CN-121725008-B - Brain vascular endothelial segmentation method based on dual-coordinate attention mechanism

CN121725008BCN 121725008 BCN121725008 BCN 121725008BCN-121725008-B

Abstract

Aiming at the problems of endothelial boundary blurring, low contrast, obvious geometric deformation and the like in an optical coherence tomography image, the invention provides a dual-path feature extraction frame integrating Cartesian and polar coordinate characterization, which comprises the steps of realizing coordinate system self-adaptive conversion through a micropolar coordinate conversion module, designing a dual-path coordinate attention module, extracting one-dimensional direction descriptors along horizontal, vertical, radial and angular directions respectively, inputting an optimized feature map into a subsequent segmentation network, generating a four-path attention weight map through cross-domain fusion, realizing feature recalibration, and finally segmenting the image of the cerebral vascular endothelium. According to the invention, modeling of the brain vascular endothelial boundary blurring is realized by constructing the Cartesian-polar coordinate dual-path feature extraction frame and the cross-domain cooperative dual-path coordinate attention mechanism, and the segmentation precision of the endothelial structure in the OCT image is improved.

Inventors

LAN QUAN
DING HUINA
HUANG RONG
WU ZHAOYE
HUANG CHENXI
ZHANG LONGHUI

Assignees

厦门大学附属第一医院(厦门市第一医院、厦门市红十字医院、厦门市糖尿病研究所)
厦门智融合科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260225

Claims (4)

1. The brain vascular endothelial segmentation method based on the dual-coordinate attention mechanism is characterized by comprising the following steps of: s1, acquiring a cerebral blood vessel optical coherence tomography image, and performing center alignment, intensity normalization and data enhancement on the cerebral blood vessel optical coherence tomography image to obtain a preprocessed image; S2, synchronously inputting the preprocessed image into a Cartesian coordinate branch and a polar coordinate branch; the Cartesian coordinate branch is used for converting the preprocessed image from the Cartesian coordinate system to the polar coordinate system through a micro polar coordinate conversion module, extracting the features from the polar coordinate image through convolution operation and outputting a second feature map; s3, the first feature map and the second feature map are input to a dual-path coordinate attention module together, one-dimensional direction feature descriptors are extracted along the horizontal direction of a Cartesian coordinate system, the vertical direction of the Cartesian coordinate system, the radial direction of a polar coordinate system and the angular direction of the polar coordinate system respectively, and feature fusion crossing coordinate domains is carried out to generate attention weight maps corresponding to the four directions; the dual-path coordinate attention module is deployed at the encoder stage of the U-Net split network and inserted after the jump connection; s4, carrying out element-by-element multiplication on the first feature map and the second feature map based on the attention weight map to obtain an enhanced first output feature map and an enhanced second output feature map; And (3) carrying out element-by-element multiplication on the first feature map and the second feature map based on the attention weight map to calibrate the attention weight, wherein the calculation formula is as follows: ; ; Wherein, the And Respectively enhancing the first output characteristic diagram and the second output characteristic diagram; And The first characteristic diagram and the second characteristic diagram are respectively; 、、 And A row-direction attention view, a column-direction attention view, a radial attention view, and an angular attention view, respectively; s5, inputting the enhanced first output feature map and the enhanced second output feature map into a segmentation network for training and deducing, and finally outputting a segmentation result of the cerebral vascular endothelium; Quantitatively evaluating the segmentation result by using the regional overlapping degree and the boundary distance evaluation index, wherein the calculation formula of the regional overlapping degree is as follows: ; ; Wherein, the Dice represents the coincidence ratio of prediction and real labeling, P is a prediction segmentation pixel set, and G is a real labeling pixel set; a boundary point set representing a prediction segmentation result; Representing a set of truly annotated boundary points; Representation of A boundary pixel point in the image; Representation of A boundary pixel point in the image; representing euclidean distances between pixel points; after the nearest distances of all boundary points are sequenced, the distance value corresponding to the 95 th percentile is taken; indicating how aligned the prediction boundary is with the real boundary.
2. The brain vascular endothelial segmentation method based on the dual-coordinate attention mechanism according to claim 1, wherein in S2, the polar coordinate branches transform the preprocessed image from the cartesian coordinate system to the polar coordinate system by a micro polar coordinate transformation module, and the specific transformation manner is as follows: Determining the central position and the maximum radius of the preprocessed image in a Cartesian coordinate system, uniformly sampling in the radial direction by taking the central position as a pole and uniformly sampling in the angular direction by taking the range not exceeding the maximum radius in a polar coordinate domain, constructing a polar coordinate sampling grid, mapping each sampling point in the polar coordinate sampling grid back to the corresponding coordinate position in the Cartesian coordinate system according to the radial distance and the angle of each sampling point, and sampling from the corresponding coordinate position in the preprocessed image by bilinear interpolation to obtain a polar coordinate domain feature map.
3. The brain vascular endothelial segmentation method based on the dual-coordinate attention mechanism according to claim 1, wherein in S3, extracting a one-dimensional direction feature descriptor specifically comprises: carrying out global average pooling on the first feature map along the horizontal direction and the vertical direction of the first feature map to obtain a row direction descriptor and a column direction descriptor; And carrying out global average pooling on the second feature map along the radial direction and the angular direction of the second feature map respectively to obtain radial descriptors and angular descriptors.
4. The brain vascular endothelial segmentation method based on the dual-coordinate attention mechanism according to claim 3, wherein in S3, generating an attention weight map corresponding to four directions specifically comprises: splicing the row direction descriptor, the column direction descriptor and the radial descriptor, generating a first fusion characteristic through first lightweight bottleneck network processing, and splitting the fusion characteristic into a row direction attention force diagram and a column direction attention force diagram; And splicing the row direction descriptor, the column direction descriptor and the angular direction descriptor, and generating a second fusion characteristic through processing of a second lightweight bottleneck network, so that the fusion characteristic is split into a radial attention force diagram and an angular attention force diagram.

Description

Brain vascular endothelial segmentation method based on dual-coordinate attention mechanism Technical Field The invention relates to the technical field of image processing, in particular to a brain vascular endothelial segmentation method based on a dual-coordinate attention mechanism. Background Cerebrovascular disease is one of the leading causes of death and disability worldwide, with ischemic stroke accounting for up to 60% -80%. Accurate assessment of pathological changes such as vascular lumen stenosis is critical to early diagnosis and treatment of disease. Optical Coherence Tomography (OCT) is used as a high-resolution imaging technology, can clearly present the layered structure of the vascular wall with micron-scale precision, and provides possibility for fine observation of the vascular endothelium of the brain. However, endothelium in OCT images often appears thin, blurred in boundaries, and is susceptible to speckle noise, which is time-consuming and labor-consuming for manual segmentation and subject to subjective differences; currently, medical image segmentation methods based on deep learning, such as U-Net and its variants, have been successful in a variety of tasks. However, the direct application of these mainstream models to OCT brain vascular endothelial segmentation still faces challenges, firstly, most networks are designed in cartesian coordinate space, it is difficult to effectively fit the inherent radial symmetry and annular features of vascular structures, limiting the modeling capability of vascular wall continuity and directionality, and secondly, the existing attention mechanisms focus on feature enhancement of channels or single spatial domain (such as cartesian space) and lack explicit modeling of radial and angular features in polar coordinate domain, failing to fully mine complementary geometric information in two coordinate systems. Although research has been attempted to introduce polar transformation or channel attention to improve performance, how to cooperatively utilize the dual domain features and perform efficient cross-coordinate attention interactions remains a challenge. Therefore, a new technical scheme is urgently needed, the characterization advantages of Cartesian coordinates and polar coordinates can be deeply fused, and the accurate and robust segmentation of the cerebral vascular endothelium in the OCT image is realized through a more intelligent attention mechanism, so that reliable technical support is provided for clinical evaluation. Disclosure of Invention In order to solve the problems, the invention provides a brain vascular endothelial segmentation method based on a dual-coordinate attention mechanism, which realizes robust modeling of challenges such as brain vascular endothelial boundary blurring, low contrast, geometric deformation and the like by constructing a Cartesian-polar dual-path feature extraction frame and a cross-domain cooperative dual-path coordinate attention mechanism, and remarkably improves segmentation accuracy and boundary positioning accuracy of endothelial structures in OCT images. A brain vascular endothelial segmentation method based on a dual-coordinate attention mechanism comprises the following steps: s1, acquiring a cerebral blood vessel optical coherence tomography image, and performing center alignment, intensity normalization and data enhancement on the cerebral blood vessel optical coherence tomography image to obtain a preprocessed image; S2, synchronously inputting the preprocessed image into a Cartesian coordinate branch and a polar coordinate branch; the Cartesian coordinate branch is used for converting the preprocessed image from the Cartesian coordinate system to the polar coordinate system through a micro polar coordinate conversion module, extracting the features from the polar coordinate image through convolution operation and outputting a second feature map; s3, the first feature map and the second feature map are input to a dual-path coordinate attention module together, one-dimensional direction feature descriptors are extracted along the horizontal direction of a Cartesian coordinate system, the vertical direction of the Cartesian coordinate system, the radial direction of a polar coordinate system and the angular direction of the polar coordinate system respectively, and feature fusion crossing coordinate domains is carried out to generate attention weight maps corresponding to the four directions; s4, carrying out element-by-element multiplication on the first feature map and the second feature map based on the attention weight map to obtain an enhanced first output feature map and an enhanced second output feature map; S5, inputting the enhanced first output feature map and the enhanced second output feature map into a segmentation network for training and deducing, and finally outputting a segmentation result of the cerebral vascular endothelium. Further, in S2, the polar coordinate branch conve