Automatic Segmentation Of Vessels In Retinal OCTA Images Based On Semi-Supervised Learning

Optical Coherence Tomography Angiography(OCTA)is an advanced non-invasive vascular imaging technology.The automatic segmentation of retinal OCTA image vascular network plays an important role in the early diagnosis and progression assessment of many vision-related diseases.However,most of the existing methods for automatic segmentation of blood vessels in OCTA images rely heavily on pixel-level labeled data,while OCTA images usually have low contrast and it is very difficult to annotate rich retinal blood vessels.Therefore,blood vessel labeling in retinal OCTA images is expensive and very scarce.In addition,due to the complex morphological structure of retinal vascular network,different sizes and overlapping,the segmented results tend to be severely missing and broken,which greatly harms the structural integrity and segmentation accuracy.To solve these mentioned problems,this thesis mainly studies how to reduce the dependence on labeled data and enhance the connectivity of segmented vessels to maintain the integrity of vascular network topology.The main contents of this research are as follows.(1)For the problem of difficulty in annotating and scarce labeled data of OCTA images,a dual-consistency semi-supervised OCTA vessel segmentation method DCSS-Net is proposed combined with self-supervised tasks.This method uses a small amount of labeled data and a large amount of unlabeled data to properly train the network,which can take full advantages of unlabeled data and reduce the dependence on labeled data.Specifically,combining self-supervised and semi-supervised,a novel multi-scale self-supervised jigsaw task is designed to facilitate semi-supervised models to learn better feature representations.Then,to avoid network overfitting,a dual-consistency regularization strategy based on image transformation consistency and feature perturbation consistency is proposed to obtain better generalization ability and generate more accurate segmentation predictions.Experiments on two OCTA datasets demonstrate the superiority of DCSS-Net,which outperforms other semi-supervised segmentation methods.(2)For the problem of inaccurate segmentation caused by poor connectivity of segmented vessels,a skeleton graph-based dual-task consistent OCTA semi-supervised vessel segmentation method ETSS-Net is proposed.ETSS-Net properly combines deep learning with traditional algorithms,using deep neural networks to achieve the overall vessel segmentation task,and traditional fast refinement algorithms to achieve the task of skeleton extraction.Then,the consistency loss based on the overall vessel segmentation task and the consistency loss based on the vessel skeleton extraction task are used to jointly optimize the network.While paying more attention to the vessel structure,double-consistency training is performed on the model to make full use of unlabeled data to improve the connectivity and accuracy of vessel segmentation.Experiments on two OCTA datasets show that the introduction of a vascular skeleton significantly improves the performance of the model,and the connectivity of blood vessels is greatly improved.Compared to other semi-supervised methods,ETSS-Net achieves the best performance and is comparable to fully supervised methods.