segmentation method and system based on weak supervision learning. The method comprises the steps of obtaining an ultrasonic image of the breast to be screened, adjusting resolution, and inputting the ultrasonic image to an M2A-U-Net network model. The model consists of an encoder, a bottleneck layer, a decoder and an output layer. The encoder performs feature extraction through a multi-scale cavity convolution fusion module and a fuzzy pooling layer, the bottleneck layer performs cross-dimensional feature reinforcement through a triple attention mechanism module, the decoder embeds a high-efficiency multi-scale attention module in jump connection to fuse features, and the output layer outputs a pixel-level focus segmentation result after being optimized by a compression-excitation attention module. In the training stage, a weak supervision segmentation strategy is introduced, and the network model is jointly trained based on the marked and unmarked breast ultrasonic image samples. The problems of noise interference of the ultrasonic image and fuzzy focus boundary are effectively solved, the manual labeling cost is reduced, and the segmentation precision and the robustness of breast cancer screening are improved.

Inventors

Bian Xuesheng
SUN YUHAN
WANG LINXUAN
TANG NAN
XU SEN
ZHU JINXIN
SHAO JUN

Assignees

盐城工学院

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (10)

1. The breast cancer ultrasonic image intelligent segmentation method based on weak supervision learning is characterized by comprising the following steps of: Acquiring a breast ultrasonic image to be screened, and adjusting the resolution of the breast ultrasonic image by adopting a bilinear interpolation method to obtain a standard input image with preset resolution; Constructing an M2A-U-Net network model, wherein the M2A-U-Net network model comprises an encoder, a bottleneck layer, a decoder and an output layer which are sequentially connected; Inputting the standard input image into a pre-trained M2A-U-Net network model, sequentially extracting multi-scale features through the encoder, strengthening cross-dimension features of the bottleneck layer, fusing and reconstructing cross-scale features of the decoder, optimizing channels of the output layer, and outputting a pixel-level focus segmentation result; The encoder comprises a plurality of multi-scale cavity convolution fusion modules connected in series, and the output end of each multi-scale cavity convolution fusion module is connected with a fuzzy pooling layer; the bottleneck layer comprises a triple attention mechanism module which is used for carrying out cross-dimension dependency modeling on input features through parallel branches; The decoder is embedded with a high-efficiency multi-scale attention module in a jump connection path corresponding to the encoder; The output layer includes a compression-stimulus attention module; The training process of the M2A-U-Net network model comprises a weak supervision segmentation training step, wherein the weak supervision segmentation training step is based on a marked breast ultrasonic image sample and a non-marked breast ultrasonic image sample to perform joint training optimization on the network model.
2. The breast cancer ultrasonic image intelligent segmentation method based on weak supervision learning according to claim 1, wherein the multi-scale cavity convolution fusion module comprises four convolution branches with different expansion rates, and the characteristic extraction calculation formula is as follows: ; Wherein, the A feature map representing the input to the module; index sequence number representing convolution branch, and the value range is an integer from 1 to 4; Represent the first Feature graphs output by the convolution branches; Represent the first Feature graphs output by the convolution branches; representing the expansion rate of the cavity convolution; indicating that the convolution kernel is of size And the expansion rate is The expansion rate of four convolution branches Respectively setting 1, 2, 4 and 8; The characteristic fusion formula of the multi-scale cavity convolution fusion module is as follows: ; Wherein, the Representing the fused feature map; where corresponds to the first Feature map of individual convolved branch outputs ; Represent the first The weight vectors of the scales are obtained by global average pooling and Softmax function calculation of each scale feature map; Representing a channel-by-channel multiplication operation; representing a pixel-by-pixel summing operation; The fuzzy pooling layer is used for fusing the feature images First perform standard deviation A kind of electronic device Gaussian filtering and downsampling by a factor of 2.
3. The method for intelligent segmentation of breast cancer ultrasound images based on weakly supervised learning as set forth in claim 1, wherein the triple attention mechanism module comprises three parallel cross-dimensional attention branches of channel-width, channel-height, height-width, and input feature maps for each branch And executing a Z-Pooling operation, wherein the formula is as follows: ; Wherein, the Representing an input feature map; Representing the pooled two-dimensional feature vector; Representing a global max pooling operation; Representing a global average pooling operation; Representing a splice operation performed in a channel dimension; Output feature map of the triple attention mechanism module The calculation formula of (2) is as follows: ; Wherein, the Representing the reinforced output characteristic diagram; representing an element-wise multiplication operation; 、、 attention weight graphs generated by the channel-width branch, the channel-height branch and the height-width branch are respectively represented, and the attention weight graphs are calculated by the corresponding Z-Pooling result through convolution and Sigmoid activation function.
4. The method for intelligently segmenting the breast cancer ultrasonic image based on weak supervised learning according to claim 1, wherein the efficient multi-scale attention module is arranged in a jump connection path between the decoder and the encoder, and the processing procedure comprises the following steps: Dividing an input feature map evenly along a channel dimension into Sub-feature group, wherein the number of packets ; Three branching operations are performed in parallel for each group of sub-feature groups, a first branching performing horizontal-direction global average pooling to generate a horizontal-direction channel attention pattern, a second branching performing vertical-direction global average pooling to generate a vertical-direction channel attention pattern, a third branching performing Convolving to capture local multi-scale features; and fusing and re-weighting the outputs of the three branches through a cross-space information aggregation mechanism, and finally splicing all the sub-feature groups to obtain the enhanced feature map.
5. The method for intelligently segmenting breast cancer ultrasound images based on weak supervised learning as set forth in claim 1, wherein the compression-excitation attention module is configured to compare a feature map with a feature map The channel of (2) is recalibrated, and the calculation formula is as follows: ; ; ; Wherein, the Represent the first Global statistics of individual channels; And The height and width of the feature map are represented respectively; Represent the first The individual channels are in coordinates Pixel values at; Is composed of all A composed channel description vector; representing the generated channel weight vector; And Respectively representing weight matrixes of the first full connection layer and the second full connection layer; representing a ReLU activation function; representing a Sigmoid activation function; Representing channel weight vectors Corresponds to the first The weight value of each channel; Representing the calibrated first A channel feature map; Representing scalar multiplication; And the output layer processes the calibrated feature map through a Sigmoid activation function to generate a pixel-level focus probability map, and judges that the pixels with probability larger than 0.5 are focuses.
6. The method for intelligently segmenting breast cancer ultrasonic images based on weak supervision learning according to claim 1, wherein the training process of the M2A-U-Net network model adopts a mixed loss function The calculation formula is as follows: ; ; ; Wherein, the Representing a mixing loss function value; Representing a binary cross entropy loss function value; Representing a Dice loss function value; Representing the total number of pixels of a single image; Representing a pixel index; Represent the first The true label of each pixel has a value of 0 or 1,1 represents a focus, and 0 represents a background; representation model prediction No The individual pixels are probability values of the lesion; representing a natural log operation.
7. The method for intelligently segmenting breast cancer ultrasound images based on weak supervision learning according to claim 1, wherein the weak supervision segmentation training step comprises the following steps: Constructing at least two M2A-U-Net split networks which have the same structure and are related to parameters, and respectively serving as a first split network and a second split network; Inputting the marked breast ultrasonic image sample into the first segmentation network, calculating the supervised segmentation loss according to the corresponding real marking information, and updating the parameters of the first segmentation network; Respectively applying data disturbance of different intensities to the non-marked breast ultrasonic image samples, and then inputting the data disturbance to the first segmentation network and the second segmentation network to obtain corresponding segmentation prediction results, wherein the data disturbance comprises random horizontal/vertical overturn, 10% random scaling and Gaussian noise, and the disturbance intensities of different networks are randomly distributed to ensure the diversity of input differences; Constructing a prediction consistency constraint based on a segmentation prediction result of the first segmentation network and the second segmentation network on the same non-labeling breast ultrasound image sample so as to restrict the segmentation output of the first segmentation network on the non-labeling sample to be consistent with the second segmentation network; in the weak supervision segmentation training process, the method further comprises the following steps: extracting corresponding intermediate feature representations in intermediate feature layers of the first segmentation network and the second segmentation network; Based on intermediate feature representations obtained by the non-labeled breast ultrasound image sample under different data disturbance conditions, feature consistency constraint is constructed to constrain feature responses of different segmentation networks to the same non-labeled sample to be consistent, so that robust characterization capability of the segmentation networks to breast focus structural features is enhanced.
8. The method for intelligently segmenting breast cancer ultrasound images based on weak supervision learning according to claim 1, wherein in the weak supervision segmentation training process, the method further comprises the following steps: Extracting intermediate characteristic representations corresponding to the non-labeling breast ultrasound image samples from intermediate characteristic layers of the first segmentation network and the second segmentation network; Based on the segmentation prediction result of the non-labeling breast ultrasound image sample, performing region-level feature aggregation operation on a focus region and a background region in the intermediate feature representation respectively to obtain a corresponding region-level feature representation; a class-level feature memory unit is constructed and used for storing area-level feature representations corresponding to different classes and carrying out iterative update on the class-level feature representations by adopting a smooth update strategy; Based on the similarity relation between the regional level feature representation and the corresponding features in the category level feature memory unit, a regional level comparison constraint is constructed to constrain the same-class regional features to be consistent in the feature space and the different-class regional features to be distinguished in the feature space, so that the distinguishing capability and the structural stability of the segmentation network on the unlabeled samples are improved.
9. The method for intelligently segmenting the breast cancer ultrasonic image based on weak supervision learning according to claim 1, wherein the step of inputting the standard input image into a pre-trained M2A-U-Net network model and outputting a pixel-level focus segmentation result specifically adopts a multi-view consistent fusion reasoning strategy, comprises the following steps: constructing a transformation view set of the standard input image, wherein the transformation view set at least comprises the standard input image and an auxiliary reasoning image generated after horizontal overturn geometric transformation and vertical overturn geometric transformation are respectively carried out on the standard input image; Each image in the transformation view set is respectively input into the M2A-U-Net network model in parallel to obtain a plurality of corresponding initial probability feature images; Respectively executing space mapping operation which is reciprocal to the input geometric transformation logic of the initial probability feature images to the initial probability feature images so that the pixel coordinate systems of all the initial probability feature images are regressed to be consistent with the standard input images; And calculating probability average values of the multiple initial probability feature graphs at corresponding pixel positions after regression, and generating a fused anti-noise probability distribution map so as to inhibit random interference of ultrasonic speckle noise on a prediction result.
10. Breast cancer ultrasonic image intelligent segmentation system based on weak supervision study, which is characterized by comprising: A memory for storing a computer program; A processor for executing the computer program to implement the steps of a breast cancer ultrasound image intelligent segmentation method based on weak supervised learning as set forth in any one of claims 1-9.

Description

Breast cancer ultrasonic image intelligent segmentation method and system based on weak supervision learning Technical Field The invention relates to the technical fields of medical image processing, artificial intelligence and computer vision, in particular to an intelligent breast cancer ultrasonic image segmentation method and system based on weak supervision learning. Background Breast cancer is the malignant tumor with the highest incidence of women worldwide, and early screening is important for improving survival rate. Breast ultrasound is widely used because of its advantages of no radiation, low cost, etc. However, breast ultrasound images are deeply affected by speckle noise, low contrast, and foci boundary blurring. As shown in fig. 1, significant morphological differences (e.g., varying size, irregular shape) and blurred boundary features (e.g., yellow outline) of breast cancer lesions between individuals are demonstrated. When the existing deep learning segmentation method (such as U-Net) processes such complex images, the existing deep learning segmentation method is sensitive to noise, local details and global contours are difficult to consider, cross-level features are not sufficiently fused, and fuzzy boundary information is easy to lose. In addition, the deep learning segmentation method is mostly based on a large number of high-quality pixel-level manual labeling samples for supervision training, and the fine labeling process of the breast ultrasound images is highly based on imaging doctors with abundant clinical experience, so that the labeling cost is high, the period is long, and sufficient labeling data are difficult to acquire in an actual clinical scene. Under the condition that the number of marked samples is limited, the model is easy to be subjected to fitting, the generalization capability is insufficient, and the improvement of the segmentation performance is further restricted. Therefore, an intelligent segmentation method which can fully utilize the non-marked breast ultrasonic image information under the limited marking condition, has the multi-scale focus modeling and fuzzy boundary describing capability and can effectively inhibit the interference of ultrasonic noise is needed to improve the accuracy and the robustness of breast cancer ultrasonic screening. Disclosure of Invention The invention aims to provide a breast cancer ultrasonic image intelligent segmentation method and system based on weak supervision learning, which are characterized in that a weak supervision segmentation training strategy is introduced while improving the characteristic capacity of a network to ultrasonic image multi-scale focus features and fuzzy boundary characterization so as to relieve the dependence on a large-scale fine labeling sample and further improve segmentation precision and model generalization capacity by constructing an M2A-U-Net network model and combining a multi-scale cavity convolution fusion (MDCF), a triple attention mechanism (TRA), a high-efficiency multi-scale attention (EMA) and a compression-excitation (SE) module. In a first aspect, an embodiment of the present invention provides a breast cancer ultrasound image intelligent segmentation method based on weak supervised learning, including the following steps: Acquiring a breast ultrasonic image to be screened, and adjusting the resolution of the breast ultrasonic image by adopting a bilinear interpolation method to obtain a standard input image with preset resolution; Constructing an M2A-U-Net network model, wherein the M2A-U-Net network model comprises an encoder, a bottleneck layer, a decoder and an output layer which are sequentially connected; Inputting the standard input image into a pre-trained M2A-U-Net network model, sequentially extracting multi-scale features through the encoder, strengthening cross-dimension features of the bottleneck layer, fusing and reconstructing cross-scale features of the decoder, optimizing channels of the output layer, and outputting a pixel-level focus segmentation result; The encoder comprises a plurality of multi-scale cavity convolution fusion modules connected in series, and the output end of each multi-scale cavity convolution fusion module is connected with a fuzzy pooling layer; the bottleneck layer comprises a triple attention mechanism module which is used for carrying out cross-dimension dependency modeling on input features through parallel branches; The decoder is embedded with a high-efficiency multi-scale attention module in a jump connection path corresponding to the encoder; The output layer includes a compression-stimulus attention module; The pre-trained M2A-U-Net network model is obtained through a weak supervision learning mode, and the training process of the pre-trained M2A-U-Net network model simultaneously utilizes marked breast ultrasonic image samples and unmarked breast ultrasonic image samples to carry out joint training. Optionally, the multi-scale cav