Research Of Weakly Supervised Retinopathy Segmentation Algorithm Based On Deep Convolution Neural Network

Visual impairment seriously affects people’s production and life,and retinopathy is one of the important causes of visual impairment.It has important clinical significance and research value to use image processing methods to automatically identify and quantify the lesion areas in Spectral Domain Optical Coherence Tomography(SD-OCT)retinal images.Automated lesion segmentation is one of the important tasks for quantitative assessment of retinal diseases in SD-OCT image.Recently,deep convolutional neural networks(CNN)have shown its promising advancements in the field of automated image segmentation,whereas always benefit from the large-scale datasets with high quality pixel-wise annotations.Unfortunately,obtaining accurate annotations is expensive in both human effort and finance.Therefore,how to reduce the cost of labeling,so as to achieve effective automatic segmentation and quantification of lesions has become an important research topic.This thesis focuses on the automatic segmentation and quantification of neurosensory retinal detachment(NRD)in SD-OCT retinopathy images based on weakly supervised learning.The main contributions of this thesis are summarized as follows:(1)We propose a weakly supervised two-stage learning architecture to detect and further segment central serous chorioretinopathy(CSC)retinal detachment with only image-level annotations.Specifically,in the first stage,a Located-CNN is designed to detect the location of lesion regions in the whole SD-OCT retinal images,and highlight the distinguishing regions.To generate available pseudo pixel-level label,conventional level set method is employed to refine the distinguishing regions.In the second stage,we customize the activecontour loss function in deep networks to achieve the effective segmentation of the lesion area.(2)We proposed fine-to-coarse-to-fine weakly supervised learning framework for volumetric SD-OCT images segmentation.Firstly,we propose a global alternate max-avg pooling(GTP)network to locate the lesion region accurately by using only image-level annotations.Secondly,we designed a network module based on the GTP network and a semantic transfer module to iteratively transfer the attention of the network to the lesion region to continuously discover and expand the target lesion region.Furthermore,pseudo-volumetric labels can be generated using the 3D gray distribution histogram method.This process leverages a classification network to generate coarse lesion region labels from the fine category.Finally,a novel 3D-level set loss function was designed to achieve shape preservation in the characteristics of medical volumetric images to perform coarse-to-fine volumetric segmentation.