CN-120411514-B - Intravascular ultrasound image segmentation method based on frequency domain-airspace cooperative enhancement

CN120411514BCN 120411514 BCN120411514 BCN 120411514BCN-120411514-B

Abstract

The embodiment of the invention discloses an intravascular ultrasound image segmentation method based on frequency domain-space domain cooperative enhancement, which comprises the steps of utilizing a layered feature extraction module to extract a coding feature image of a training image, utilizing a multi-frequency feature decoupling module to conduct wavelet transformation on the coding feature image to generate a low-frequency component feature image and a horizontal, vertical and diagonal high-frequency component feature image, fusing to generate a frequency domain enhancement feature image, utilizing a cross-domain interaction module to generate a space-frequency domain fusion feature image based on the coding feature image and the high-frequency component feature image, utilizing a dual-domain cooperative enhancement module to generate a dual-domain cooperative enhancement feature image based on the frequency domain enhancement feature image and the space-frequency domain fusion feature image, utilizing a multi-stage feature decoding module to generate a high-resolution segmentation feature image based on the dual-domain cooperative enhancement feature image, utilizing a loss optimization module to train a segmentation model to obtain an optimal segmentation model, and segmenting an input image. The invention further improves the accuracy of the intravascular ultrasound image segmentation.

Inventors

WANG YONGSHUN
LIU JINGJIN
LIU HUADONG
YIN HAN
HUANG JIECHENG

Assignees

深圳市人民医院

Dates

Publication Date: 20260508
Application Date: 20250425

Claims (10)

1. The intravascular ultrasound image segmentation method based on the frequency domain-airspace cooperative enhancement is characterized by comprising the following steps of: Step S1, constructing a layered feature extraction module, and extracting a coding feature map of an intravascular ultrasound image through a multi-level coding structure , wherein, The number of levels of the coding feature map; Step S2, constructing a multi-frequency characteristic decoupling module, and coding the characteristic diagram of each level Performing two-dimensional discrete wavelet transformation to generate low-frequency component characteristic diagram Horizontal high frequency component feature map Vertical high frequency component profile And diagonal high frequency component feature map Generating a frequency domain enhanced feature map through channel dimension splicing and convolution layer fusion ; Step S3, constructing cross-domain interaction modules, and respectively mapping the horizontal high-frequency component characteristic images Vertical high frequency component profile And diagonal high frequency component feature map Mapping to query sequence, coding feature map Mapping into key and value sequence, and generating a space-frequency domain fusion feature map by fusing airspace semantic features and frequency domain detail features through a high-frequency guidance-based cross attention module ; S4, combining the multi-frequency characteristic decoupling module and the cross-domain interaction module into a dual-domain collaborative enhancement module, and carrying out frequency domain enhancement on the characteristic map Fusing the feature map with the space-frequency domain Element-by-element addition for fusion to generate a dual-domain collaborative enhancement feature map ; S5, constructing a multi-level feature decoding module based on the dual-domain collaborative enhancement feature map Calculating to obtain a decoding characteristic diagram And based on the resulting decoded feature map Generating a high resolution segmentation feature map of a training image ; S6, forming a segmentation model based on frequency domain-space domain cooperative enhancement by the layered characteristic extraction module, the double-domain cooperative enhancement module and the multi-stage characteristic decoding module, constructing a loss optimization module, and segmenting a characteristic diagram based on high resolution of a training image Training the segmentation model by using the label image to obtain an optimal intravascular ultrasound image segmentation model; and S7, inputting an input image into the optimal intravascular ultrasound image segmentation model to obtain a segmentation result of the input image.
2. The method according to claim 1, wherein the step S1 comprises the steps of: s11, preprocessing intravascular ultrasound training images in a training set; Step S12, inputting the training image obtained after preprocessing into the layered feature extraction module to obtain a coding feature map of the training image 。
3. The method according to claim 1, wherein said step S2 comprises the steps of: step S21, coding feature diagram of the training image Performing two-dimensional discrete wavelet transformation to generate low-frequency component characteristic diagram Horizontal high frequency component feature map Vertical high frequency component profile And diagonal high frequency component feature map ; Step S22, splicing and fusing the low-frequency component, the horizontal high-frequency component, the vertical high-frequency component and the diagonal high-frequency component feature map according to the channel dimension, and passing through one The convolution layer performs dimension reduction to generate a frequency domain enhancement feature map 。
4. A method according to claim 3, wherein the two-dimensional discrete wavelet transform employs a hal wavelet basis function.
5. The method of claim 4, wherein the frequency domain enhancement profile Expressed as: , , , , , Wherein, the Representative of The layer of the convolution is formed from a layer of, Representing a channel dimension stitching operation, 、、 And Representing a haar wavelet low frequency component filter, a haar wavelet horizontal high frequency component filter, a haar wavelet vertical high frequency component filter and a haar wavelet diagonal high frequency component filter, respectively.
6. The method according to claim 1, wherein the step S3 comprises the steps of: step S31, a cross-domain interaction module is constructed, wherein the cross-domain interaction module comprises a layer standardization module, a cross attention module based on high-frequency guidance, a multi-layer perceptron and a residual error structure; Step S32, the coding feature map is processed Horizontal high frequency component feature map Vertical high frequency component profile And diagonal high frequency component feature map Input to cross-domain interaction module to generate space-frequency domain fusion feature map 。
7. The method of claim 6, wherein the space-frequency domain fusion profile Expressed as: , wherein, the Representing a cross-domain interaction module.
8. The method according to claim 1, wherein said step S5 comprises the steps of: step S51, for the final stage double-domain collaborative enhancement feature map Proceeding with Convolution processing to obtain final decoding characteristic diagram ; Step S52, decoding the feature map of the final stage Double upsampling to generate the first Hierarchical intermediate decoding feature map ; Step S53, the dual-domain collaborative enhancement feature map Double up-sampling and then Hierarchical intermediate decoding feature map Splice by channel dimension and pass Convolutional layer generation Hierarchical decoding feature map ; Step S54, repeating the steps S52 and S53 until a first-level decoding feature map is obtained Decoding the first-level decoding feature map By passing through Deconvolution of high resolution segmented feature maps with reduced vitamins in the convolutional layer 。
9. The method according to claim 1, wherein said step S6 comprises the steps of: Step S61, forming a segmentation model based on frequency domain-space domain cooperative reinforcement by the layered characteristic extraction module, the double-domain cooperative reinforcement module and the multi-stage characteristic decoding module; step S62, constructing a loss optimization module, and dividing the high resolution of the training image into feature images And the label image of the training image is input into the loss optimization module, and the segmentation model is optimized to obtain an optimal intravascular ultrasound image segmentation model.
10. The method of claim 9, wherein the loss function of the loss optimization module Expressed as: , Wherein, the Representing the total number of pixels in the training image, The number of categories is indicated and, As a display function, represent the first training image The pixel belonging to the first The true labels of the individual categories are presented, Representing the first of the training images The pixel belonging to the first Prediction probabilities for individual categories.

Description

Intravascular ultrasound image segmentation method based on frequency domain-airspace cooperative enhancement Technical Field The invention belongs to the fields of medical image processing, computer vision and artificial intelligence, and particularly relates to an intravascular ultrasound image segmentation method based on frequency domain-airspace cooperative enhancement. Background Intravascular ultrasound image segmentation plays an important role in the diagnosis and treatment of cardiovascular diseases, and can provide detailed information of blood vessel walls and plaques, thereby helping doctors to accurately evaluate and intervene. Intravascular ultrasound image segmentation is a key task in medical image analysis, wherein segmentation of the intravascular ultrasound intima and media can extract the boundaries of the intima and media from the image, facilitating diagnosis and treatment of related cardiovascular diseases. Intravascular ultrasound image segmentation can provide finer spatial distribution information than methods that provide only global image information, providing important support for clinical decisions. However, intravascular ultrasound image segmentation faces challenges such as noise interference, low image resolution, complex plaque morphology, etc. Accordingly, researchers have proposed various methods to solve these problems. These methods are mainly classified into two types, a conventional image processing-based method and a deep learning-based method. Methods based on traditional image processing rely mainly on low-level features of the image and classical image processing techniques. Jodas et al propose a full-automatic intravascular ultrasound image lumen segmentation method that can effectively identify the lumen region in the coronary intravascular ultrasound image by combining K-means clustering, circularity index and active contour model. Hammouche et al propose a 3D spiral active contour model based method for lumen segmentation of an intravascular ultrasound image that allows for identification of the lumen region in the intravascular ultrasound image by combining spiral geometry with Rayleigh distribution. However, the conventional method has a certain limitation in handling the complex situation due to the complexity of the intravascular ultrasound image and noise interference. The deep learning-based method adopts a deep learning model such as a convolutional neural network and a transducer, and can automatically learn image characteristics and perform end-to-end segmentation. Yang et al propose DPU-Net networks for intravascular ultrasound image segmentation based on the UNet structure using a multi-scale convolution layer at the feature extraction module to segment the lumen and the outer elastic membrane. Chen et al propose TransUNet a convolutional neural network in combination with a transducer, which segments images in combination with the feature that the transducer can capture global context information. Such methods are excellent in accuracy and robustness, but perform poorly in noise interference, plaque morphology complexity, and segmented edge regions. Disclosure of Invention The invention aims to excavate frequency information, enhance the difference between categories, make the boundary between each category clearer, and simultaneously combine the space domain and frequency domain information of images, enhance the capturing capability of the model on edges and details, thereby improving the segmentation effect. For this reason, the invention provides a method for segmenting an intravascular ultrasound image based on frequency domain-space domain cooperative reinforcement, which comprises the following steps: Step S1, constructing a hierarchical feature extraction module, and extracting a coding feature map F i (i=1, 2,., I) of an intravascular ultrasound image through a multi-stage coding structure, wherein I is the number of levels of the coding feature map; S2, constructing a multi-frequency characteristic decoupling module, performing two-dimensional discrete wavelet transformation on the coding characteristic map F i of each level, and generating a low-frequency component characteristic map Horizontal high frequency component feature mapVertical high frequency component feature mapAnd diagonal high frequency component feature mapGenerating a frequency domain enhancement feature map D i through channel dimension splicing and convolution layer fusion; step S3, constructing cross-domain interaction modules, and respectively mapping the horizontal high-frequency component characteristic images Vertical high frequency component feature mapAnd diagonal high frequency component feature mapMapping the code feature map F i into a key and value sequence, and generating a space-frequency domain fusion feature map S i by fusing space domain semantic features and frequency domain detail features through a high-frequency guidance-based cross attention modu