Abstract
Purpose :
The optic cup-to-disc ratio is an important indicator for the evaluation of glaucoma as a disease, and the cup-to-disc ratio is one of the fundus examinations that has positive implications for the diagnosis of fundus images and optic neuropathy. The deep learning model shows excellent performance in optic cup-to-disc segmentation, but it relies heavily on a single label. In the medical field, an image is usually labeled by multiple experts. However, traditional label fusion strategies tend to overfit the model and ignore the information expressed among multiple experts. To address these problems, this study develops an objective and quantifiable coarse-to-fine model to improve optic cup and optic disc segmentation by Probabilistic variational auto-encoder and soft attention mechanism.
Methods :
We employ the ORIGA dataset which contains 650 fundus images with six expert annotations. The model we constructed is shown in Figure 1, it includes the following three components: (1) The backbone network is Unet, and the feature of the last layer of the decoder is used as the coarse segmentation features F1. (2) Meanwhile, the original input images and labels are fed into the prior and posterior networks of the probabilistic variational self-encoder, respectively. The prior and posterior features are connected with Unet's last layer feature F1 to obtain the coarse segmentation results. (3) To obtain the refined segmentation results, a soft attention mechanism is used to focus on the uncertainty map between features and labels, thus fusing the consistent and disparity information and enabling the model to obtain a more accurate representation.
Results :
We select the sensitivity (SE), specificity (SP), accuracy (AC), AUC, and dice coefficient (Dice) to evaluate the performance of the proposed model. Extensive comparison experiments have verified that our model is superior to the state-of-the-art models. The Dice of this method in the optic cup and optic disc segmentation is 96.78% and 86.89% respectively.
Conclusions :
Our multi-label coarse-to-fine deep learning framework provides an automated tool for accurate segmentation of the optic cup and optic disc from fundus images. It helps to learn annotation information among multiple experts, improving not only the accuracy but also the generalization ability of the model.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.