Addressing inter-device variations in optical coherence tomography angiography: will image-to-image translation...

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Optical coherence tomography angiography (OCTA) provides non-invasive visualization and quantification of retinal microvasculature.
• Technical variations among OCTA devices lead to significant inter-device differences in data, limiting comparability and generalizability.
• These variations pose a domain shift problem for machine learning models trained on data from different OCTA machines.

Discussion:

• Unsupervised deep image-to-image translation methods, such as Cycle-Consistent Generative Adversarial Networks (Cycle-GANs) and Denoising Diffusion Probabilistic Models (DDPMs), offer a potential solution.
• These models can perform cross-domain translation of OCTA images by training on unpaired data from different device domains.
• Previous applications demonstrate success in medical imaging tasks like segmentation, denoising, and cross-modality translation.

Key Insights:

• Cycle-GANs and DDPMs can translate OCTA images between different device domains without requiring paired data.
• This cross-domain translation has the potential to mitigate the domain shift problem caused by OCTA device variability.
• Successful application of these techniques can improve the consistency and reliability of OCTA data analysis.

Outlook:

• Exploring image-to-image translation methods can significantly improve OCTA data comparability across different centers and devices.
• This facilitates more efficient analysis of heterogeneous OCTA datasets.
• Broader applicability of machine learning models trained on limited OCTA datasets is anticipated, advancing retinal imaging research.