A Study Of Classification And Segmentation Of Macular Optical Coherence Tomography Images Based On Deep Learning

The macula of retina takes charge of fine-grained vision.Optical Coherence Tomography(OCT)is an imaging technique which generates high-resolution images of optical scattering media by non-invasive penetration,which is widely used to generate three-dimensional images of macula for retinal pathology diagnosis.The macular OCT images are usually numerous and noisy,which makes analysing them a tedious and timeconsuming job.Therefore,fully automated macular OCT image analysis techniques are of great clinical values.In this paper,we explore deep learning based classification and segmentation methods of macular OCT images.In particular,1.For classification of 3D macular OCT images,we propose self-supervised iterative refinement learning based OCT image classification based algorithms.Selfsupervised iterative refinement learning introduces a relabeling stage to sieve suspicious 2D B-scan labels in addition to the traditional training stage,and consistently refines the 2D B-scan label set by iteratively repeating training and relabeling stages.To match up with self-supervised iterative learning,we also propose a new volumelevel inference strategy instead of using majority voting.Experimental results on two datasets show that using Res Net-101 as baseline classification model,our proposed method effectively improves the classification accuracy of the baseline model.2.For joint segmentation of retinal layers and lesions in 2D macular OCT images,we introduce deep supervision and adaptive Dice loss function aiming at better optimization for convolutional neural networks based segmentation algorithms.By adding extra output layers and loss functions on the intermediate feature layers of the network,deep supervision modules ensure that the weights of those shallow layers can be better trained.Adaptive Dice loss weights the classes adaptively based on the frequencies of them,which makes the network focus more on those rarely seen classes.Experimental results on a public dataset show that our proposed method effectively boosts the segmentation accuracy of the baseline U-Net model.