Self-attention CNN for retinal layer segmentation in OCT.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal layer structure in Optical Coherence Tomography (OCT) images is crucial for diagnosing ophthalmic diseases.
• Current U-net based methods struggle with long-range dependencies and complex diseased retinal layer morphology.
• Accurate retinal layer segmentation is essential for clinical diagnosis and disease monitoring.

Purpose of the Study:

• To develop an improved method for retinal layer segmentation in OCT images.
• To enhance the capture of both local and global contextual information for more accurate segmentation.
• To improve diagnostic capabilities for ophthalmic diseases through precise retinal layer analysis.

Main Methods:

• Proposed a U-shaped network integrating an encoder-decoder architecture with self-attention mechanisms.
• Incorporated a vertical self-attentive module at the network's base and attention in skip connections and up-sampling.
• Combined convolutional neural networks for local feature extraction with transformers for global context awareness.

Main Results:

• Achieved high accuracy in retinal layer segmentation on two public OCT datasets.
• Demonstrated superior performance compared to existing methods, with average Dice scores of 0.871 and 0.820.
• The proposed method effectively integrates global perception from self-attention with local feature extraction from CNNs.

Conclusions:

• The proposed U-shaped network with self-attention mechanisms significantly enhances retinal layer segmentation accuracy.
• This approach addresses limitations of traditional methods in capturing contextual information in complex retinal images.
• The method shows promise for improving the diagnosis and management of ophthalmic diseases.