PaperlayerView exact revision on arXiv

Source-linked AI summary

ARCOS: Zero-shot Boundary Localization for Corneal Layer Segmentation Across Optical Coherence Tomography Devices

Nuno Vivas Brás, Benjamin Memmi, Maëlle Bouhassane, Cristina Georgeon, Vincent Borderie, Karsten Plamann, Anatole Chessel

arXiv:2609.03668v1cs.CV

TL;DR

Corneal OCT interfaces can be difficult to segment consistently because they are thin, noisy, and variable across devices. ARCOS localizes boundaries from overlapping native-resolution patches, and the paper reports strong zero-shot generalization across unseen OCT systems, outperforming representative baselines.

  • Problem

    Corneal interfaces can be faint or ambiguous due to pathology, speckle, artifacts, and device differences, while generalization to entirely unseen OCT systems remains unclear.

    Source: §1 · p. 2

  • Method

    ARCOS predicts boundary heatmaps from overlapping native-resolution patches and reconstructs full-B-scan interface predictions, using multiscale features and self-conditioned refinement.

    Source: §1 · p. 2, §2.3 · p. 5, §2.5 · p. 6

  • Results

    ARCOS demonstrated strong zero-shot cross-device generalization and outperformed representative baseline architectures on unseen devices without device-specific fine-tuning.

    Source: §5 · p. 16

  • Takeaways & Limitations

    The results support boundary-focused, resolution-preserving segmentation strategies for corneal OCT analysis across heterogeneous imaging devices.

    Source: §5 · p. 16

    Future work is directed toward broader validation and adaptation to different OCT systems’ full fields of view, with uncertainty-based masking and failure-case mitigation also proposed.

    Source: §5 · p. 16

Abstract

from arXiv · show

Accurate segmentation of corneal layers in optical coherence tomography (OCT) is essential for quantitative assessment of corneal morphology, including layer thickness and structural changes associated with disease or surgery. However, automatic segmentation remains challenging because corneal interfaces are thin, affected by speckle noise, and variable across acquisition devices. In this work, we propose ARCOS, a patch-based zero-shot boundary localization framework for corneal layer segmentation in clinical anterior-segment OCT images. Rather than performing conventional region classification, the method predicts boundary heatmaps for the main corneal interfaces from overlapping native-resolution patches. Patch-level predictions are stitched across the full B-scan and converted into boundary locations to obtain continuous, anatomically ordered layer segmentations. The network combines multi-scale feature fusion with a self-conditioned refinement module that uses intermediate boundary information to improve local heatmap predictions while preserving spatial detail. The method was evaluated on clinical OCT images acquired from multiple devices and compared with representative segmentation baselines using boundary localization and derived thickness metrics. The proposed method achieved an off-by-one boundary localization accuracy of 95.1% and a mean absolute boundary error of 0.514 pixels on the matched-device test set. In zero-shot cross-device evaluation, it maintained an average off-by-one accuracy of 84.3% and a mean absolute boundary error of 0.855 pixels across unseen acquisition devices, outperforming the baseline models. Thickness estimates derived from the predicted boundaries showed low error across corneal regions, supporting the method's use for quantitative corneal OCT analysis.

1 Introduction

AS-OCT can reveal corneal layers and support measurements of morphology, but faint and variable interfaces make consistent automated segmentation difficult. ARCOS treats segmentation as boundary localization and uses native-resolution patches to predict corneal interfaces across devices.

1 Introduction

  • Corneal layer delineation supports thickness maps, geometry descriptors, and structural biomarkers for diagnosis, monitoring, and surgical planning.The measurements are relevant to conditions including keratoconus, Fuchs’ endothelial corneal dystrophy, infectious keratitis, and post-surgical remodeling.Source: §1 · pp. 1–2
  • Faint, noisy, or ambiguous interfaces arise with pathology-related deformation, speckle, acquisition artifacts, and device differences, while manual annotation is time-consuming and difficult to perform consistently.Automated methods are sought to improve reproducibility, reduce analysis time, and support objective imaging biomarkers.Source: §1 · p. 2
  • Prior methods were mainly evaluated under consistent imaging conditions, leaving generalization to entirely unseen OCT systems unclear; limited public layer-annotated datasets also complicate reproducible benchmarking.Unseen systems may differ in resolution, contrast, noise, and scanner-specific appearance.Source: §1 · p. 2
  • ARCOS processes overlapping full-resolution patches, predicts interface heatmaps, and combines patch outputs into full-resolution boundary estimates across the B-scan.The patch-based design avoids resizing the full image to a fixed input resolution and accommodates variation in image dimensions, pixel spacing, and field of view.Source: §1 · pp. 2–3
  • ARCOS frames segmentation as boundary localization because clinically relevant measurements depend on thin, continuous, anatomically ordered interfaces rather than large object-like regions.The method is designed to improve interface localization, reduce dependence on user-defined parameter tuning, and support segmentation across devices and pathological cases.Source: §1 · p. 3

2 Materials and Methods

The methods evaluate ARCOS across heterogeneous OCT systems and clinical compositions, using manual interface annotations and a patch-based heatmap pipeline. The network combines multiscale features and self-conditioned refinement, while training uses progressively strengthened augmentations.

2.1 Devices and Datasets

  • The study includes clinical and public datasets from five AS-OCT systems, with differences in modality, resolution, field of view, and pixel spacing; evaluation also spans healthy, post-surgical PRK, Fuchs’ dystrophy, and infectious keratitis cases.The device set comprises Avanti, Solix, MS-39, TowardPi, and CASIA SS-1000.Source: §2.1 · p. 3

2.2 Annotation Protocol

  • Five visually and clinically distinguishable interfaces were manually delineated, and inter-operator disagreement averaged 1.06 ± 0.56 pixels in pooled pairwise comparisons.Larger deviations occurred mainly in lateral image regions where signal quality decreases and delineation is less reliable.Source: §2.2 · p. 4

2.3 Method Overview

  • ARCOS predicts heatmaps for five interfaces from native-resolution overlapping patches, then maps and Gaussian-weight-stitches patch predictions into continuous full-B-scan boundary profiles.Central patch regions receive greater weight, attenuating edge-related artifacts; avoiding full-image resizing reduces dependence on image dimensions, field of view, and scan geometry.Source: §2.3 · p. 5, Figure 2 · p. 6

2.4 Patch Extraction

  • Patches are 384 × 384 pixels with a target 50% lateral overlap, and their vertical positions adapt to estimated corneal surfaces to keep the tissue visible.The patch height was selected to contain the full corneal thickness while providing lateral anatomical context.Source: §2.4 · p. 5

2.5 Network Architecture and Heatmap Prediction

  • A ConvNeXt-Small encoder and attention-gated multiscale decoder produce initial heatmaps, which a lightweight FiLM-based residual branch refines using a patch-level embedding.The embedding is obtained by global average pooling followed by an MLP; the residual formulation preserves the main prediction pathway while allowing patch appearance to guide local refinement.Source: §2.5 · pp. 6–7, Figure 3 · p. 7

2.6 Training Strategy

  • Training uses Adaptive Wing loss and progressively increases spatial, photometric, and OCT-specific frequency and intensity augmentations to expose the model to acquisition and appearance variability.Fourier amplitude perturbation and histogram warping add scanner-like frequency, texture, contrast, and intensity variation without requiring target-device images.Source: §2.6 · pp. 7–8, Figure 4 · p. 8

3 Experiments

Experiments train ARCOS on Avanti data and reserve other OCT systems for inference-only cross-device evaluation. The protocol standardizes preprocessing and compares representative baselines using boundary-based metrics and device-specific paired statistical tests.

3.1 Data Splits

  • Avanti supplies training and validation data, while Solix, MS-39, TowardPi, and CASIA SS-1000 are reserved exclusively for inference-time evaluation and cross-device generalization assessment.The Avanti test set holds out hyperopic PRK cases and 10% of Fuchs’ dystrophy cases.Source: §3.1 · p. 8

3.2 Preprocessing and Standardization

  • All systems use a common standardization pipeline with linear intensity scaling, and evaluation is restricted to a common central 6 mm ROI.The ROI limits the influence of peripheral signal degradation and annotation uncertainty, particularly in some CASIA SS-1000 scans.Source: §3.2 · p. 9

3.3 Baseline Models

  • Comparators include an ablated Basic Model, a full-width version of ARCOS, a region-based CorneaNet implementation, and a resized full-image ScLNet-based model.Baselines share dataset splits, preprocessing, and metrics; masks are converted to boundary coordinates for boundary-based evaluation.Source: §3.3 · p. 9

3.4 Evaluation Metrics

  • Off-by-one accuracy measures predictions within ±1 pixel, MAE measures average boundary deviation, and HD and HD95 assess geometric consistency and outliers.Device-specific paired Wilcoxon tests use image-level observations, with Benjamini–Hochberg correction and adjusted p < 0.05 as the significance threshold.Source: §3.4 · pp. 9–10

3.5 Implementation Details

  • ARCOS variants use AdamW with an initial learning rate of 10^-3, weight decay of 10^-4, and augmentation strength ramped to its maximum by 60% of training.Models use batches of 12 for up to 150 epochs; CorneaNet- and ScLNet-based baselines use optimization settings from their original publications.Source: §3.5 · p. 10

4 Results and Discussion

ARCOS localizes corneal boundaries accurately on the training device and retains the strongest, most consistent localization across unseen OCT systems. Its performance is stable with larger inference patches, depends on the combined model components, and has specific edge and artifact failure modes alongside few-second full-scan inference.

4.1 Training device results

  • On the matched-device test set, ARCOS achieved 95.1% accuracy, 0.514-pixel MAE, and 1.22-pixel HD95, the best overall localization performance.The Basic Model was the closest competitor, with a small performance gap.Source: §4.1 · p. 10, Table 4 · p. 10
  • Lateral MAE stayed below 1 pixel for every boundary, remaining low centrally and increasing moderately toward the periphery.Challenging cases included keratoconus, FECD, low-signal scans, and central artifacts; predicted boundaries remained smooth and anatomically ordered.Source: §4.1 · pp. 10–11, Figure 5 · p. 11, Figure 6 · p. 12

4.2 Cross-Device Generalization

  • ARCOS achieved the highest accuracy and lowest MAE on all five unseen OCT systems, and the lowest HD and HD95 on four.The largest gains over baselines were on MS-39, TowardPi, and CASIA SS-1000.Source: §4.2 · p. 11, Table 5 · p. 13

4.3 Model Robustness

  • Performance remained close to the 384 × 384 reference with 512 × 512 inference patches, but decreased with 256 × 256 patches.The authors suggest very small patches may omit corneal structure relevant to localization.Source: §4.3 · p. 12, Table 6 · p. 13

4.4 Ablation

  • The full model had the strongest overall performance; replacing Adaptive Wing heatmap supervision with mean squared error increased MAE from 0.79 to 1.48 pixels.Removing attention gates, FiLM-based refinement, or multi-scale fusion also reduced performance, with larger drops when components were removed jointly; combined gains were not attributable to one component alone.Source: §4.4 · p. 13, Table 7 · p. 15

4.5 Failure Cases

  • Failure modes occurred in approximately 7% of test cases, mainly at scan edges or with strong artifacts that could impair posterior-boundary localization.Evaluation and thickness measurements were restricted to the central 6 mm region; edge predictions could continue beyond visible anatomy.Source: §4.5 · pp. 14, 16

4.6 Running Time

  • Full-pipeline inference took an average of 0.52 s per B-scan on an NVIDIA Quadro P2000 and approximately 3.8 s on an Intel Xeon.Timing included patch extraction, network prediction, heatmap reconstruction, and boundary stitching.Source: §4.6 · p. 16

5 Conclusion

ARCOS uses native-resolution overlapping patches and boundary heatmap regression, generalizing across heterogeneous OCT devices without device-specific fine-tuning and outperforming representative baselines on unseen devices. Broader validation and full-field adaptation remain future work.

5 Conclusion

  • ARCOS outperformed representative baseline architectures on unseen OCT devices without device-specific fine-tuning.The conclusion reports accurate boundary localization and strong zero-shot cross-device generalisation across multiple OCT systems.Source: §5 · p. 16
  • ARCOS combines native-resolution overlapping patches with boundary heatmap regression to preserve local detail and target ordered corneal interfaces.Source: §5 · p. 16
  • Future work will broaden validation, adapt the method to different systems’ full fields of view, and address poorly visible boundaries and failure cases.Source: §5 · p. 16
Loading 2609.03668v1…