Joint Optic Cup And Disc Segmentation Algorithm Based On An Improved Swin-Unet Model

Glaucoma is a common eye disease whose severity corresponds directly to the physiological burden on patients,possibly even resulting in vision loss.Clinically diagnosing glaucoma patients through segmentation of the optic disc and cup regions in fundus images can provide physicians with various physiological indicators such as the cup-to-disc ratio and nerve fiber layer thickness,which can aid in more accurate diagnosis and treatment.Therefore,this work has important clinical value and social significance for the early diagnosis and grading of glaucoma.This dissertation proposes an algorithm named Swin-AC-Unet based on an improved Swin-Unet model architecture.This algorithm uses Swin-Unet as its main framework model while incorporating ACmix modules,which integrate convolutions and self-attention to retain the benefits of both while improving segmentation accuracy.Furthermore,this dissertation uses the Lion optimizer instead of the conventional Adam W optimizer,which only tracks momentum and employs sign operations to ensure consistent update magnitude for each parameter,resulting in fewer parameters and faster training speed while achieving higher segmentation accuracy.Finally,this dissertation employs a hybrid loss function to train the model,which considers both cross-entropy loss(CELoss)and Lovasz Loss,endowing the model with better stability and robustness that improve its segmentation accuracy.Experimental results demonstrate that the proposed model performs well in segmentation accuracy,which is of crucial clinical value in assisting physicians with early screening and clinical diagnosis of glaucoma.In the experimental section,this dissertation uses the Refuge dataset and Rim-one-r3 dataset,evaluating the model in terms of the Dice coefficient and Io U metrics.The Refuge dataset includes 1200 fundus images,out of which 1080 are normal images and 120 are glaucomatous images.The Rim-one-r3 dataset includes159 fundus images,out of which 85 are normal images and 74 are glaucomatous images.On the Refuge dataset,the proposed model achieves an optic disc Dice coefficient of 0.984 and Io U of 0.967,and a cup Dice coefficient of 0.918 and Io U of0.849.On the Rim-one-r3 dataset,the model achieves an optic disc Dice coefficient of0.983 and Io U of 0.966.The cup’s Dice coefficient reaches 0.936 and Io U 0.87.These further indicate that the proposed model achieved good performance in cup and disc segmentation tasks.In summary,this dissertation proposes an algorithm for joint cup and disc segmentation based on an improved Swin-Unet architecture.The algorithm performs well on the Refuge and Rim-one-r3 datasets,providing effective solutions for the early diagnosis and treatment of glaucomatous eye disease.