MLGF-GAN: a multi-level local-global feature fusion GAN for OCT image super-resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Optical Coherence Tomography (OCT) is vital in cardiology and ophthalmology but limited by resolution.
• Current super-resolution methods often prioritize reconstruction accuracy over perceptual quality, impacting diagnostic effectiveness.
• Enhancing perceptual quality is crucial for improving human visual recognition and diagnostic utility in OCT imaging.

Purpose of the Study:

• To develop a novel super-resolution method for OCT images that enhances perceptual quality while maintaining reconstruction accuracy.
• To address the limitations of existing methods by integrating local details, global context, and multi-level features.
• To improve the diagnostic utility of OCT through superior image quality.

Main Methods:

• Proposed a Multi-level Local-Global feature Fusion Generative Adversarial Network (MLGF-GAN).
• Employed a Local Feature Extractor (LFE) with Coordinate Attention-enhanced CNN for local refinement.
• Utilized a Global Feature Extractor (GFE) with shifted-window Transformers for long-range dependencies and a Multi-level Feature Fusion Structure (MFFS) for hierarchical aggregation.

Main Results:

• The MLGF-GAN achieved highly competitive perceptual quality across multi-scale evaluations (×2, ×4, ×8) on coronary and retinal OCT datasets.
• Generated OCT super-resolution images demonstrated superior texture detail restoration and spectral consistency.
• The model maintained reconstruction accuracy while significantly improving perceptual quality.

Conclusions:

• The proposed MLGF-GAN effectively enhances perceptual quality and diagnostic reliability of OCT images.
• The model shows excellent generalization capability across different pathologies, as confirmed by cross-pathology experiments.
• This approach offers a promising solution for improving OCT-based clinical assessment.