CN-121999275-A - Ischemic cerebral apoplexy fundus image classification method based on multi-scale blood vessel characteristics

CN121999275ACN 121999275 ACN121999275 ACN 121999275ACN-121999275-A

Abstract

The invention discloses an ischemic cerebral apoplexy fundus image classification method based on multi-scale blood vessel features, which comprises the steps of gradually extracting the multi-scale blood vessel features of fundus images through a feature extraction module, gradually fusing the blood vessel features of different scales through a feature fusion module, and realizing normal and ischemic cerebral apoplexy fundus image classification according to the fused multi-scale blood vessel features through a classification recognition module. According to the invention, the capability of extracting the blood vessel micro morphological characteristics of the fundus image is enhanced through the asymmetric convolution block, the vascular lesion area is positioned on the channel and space dimensions by utilizing the coordinate attention mechanism, and the deep semantic characteristics and the shallow detail characteristics are effectively combined by combining a multi-level characteristic fusion strategy, so that the more accurate ischemic cerebral apoplexy fundus image classification is realized.

Inventors

QIU CHENG
LIANG WEI
LI YUNTAO
LI YONGJIE
TAN YING
MA HONGGANG
TANG HAIYAN

Assignees

湖州市中心医院
电子科技大学长三角研究院(湖州)

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (5)

1. A method for classifying ischemic cerebral apoplexy fundus images based on multi-scale blood vessel characteristics is characterized by comprising the following steps: Step S1, gradually extracting blood vessel features of a plurality of scales of fundus images through a feature extraction module, wherein the feature extraction module is integrated with an Asymmetric Convolution Block (ACB) and is used for enhancing the extraction capability of micro morphological features of fundus blood vessels; Step S2, gradually fusing the blood vessel features with different scales extracted in the step S1 through a feature fusion module, wherein the feature fusion adopts a multi-level feature fusion strategy for effectively combining deep semantic features with shallow detail features; S3, dimension reduction is carried out on the multi-scale blood vessel characteristics obtained in the step S2; and S4, classifying the normal and ischemic cerebral apoplexy fundus images through a classification and identification module according to the multi-scale blood vessel characteristics obtained in the step S3.
2. The method for classifying ischemic cerebral apoplexy fundus images based on the fundus image multi-scale blood vessel features of claim 1, wherein in the step S1, the feature extraction module comprises four levels of feature extraction submodules, each level of feature extraction submodule sequentially extracts shallow detail blood vessel features, middle shallow structure blood vessel features, middle layer structure blood vessel features and deep semantic blood vessel features of the fundus image, each level of feature extraction submodule comprises an Asymmetric Convolution Block (ACB), a coordinate attention mechanism and downsampling, and the three are sequentially connected in series.
3. The method for classifying ischemic stroke fundus images based on the multi-scale vascular features of the fundus images according to claim 2, wherein the Asymmetric Convolution Block (ACB) comprises an input feature coding branch, three parallel convolution branches, an output feature fusion branch and a residual connection branch, and is specifically implemented by the following steps: (1) Input feature coding, namely 1×1 convolutional coding is carried out on input, and the coding formula is as follows: x entry ＝Dropout(Act(BN(Conv1(x in )))) Wherein x in is ACB module input, conv represents convolution operation, BN represents normalization operation, act represents ReLU activation function, dropout represents regularization operation, and x entry is a coded feature map; (2) Parallel convolution extraction, namely three parallel convolution branches are respectively: the asymmetric convolution branches of the 'first 1×n convolution and then n×1 convolution' are given by the formula: b 1 ＝Act(BN(Conv n×1 (Act(BN(Conv 1×n (x entry )))))) the asymmetric convolution branches of 'n x1 convolution followed by 1 x n convolution' have the formula: b 2 ＝Act(BN(Conv 1×n (Act(BN(Conv n×1 (x entry )))))) the standard convolution branch of "n×n convolution" has the formula: b 3 ＝Act(BN(Conv n×n (x entry )))) wherein n is an integer greater than 1; (3) Output feature fusion, namely splicing three branch output feature graphs, and then fusing by 1X 1 convolution, wherein the formula is as follows: x exit ＝Dropout(BN(Conv1(Concat(b 1 ,b 2 ,b 3 )))) (4) And residual connection, namely adding the residual branch with the fusion feature map, wherein the formula is as follows: x acb ＝Act(x exit +x res );x res ＝BN(Conv1(x in )) Wherein x res is a residual branch output feature map, and x acb is an ACB module final output feature map.
4. The ischemic cerebral apoplexy fundus image classification method based on the fundus image multi-scale blood vessel features of claim 2 is characterized in that the coordinate attention mechanism is connected in series behind an Asymmetric Convolution Block (ACB) and is used for positioning a blood vessel lesion region in channel and space dimensions and enhancing corresponding features, and specifically comprises the steps of generating attention weights in height dimension and width dimension for a feature map output by the Asymmetric Convolution Block (ACB), multiplying the attention weights by original feature map element by element to obtain an output feature map for enhancing the feature of the lesion region, wherein the formula is as follows: x enhanced ＝x raw ⊙A h ⊙A w Wherein x raw is the original feature map output by the Asymmetric Convolution Block (ACB), a h 、A w is the attention weight of the height and width dimensions, and x enhanced is the enhanced feature map.
5. The method for classifying the ischemic cerebral apoplexy fundus image based on the fundus image multi-scale blood vessel features of claim 1, wherein the multi-level feature fusion strategy of step S2 is characterized in that the feature images are sequentially subjected to bilinear interpolation and up-sampling from deep layers to shallow layers and spliced with the corresponding shallow layer feature images, and finally a fused multi-scale blood vessel feature image is obtained, wherein a feature fusion formula is as follows: F k ＝Concat(Up(F k+1 ),feat k ) Wherein, F k+1 is a feature map to be upsampled (deep feature) of the k+1th layer, feat k is a shallow feature map of the k layer, concat represents feature stitching operation, F k is a feature map after the k-th level fusion, and Up represents bilinear interpolation upsampling operation.

Description

Ischemic cerebral apoplexy fundus image classification method based on multi-scale blood vessel characteristics Technical Field The invention belongs to the technical field of image recognition and classification, and particularly relates to an ischemic cerebral apoplexy fundus image classification method based on multi-scale blood vessel characteristics. Background Ischemic stroke is a syndrome of ischemic and hypoxic necrosis of brain tissue caused by brain local blood supply disorder, and has the characteristics of high morbidity, high recurrence rate, high disability rate and high death rate. The retinal blood vessel and the cerebral blood vessel are homologous, are the only micro blood vessels which can be observed on living bodies, the morphological and functional changes and the cardiovascular and cerebrovascular diseases can be mutually proved, the morphological and structural changes of the retinal blood vessel (such as vascular stenosis, tortuosity, microaneurysms and the like) are closely related to the pathogenesis of the ischemic cerebral apoplexy, and can be used as the important characteristic basis of the ocular fundus image classification of the ischemic cerebral apoplexy. The existing ischemic cerebral apoplexy fundus image classification method is mostly dependent on a traditional image processing or conventional deep learning model, and has the technical defects that fundus blood vessel scale difference is obvious, micro capillaries with diameters of only 1-2 pixels are included, a traditional convolution module is limited in receptive field, the micro morphological characteristics are difficult to fully extract, the leak of an ischemic cerebral apoplexy related fundus image is caused, the characteristic area positioning accuracy is insufficient, the conventional attention mechanism is easy to lose space position information, the vascular characteristic area related to the ischemic cerebral apoplexy cannot be focused accurately, the interference of background noise is large, the multi-scale characteristic fusion strategy is imperfect, shallow detail characteristics (such as blood vessel edges and branches) and deep semantic characteristics (such as an integral characteristic mode related to diseases) are difficult to be combined effectively, and the classification accuracy is limited. Disclosure of Invention In order to solve the defects in the prior art and achieve the aim of classifying the ocular fundus images of the ischemic cerebral apoplexy more accurately, the invention adopts the following technical scheme: an ischemic cerebral apoplexy fundus image classification method based on multi-scale blood vessel characteristics comprises the following steps: Step S1, gradually extracting vascular features of multiple scales of fundus images through a feature extraction module, wherein the feature extraction module is integrated with an Asymmetric Convolution Block (ACB) and used for enhancing the extraction capability of micro morphological features of fundus blood vessels, and the feature extraction module is used for positioning a vascular lesion region in channel and space dimensions by utilizing a coordinate attention mechanism. Step S2, gradually fusing the blood vessel features with different scales extracted in the step S1 through a feature fusion module, wherein the feature fusion adopts a multi-level feature fusion strategy for effectively combining deep semantic features with shallow detail features; And S3, performing dimension reduction on the multi-scale blood vessel characteristics obtained in the step S2. And S4, classifying the normal and ischemic cerebral apoplexy fundus images through a classification and identification module according to the multi-scale blood vessel characteristics obtained in the step S3. Further, in the step S1, the feature extraction module includes four-stage feature extraction sub-modules, each stage feature extraction sub-module sequentially extracts shallow detail blood vessel features, middle shallow structure blood vessel features, middle structure blood vessel features and deep semantic blood vessel features of the fundus image, each stage feature extraction sub-module includes an Asymmetric Convolution Block (ACB), a coordinate attention mechanism and downsampling, and the three are sequentially connected in series. Further, the Asymmetric Convolution Block (ACB) includes an input feature encoding branch, three parallel convolution branches, an output feature fusion branch, and a residual connection branch, and the specific implementation process is as follows: (1) Input feature coding, namely 1×1 convolutional coding is carried out on input, and the coding formula is as follows: xentry＝Dropout(Act(BN(Conv1(xin)))) Wherein x in is ACB module input, conv represents convolution operation, BN represents normalization operation, act represents ReLU activation function, dropout represents regularization operation, and x entry is a coded feature