Research On Segmentation Algorithm In Pathological Retinal Optical Coherence Tomography

Diabetes is a chronic condition that can cause deformation of the retina and accumulation of fluids and increase the risk of blindness.Optical coherence tomography(OCT)is a high-resolution,non-invasive imaging modality that has become one of the most commonly used diagnostic imaging techniques for ophthalmology.Therefore,it is important to monitor the retinal morphological changes and fluid accumulation in diabetic patients.In this problem,the segmentation of the retinal layer and the liquid region is an important step in detecting changes.However,manual segmentation is often a timeconsuming and subjective process.For this problem,researchers have proposed many semi-automated,automated segmentation methods.Segmentation of the retinal layer and the liquid region in the OCT image is a challenging task due to deformation of the retinal layer and accumulation of liquid.This thesis studies the retinal layer structure and fluid region segmentation in OCT images.In general,the lesions deforme the retina structure and lead to inaccurate segmentation.In this thesis,we propose an improved shortest path algorithm with backtracking to segment the layer structures affected by diseases.The three weights of gray,gradient and Gabor information are combined to calculate the weight map,and the dynamic direction consistent loss is defined to determine the propagation direction of the algorithm.Finally,the backtracking idea is used to replace the traditional shortest path algorithm to avoid the short path problem.The experimental results show that our method has achieved good results.Some methods have been proposed for the segmentation of fluid regions in the retina with lesions,but they often have problems such as inaccurate segmentation or excessive computation.In this thesis,we conducted some studies of fluid region segmentation techniques.Firstly,a patch-based convolutional neural network segmentation method is proposed.This method transforms the segmentation problem into a two-class problem and trains a patch-based convolutional neural network for pixel classification.The segmentation result is better than the traditional threshold segmentation method and Kclustering method.In the above method,there is a problem of connecting several separated fluid regions into one whole,and therefore a second U-NET based on segmentation method is proposed.In this method,the attention mechanism is introduced to automatically locate the fluid region,which avoids the problem of excessive calculation of multiple networks.At the same time,the use of dense cross-connections to combine high-level and low-level features makes the segmentation result more accurate.Joint losses are used in optimaization,including cross entropy loss,dice loss,and regression loss,where regression losses are used to avoid the problem of connection several separated fluid regions into one.The experimental results show that the proposed method can adapt to OCT images from different decices,and the proposed method can outperform in segmentation accuracy than other state-of-the-art methods.