Frequency-aware optical coherence tomography image super-resolution via conditional generative adversarial neural...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Optical coherence tomography (OCT) is vital for medical diagnosis, but resolution limits hinder detailed morphological analysis.
• Deep learning super-resolution (SR) enhances OCT images but often neglects frequency fidelity, causing bias.
• Existing methods struggle to accurately reconstruct high-frequency details crucial for precise diagnosis.

Approach:

• A novel frequency-aware super-resolution framework is proposed, integrating frequency transformation, skip connections, and alignment modules.
• The framework utilizes a conditional generative adversarial network (cGAN) architecture.
• A frequency-based loss function is incorporated to ensure frequency fidelity during image reconstruction.

Key Points:

• The proposed framework significantly improves super-resolution of OCT images compared to existing deep learning methods.
• Demonstrated superiority on a coronary OCT dataset, highlighting enhanced morphological detail.
• Validated generalizability across different biological tissues, including fish corneal and rat retinal images.

Conclusions:

• The frequency-aware framework effectively addresses the frequency bias in deep learning-based OCT super-resolution.
• Achieves superior reconstruction of fine morphological structures in medical images.
• Offers a promising tool for advancing OCT-based diagnosis and treatment in cardiology and ophthalmology.