Optic Disc Segmentation Using Attention-Based U-Net and the Improved Cross-Entropy Convolutional Neural Network.

Area of Science:

• Medical image analysis
• Deep learning in ophthalmology
• Glaucoma auxiliary diagnosis

Background:

• Medical image segmentation is vital for analysis, with deep learning showing success.
• U-Net based methods are common for optic disc segmentation but ignore feature channel dependence and struggle with small areas.
• Accurate optic disc segmentation aids glaucoma diagnosis.

Purpose of the Study:

• To propose a novel aggregation channel attention network for improved optic disc segmentation in fundus images.
• To address limitations of existing methods by exploiting channel dependencies and multi-scale information.
• To enhance segmentation performance, particularly in small regions, for better glaucoma diagnosis support.

Main Methods:

• Developed a new aggregation channel attention network integrating channel dependencies and multi-scale information.
• Improved the loss function by combining Dice coefficient and cross-entropy for balanced contribution.
• Applied the network to fundus optic disc segmentation tasks.

Main Results:

• The proposed network demonstrated improved segmentation performance on the Messidor and RIM-ONE datasets.
• It enhanced prediction accuracy compared to base architectures while maintaining computational efficiency.
• Achieved a 0.0469 overlapping error on the Messidor dataset, indicating superior segmentation.

Conclusions:

• The novel aggregation channel attention network effectively improves optic disc segmentation in fundus images.
• The method enhances feature utilization and segmentation accuracy, especially for small areas.
• This approach offers a more robust tool for computer-aided glaucoma diagnosis.