CN-121999522-A - Retinal vessel segmentation method and device for fundus image System, storage medium

CN121999522ACN 121999522 ACN121999522 ACN 121999522ACN-121999522-A

Abstract

The invention discloses a fundus image retinal vessel segmentation method, a device, a system and a storage medium, which comprise the steps of S1, obtaining fundus image data to be segmented, preprocessing, S2, constructing a segmentation model based on receptive field attention and cross attention decoding according to the preprocessed fundus image data, and carrying out end-to-end feature extraction and pixel level classification on an image by utilizing the segmentation model, wherein the segmentation model comprises a RFAConv self-adaptive encoder, a self-attention bottleneck module, a feature enhanced jump connection module and a cross attention-based decoding module. By adopting the technical scheme of the invention, the defects of the prior art in the aspects of feature extraction flexibility and feature fusion effectiveness are overcome, and the edge accuracy and robustness of retinal vessel segmentation of fundus images are improved.

Inventors

LIU QIAOHONG
WANG JUNSHI
LIN MIN

Assignees

上海健康医学院
上海市同济医院

Dates

Publication Date: 20260508
Application Date: 20260202

Claims (8)

1. A retinal vessel segmentation method for fundus images, comprising: S1, acquiring fundus image data to be segmented, and preprocessing; And S2, constructing a segmentation model based on the decoding of receptive field attention and cross attention according to the preprocessed fundus image data, and carrying out end-to-end feature extraction and pixel level classification on the image by utilizing the segmentation model, wherein the segmentation model comprises a RFAConv self-adaptive encoder, a self-attention bottleneck module, a feature enhancement jump connection module and a decoder based on cross attention.
2. The fundus image retinal vessel segmentation method according to claim 1, wherein in step S1, the input medical image is set as , wherein, In order to provide the number of channels, , Height and width; generating two downsampled copies by bilinear interpolation during the preprocessing stage And : ; 。
3. The fundus image retinal vessel segmentation method according to claim 2, wherein in step S2, a joint loss function is employed Optimizing the segmentation model, and setting For a real tag, the total loss is a weighted sum of three scale losses: ; Wherein, the For the two-value cross entropy loss, In order to achieve the loss of the cross-over ratio, For the weight coefficient of each scale, Is a predictive probability map.
4. A retinal vessel segmentation apparatus for fundus images, comprising: The first processing module is used for acquiring fundus image data to be segmented and preprocessing the fundus image data; The second processing module is used for constructing a segmentation model based on the decoding of receptive field attention and cross attention according to the preprocessed fundus image data, and carrying out end-to-end feature extraction and pixel level classification on the image by utilizing the segmentation model, wherein the segmentation model comprises a RFAConv self-adaptive encoder, a self-attention bottleneck module, a feature enhancement jump connection module and a decoder based on cross attention.
5. The fundus image retinal vessel segmentation device according to claim 4, wherein the first processing module is configured to input a medical image as , wherein, In order to provide the number of channels, , Height and width; generating two downsampled copies by bilinear interpolation during the preprocessing stage And : ; 。
6. The fundus image retinal vessel segmentation device of claim 5 wherein the second processing module employs a joint loss function Optimizing the segmentation model, and setting For a real tag, the total loss is a weighted sum of three scale losses: ; Wherein, the For the two-value cross entropy loss, In order to achieve the loss of the cross-over ratio, For the weight coefficient of each scale, Is a predictive probability map.
7. A fundus image retinal vessel segmentation system comprising a memory and a processor, the memory having stored thereon a computer program for execution by the processor, the computer program when executed by the processor performing the fundus image retinal vessel segmentation method of any of claims 1-3.
8. A storage medium having stored thereon a computer program which, when run, performs the fundus image retinal vessel segmentation method according to any one of claims 1-3.

Description

Retinal vessel segmentation method and device for fundus image System, storage medium Technical Field The invention belongs to the technical field of image processing, and particularly relates to a retinal vessel segmentation method, a retinal vessel segmentation device, a retinal vessel segmentation system and a retinal vessel segmentation storage medium based on a receptive field attention convolution (RECEPTIVE-FieldAttentionConvolution, RFAConv) and jump connection characteristics, which are suitable for automatic segmentation and extraction of fine targets such as retinal vessels and lesion areas thereof. Background In recent years, a Convolutional Neural Network (CNN) of a codec (Encoder-Decoder) structure typified by U-Net has achieved remarkable results in fundus medical image segmentation tasks. However, existing segmentation methods still face serious challenges in treating fundus retinal blood vessels and their microscopic tumors. First, the convolution kernel parameters of the conventional convolution operation are shared in the whole image range during feature extraction, which means that the model adopts the same processing mode for features at different positions in the fundus medical image. However, fundus medical images tend to have a high degree of non-stationarity, with large differences in texture and shape from region to region, such as the distal bifurcation of retinal blood vessels and the thickness of the trunk. Standard convolution has difficulty in adaptively adjusting receptive field weights based on spatial variations of input features, resulting in insufficient capture capability for fine structures. Second, the U-Net architecture relies primarily on a jump connection (SkipConnection) to splice the shallow features of the encoder directly to the decoder. Existing jump connections typically employ simple channel splicing (Concatenation) or element-by-element addition. This approach, while introducing spatial information, also introduces background noise and redundancy features of the encoding stage. The importance of the foreground and background cannot be distinguished by a simple splicing operation, which results in a decoder that is prone to edge blurring or artifacts when resolution is restored. Furthermore, the upsampling process of existing models often lacks efficient semantic guidance, and it is difficult to maintain semantic consistency while restoring spatial resolution. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a fundus image retinal vessel segmentation method, a fundus image retinal vessel segmentation device, a fundus image retinal vessel segmentation system and a fundus image retinal vessel segmentation storage medium, which solve the defects in the aspects of feature extraction flexibility and feature fusion effectiveness in the prior art and improve the edge accuracy and robustness of fundus image retinal vessel segmentation. In order to achieve the above object, the present invention provides the following solutions: A retinal vessel segmentation method of fundus images, comprising: S1, acquiring fundus image data to be segmented, and preprocessing; And S2, constructing a segmentation model based on the decoding of receptive field attention and cross attention according to the preprocessed fundus image data, and carrying out end-to-end feature extraction and pixel level classification on the image by utilizing the segmentation model, wherein the segmentation model comprises a RFAConv self-adaptive encoder, a self-attention bottleneck module, a feature enhancement jump connection module and a decoder based on cross attention. Preferably, in step S1, the input medical image is set as, wherein,In order to provide the number of channels,,Height and width; generating two downsampled copies by bilinear interpolation during the preprocessing stage And: ; ; Preferably, in step S2, a joint loss function is usedOptimizing the segmentation model, and settingFor a real tag, the total loss is a weighted sum of three scale losses: ; Wherein, the For the two-value cross entropy loss,In order to achieve the loss of the cross-over ratio,For the weight coefficient of each scale,Is a predictive probability map. The invention also provides a retina blood vessel segmentation device of the fundus image, which comprises: The first processing module is used for acquiring fundus image data to be segmented and preprocessing the fundus image data; The second processing module is used for constructing a segmentation model based on the decoding of receptive field attention and cross attention according to the preprocessed fundus image data, and carrying out end-to-end feature extraction and pixel level classification on the image by utilizing the segmentation model, wherein the segmentation model comprises a RFAConv self-adaptive encoder, a self-attention bottleneck module, a feature enhancement jump connection module and a decoder based