Abstract
Purpose :
Accurate segmentation of the fovea and the arteriovenous vein in the optic disc centre plays a vital role in the diagnostic system for diabetic retinopathy. In recent years, deep learning has shown its powerful performance in medical image analysis. Accordingly, this study presents a fundus image segmentation method based on self-attention and mixed-loss adversarial networks. Extensive experiments were performed on the fovea and the optic disc's central fine arteriovenous vein in two public datasets.
Methods :
As shown in Fig. 1, the fundus images were preprocessed by using data augmentation operations such as rotation and symmetry and sent to the generator based on the separable convolutional U-Net. It adopts a self-attention mechanism to adjust the feature weights between the sampling points to guarantee the detection accuracy. Second, we leverage the separable convolutional U-Net to implement the segmentation by adding the adversarial training framework. The generator's outcome is improved by the output of the judge and the loss return.
Results :
The proposed method outperforms the state-of-the-art algorithms in accuracy and execution efficiency. In general, the average sensitivity is 95.56%, and the detection performance is 97.21% over the data samples.
Conclusions :
The proposed method has low computational complexity while achieving promising accuracy. The segmentation results obtained are close to manual delineation. It is therefore a potential instrument in clinical applications.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.