CN-122023435-A - Automatic segmentation method, system, storage medium and equipment for breast cancer postoperative targets based on Mamba feature correction

CN122023435ACN 122023435 ACN122023435 ACN 122023435ACN-122023435-A

Abstract

The invention discloses a method, a system, a storage medium and equipment for automatically segmenting a breast cancer postoperative target based on Mamba feature correction, which belong to the technical field of intelligent medical treatment, and comprise the following steps of S1, collecting CT images of screened breast cancer breast-preserving postoperative radiotherapy patients and corresponding manual sketching outline data; S2, carrying out standardized pretreatment on the CT image, S3, constructing and training a target segmentation model, embedding a selective Mamba characteristic corrector module in jump connection of a nnU-Net frame in the target segmentation model, S4, inputting the pretreated CT image into the trained target segmentation model, and outputting a prediction probability map for delineating a clinical target area and a tumor bed. The invention realizes the active correction and fusion of the local characteristics of the encoder, can effectively inhibit the image noise and artifact interference, enhance the boundary characteristics of the target area, improve the segmentation precision and consistency, and is suitable for the automatic and high-precision sketching of the radiotherapy target area after breast cancer operation.

Inventors

YUAN HANGYU
LI XIAOYU
FENG MEILING
ZHOU QIUYU
WANG JIANSHU
ZHENG DESHENG
Tian Xuwei

Assignees

电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260209

Claims (10)

1. The automatic segmentation method for the breast cancer postoperative target based on Mamba feature correction is characterized by comprising the following steps of: s1, collecting CT images of screened breast cancer postoperative radiotherapy patients and corresponding manual delineation outline data; s2, carrying out standardized pretreatment on the CT image; s3, constructing and training a target segmentation model, wherein the target segmentation model is embedded with a selective Mamba characteristic corrector module in jump connection of nnU-Net frames; s4, inputting the preprocessed CT image into a trained target segmentation model, and outputting a prediction probability map for delineating a clinical target area and a tumor bed.
2. The method for automatically segmenting a target after breast cancer surgery based on Mamba feature correction according to claim 1, wherein the target segmentation model comprises: An input layer for inputting the preprocessed CT image; The encoder is used for extracting multi-scale characteristics of the CT image layer by layer and realizing space downsampling and comprises a plurality of cascaded encoder blocks, wherein each encoder block consists of two continuous 3X3 convolution layers, a normalization layer, an activation function and a 2X 2 max pooling layer; A decoder including a plurality of concatenated decoder blocks, each decoder block spatially upsampling a low resolution feature map from a previous layer by 2 x 2 transpose convolution; mamba feature corrector module, comprising: The global context extraction unit is used for carrying out global average pooling on the decoder feature map to obtain a channel-level global feature vector; The feature correction unit is used for multiplying the global feature vector with the feature map extracted by the encoder element by element to obtain a preliminary correction feature; A boundary enhancement unit for generating a boundary weight map based on the spatial gradient of the feature map extracted by the decoder and fusing with the preliminary correction feature; an output layer consisting of a1 x1 convolutional layer and a Softmax activation function, wherein, the number of characteristic channels is mapped to the number of classes of target region by a1 x1 convolution layer, and pass through The function outputs a predictive probability map.
3. The method for automatically segmenting a post-operative target for breast cancer based on Mamba feature correction as claimed in claim 2, wherein the Mamba feature corrector module further comprises: and the self-adaptive weighted fusion module is used for carrying out self-adaptive weighted fusion on the characteristics generated by the boundary enhancement unit according to the quality of the characteristics.
4. The method for automatically segmenting a breast cancer postoperative target based on Mamba feature correction according to claim 1, wherein the target segmentation model uses a combined loss function of a Dice loss and a cross entropy loss.
5. The method for automatically segmenting a breast cancer postoperative target based on Mamba feature correction according to claim 4, wherein the combined loss function is as follows: where α is a weight coefficient, and Dice represents a Dice loss function: , the CE is a cross-loss function: Is the true label for the i-th pixel, The probability that the pixel belongs to the target region is predicted for the model, and N is the total number of pixels of the image.
6. The method for automatically segmenting a breast cancer postoperative target based on Mamba feature correction according to claim 2, wherein the generating of the boundary weight map includes: calculating the three-dimensional spatial gradient of the feature map extracted by the decoder, and extracting a boundary region to obtain a spatial gradient map; and carrying out normalization processing on the spatial gradient map to obtain a boundary weight map.
7. The method for automatically segmenting a target after breast cancer surgery based on Mamba feature correction as claimed in claim 1, wherein the standardized pretreatment comprises: And carrying out voxel space normalization and pixel value adjustment on the CT image.
8. Automatic segmentation system of breast cancer postoperative target based on Mamba feature correction, characterized by comprising: The data acquisition module is used for collecting CT images of the screened breast cancer breast-preserving postoperative radiotherapy patients and corresponding manual delineating outline data; the data preprocessing module is used for carrying out standardized preprocessing on the CT image; the model construction and training module is used for constructing and training a target segmentation model, wherein the target segmentation model is embedded with a selective Mamba characteristic corrector module in jump connection of nnU-Net frames; And the model prediction module is used for inputting the preprocessed CT image into a trained target segmentation model and outputting a prediction probability map for outlining a clinical target area and a tumor bed.
9. A computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method for automatic segmentation of breast cancer postoperative targets according to any one of claims 1-7.
10. An electronic device comprising a memory and a processor, the memory having stored thereon computer instructions executable on the processor, wherein the processor executes the method for automatically segmenting a target post-operative breast cancer of any one of claims 1-7 when the processor executes the computer instructions.

Description

Automatic segmentation method, system, storage medium and equipment for breast cancer postoperative targets based on Mamba feature correction Technical Field The invention relates to the technical field of intelligent medical treatment, in particular to a method, a system, a storage medium and equipment for automatically segmenting breast cancer postoperative targets based on Mamba feature correction. Background Breast cancer is one of the most common malignant tumors in women worldwide. Post-operative radiation therapy is a recognized standard treatment after undergoing breast-conserving surgery. Accurately delineating a clinical target Volume (CLINICAL TARGET Volume, CTV) and a Tumor Bed (Tumor Bed, TB) is a key step in ensuring success of radiotherapy, directly influences the distribution of radiation dose, and determines effectiveness and safety of treatment. Currently, the clinic generally relies on the radiation oncologist to manually delineate the target area, and this process has the following drawbacks: 1. the manual sketching of the target area not only consumes a great deal of time and energy, but also easily causes variability among observers due to the difference of experiences of different doctors, so that standardization and consistency of sketching results are difficult to ensure. 2. The soft tissue contrast in the postoperative computed tomography (Computed Tomography, CT) image is low, and the boundary between the tumor bed and the target area and surrounding normal breast tissue is blurred, further exacerbating the difficulty of delineating. In addition, the artifact generated by the metal titanium clamp for surgical implantation greatly interferes with accurate judgment of the doctor on the image. In addition, although the deep learning technology has made great progress in the field of medical image segmentation, various automatic segmentation methods have been proposed successively, in practical application, particularly when complex post-operation CT images are processed, the segmentation accuracy of the existing model still has difficulty in reaching a clinically required high level, and when facing a tumor bed with irregular morphology and blurred boundary, the boundary of a target region cannot be accurately depicted, so that the requirement of clinical high-accuracy segmentation is difficult to meet. Therefore, in the actual clinical radiotherapy plan, the doctor is difficult to rely on the automatic segmentation result completely, and a great deal of time is still required for manual correction and adjustment, so that the popularization and application of the automatic segmentation technology in clinic are greatly limited. Therefore, there is a need for an automatic segmentation technique for breast cancer postoperative targets, which can adapt to the characteristics of postoperative CT images, has high accuracy and consistency, and has strong generalization capability. Disclosure of Invention The invention aims to overcome the problems in the prior art and provides a method, a system, a storage medium and equipment for automatically segmenting a breast cancer postoperative target based on Mamba feature correction, wherein the core is to construct a fusion model (Mamba-Refined UNet, MR-UNet), the model is based on an automatic nnU-Net framework, and a selective Mamba feature corrector (SELECTIVE MAMBA FEATURE CORRECTOR, SMFC) is embedded in the traditional jump connection to realize active and semantic guided feature fusion. The aim of the invention is realized by the following technical scheme: In a first aspect, an automatic segmentation method for a breast cancer postoperative target based on Mamba feature correction is provided, including the following steps: s1, collecting CT images of screened breast cancer postoperative radiotherapy patients and corresponding manual delineation outline data; s2, carrying out standardized pretreatment on the CT image; s3, constructing and training a target segmentation model, wherein the target segmentation model is embedded with a selective Mamba characteristic corrector module in jump connection of nnU-Net frames; s4, inputting the preprocessed CT image into a trained target segmentation model, and outputting a prediction probability map for delineating a clinical target area and a tumor bed. In some embodiments, the target segmentation model comprises: An input layer for inputting the preprocessed CT image; The encoder is used for extracting multi-scale characteristics of the CT image layer by layer and realizing space downsampling and comprises a plurality of cascaded encoder blocks, wherein each encoder block consists of two continuous 3X3 convolution layers, a normalization layer, an activation function and a 2X 2 max pooling layer; A decoder including a plurality of concatenated decoder blocks, each decoder block spatially upsampling a low resolution feature map from a previous layer by 2 x 2 transpose convolution; mamba feature corr