PaperlayerView exact revision on arXiv

Source-linked AI summary

ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model

Yuhui Ma, Huaying Hao, Huazhu Fu, Jiong Zhang, Jianlong Yang, Jiang Liu, Yalin Zheng, Yitian Zhao

arXiv:2007.05201v2eess.IVcs.CV

TL;DR

OCTA vessel segmentation is hindered by difficult microvascular imaging and the absence of a public manually annotated benchmark. The paper introduces the ROSE dataset and a split-based coarse-to-fine network, then evaluates segmentation methods on ROSE. The authors report that their approach outperforms other state-of-the-art methods and shows significant improvements for small-capillary segmentation.

  • Problem

    OCTA vessel segmentation is challenging because capillaries have low visibility amid noise and complex artifacts, while no public manually annotated OCTA dataset supports unified validation.

  • Method

    The paper constructs ROSE with centerline- and pixel-level annotations and proposes split-based coarse-to-fine segmentation using separate branches for vessel characteristics and annotation types.

  • Results

    The proposed vessel segmentation approach outperforms other state-of-the-art methods on ROSE, with significant improvements reported for small-capillary metrics.

  • Takeaways & Limitations

    ROSE provides a public benchmark for OCTA vessel segmentation, while the results support further analysis of retinal microvasculature, including potential Alzheimer’s disease studies.

Abstract

from arXiv · show

Optical Coherence Tomography Angiography (OCT-A) is a non-invasive imaging technique, and has been increasingly used to image the retinal vasculature at capillary level resolution. However, automated segmentation of retinal vessels in OCT-A has been under-studied due to various challenges such as low capillary visibility and high vessel complexity, despite its significance in understanding many eye-related diseases. In addition, there is no publicly available OCT-A dataset with manually graded vessels for training and validation. To address these issues, for the first time in the field of retinal image analysis we construct a dedicated Retinal OCT-A SEgmentation dataset (ROSE), which consists of 229 OCT-A images with vessel annotations at either centerline-level or pixel level. This dataset has been released for public access to assist researchers in the community in undertaking research in related topics. Secondly, we propose a novel Split-based Coarse-to-Fine vessel segmentation network (SCF-Net), with the ability to detect thick and thin vessels separately. In the SCF-Net, a split-based coarse segmentation (SCS) module is first introduced to produce a preliminary confidence map of vessels, and a split-based refinement (SRN) module is then used to optimize the shape/contour of the retinal microvasculature. Thirdly, we perform a thorough evaluation of the state-of-the-art vessel segmentation models and our SCF-Net on the proposed ROSE dataset. The experimental results demonstrate that our SCF-Net yields better vessel segmentation performance in OCT-A than both traditional methods and other deep learning methods.

I. INTRODUCTION

OCTA provides non-invasive, high-resolution views of retinal microvasculature, but vessel segmentation remains difficult and lacks a public manually annotated benchmark. The paper motivates ROSE around detailed 3 × 3 mm2 capillary imaging.

  • I. INTRODUCTION: Retinal vascular morphology supports disease assessment, progression understanding, and evaluation of therapeutic effects.
  • I. INTRODUCTION: OCTA non-invasively produces high-resolution 3D retinal vascular images and en face projections across depth layers.These include the superficial vascular complex, deep vascular complex, and combined inner retinal plexus.
  • I. INTRODUCTION: No publicly available OCTA dataset had manual vessel annotations, while prior deep-learning studies used private datasets that prevented unified evaluation.Existing methods therefore lacked a common public benchmark for validation.
  • I. INTRODUCTION: 3 × 3 mm2 scans provide higher resolution and clearer delineation of the FAZ and capillaries than 6 × 6 mm2 scans.The 6 × 6 mm2 mode offers wider coverage but poorer microvascular delineation.
  • I. INTRODUCTION: Low SNR, depth-dependent appearance, projection, motion, shadow artifacts, pathology, and dense capillaries challenge accurate OCTA vessel segmentation.These factors can produce discontinuous segmentations, especially in complex microvascular networks.

A. Contributions

The paper addresses missing public OCTA benchmarks and limited microvascular segmentation by releasing ROSE, proposing a split-based coarse-to-fine network, and benchmarking segmentation methods.

  • A. Contributions: ROSE is presented as a publicly available retinal OCTA dataset with precise manual annotations of retinal microvascular networks.The authors release code, baseline models, and evaluation tools alongside the dataset.
  • A. Contributions: The proposed split-based coarse-to-fine network separately targets thick and thin vessels through coarse confidence estimation and refined segmentation.Its modules are described as split-based coarse segmentation and split-based refined segmentation.
  • A. Contributions: The work benchmarks OCTA microvascular segmentation quantitatively and qualitatively across selected traditional and deep-learning methods.The contribution statement reports evaluation on different retinal image types.
  • II. RELATED WORKS: Earlier OCTA vessel segmentation was relatively unexplored and relied mainly on thresholding schemes.The related work describes hybrid Hessian/intensity, reflectance-based, and adaptive-thresholding approaches.
  • II. RELATED WORKS: Threshold-based methods are sensitive to noise and perform poorly where intensity differences are not significant.

III. DATASET

ROSE contains two device-specific OCTA subsets totaling 229 images, with expert annotations at centerline and/or pixel level. ROSE-1 emphasizes layered 3 × 3 mm2 microvascular analysis, while ROSE-2 covers diseased maculae.

  • III. DATASET: ROSE comprises ROSE-1 and ROSE-2, acquired using two different OCTA devices under approved studies with informed consent.The dataset includes each subject’s diagnosis result.
  • A. ROSE-1: ROSE-1 contains 117 images from 39 subjects, including 26 with Alzheimer’s disease and 13 healthy controls.
  • A. ROSE-1: ROSE-1 provides en face images of the SVC, DVC, and combined SVC+DVC with centerline-level and pixel-level manual annotations.Figure 3 organizes layers by rows and image or annotation types by columns.
  • A. ROSE-1: Small macular capillaries were graded at centerline level because experts could not reliably perceive their diameters.These labels were combined with complete pixel-level microvascular grading to form the final pixel-level annotation.

B. ROSE-2

ROSE-2 contains 112 OCTA images from eyes with various macular diseases, and OCTA-Net processes vessel segmentation through coarse confidence estimation followed by fine fusion refinement.

  • B. ROSE-2: ROSE-2 contains 112 OCTA images from 112 eyes with various macular diseases, acquired using a Heidelberg OCT2 system.The images are en face angiograms of the superficial vascular complex within a 3×3 mm2 fovea-centered area.
  • B. ROSE-2: OCTA-Net uses two stages: SCS produces preliminary vessel confidence maps, and SRS fuses them to generate final optimized segmentations.The coarse-to-fine pipeline is illustrated in Fig. 5.

A. Coarse Stage: Split-based Coarse Segmentation Module

The coarse stage uses split representations and attention to segment vessels, with separate branches when both pixel-level and centerline-level annotations are available.

  • A. Coarse Stage: Split-based Coarse Segmentation Module: The SCS module uses a partially shared encoder and two decoder branches for pixel-level and centerline-level vessel segmentation in ROSE-1.For ROSE-2 and ROSE-1 DVC images, only one encoder-decoder path is used because only centerline-level annotations are available.
  • A. Coarse Stage: Split-based Coarse Segmentation Module: Each pixel-level segmentation path is a U-shaped network with five encoder layers and symmetric decoder layers using ResNeSt blocks.ResNeSt blocks use split feature groups and channel-wise attention.
  • A. Coarse Stage: Split-based Coarse Segmentation Module: ResNeSt blocks split representations into feature groups, process parallel branches, and integrate them with channel-wise attention.The input is divided into cardinal groups and then further split into parallel branches.
  • A. Coarse Stage: Split-based Coarse Segmentation Module: The split attention module sums branch features, applies global pooling, and uses fully connected layers with softmax to obtain channel-wise weights.Weighted branch outputs are summed, concatenated across cardinal blocks, transformed by a 1 × 1 convolution, and combined with a shortcut.
  • A. Coarse Stage: Split-based Coarse Segmentation Module: Centerline-level segmentation targets vessels in poor-contrast, complex, and relatively small-diameter regions, requiring attention to low-level features.The method therefore shares some extracted features while using a dedicated centerline-level pathway.

B. Fine Stage: Split-based Refined Segmentation Module

The fine stage refines coarse vessel predictions by combining the original OCTA image with preliminary maps and propagating neighbor information through learned local coefficients.

  • B. Fine Stage: Split-based Refined Segmentation Module: The SRS module concatenates the original single-channel OCTA image with pixel-level and centerline-level coarse maps as a three-channel input.This design integrates both vessel representations before refinement.
  • B. Fine Stage: Split-based Refined Segmentation Module: SRS produces normalized m × m local propagation coefficient maps at every position for vessel refinement.The coefficients model relationships between each position and its local neighbors.
  • B. Fine Stage: Split-based Refined Segmentation Module: The learned coefficient vector at each position is multiplied by coarse confidence maps for thick or thin vessels and aggregated at the center point.The resulting vector is the refined vessel prediction.
  • B. Fine Stage: Split-based Refined Segmentation Module: Propagation coefficients learn spatial relationships among neighboring positions while encouraging refined maps to remain similar to pre-refinement predictions.Initialization makes the center position coefficient larger than those of its neighbors.

V. EXPERIMENTS

The experiments train OCTA-Net on ROSE subsets with separate data splits and evaluate segmentation using visual comparisons, loss settings, and dataset-specific annotation handling.

  • V. EXPERIMENTS: Both coarse and fine stages were trained for 200 epochs using Adam with learning rate 0.0005, batch size 2, and weight decay 0.0001.A poly learning-rate policy with power 0.9 was used for more stable training.
  • V. EXPERIMENTS: The coarse stage uses mean square error loss, with N denoting the total number of pixels and p_i and g_i denoting prediction and ground-truth pixels.The implementation sets the split-attention reduction ratio to r = 16.
  • V. EXPERIMENTS: OCTA-Net produced more visually informative results than U-Net, CE-Net, CS-Net, and the ResNeSt backbone in qualitative comparisons.U-Net tended to over-segment dense regions, whereas CE-Net and CS-Net did not preserve fine capillaries well.
  • V. EXPERIMENTS: The fine stage uses Dice coefficient loss instead of MSE because vessel regions are highly imbalanced against background regions.The aggregation-neighbor size is set to m = 3.
  • V. EXPERIMENTS: ROSE-1 was split into 90 training and 27 testing images, while ROSE-2 was split into 90 training and 22 testing images.Random rotations between −10° and 10° were used for augmentation during all training stages.
  • V. EXPERIMENTS: OCTA-Net was trained separately on ROSE-1 and ROSE-2 so annotation differences would not affect evaluation reliability.ROSE-1 training used both pixel-level and centerline-level manual annotations.

B. Evaluation Metrics

The evaluation uses complementary metrics for classification agreement, vessel detection, error rates, and overall segmentation quality, with a three-pixel tolerance for centerline annotations.

  • The study compares AUC, sensitivity, specificity, accuracy, Kappa, false discovery rate, G-mean, and Dice coefficient.These metrics cover discrimination, detection, agreement, error rates, and overlap-based segmentation quality.
  • Sensitivity, specificity, and accuracy are defined from true-positive, false-positive, true-negative, and false-negative counts.Sensitivity is TP/(TP + FN), specificity is TN/(TN + FP), and accuracy is (TP + TN)/(TP + TN + FP + FN).
  • Kappa measures agreement beyond opportunity consistency, while FDR measures the fraction of predicted positives that are false.Kappa uses pe for opportunity consistency, and FDR is FP/(FP + TP).
  • Sensitivity and specificity alone are inadequate because over-segmentation can preserve sensitivity while most image pixels are non-vessel.This motivates using additional metrics for the highly imbalanced segmentation task.
  • For centerline-level detection, pixels within a three-pixel tolerance region around manually traced centerlines count as true positives.This rule applies to DVC images in ROSE-1 and all images in ROSE-2.

C. Performance Comparison and Analysis

The proposed two-stage method is compared with conventional and deep learning vessel-segmentation approaches across ROSE-1 and ROSE-2. It generally achieves the strongest reported performance, especially for thin capillaries and challenging vessel structures, while retaining relatively short inference time.

  • Comparison methods: The evaluation compares the proposed method with three conventional and six deep learning segmentation approaches on ROSE-1 and ROSE-2.The conventional methods are IPAC, trainable COSFIRE, and COOF; deep learning methods include U-Net, ResU-Net, and other state-of-the-art networks.
  • Subjective comparisons: The proposed method identifies more complete and thinner vessels, with more uniform responses across thick and thin vessels and improved capillary sensitivity.These visual advantages are reported particularly for ROSE-1 DVC and SVC+DVC angiograms.
  • Performance on the DVC layer in ROSE-1: About 12.0% higher Kappa, 11.9% higher Dice, and 13.2% lower FDR than CS-Net are reported for ROSE-1 DVC images.The AUC improvement over CS-Net and DUNet is not significant, but the proposed network outperforms the other compared approaches overall.
  • Performance on SVC+DVC angiograms: The proposed method achieves overall best performance on ROSE-1 SVC+DVC angiograms, while conventional methods have lower detection rates than deep learning methods.All reported statistical tests for the proposed method have p ≤ 0.001; Azzopardi et al. obtains FDR 0.0988.
  • Performance on ROSE-2: On ROSE-2, the method achieves the best AUC, accuracy, Kappa, and Dice, with all comparisons significant at p < 0.001 except against the three-stage model.ROSE-2 contains centerline-level annotations for en face SVC images.
  • Inference time: The two-stage framework achieves better segmentation performance within relatively shorter inference time on both ROSE-1 and ROSE-2.Inference-time comparisons among deep learning models were conducted using PyTorch for fairness.
  • ROC-curve analysis: ROC curves show the proposed two-stage architecture achieving the best curve in the displayed comparisons, attributed to ResNeSt feature extraction and fine-stage local refinement.The figure shows SVC+DVC results for ROSE-1 and also compares SVC results for ROSE-2.

A. Ablation Studies

Ablation studies examine the ResNeSt backbone, joint pixel- and centerline-level learning, and two-stage training. The reported analyses support contributions from the backbone and joint learning, while the supplied passage only introduces the two-stage analysis.

  • Ablation design: The ablation studies evaluate the ResNeSt backbone, joint pixel- and centerline-level learning, and two-stage training against a U-Net baseline.The paper gradually evaluates how these components affect segmentation results.
  • Ablation for ResNeSt backbone: The ResNeSt encoder-decoder achieves the best AUC, accuracy, Kappa, Dice, and FDR on both ROSE-1 and ROSE-2 compared with U-Net and ResU-Net.This comparison isolates the effect of the ResNeSt backbone within the encoder-decoder architecture.
  • Ablation for joint learning: Joint pixel- and centerline-level learning produces higher AUC, accuracy, Kappa, and Dice than a single segmentation branch.The authors associate this improvement with highlighting topological distinctions between pixel-level and centerline-level vessels.
  • Ablation for two-stage training: The two-stage ablation treats coarse-stage outputs as preliminary segmentations and uses them with the original image at the fine stage to produce final segmentations.The supplied passage describes the procedure for ROSE-1 and ROSE-2 but does not report its quantitative outcome.
  • Downstream analysis: ROSE-1 contains normal and Alzheimer’s disease subjects whose SVC, DVC, and SVC+DVC segmentations are used for fractal-dimension analysis.The analysis includes 39 normal images and 78 Alzheimer’s disease images, with three angiograms per subject.

C. Projection artefacts

Projection artefacts are a major OCTA limitation that may persist after proprietary preprocessing, while the paper evaluates a manually annotated dataset and coarse-to-fine segmentation approach.

  • C. Projection artefacts: Projection artefacts remain a main OCTA limitation and may require further technological improvements beyond this paper’s focus.
  • D. Conclusions: ROSE is a manually annotated OCTA dataset containing centerline-level and/or pixel-level vessel labels from two imaging devices.The dataset was released for vessel-segmentation research and protects patient privacy by removing identities.
  • D. Conclusions: The proposed two-stage framework first produces preliminary segmentations, then refines them using the original images and coarse-stage results.The coarse stage uses a split-based coarse segmentation module, while the fine stage uses a split-based refined segmentation module.
  • D. Conclusions: The approach outperforms other state-of-the-art methods on ROSE, with particularly significant improvements reported for small capillaries.These capillaries are often only 2–4 pixels wide, and extreme foreground–background imbalance makes small metric gains meaningful.
Loading 2007.05201v2…