Vessel-promoted OCT to OCTA image translation by heuristic contextual constraints

Shuhan Li, Dong Zhang, Xiaomeng Li, Chubin Ou, Lin An, Yanwu Xu, Weihua Yang, Yanchun Zhang, Kwang-Ting Cheng

Introduction

In modern ophthalmology, Optical Coherence Tomography (OCT) is the gold standard for visualizing retinal structure. However, assessing microvasculature requires Optical Coherence Tomography Angiography (OCTA), a modality that is often expensive and technically demanding to acquire due to the need for specialized sensors and precise motion tracking. To bridge this gap, researchers have developed TransPro, a deep learning framework designed to translate standard 3D OCT scans into high-quality OCTA images using a generative adversarial network (GAN).

The Challenge of OCT-to-OCTA Translation

Previous attempts to use AI for OCT-to-OCTA translation faced significant hurdles. Most models focused on 2D slices or projection maps, failing to maintain the 3D continuity of blood vessels. Furthermore, standard training losses often led to "vessel neglect," where the AI prioritized the background over the sparse vascular structures, resulting in blurry or disconnected maps.

Comparison of 3D volumes, 2D B-scans, and 2D projection maps for OCT and OCTA modalities.

The TransPro Framework

TransPro utilizes a 3D Conditional Generative Adversarial Network (cGAN) enhanced by two novel guidance modules:

  1. Heuristic Contextual Guidance (HCG): This module enforces consistency between the generated 3D volume and its 2D projection map. By leveraging a pre-trained 2D model, it provides contextual hints that help the 3D generator maintain vessel continuity despite noise in the original scan.
  2. Vessel Promoted Guidance (VPG): To ensure sharp and anatomically correct vasculature, a pre-trained segmentation model is integrated during training. This forces the generator to pay extra attention to vascular regions, preventing the model from ignoring small vessels.

Notably, these modules are only required during the training phase, allowing the final model to remain lightweight for clinical use.

Experimental Setup and Data

The framework was evaluated using the OCTA-500 dataset, which includes 500 pairs of 3D OCT and OCTA volumes. The dataset is divided into two subsets: OCTA-3M (3mm × 3mm high-detail scans) and OCTA-6M (6mm × 6mm wide-area scans).

Beyond standard metrics like PSNR, the authors introduced Vessel Density Error (VDE) and Vessel Density Correlation (VDC). These metrics specifically evaluate the model's ability to capture the density and local distribution of blood vessels—features critical for clinical diagnosis.

Results and Clinical Impact

TransPro demonstrated superior performance over existing state-of-the-art methods like Pix2pix 3D and AdjacentGAN. Key results include:

  • An 11.4% improvement in Mean Absolute Error (MAE).
  • A 40% reduction in Vessel Density Error (VDE), highlighting the success of the vessel-promoted guidance.
  • A 9.1% boost in Vessel Density Correlation (VDC).

In a blind clinical test, an experienced ophthalmologist achieved 100% accuracy in identifying diseases such as Diabetic Retinopathy (DR) using TransPro-generated images, matching the performance achieved with real OCTA scans.

Conclusion

TransPro represents a significant advancement in medical image translation, offering a cost-effective way to generate high-quality vascular maps from standard OCT scans. While it currently serves as a powerful screening tool rather than a total replacement for specialized hardware in complex surgical cases, its ability to maintain 3D continuity and vessel sharpness marks a major step toward wider accessibility of advanced retinal diagnostics.

Summary updated

Abstract

Highlights•We identify a novel OCT-OCTA translation motivation, which is to maintain the consistency of OCTA image views.•We introduce a novel Heuristic Contextual Guidance (HCG) to enhance OCTA projection map quality.•We propose a novel Vessel Promoted Guidance (VPG) module to enhance vessel quality in OCTA projection maps.•We evaluate our TransPro on OCTA B-scan images and projection maps using common and proposed vessel-specific metrics.•Our TransPro surpasses SOTA methods on OCTA-3M and OCTA-6M datasets, demonstrating its quantative and qualitive effectiveness.
BibTeX
@article{li2024vessel,
  author = {Shuhan Li and Dong Zhang and Xiaomeng Li and Chubin Ou and Lin An and Yanwu Xu and Weihua Yang and Yanchun Zhang and Kwang-Ting Cheng},
  title = {{Vessel-promoted OCT to OCTA image translation by heuristic contextual constraints}},
  journal = {Medical Image Analysis},
  year = {2024},
  doi = {10.1016/j.media.2024.103311},
}
Figure