Introduction
Optical Coherence Tomography Angiography (OCTA) has revolutionized ophthalmology by providing non-invasive, high-resolution visualization of the retinal and choroidal microvasculature. However, clinical utility is often hampered by low signal-to-noise ratios (SNR) and "tail artifacts"—vertical streaks that obscure deeper vascular layers. Current gold-standard imaging requires merging approximately ten repeated scans to achieve sufficient quality, a process that is time-consuming and difficult for many patients. Freqformer addresses these challenges by reconstructing high-definition 3-D volumes from a single, noisy scan.
The Freqformer Architecture
The core innovation of this work is the Freqformer block, a dual-branch architecture designed to balance global spatial context with fine-grained detail preservation. Unlike traditional Transformers that can be computationally prohibitive for high-resolution medical data, Freqformer employs a "slice-wise" 2-D processing strategy that is significantly faster than 3-D patch-based methods.
The architecture consists of two primary modules:
- Simplified Multi-Head Attention (Sim-MHA): This module utilizes cross-covariance attention to compute interactions across channels rather than spatial tokens. This shift reduces computational complexity from quadratic to linear, allowing the model to handle large medical volumes efficiently.
- Complex-valued Frequency-Domain Module (CFDM): Transformers often act as low-pass filters, which can blur delicate capillary structures. The CFDM branch operates in the Fourier domain to adaptively boost high-frequency components, ensuring that the reconstructed images retain the sharpness necessary for clinical quantification.
Artifact Suppression and Layer Visualization
Before the data enters the Transformer blocks, it undergoes a Moving Average Subtraction Filter (MASF) step. This preprocessing stage is critical for removing tail artifacts by identifying and suppressing the weaker de-correlation values associated with these vertical streaks.
As shown in Figure 1, the removal of these artifacts significantly clarifies the visualization of the Inner Capillary Plexus (ICP) and Deep Capillary Plexus (DCP), which are often obscured by the shadows of larger vessels in the Superficial Capillary Plexus (SCP). This allows for more accurate depth-resolved analysis of the retinal circulation.
Key Results and Clinical Implications
Freqformer demonstrates superior performance across both qualitative and quantitative benchmarks. It achieved a Peak Signal-to-Noise Ratio (PSNR) of 26.90 dB and a 3D-SSIM of 0.741, outperforming state-of-the-art CNNs (like NAFNet) and other Transformer models (like Restormer).
Crucially, the study emphasizes that the enhanced images are not just visually appealing but also quantitatively accurate. Metrics such as vessel density and flow index derived from a single Freqformer-enhanced scan were found to be statistically indistinguishable from those obtained from 10-scan merged ground truths. Furthermore, the model is highly efficient, processing a full volume in approximately 45 seconds—a nearly 10x speedup over 3-D patch-based approaches.
Conclusion and Future Work
Freqformer represents a significant step forward in retinal imaging, offering a path toward high-quality 3-D vascular quantification without the burden of prolonged scan times. While future work is needed to validate the model on pathological cases (e.g., diabetic retinopathy) and different hardware platforms, its ability to "sharpen" the vascular network through frequency-domain learning provides a robust solution for modern ophthalmic diagnostics.
Summary updated