Date

2022-2-9

Author

Çontar, Elif K

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

315
views

197
downloads

Cite This

Diabetic retinopathy is a leading cause of preventable blindness among adults. Detection of diabetic retinopathy-related retinal lesions is essential for automatic detection of DR. There are different kinds of lesions related to the disease, namely microaneurysm, hemorrhage, hard exudate, and soft exudate. Each lesion has different characteristics: color, size, and shape. In the literature, the detection of retinal lesions has been examined as a localization or segmentation problem. Besides traditional image processing methods, machine learning-based and neural network-based methods have been proposed widely in the last years. Most of the works focused on detecting only one type of lesion. These methods can not be transferred to detect another kind of lesion because of the different characteristics of the lesions. Additionally, segmentation of retinal lesion task is an imbalanced classification problem. Task includes both foreground-background imbalance and imbalance between positive classes. In this study, we developed a new instance-based intersection over union(IB_IoU) objective function to segment multi-class retinal lesions from fundus images. The loss has targeted the following two problems. Firstly, it aims to solve the imbalance problem by averaging intersection over union(IoU) scores across the classes. Secondly, IoU score is calculated separately for every instance with a closed contoured shape. The aim is to improve the detection performance of lesions with small pixel areas. The connected component analysis is applied to find instances on a union of prediction and ground truth labels. The results show that the proposed algorithm is comparable to state-of-the-art methods focused on detecting single lesions. Additionally, the proposed loss function has improved detection performance of microaneurysm and exudate lesions over other loss functions used in multi-class retinal lesion segmentation.

Subject Keywords

diabetic retinopathy, segmentation, retinal lesion, deep learning, fundus image

URI

https://hdl.handle.net/11511/96686

Collections

Graduate School of Natural and Applied Sciences, Thesis