Multi-depth transfer learning-based approaches via generative models for foveal avascular zone segmentation in OCTA images

Abstract

[Abstract]: The automatic identification and segmentation of the Foveal Avascular Zone (FAZ) is crucial for diagnosing and monitoring retinal diseases. However, the limited availability of Optical Coherence Tomography Angiography (OCTA) images with ground truth annotations poses a significant challenge for developing robust deep learning models. Traditional transfer learning techniques, such as ImageNet-based pre-training, require large datasets and struggle to adapt to domain-specific tasks in medical imaging. To address this issue, we propose a novel multi-deep transfer learning framework that leverages pretrained models on a generation task using both superficial (SCP) and deep capillary plexuses (DCP) of the retina. To the best of our knowledge, this is the first approach that integrates depth information from multiple vascular plexuses, allowing the model to capture cross-plexus structural relationships and enhance its ability to learn domain-specific vascular features while mitigating data scarcity constraints. We validated our framework through extensive experiments, demonstrating competitive or even superior segmentation performance compared to state-of-the-art pretraining models, despite using significantly smaller datasets. Our approach achieved a Dice coefficient score of 0.8238 ± 0.0263, a Jaccard Index of 0.7189 ± 0.0293, and a Correlation Index of 0.8328 ± 0.0229, highlighting the effectiveness of our multi-depth feature learning strategy in improving segmentation precision and generalization, even with limited data availability.