Vascular Information-Guided Automated Conversion Between OCT and OCTA

Renzhi Tian, Jinjie Wang, Hua Shi, Weizhen Li, Haoran Chen, Pan Chengchang, Honggang Qi

Keywords: Optical Coherence Tomography (OCT), Optical Coherence Tomography Angiography (OCTA), Image Translation, Modality Translation

TL;DR: We develop a bidirectional framework for OCT-OCTA conversion that enables OCTA-equivalent vascular imaging using standard OCT equipment, improving diagnostic accessibility in clinical settings.

Abstract:

Optical Coherence Tomography (OCT) and Optical Coherence Tomography Angiography (OCTA) provide complementary perspectives for retinal disease assessment—OCT captures structural layers while OCTA visualizes microvascular networks. However, their synergistic use is limited by OCTA's specialized equipment requirements and higher costs, particularly in resource-constrained settings. Moreover, clinical scenarios often require cross-validation between modalities to verify vascular-structural correlations or to compensate for artifacts and missing data.

We propose a bidirectional conversion framework that enables seamless transformation between OCT and OCTA modalities, facilitating comprehensive multi-modal analysis from single-modality inputs. Our approach integrates generative adversarial networks with wavelet decomposition and attention mechanisms to preserve both morphological coherence and diagnostic fidelity. The framework consists of three synergistic components: a 3D Cross-Modal Transformer for volumetric synthesis, a Vessel Structure Matcher for vascular topology preservation, and Hierarchical Feature Calibration for layer-specific refinement.

Extensive validation on the OCTA-500 dataset demonstrates superior performance with PSNR of 30.58, SSIM of 90.64\%, and MAE of 0.0199 for OCT-to-OCTA synthesis. Clinical disease classification experiments show that combining real OCT with synthesized OCTA achieves 75.06\% accuracy, surpassing single-modality baselines by up to 29\%. This bidirectional capability not only addresses accessibility barriers but also enables cross-modal validation essential for artifact disambiguation and comprehensive diagnostic assessment, ultimately advancing precision medicine in ophthalmology.The code is available in the supplementary materials.

Primary Area: generative models

Submission Number: 7358