Automatic And Interactive OCT Image Segmentation Of Choroidal Neovascularization Based On Deep Learning

Choroidal neovascularization(CNV)is a typical clinical manifestation of agerelated macular degeneration(AMD),and it is also one of the main causes of vision loss or blindness in elderly groups all over the world.Optical coherence tomography(OCT)is an important fundus retinal imaging technology.Due to its imaging characteristics of non-invasive,non-contact and high-resolution,OCT has become a common tool of evaluating CNV for researchers and clinicians.The segmentation and quantification of CNV lesions help ophthalmologists discriminate the expansion or contraction of CNV and estimate the development trend of CNV.It has important reference significance for clinical judgment of activity,recurrence rate and curative effect of CNV.In order to obtain the quantitative results of CNV lesions,doctors need to outline the lesions manually.However,manual labeling is subjective and variable,and it costs doctors a lot of time and energy.Therefore,it is particularly important to develop a reliable computeraided segmentation method.Then,for the CNV with fuzzy or invisible boundary,large difference in shape,size and position,the existing CNV segmentation algorithm is difficult to meet the needs of clinical accuracy.Therefore,to further improve the accuracy and robustness of CNV segmentation,the following research work is carried out in this thesis:(1)Research on automatic CNV segmentation algorithm based on mixed multiscale feature fusion and attention.The network architecture of the algorithm is based on the"encoder-decoder" structure,and two optimization strategies are proposed to improve the segmentation performance of the model:mixed multiscale feature fusion module and attention gate module.The mixed multi-scale feature fusion module fully integrates different levels of complementary information by using multi-scale spatial information fusion structure and multi-scale channel information fusion structure respectively.The attention module is used to refine the multi-scale layer features,further enhance the recognition ability of the network,highlight the salient features and suppress the feature response of irrelevant regions.Experimental results show that the algorithm can effectively solve the difficulties of CNV segmentation.(2)Research on interactive CNV segmentation algorithm guided by extreme points.For the blurred or even disappeared lower boundary,slender or small CNV region,the segmentation results of automatic prediction will have the problems of overflow,false segmentation or under segmentation.In order to solve these problem,extreme point guidance is added to the segmentation network,and the region of interest is cropped according to the rectangular box surrounded by extreme points.Finally,the cropped OCT image is connected with the guidance signal channel and input them into the segmentation network.The experimental results show that the manual intervention effectively solves the problems encountered in automatic segmentation and greatly improves the accuracy and robustness of segmentation.(3)Research on improved interactive CNV segmentation algorithm based on Euclidean distance mapping.The Euclidean distance map is constructed by using the extreme box formed by the connection of extreme points,and connected with OCT image and Gaussian heatmap to form the three channel input of the network,which makes full use of the context information of user interaction.Experiments show that the improved interactive CNV segmentation algorithm achieves higher segmentation accuracy under limited user interaction.The algorithms proposed in this thesis is experimentally verified on a 487 SDOCT data-set obtained by the spectral OCT instrument(Heidelberg engineering,Germany)provided by the ophthalmology center of Jinan Central Hospital and Qilu Medical College of Shandong University.Through numerical analysis and visual effect analysis,it is comprehensively verified that the automatic CNV segmentation algorithm based on mixed multi-scale feature fusion and attention,the interactive CNV segmentation algorithm guided by extreme points and the improved interactive CNV segmentation algorithm based on distance mapping are better than the existing CNV segmentation algorithm and some other advanced segmentation algorithms.