Quantitative Analysis And Applicantional Research On Age-Related Macular Degeneration Based On Deep Learning

Age-related macular degeneration(AMD)has become the main cause of irreversible blindness among middle-aged and elderly people in developing or developed countries.AMD is responsible for about 80 percent of severe visual loss.Visual loss leads to poor selfcare ability and mood depression,seriously affecting the quality of life of middle-aged and elderly people.Geographic atrophy(GA)and choroidal neovascularization(CNV)are the two main manifestations of the advanced AMD.Understanding the pathological features and symptoms of AMD can help guide patients to seek medical attention and prevent further vision loss.At present,spectral domain optical coherence tomography(SD-OCT)and optical coherence tomographic angiography(OCTA)can image the static tissue structure and dynamic blood flow signals in the retina respectively.Due to the characteristics of real-time,non-invasive,high sensitivity and high resolution,they have been widely used in clinical examination,diagnosis and treatment of the retina.Using machine learning and image processing methods to perform quantitatively analysis and applicational study of AMD in OCT imaging,and realize the auxiliary diagnosis and treatment of AMD,has important clinical significance.Based on SD-OCT images and OCTA images,focusing on quantitative analysis and practical clinical application of AMD,this paper studies the key issues of retinal layer segmentation,advanced AMD segmentation,GA progression prediction,CNV treatment response prediction and semi-supervised diseases screening.The main work of this paper is summarized as follows:(1)Multi-modal OCT based robust retinal layer segmentation model is proposed.In the process of region coding,the proposed model breaks through the limitation of using only SD-OCT images,firstly introduces OCTA images to assist region coding,and successfully improves the layer segmentation performance.In this model,a multitask layer-wise recoding module is proposed to weaken the sensitivity of the model to retinal diseases,and a layer surface coding module is designed to completely remove the missegmentation and protect the inherent properties of the retinal layer structure.Validation results on data sets with complex retinopathy of various types show that the proposed model can achieve high precision and robust layer segmentation results under the influence of various retinopathy.(2)Multi-scale fusion and boundary gradient protection based AMD segmentation model is proposed.For the problems that AMD have large scale variation and boundary is hard to define,the proposed model uses multiple parallel multi-scale branch to extract high-dimensional image features with different scales from SD-OCT image.In the process,multi-scale information of adjacent branch is borrowed to raise their power of feature extraction.Finally,all multi-scale features are cascade for AMD segmentation.The proposed model designs boundary gradient maximum constraint to restrict the AMD boundary.The proposed model performs well both on CNV segmentation and GA segmentation.(3)Neighborhood dependency based time adaption GA progression prediction model is proposed.The proposed model use two follow-up visits of the same patient to predict any future GA locations by adjusting time factor.Firstly,SD-OCT cubes are aligned to the first follow-up visits.Then bi-directional long short-term memory network(Bi LSTM)models each A-scan with its adjacent A-scan and time factor is used to control the GA risk.Finally,simulated GA growth maps and 3D-UNet refine the GA prediction results to obtain the accurate future GA location.The basic performance evaluation and generalization evaluation verified the effectiveness of the proposed model in predicting the progression of GA in the future.(4)Lesion attention maps and multi-task consistency based CNV treatment response prediction model is proposed.The proposed model directly predicts the CNV volume in the next month after intravitreal anti-vascular endothelial growth factor(anti-VEGF)injection,without CNV segmentation and CNV registration.Firstly,the proposed model generates CNV attention maps with CAM method,and then extracts and fuses multi-scale global and local features to predict the treatment response.The proposed model also introduces the concept of multi-task learning that uses two paths to perform regression based CNV volume prediction and classification based CNV volume change trend prediction.Finally,we prepose a multi-task consistency-based loss to force the restrict the prediction results of two paths.The experimental results show that the proposed model can be effectively applied to the prediction of CNV treatment response,and the influence of different OCT devices and different types of CNV on the model performance is verified.(5)Multi-task self-supervision and contrastive learning based semi-supervised retinal diseases screening model is proposed.In view of the large amount of SD-OCT images and the annotation difficulty,the proposed model can learn effective image feature representations from a few-shot labeled SD-OCT images and a large number of unlabeled SD-OCT images for retinal diseases screening.The proposed model explores the effectiveness of the combination of multiple self-supervised strategies and presents category-wise contrastive learning to further improve the model performance.The proposed model achieves better classicaition accuracy than comparative methods using only 10% and 1% labels.The proposed model can be trained end-to-end.Selfsupervised strategies in model can be replaced and the model is compatible with other semi-supervised approaches that do not limit the model architecture.