Optical coherence tomography (OCT) produces 3D images of the retina from coherent light. Every OCT volume is covered in speckle, a grainy pattern of bright and dark spots created when light scattered by tissue interferes with itself. Most segmentation networks treat speckle as noise. One recent result shows it is also a tissue signature: SpecstatOR fits a Gamma distribution to small patches of intraoperative OCT and trains a U-Net on the fitted parameters instead of raw brightness. On an unseen glass needle, that U-Net reached a mean Dice of 0.90, while a raw-intensity U-Net reached 0.10.
There is a catch that free-space optical communication worked out long ago. Speckle statistics are not a property of the tissue alone. They also depend on how many independent speckle "cells" the detector integrates over. That count is set by the collecting aperture, the scanner's resolution, any repeated-frame averaging and the analysis window. This effect is called aperture averaging, and it is why larger telescope receivers see weaker scintillation. For OCT it means that a Gamma parameter fitted over a fixed 7×7 window measures a mix of tissue and device. Our proposal: estimate each volume's 3D speckle correlation size directly from the data, then divide out the aperture-averaging factor so that the resulting statistics describe tissue rather than scanner. We test whether this makes 3D segmentation transfer across OCT vendors.
Core Claim
Fitted speckle statistics carry both tissue information and a device-specific averaging factor that can be predicted. Removing that factor gives segmentation inputs whose statistics line up across scanners. A 3D network trained on these inputs from one vendor should lose less accuracy on other vendors than two alternatives: networks trained on raw intensity, and networks trained on SpecstatOR-style fixed-window statistics.
The physics comes from the communications literature. Performance of Integrated Ground-Air-Space FSO Links Over Various Turbulent Environments and On the Ergodic Capacity of MIMO Free-Space Optical Systems over Turbulence Channels show that the strength of intensity fluctuation falls as the receiver spans more independent speckles. We borrow that relation as a normalizer. We do not borrow the hardware.
The standard multi-look model states the relation precisely. Suppose intensity inside a window $W$ averages $M$ independent speckle cells, each of which follows a Gamma law with intrinsic shape $k_0$. Then the fitted shape is approximately
$$k_W \approx M\,k_0, \qquad M \approx |W| / V_c,$$
where $|W|$ is the window volume in voxels and $V_c$ is the speckle correlation volume. We estimate $V_c$ per volume from the width of the 3D intensity autocorrelation in homogeneous tissue.
The normalized input is $\tilde{k} = k_W / \hat{M}$, which estimates the tissue's intrinsic shape. It is paired with the local mean $\mu_W$. The squared speckle contrast $1/k_W$ plays the role of the scintillation index, which aperture averaging reduces by the factor $M$.
Multilook compressive sensing in the presence of speckle noise uses the same multi-look model to reconstruct images from 2D radar and holography data. We use the model differently: to standardize features for segmentation, with no reconstruction step.
Relation to Prior Work
The closest work is SpecstatOR. It supplies the core idea that speckle parameters are good segmentation inputs, and it motivates this proposal through its invariance result. It differs from our method in three ways. It works on 2D B-scan patches. Its window size is fixed rather than set by the physics. It was tested on a single intraoperative system, so it never faced the question of whether its parameters carry over between devices.
Multi-Aperture Transformers for 3D (MAT3D) uses "aperture" to mean a multi-scale receptive field inside the network. It works only on reconstructed intensities and never models detection physics. It is a strong 3D architecture we can reuse, not a competing idea.
Universal digital filtering for denoising volumetric retinal OCT and OCTA in 3D shearlet domain shows that 3D context separates speckle from anatomy better than 2D. That supports working in 3D. However, it removes speckle, which is the signal our method keeps.
On the application side, two papers document the cross-vendor problem we target on RETOUCH. RetiFluidNet relies on attention and multi-kernel pooling. Pathological Retinal Region Segmentation From OCT Images Using Geometric Relation Based Augmentation (GeoGAN) relies on shape-aware generative augmentation. Both treat vendor shift as something to handle in software on intensity images. Neither asks whether a physical quantity explains the shift.
Research Program
The work has three stages.
Stage 1: can the device factor be measured? This stage needs no learning. On RETOUCH volumes from each of its three vendors (Cirrus, Spectralis, Topcon), we first undo the display log-compression where possible. We then estimate $V_c$ from the 3D autocorrelation inside a homogeneous outer-retina region defined by the reference masks. Next we fit Gamma parameters over fixed windows and over windows sized to a fixed $\hat{M}$. The hypothesis is that raw $k_W$ in matched tissue classes differs across vendors, and that $\tilde{k}$ narrows the gap. We measure the gap as the distance between per-vendor histograms. If normalization does not narrow it, the premise fails and we stop.
Stage 2: segmentation. We compute a small stack of 3D statistics channels at two or three values of $\hat{M}$, with window extents set per volume from $\hat{V}_c$. The channels are $\tilde{k}$, $\log \mu_W$ and fit uncertainty. These feed a standard 3D U-Net, and optionally the intensity channel is added. The same stack is also fed to a MAT3D-style backbone, which checks that any gain comes from the inputs rather than the architecture.
Stage 3: stress test. Within a single vendor, we synthetically average $N$ adjacent co-registered B-scans or apply spatial low-pass averaging. This changes $M$ in a known way. We then measure how each model's Dice changes as $N$ varies.
Baselines
The naive reference point is a 3D U-Net trained on raw log-intensity volumes of one vendor and tested on the others. It uses the same backbone, preprocessing and tuning budget as our model. SpecstatOR's 2D raw U-Net result (0.10 versus 0.90) warns that this baseline can fail badly under shift.
The strongest reported competitors on RETOUCH are two published results. GeoGAN reports a test Dice of 0.906 and a cross-device average Dice of 0.893. RetiFluidNet reports a mean Dice of 86.6%. Their protocols differ from ours: GeoGAN's cross-device numbers do not appear to be strict leave-one-vendor-out. So we will retrain RetiFluidNet, and GeoGAN-style augmentation if the code is obtainable, under our exact splits, with matched epochs and the same hyperparameter-search budget. The published numbers serve only as sanity anchors.
Beating these baselines means a statistically significant gain in mean leave-one-vendor-out Dice on fluid classes (IRF, SRF, PED) over the retrained strongest baseline, across seeds and patients. It also means a smaller gap between in-vendor and out-of-vendor Dice, since that gap is the quantity we claim to shrink. We also report the 95th-percentile Hausdorff distance (HD95).
The key ablation isolates normalization. It uses a SpecstatOR-style fixed 7×7 2D Gamma fit, a fixed-size 3D Gamma fit without $\hat{M}$ normalization, and our normalized 3D fit, all with an identical network. If the fixed 3D fit matches ours, the aperture-averaging correction adds nothing and the contribution collapses to "use 3D speckle features."
Why not just…?
Why not just train a strong intensity network with heavy augmentation? Networks like RetiFluidNet already fit RETOUCH well. Vendor shift, however, is exactly where augmentation needed geometry-specific machinery: GeoGAN without its shape loss dropped to 0.830 Dice. Retraining these networks under strict leave-one-vendor-out splits is the direct test.
Why not just denoise first with 3D shearlet or BM4D filtering? Denoising removes the scatterer statistics that SpecstatOR found informative, and BM4D takes 332 seconds per volume according to the shearlet paper. We will include a shearlet-denoised 3D U-Net baseline. If it matches our model, the speckle-as-signal argument is weakened for fluid segmentation.
Why not just augment with random multi-look speckle? This would mean re-averaging looks or multiplying by Gamma noise with random shape during training. It is the serious threat, because it targets the same nuisance factor by teaching invariance instead of dividing it out. We will run it with the $N$ range matched to Stage 1 measurements. Whichever method wins in Stage 3, and whether combining them helps, settles the question honestly. If augmentation matches normalization, we will report that the physics is better used as an augmentation prior.
Decisive Checks and Risks
The proposal stands or falls on three checks:
- whether normalization collapses cross-vendor statistic histograms in Stage 1;
- whether the normalized-versus-fixed ablation shows a leave-one-vendor-out gain;
- whether Dice for the normalized model stays flat as the synthetic look count $N$ varies while baselines degrade.
The main risks are practical. Vendor exports apply unknown nonlinear processing that may not invert cleanly. Fluid regions are dark, so their statistics are noise-dominated and the gain may appear mainly at boundaries. The large statistics windows may blur thin structures, which HD95 will expose.
As a secondary target, OCHID choroidal vessels (MFGNet reports Dice 0.762) will test whether normalized statistics help low-contrast vessel boundaries within a single device. The Duke and OPTIMA datasets extend the vendor range.
References
- SpecstatOR: speckle statistics-based iOCT segmentation network for ophthalmic surgery
- Performance of Integrated Ground-Air-Space FSO Links Over Various Turbulent Environments
- On the Ergodic Capacity of MIMO Free-Space Optical Systems Over Turbulence Channels
- Multilook compressive sensing in the presence of speckle noise
- Multi-Aperture Transformers for 3D (MAT3D) Segmentation of Clinical and Microscopic Images
- Universal digital filtering for denoising volumetric retinal OCT and OCT angiography in 3D shearlet domain
- RetiFluidNet: A Self-Adaptive and Multi-Attention Deep Convolutional Network for Retinal OCT Fluid Segmentation
- Pathological Retinal Region Segmentation From OCT Images Using Geometric Relation Based Augmentation
- A novel multi-scale and fine-grained network for large choroidal vessels segmentation in OCT