Feature fusion-based multi-module unsupervised learning retinal vessel segmentation system

Abstract

The invention relates to a feature fusion-based multi-module unsupervised learning retinal vessel segmentation system. The system comprises an image denoising and enhancing module, a feature extraction and fusion module, a multi-module learning module and a synthesis and result analysis module, wherein the image denoising and enhancing module is used for denoising color fundus images and enhancingcontrasts of the color fundus images; the feature extraction and fusion module is used for extracting invariant moment features, Hessian moment features, Gabor wavelet features, phase equalization features and Candy edge operator features of pixels of the color fundus images, and fusing the features into feature vectors; the multi-module learning module is used for segmenting the feature vectorsof the pixels of the color fundus images into a plurality of modules and respectively clustering the modules; and the synthesis and result analysis module is used for synthesizing and comparing clustering results. The system disclosed by the invention is capable of covering the shortages that the training samples are difficult to obtain and the training time is long as supervised retinal vessel segmentation methods require experts to manually mark vessels.

Description

technical field [0001] The invention relates to the technical field of fundus image analysis, in particular to a multi-module unsupervised learning retinal vessel segmentation system based on feature fusion. Background technique [0002] Retinal vascular lesions provide important information for systemic diseases such as diabetes, hypertension, and cardiovascular diseases, which often cause changes such as branching, bifurcation, and bending of retinal blood vessels. For example, in hypertensive patients, retinal arteries cause intermittent constriction; diabetic retinopathy causes retinal vascular dilation and atresia, retinal edema, hemorrhage, exudation, and ischemic symptoms. It is of great significance to segment retinal vessels from color fundus images for early screening and auxiliary clinical diagnosis. [0003] At present, retinal vessel segmentation methods are mainly divided into two categories: one is a feature-based supervised learning method, such as using k-n...

AI Technical Summary

Problems solved by technology

Unsupervised learning methods are fast and simple, but are generally less accurate for retinal vessel segmentation than supervised methods

Images