Fundus Image Classification And Segmentation Based On Deep Learning

Glaucoma and diabetic retinopathy(DR)have become the leading causes of visually impaired and blindness in hunmans.The fundus images taken by fundus cameras contain rich information,which is important for doctors to diagnose fundus diseases.Currently,the screening of fundus diseases mainly rely on the well-trained ophthalmologists,which is expensive due to the scarity of opthalmologists.With the rapid development of computer and artificial intelligence,computer aided diagnosis has been widely applied in clinical diagnosis,which can reduce the workload of doctors and promoting large-scale fundus screeing.Therefore,it is important to apply artificial intelligence techniques to process and analyze fundus images.In this paper,in order to lay the foundation for glaucoma and DR computer-aided screening system,image processing and deep learning techniques are used to identify glaucoma and detect the lesions of DR.This paper mainly studies the following aspects:(1)For glaucoma diagnosis,we study the domain difference between differet datasets,and propose a conditional adversarial transfer(CAT)method to imporve the performance of glaucoma diagnosis in the dataset of interest.Different from the most existing works of deep transfer learning for marginal distribution matching,we match the label conditional distributions of the source and target domains by adversarial learning.We conduct experiments on three glaucoma datasets,and the results demonstrate the effectiveness of CAT in terms of multiple metrics.(2)For the pre-processing stage of the lesions segmentation,we extract the green channel of fundus images,and apply the median filtering and CLAHE to denoise and enhance the fundus images In addition,since the optic disc and lesions are similar in fundus images,we detect and eliminate the optic disc structure to improve the detectability of lesions region.(3)For the DR lesion segmentation,in order to extract and fuse the global coarse-grained fetures and local detail features,we propose to introduce the multi-scale approach into the UNet model.We conduct experiments on the IDRiD dataset,and the results demonstrate that the fusion of multi-scale features is beneficial for imporving the segmentation performance.