Research On Optic Disc Localization And Segmentation Based On Deep Learning

The formation of some retinal fundus diseases is usually accompanied by changes in the optic disc and the optic cup.Therefore,accurate measurement of various parameters of the optic disc and optic cup is an important indicator for the diagnosis of fundus diseases,which is of great significance for the early screening,prevention and containment of fundus diseases.The traditional methods of localization and segmentation of the optic disc mainly by extracting the characteristics of the optic cup and optic disc,such as shape,brightness,and blood vessel orientation,and designing corresponding algorithms.These methods rely on the experience of relevant professionals to a certain extent,are affected by human factors,take a long time for data processing,and the accuracy of localization and segmentation is not ideal.This paper proposes a method of disc localization and segmentation based on deep learning,which is optimized in performance to a large extent.1.Combined with the characteristics of the region of interest of the disc and the actual task scenario,the basic network,multi-scale feature mapping and default frame matching strategy of the traditional SSD(Single Shot Multibox Detector)algorithm are improved,a lightweight and efficient improved SSD algorithm is proposed.Experiments show that the improved algorithm has achieved good experimental results on the task of localizing the region of interest of the optical disc: the localization accuracy on the two test sets are 100% and 99.6%,respectively,which are 1% and 6.6%higher than the traditional SSD algorithm.The AP value on the verification set is 0.81,which is 0.08 higher than the traditional SSD algorithm.The average training time on the training set are 0.35 s,respectively,which is 6.75 s less than the traditional SSD algorithm.2.An improved U-Net network model combining residual structure,attention mechanism and feature fusion is proposed,which realizes the joint segmentation of optic cup and optic disc.In the process of up and down sampling,the Residual block is used as the main body,and the SE Net module is used to reweight the feature channels after feature fusion,and the feature information of the upper and lower layers is fused in the decoding part to further accelerate the model convergence and integration.Improve the segmentation performance of the model.Finally,through experiments on the Drishti-GS fundus image data set,the experiment shows that the improved U-Net network has achieved good experimental results on the joint segmentation task of the cup and optic disc: the IOU value and F1-Score on the test set are respectively 0.8545 and 0.9209,0.069 and 0.0368 higher than U-Net network respectively.