Wavelet attention network for the segmentation of layer structures on OCT images

Area of Science:

• Medical Imaging
• Deep Learning
• Biomedical Engineering

Background:

• Accurate segmentation of layered tissues in optical coherence tomography (OCT) is crucial for clinical analysis.
• Existing deep learning methods for OCT segmentation often exhibit topological errors like outlier prediction and disconnected labels.
• Current attention mechanisms, while effective, rely on global average pooling (GAP), limiting their analysis to low-frequency components.

Purpose of the Study:

• To introduce the Wavelet Attention Network (WATNet), a novel deep learning model for enhanced tissue layer segmentation in OCT images.
• To address the limitations of existing attention mechanisms by incorporating discrete wavelet transform (DWT) for multi-spectral information extraction.
• To improve the robustness and accuracy of OCT image segmentation, mitigating topological errors.

Main Methods:

• Proposed the Wavelet Attention Network (WATNet) integrating discrete wavelet transform (DWT) for multi-spectral feature extraction.
• Developed a DWT-based attention mechanism that analyzes multiple frequency components without complex selection.
• Embedded the DWT attention mechanism into existing deep learning frameworks for adaptability.

Main Results:

• WATNet demonstrated superior performance in tissue layer segmentation compared to established deep learning networks.
• Experiments conducted on esophageal and retinal OCT datasets validated the effectiveness of the proposed method.
• The DWT-based attention mechanism successfully extracted relevant multi-spectral information, improving segmentation accuracy.

Conclusions:

• The Wavelet Attention Network (WATNet) offers a robust and adaptable solution for accurate OCT tissue segmentation.
• The integration of DWT into attention mechanisms provides a powerful tool for analyzing multi-spectral image data.
• WATNet significantly outperforms existing methods, highlighting its potential for clinical OCT image analysis.