Abstract
Purpose :
Glaucoma is a chronic ophthalmic disease that has irreversible damage to vision. The screening criteria for glaucoma are usually structural changes in the optic nerve head (ONH), which are performed by image segmentation of the optic disc (OD) and optic cup (OC), The vertical ratio of the divided optic cup to the optic disc is an important indicator for the diagnosis of glaucoma. Fundus images usually need to be annotated by multiple experts, and in order to explore the agreement information between expert annotations, we designed a fundus image segmentation model based on an information bottleneck approach.
Methods :
We used three public datasets, which included MESSIDOR (460 images), BinRushed (195 images) and Magrabia (95 images). The model we constructed is shown in Figure 1, it contains the following three components: (1) A hard parameter sharing module is designed, in which the encoder has three shared layers, and then multiple specific layers extract the corresponding individual feature mapGiping. The feature maps of the specific layers are fed into the Multilayer Perceptron to obtain the representations. (2) Then, a Multi-Rater Agreement Information Module is inserted in the bottleneck layer to discard all information in the representation that is not shared by those views, as this information is guaranteed to be redundant. (3) The extracted multiple independent representations are concatenated to form the joint representation sent to the final decoder.
Results :
We use representative segmentation metrics evaluate the performance of proposed model. The experimental results show that our model exceeds the best model in terms of view cup segmentation by 2%, reaching 95.5%; in terms of view disc segmentation, our model achieves the best results in all indexes, with a dice of 83%.
Conclusions :
In this research, we focus on extracting consistent information from multi-rater annotations that reflect consistent agreement among all raters. We propose to use a multi-view information bottleneck to obtain a concise representation of each view in a supervised setting and an unsupervised approach to maximize consistent information while eliminating information that is not shared between views. Extensive empirical experiments demonstrate the overall superior performance of our model on a range of fundus image segmentation tasks.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.