Introduction
Optical Coherence Tomography (OCT) is the gold standard for non-invasive imaging of the eye's posterior segment. However, the choroid—the vital vascular layer beneath the retina—remains difficult to map accurately. Specifically, the "Haller’s layer," which contains the largest choroidal vessels, is critical for understanding diseases like high myopia (HM), age-related macular degeneration (AMD), and diabetic retinopathy.
Automated segmentation of these vessels is challenging due to low contrast at vessel boundaries and significant scale variance. Existing deep learning models often struggle to balance the need for global context (captured well by Transformers) with the need for fine-grained local detail (captured well by CNNs).
Methodology: The MFGNet Framework
To bridge this gap, the authors propose MFGNet, a hybrid neural network designed for precise segmentation of large choroidal vessels. The architecture introduces two key innovations:
Large Kernel and Multi-scale Attention (LKSA)
The LKSA module is placed at the beginning (stem) of the network. By using large $7 \times 7$ depth-wise convolutions, it captures a wider field of view early on. This helps the model "see" through the blurred boundaries of large vessels that smaller, standard kernels might miss.
Dual-branch Fine-grained Feature Extraction (DFG)
The DFG module is the core of the network, running two branches in parallel:
- Transformer Branch: Captures long-range dependencies to maintain global vessel patterns.
- Nested Convolutional Branch: Extracts fine-grained local features using a mini U-Net structure.
Information exchange between these branches ensures that the global shape of the vascular network aligns perfectly with local edge details.
Key Results and Performance
MFGNet was evaluated on the OCHID (Optical Choroidal Haller-layer Image Dataset), consisting of 800 manually annotated OCT images. The model achieved state-of-the-art performance, outperforming 11 other major models, including U-Net, TransUNet, and SegFormer.
Quantitatively, MFGNet reached a Dice score of 0.762, an IoU of 0.619, and a Precision of 0.751. As shown in the visual comparison, MFGNet produces segmentations that are more anatomically consistent and closer to the ground truth (GT) than its predecessors.
Clinical Insights from 3D Analysis
The true value of MFGNet lies in its ability to enable 3D morphological analysis. By reconstructing the vascular network from 2D slices, researchers identified significant structural differences between healthy eyes and those with high myopia.
The study utilized advanced metrics such as 3D Fractal Dimension (3D-FD) and 3D Vascular Density (3D-VD). The results showed that high myopia patients had significantly lower 3D-FD and 3D-VD ($p < 0.05$), confirming a "simplification" and thinning of the vascular network associated with the disease. Interestingly, the 3D Vascular Curvature remained stable, suggesting that while vessels atrophy, their fundamental "twistiness" is preserved.
Conclusion
MFGNet represents a significant advancement in automated ocular imaging. By successfully hybridizing Transformer and CNN architectures, it provides a robust tool for the quantitative analysis of choroidal diseases, offering clinicians new biomarkers for diagnosis and treatment monitoring.
Summary updated