Classification Technology Of Retinal OCT Images Based On Feature Learning

Optical coherence tomography(OCT)is a non-contact,noninvasive and high-resolution imaging technology,which becomes one of the main imaging technology for ophthalmic diagnosis at present.However,in front of the increasing requirement for OCT imaging,many ophthalmologists have always been stuck in an analysis of a big mass of image data,which is not only demanding for expert graders,but is also extremely time-consuming.In this case,the labor cost of ophthalmic diagnosis is greatly increased,and the patients cannot receive effective treatment timely.Therefore,it is necessary to develop an automatic tool for the retinal OCT image classification.There exist two key components for the classification technology approach:feature extraction and classifier design.In this dissertation,we focus on how to extract representative information from the OCT images for classification.Specially,we present a feature learning based classification method,termed as the principal component analysis network and support vector machine,for the classification of retinal OCT images.The main contents of this dissertation are summarized as follows:1.Classification technology of retinal OCT B-scans based on PCANet feature learning.Firstly,each 2D B-scan slice is donoised,flattened and cropped.Secondly,the PCANet is utilized to learn features from the training samples and the learned features are used to train SVM classifier.Finally,each OCT B-scan is identified by SVM classifier.2.Classification technology of retinal OCT volumes based on 3D-PCANet feature learning.Since the clinically acquired OCT image is usually a big volumetric data,it is reasonable and meaningful to classify the whole OCT volumes.To be exactly,the approach presents a modified 3D-PCANet on the basis of PCANet to learn high-level,effective features from OCT volumes.In addition,integrated the SVM based on composite kernel to classify OCT volumes.3.Classification software system of retinal OCT images.The software system is programmed based on MATLAB 2014b programming platform,using GUI application framework.The software system can achieve the classification of retinal OCT B-scans and OCT volumes based on PCANet and 3D-PCANet,respectively.To demonstrate the effectiveness of the proposed techniques,experiments are performed on the Duke and HUCM(Hunan University of Chinese Medicine)dataset.The experimental results demonstrate that the proposed approach can effectively extract more representative information from OCT images,especially,the superiority of the modified 3D-PCANet feature learning in OCT volumes classification.