Technical Summary: Test-Time Adaptation in OCT Using Trajectory-Aligned Time-Independent Flow

Problem Statement

Optical Coherence Tomography (OCT) is a cornerstone of ophthalmology, yet automated analysis via deep learning is hindered by inconsistent image quality, particularly in low-cost scanning devices. These devices introduce severe noise characteristics and distribution shifts that are rarely represented in standard training datasets. Consequently, model performance deteriorates significantly when encountering these domain gaps. While Test-Time Adaptation (TTA) methods have been proposed to address this, existing approaches face limitations:

• Weight-based TTA: Methods like TENT and CoTTA rely on prediction entropy or pseudo-labels, which often yield inconsistent improvements in segmentation tasks due to the difficulty of designing robust loss functions without access to the training distribution.
• Generative TTA: Diffusion-based methods treat test images as intermediate points on a diffusion trajectory. However, they often assume ideal, fixed-magnitude Gaussian noise. Real-world noise deviates from this, and attempts to compensate (e.g., re-noising or downsampling) risk blurring critical small lesions or failing to align with actual noise distributions.

Methodology: TTA-Flow

The authors propose TTA-Flow, a flow-matching-based TTA framework designed to generate high-quality surrogate images from noisy inputs by aligning them with the training data manifold. The framework introduces two critical modifications to standard flow matching to handle real-world noise deviations:

1. Trajectory Alignment via Histogram Matching

To bridge the gap between theoretical noise distributions and actual test data, the method employs a reference trajectory generation process:

• Reference Generation: The trained flow model simulates $n$ sample trajectories from pure noise to target data. From these, the authors compute the mean target intensity histograms ($\bar{H}_s$) for each discrete time step $s$.
• Inference Alignment: During inference, an incoming noisy test sample $\zeta$ undergoes histogram matching against a specific reference histogram $\bar{H}_{s_{target}}$. This preprocessing step aligns the sample's intensity distribution with the expected distribution of the reference trajectories at a predetermined time step $s_{target}$.
• Reconstruction: The aligned sample is used to initialize the Ordinary Differential Equation (ODE) solver for the remaining steps, effectively mapping the image into the target distribution without requiring explicit data-fidelity terms.

2. Time-Independent Flow

Standard flow matching networks typically condition on time (or noise level) to guide the denoising process. The authors argue that because real-world noise levels deviate from the reference trajectory, forcing a specific time condition can lead to inaccurate assumptions.

• Approach: The network intentionally omits explicit time conditioning.
• Rationale: By removing this constraint, the network is forced to implicitly manage unknown real-world noise levels, a strategy shown to yield superior results when the true noise level is unknown.

Key Contributions

1. Novel TTA Framework: Introduction of TTA-Flow, which utilizes flow matching to adapt OCT images from low-cost devices (Topcon, Cirrus) to high-quality reference distributions (Spectralis).
2. Trajectory Alignment: A method to align test image histograms with synthetic reference trajectories, effectively bridging the domain gap between theoretical and actual noise distributions.
3. Time-Independent Architecture: The removal of time conditioning from the flow network to better accommodate deviations in real-world noise distributions, demonstrating that implicit noise management outperforms explicit conditioning in this context.
4. Data-Fidelity Independence: The approach achieves robust performance without incorporating explicit data-fidelity terms, which are often sensitive to hyperparameters and can risk hallucinating structures.

Experimental Results

The method was evaluated on two datasets: the public RETOUCH dataset (focusing on Intraretinal Fluid, Subretinal Fluid, and Pigment Epithelial Detachment) and an in-house dataset of Geographic Atrophy (GA) in Age-related Macular Degeneration (AMD).

• Segmentation Performance:
  • Cirrus $\to$ Spectralis: TTA-Flow (unconditional) achieved a state-of-the-art mean Dice Similarity Coefficient (DSC) of 58.6, significantly outperforming baselines (e.g., standard Flow Matching at 50.8, DDA at 40.3) and approaching the supervised upper bound (61.6).
  • Topcon $\to$ Spectralis: The method maintained the highest mean DSC (51.0) on the more challenging Topcon dataset, outperforming diffusion-based baselines like CPDM and DDA.
  • GA Segmentation: The method reduced segmentation errors significantly compared to baselines, achieving a DSC of 56.1 (unconditional) on the in-house dataset.
• Ablation Studies:
  • Histogram Matching: Integrating histogram matching improved mean DSC from 50.8 to 58.6 (Cirrus) and 49.4 to 56.1 (In-house), confirming its efficacy.
  • Time Conditioning: The unconditional network consistently outperformed the conditional version in fluid segmentation and produced superior perceptual quality (FID reduced from 85.4 to 32.2). The unconditional approach also showed low sensitivity to the choice of the starting time point $s_{target}$.
• Qualitative Results: Reconstructions preserved original anatomy while visibly reducing noise and maintaining high contrast, leading to downstream segmentations that aligned well with ground-truth annotations.

Significance and Claims

The paper claims that TTA-Flow offers a simple yet highly effective solution for OCT domain adaptation. By leveraging flow matching with trajectory alignment and removing time conditioning, the method successfully generates high-quality surrogate images that enable state-of-the-art segmentation performance on low-cost devices.

The authors emphasize that:

• Histogram matching is a computationally efficient preprocessing technique that aligns test data with network expectations.
• Removing time conditioning enhances denoising performance and visual fidelity compared to standard conditional setups.
• The method outperforms multiple baselines, including those using explicit data-fidelity terms, without the risk of hallucinating structures observed in practice.
• The framework is robust across different low-cost OCT devices and clinical biomarkers (fluids and GA), demonstrating its potential to improve automated analysis in resource-constrained settings.

Future work is noted to focus on dynamically estimating noise levels directly from input images to automate the selection of the optimal histogram-matching target and further speed up the adaptation process.