Generative model-based fundus photography translation for enhanced cross-device consistency

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Domain gap exists between conventional fundus photography and confocal scanning laser ophthalmoscopy (cSLO).
• This gap limits the comprehensive analysis of ocular structures.
• Bridging this gap can enhance diagnostic capabilities in ophthalmology.

Purpose of the Study:

• To develop a novel image translation framework for converting fundus images to cSLO-like images.
• To address the clinical need for improved visualization of ocular anatomy.
• To enhance diagnostic accuracy by providing cSLO-quality images from standard fundus photographs.

Main Methods:

• Proposed a novel image translation framework utilizing self-attention modules.
• Employed joint optimization of structural similarity and gradient variance losses.
• Constructed a high-quality paired dataset of fundus and cSLO images from the same patients.
• Introduced a new metric, Feature Matching Success Rate (FMSR), using AKAZE descriptors for quantitative assessment.

Main Results:

• Achieved state-of-the-art performance in perceptual realism and structural accuracy.
• Demonstrated effective capture of long-range dependencies via self-attention.
• Preserved anatomical fidelity and fine details through optimized loss functions.
• Validated the clinical relevance and diagnostic utility of the generated images.

Conclusions:

• The proposed framework successfully bridges the domain gap between fundus images and cSLO.
• The method enhances anatomical accuracy and preserves fine details, aiding clinical diagnosis.
• The novel FMSR metric provides a robust quantitative measure for cross-modality anatomical consistency.