Strong generalization eye fundus image segmentation method based on semi-supervised learning

Abstract

The invention provides a strong generalization eye fundus image segmentation method based on semi-supervised learning, and the method comprises the steps: firstly carrying out the preheating training of a segmentation main network through employing a number with a label, enabling the two segmentation main networks to carry out the segmentation through employing an unsupervised method, and carrying out the comparison of the two segmentation main networks, the two segmentation networks segment the original fundus image and the disturbed image respectively without mutual communication, but the original image and the disturbed image have structural consistency. Therefore, the consistency criterion is taken as a constraint, loss errors of boundary diagrams and entropy diagrams of the two segmented main networks are solved each time, iteration is continuously carried out until the difference between segmentation results of the two segmented main networks is smaller than a specified iteration condition, and the structures of the two segmented main networks are approximately consistent. Through such iterative training, the recognition effect of the model on the interested segmentation region of the fundus image can be effectively improved, so that a high-precision segmentation model for segmenting the interested region of the fundus image is obtained.

Description

technical field [0001] The invention belongs to the technical field of computer vision and the technical field of medical imaging, and in particular relates to a method for segmenting a fundus image with strong generalization based on semi-supervised learning. Background technique [0002] Many systemic diseases such as hypertension, diabetes, etc. can occur fundus lesions, so fundus images are an important diagnostic data. Structural variation in retinal fundus images is one of the important indicators of certain diseases and is crucial for clinical diagnosis. For example, glaucoma disease can be detected by the ratio of optic cup (OC) to optic disc (OD); the segmentation of microaneurysm (MA), hemorrhage (HE), soft exudate (SE), and hard exudate (EX) can Grading identification of diabetic retinopathy and macular edema. [0003] In recent years, with the vigorous development of deep learning, especially the excellent performance of convolutional neural networks in pattern...

AI Technical Summary

Problems solved by technology

These distributional differences can cause deep networks to overfit on training datasets and lack generalization ability on unseen test datasets

The second is that these high-precision detection results rely on fine, high-quality pixel-level labels. In medical image analysis, obtaining high-quality labels for data is laborious and expensive, because accurate labeling of medical images requires the expertise of clinicians.

Images