Introduction
Optical Coherence Tomography (OCT) is a vital non-invasive imaging modality in ophthalmology, primarily used for diagnosing and monitoring retinal diseases such as age-related macular degeneration (AMD) and diabetic retinopathy. A key clinical requirement is the precise segmentation of retinal layers and intraretinal fluid (IRF). However, OCT images are inherently corrupted by "speckle noise"—a high-frequency granular interference that complicates automated analysis.
Traditional segmentation models, such as the U-Net, operate primarily in the spatial domain. These networks often struggle with two main issues: a limited receptive field that misses global structural context, and an inability to distinguish between high-frequency anatomical details and high-frequency noise. This paper introduces Υ-Net, a novel dual-encoder architecture designed to overcome these limitations by processing information in both the spatial and spectral domains.
Methods: The Spatiospectral Architecture
The core innovation of Υ-Net is its dual-encoder design, which feeds into a shared spatial decoder. This structure allows the model to capture both local textures and global frequency-based patterns simultaneously.
Dual-Encoder Design
- Spatial Encoder: A standard convolutional branch that extracts local features like edges and textures using traditional convolutional blocks.
- Spectral Encoder: This branch utilizes Fast Fourier Convolutional (FFC) blocks. Within these blocks, features are transformed into the frequency domain via a Fast Fourier Transform (FFT).
- Fourier Unit: In the spectral domain, the model applies learnable 1x1 convolutions. Because a single point in the frequency domain represents information from the entire spatial image, these operations provide a global receptive field. The data is then returned to the spatial domain using an Inverse FFT (IFFT).
The features from both encoders are concatenated and passed to a spatial decoder to produce the final segmentation map.
Experimental Results
The model was evaluated on the Duke OCT dataset (10 patients, expert-annotated) using a standard 6-2-2 split. The results demonstrate that Υ-Net significantly outperforms existing state-of-the-art models, particularly in the challenging task of fluid segmentation.
Υ-Net achieved a 0.93 Dice score for fluid segmentation, representing a 13% improvement over the industry-standard U-Net (0.80). The mean Dice score across all retinal layers also saw a 1.9% absolute improvement. Notably, Υ-Net achieves these results with fewer parameters (7.46M) than the standard U-Net (7.76M), indicating that spectral features provide a more efficient representation of retinal structures.
Discussion and Insights
The success of Υ-Net stems from its ability to "filter" noise and capture global patterns. Fluid pockets often exhibit distinct global structural signatures that are more easily identified in the frequency domain. Ablation studies revealed that clipping the frequency range to a specific band (between -10 and 10) further optimized performance, suggesting that critical anatomical information resides within a narrow spectral window.
While the addition of skip connections from the spectral encoder to the decoder did not yield improvements—likely due to the misalignment between global spectral features and local spatial maps—the overall architecture proves highly effective for noisy medical imaging. Future research may explore the application of spatiospectral networks to other modalities like ultrasound or MRI, where global context and noise suppression are equally critical.
Summary updated