This code is used to generate results of my thesis work 'Segmentation of Multi Class Retinal Lesions from Fundus Images' and paper 'Multi Label Lesion Segmentation from Retinal Images'
Abstract—Diabetic retinopathy(DR) is a leading cause of preventable blindness in working-age adults. DR causes abnormal retina patterns such as exudate, microaneurysm,and hemorrhage. Segmentation of retinal lesion is an imbalanced classification problem. Task includes both background-foreground imbalance and the imbalance between lesion classes. Additionally, lesion sizes varies in a large range. In this paper, we propose an instance based IoU loss function and two-stage model to solve these size and imbalance problems. IB IoU loss calculates IoU score for each instance separately to increase the effect of the small lesions whereas the two-stage model is proposed to decrease the background-foreground imbalance effect on training. Experiments show that, proposed method has performance improvement in small lesions.
You can reach full version of thesis from the link https://open.metu.edu.tr/handle/11511/96686
![Screenshot 2024-01-27 at 22 15 32](https://private-user-images.githubusercontent.com/29564208/300208075-1c4d01e7-77d6-4fae-8685-f4525b28f606.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.H-3pCHIGVPdYTUWHBSCArDe4k1TPuAjD0kHeLPwdO4I)