Research On Retinal Layer Segmentation In Optical Coherence Tomography Images

Optical Coherence Tomography(OCT)imaging can obtain high-resolution cross-sectional scans of the retina,and can be used to evaluate changes in retinal layer thickness.In recent years,automatic retinal layer segmentation method based on deep learning have been proposed to assist clinical diagnosis.However,most of these methods cannot handle retinal tissues affected by severe lesions well.In addition,whether it is based on traditional machine learning or deep learning-based layer segmentation methods,a large number of labeled images are required as training samples.The manual annotation of retinal OCT images requires a wealth of time,which poses a huge challenge to the construction of layer segmentation algorithms.This thesis proposes two different deep learning-based segmentation methods to solve the above problems.(1)Aiming at the topology errors occur in the deep learning-based segmentation methods,this thesis proposes the use of the distance map of the layer boundary to convert the segmentation task into a multi-task problem with classification and regression,and obtains the final topology-preserving result through a fusion module.To deal with the different difficulty of segmenting various layers,a confidence network is introduced as a hard-or-easy recognizer to provide the online training guidance for the segmentation network.In addition,inspired by the Non-local neural network,this thesis tailored a novel focus-column module for retinal layer segmentation to capture the long-range dependence between layers.The proposed method is investigated on an OCT dataset with slight deformation and two datasets with severe deformation.The experimental results show that the proposed method achieves decent performance on OCT B-scans.(2)In order to overcome the shortage of labeled OCT images and the problem that the trained model cannot handle new types of images well,this thesis proposes a few-shot learning-based layer segmentation method to predict the layers in OCT images affected by different lesions.In view of the fact that the layer distribution of multi-class OCT images is consistent,the cosine similarity loss is used to align the low-level features in the encoders of two arms,so that the segmenter can better integrate the task representation in conditioner.Also,a feature-fusion module is designed to select the effective part of the task representation adaptively.Finally,to make the model pay more attention to the deformation area caused by retinopathy,this thesis introduces a weighted guidance based on the class activate map to selectively focus on the deep semantic features in segmenter.The proposed few-shot learning-based method is verified on a public dataset with multiple types of lesions.The experimental results show that when the new category of images has only a few annotations,the proposed method can also achieve decent performance.