Advanced deep learning technique enhances choroidal vessel visualization

Researchers from Peking University have developed a novel noninvasive choroidal angiography method that enables layer-wise visualization and evaluation of choroidal vessels using deep learning. This new approach, published in Health Data Science, employs an advanced segmentation model that can handle varying quality of optical coherence tomography (OCT) B-scans, making it a promising tool for clinical applications in diagnosing retinal diseases.

Choroidal Optical Coherence Tomography Angiography (C-OCTA) offers a significant improvement in the analysis of choroidal vessels, a critical component in the study of retinal diseases such as age-related macular degeneration and central serous chorioretinopathy. Traditional methods, such as indocyanine green angiography (ICGA), are invasive and unable to provide volumetric information necessary for detailed choroidal analysis. The new method proposed by Lei Zhu, a Ph.D. student at the Department of Biomedical Engineering, Peking University, and co-researcher Associate Professor Yanye Lu from the Institute of Medical Technology at Peking University Health Science Center, leverages a deep learning framework to noninvasively capture three-dimensional choroidal vascular information from OCT B-scans.

Our approach employs a segmentation model to extract choroidal vessels from OCT B-scans, trained on high-quality scans but also effective on lower-quality scans often collected in clinical settings. This allows for a more accurate reconstruction of choroidal structures. It paves the way for better analysis of choroidal indexes in various retinal diseases."

Lei Zhu, Ph.D. student at the Department of Biomedical Engineering, Peking University

The proposed framework treats the task as a cross-domain segmentation challenge. It employs an ensemble discriminative mean teacher structure to reduce noise and improve adaptation between high-quality and low-quality B-scans. The method was tested on extensive datasets, achieving a dice score of 77.28 for choroidal vessel segmentation, demonstrating its effectiveness in reconstructing choroidal vessel distributions.

Through this framework, the team demonstrated a significant reduction in vascular indexes among patients with central serous chorioretinopathy compared to healthy individuals, particularly in the regions beyond the central macula fovea (P < 0.05). This research highlights the method's potential for noninvasive clinical analysis and diagnosis of choroidal diseases.

Looking forward, the research team aims to apply this method to further analyze choroidal indexes across a wider range of retinal diseases, potentially facilitating more accurate clinical applications and improving patient outcomes. "Our goal is to continue refining this model for broader clinical use, offering a more efficient and less invasive alternative for choroidal vessel analysis," added Lu.