The Segmentation Of Edema Region In Retinal OCT Image Based On CNN

Retinal edema is a common retinal disease and a complication of many ocular fundus diseases,including subretinal fluid(SRF),pigment epithelium detachment(PED),etc.Optical coherence tomography(OCT)technology is non-invasive and widely used in the ophthalmic clinic.Therefore,the automatic segmentation of retina edema area(REA)based on OCT images,including SRF and PED lesions,has significant impact for the clinical analysis,diagnosis and treatment of the related fundus diseasesIn this thesis,the convolutional neural network(CNN)with U-shaped structure based segmentation algorithms are explored and improved.The segmentation networks based on the deep supervised attention mechanism and the context pyramid fusion mechanism are respectively proposed,to realize the joint segmentation of REA,SRF and PED lesions in retinal OCT images.In the segmentation network based on the deep-supervised attention mechanism,inspired by the non-local operation,a deep-supervised attention module is designed and inserted to model the correlation between pixels and regional objects by connecting the rough segmentation results with the high-level semantic features,which can enhance the representation ability of pixels.In the segmentation network based on the context pyramid fusion mechanism,in order to solve the shortcoming of the lack of global context information caused by the common skip-connection and the lack of multi-scale context information capture ability in the encoder-decoder network,a global pyramid guidance(GPG)module is designed to reconstruct the skip-connection.A scale-aware pyramid fusion(SAPF)module is designed to dynamically capture multi-scale information,which can improve the context information understanding ability of the networkIn this thesis,the public dataset of retinal edema on the AI-Challenger website was used for algorithm verification and performance evaluation.This dataset contains 83 three-dimensional retinal OCT images with retinal edema.Each OCT image contains 128 two-dimensional B-scan images,totaling 10,624 two-dimensional images.The proposed segmentation networks were trained on two-dimensional images and tested on three-dimensional images.Dice coefficient,sensitivity,specificity and pixel accuracy were compared to evaluate the performance of the proposed segmentation networks.Experiments show that the two segmentation networks proposed in this thesis have achieved good segmentation performance.Among them,the model based on deep-supervised mechanism can promote the model optimization and improve the segmentation performance.The main evaluation index Dice coefficients of REA,SRF and PED were 81.04%,79.96%and 73.07%respectively.The model based on context pyramid fusion can focus on the segmentation targets and capture effective features,and the Dice coefficients of REA,SRF and PED reached 81.34%,83.49%and 74.72%respectively...