Research On Layer Segmentation Algorithms In Retinal Optical Tomography Images

Optical coherence tomography(OCT)is a high-resolution and non-invasive imaging modality that has become one of the most prevalent techniques for ophthalmic diagnosis.Many diseases,such as age-related macular degeneration(AMD),Stargardt disease(SD),retinitis pigmentosa(RP),and diabetic macular edema(DME),may cause anatomical changes in the retinal layer.Without treatment,these diseases might increase the risk of blindness,thus it is important to monitor the morphology change of the retinal layer and fluid accumulation for diabetes patients.In this problem,the segmentation of the retinal layer and the fluid region is an important step.However,manual segmentation is often a time-consuming and subjective process.For this problem,researchers have proposed many semi-automated and automated segmentation methods.Due to the existence of deformation and fluid accumulation,retinal layer and fluid region segmentation in OCT image is a challenging task.Based on this background,the thesis proposes a deep learning based algorithm for segmenting retinal layer and a deep learning based algorithm for jointly segmenting retinal layer and fluid region respectively.In general,OCT B scan images may contain noise,and the contrast between different layers may not be significant.Furthermore,the presence of retinal layer deformation makes it difficult to design hand-designed features.Therefore,for the segmentation of the retinal layer with lesions,this thesis uses the deep residual network to extract the discriminative features of the image,and combine with the hand-designed features as the final features.These features are used to train a computationally efficient structured random forest to classify the retinal edge and background,and finally the shortest path is adopted to achieve the final layer segmentation.The experimental results show that our method has achieved good results.In respect of the joint segmentation of the retinal layer and the fluid region,researchers have proposed many machine learning-based segmentation methods which rely on a large number of pixel-level annotation data.Although it is relatively simple to collect a large amount of image data,it is difficult to obtain manual annotations.In this thesis,we propose a new semi-supervised algorithm based on a full convolutional neural network to segment the retinal layer and fluid regions in the retinal OCT B-scan image.The method utilizes unlabeled data through a adversarial learning strategy,and the whole system is optimized by a new combined loss function.The proposed method was studied on the Diabetic Macular Edema(DME)dataset published by Duke University and the POne dataset.The experimental results show that the use of differential network and unlabeled data can improve the performance of segmentation,and the method performs better than other state-of-the-art Layer and fluid segmentation methods.