Diffusion model based OCT to OCTA translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ophthalmology

Background:

• Optical Coherence Tomography (OCT) is crucial for retinal imaging.
• OCT Angiography (OCTA) provides vital vascular information but requires specialized equipment.
• Translating OCT to OCTA can improve accessibility and reduce costs for diagnosing retinal diseases.

Purpose of the Study:

• To introduce a novel conditional diffusion-based approach, the Brown Bridge diffusion model (BBDM), for OCT-to-OCTA image translation.
• To address limitations of traditional Generative Adversarial Networks (GANs) in generalization and structural fidelity for this translation task.
• To evaluate the clinical utility and performance of BBDM in generating OCTA from OCT images.

Main Methods:

• Developed and implemented the Brown Bridge diffusion model (BBDM) using a bidirectional stochastic process.
• Integrated BBDM within the latent space of VQGAN for training.
• Trained the model on the OCT500 dataset and validated on a clinical dataset of diabetic retinopathy patients from UIC.

Main Results:

• BBDM demonstrated superior performance in structural similarity index (SSIM) and perceptual contrast quality index (PCQI) compared to GANs, especially in larger field-of-view scans.
• The model maintained anatomical trends consistent with ground-truth OCTA and preserved clinically relevant vascular features (BVC, BVT, VPI).
• While showing minor deviations in some metrics like FID, BBDM offered advantages in computational simplicity, training stability, and reduced image hallucinations.

Conclusions:

• BBDM represents the first diffusion-based framework for OCT-to-OCTA translation.
• The model successfully generates clinically meaningful OCTA from standard OCT images.
• This approach supports more accessible and cost-effective diagnostics for retinal diseases.