Resolution Enhancement And Quantitative Analysis Methods For Optical Coherence Tomography

Many diseases including diabetes,malignant tumors,glaucoma have remarkable changes on microvascular status.By observing the high-resolution images of diseases,doctors can take steps to avoid the disease becoming more serious.Optical coherence tomography(OCT)employs the interference characteristics of optical signals to generate 2D or 3D structure images of tissues.Because of the advantages of noninvasive,non-contact,non-destructive detection,high resolution,low cost,fast imaging speed,OCT has been applied in the diagnosis of cancer,retina,skin-related diseases.The high-resolution OCT image depends on the high-precision imaging equipment.The OCT image created by low precision imaging equipment is low resolution and difficult to be segmented.It is a challenge to quantitative analysis of these OCT images.In this paper,we propose OCT image enhancement,segmentation,quantitative analysis methods based on deep learning.The main contents of this paper include:(1)The OCT images of clinical diagnosis aren’t real-time and own low resolution.To solve this problem,a deep learning OCT resolution enhancement architecture for arbitrary imaging depth is introduced.The architecture is aimed to make use of the structure similarity of arbitrary imaging depth,full depth projection maps.By using residual attention in residual attention block and external attention,the architecture greatly improves the resolution of arbitrary imaging depth projection maps.As shown in the experiments of the clinical skin data set and publicly access retina OCT data set,the proposed method outperforms the traditional interpolation method and image super resolution methods based on convolutional neural network.And the estimated images outperform low resolution images on vessel quantitative micro-vascular analysis.(2)The complex vessel structure is hard to be segmented on OCT images.In this paper,a vessel segmentation architecture based on self attention is introduced.The architecture is aimed to construct global attention of vessel images.By employing the convolutional neural network(CNN)as CNN feature extractor,Transformer layers as encoders,the vessel images are accurately segmented.The proposed method is firstly tested on four publicly access fundus image data set,and has advantage on sensitivity,specificity,accuracy,auc-area,F1-score.Furthermore,the proposed method is validated on clinical skin data set and publicly access OCT dataset,and has excellent performance on vessel connectivity and classification accuracy of vessel pixels comparing.(3)The vessels on OCT images are hard to be analysed.To solve this problem,vessel micro-vascular analysis indices including vessel area density,vessel skeleton density,vessel complexity index,vessel premier index,vessel diameter index are designed.The experiments on skin disease nevus flammeus prove the reliability of quantitative indices.Visualization software is designed to integrate vessel enhancement,segmentation,quantitation algorithms.