Source-linked AI summary
An Ultra-Widefield Swept-Source OCTA Dataset and a Polar-Gated Mamba Network for Retinal Vessel Segmentation
Yang Liu, Yibing Shen, Keming Zhao, Cenk Jiang, Zhenghang Qian, Zhicheng Du, Chen Xiong, Qidong Shao, Zijun Lin, Yunqi Hu, Jingjing Zhou, Lian Zhang, Peter E. Lobie, Peiwu Qin, Chengming Yang
TL;DR
UWF SS-OCTA supports large-area retinal vascular imaging, but public vessel-segmentation benchmarks and comprehensive evaluation at this scale are limited. The paper introduces WOIVES and PG-Mamba, then reports improvements across most primary segmentation metrics and lower errors for several vascular measurements.
Problem
UWF SS-OCTA vessel segmentation lacks dedicated public benchmarks and comprehensive evaluation for quantitative vascular analysis.
Source: §I · p. 1, §VI · p. 9Method
The paper introduces WOIVES, a public UWF SS-OCTA dataset with soft vessel annotations, and PG-Mamba, which adds polar-coordinate scans and Dynamic FOV Gating.
Source: C. Polar Block and Polar-Scan SS2D · p. 5, D. Dynamic FOV Gating · p. 6, §VI · p. 9Results
Under subject-level five-fold cross-validation, PG-Mamba improved performance across five of six primary metrics and achieved the lowest annotation-derived errors for VD, FD, and VLD.
Source: §VII · p. 10Takeaways & Limitations
WOIVES and PG-Mamba provide a benchmark and methodological basis for further UWF OCTA vessel-segmentation and quantitative retinal vascular-analysis research.
Source: §VII · p. 10WOIVES was collected at one center with one SS-OCTA device, and its current annotations focus on superficial retinal vessels.
Source: §VI · p. 9
Abstract
from arXiv · showhide
Ultra-widefield (UWF) swept-source optical coherence tomography angiography (SS-OCTA) enables large-area retinal vascular imaging, yet vessel segmentation at this scale lacks dedicated public benchmarks and comprehensive evaluation for quantitative vascular analysis. We introduce WOIVES, to our knowledge the first publicly available UWF SS-OCTA vessel-segmentation dataset, comprising 206 eyes from 152 participants with a 24x20mm^2 field of view. WOIVES spans emmetropia to high myopia and provides soft probability vessel annotations. We further propose PG-Mamba, a visual state space model that enhances conventional directional scans with two complementary polar-coordinate scan orders. An auxiliary Dynamic FOV Gating module performs spatial modulation at the bottleneck. PG-Mamba outperformed seven competitive approaches on broad segmentation metrics under cross-validation. It achieved the lowest median absolute errors for vessel density, fractal dimension, and vessel length density. WOIVES is publicly available on Zenodo (DOI: 10.5281/zenodo.21904672), and the PG-Mamba code is available at https://github.com/syb1234567/PG-Mamba.
I. INTRODUCTION
The introduction motivates wider retinal imaging for myopia-related vascular assessment and identifies a lack of public UWF SS-OCTA vessel-segmentation resources. It frames WOIVES and PG-Mamba as addressing dataset, scan-design, and evaluation gaps.
I. INTRODUCTION
- UWF SS-OCTA can image central and peripheral vasculature, but reliable segmentation is needed because errors affect downstream vascular measurements.Thin vessels and uncertain boundaries are particularly relevant to vessel density, vessel length density, fractal dimension, and connectivity-related measures.Source: §I · p. 1
A. OCTA Datasets
- Public OCTA segmentation datasets cover fields no larger than 12×12 mm2 and lack the combined UWF SS-OCTA annotations and ocular biometry needed for myopia-oriented analysis.WOIVES addresses this resource gap with 24×20 mm2 imaging, soft probability maps, and biometry for a clinically characterized subset.Source: §I · p. 1, A. OCTA Datasets · p. 2
B. Retinal Vessel Segmentation
- Existing CNN, Transformer, and state-space approaches were evaluated mainly on conventional fields, while scan-order design and evaluation beyond binary overlap remain open concerns.The introduction motivates complementary scan patterns and assessment of continuous annotation agreement, vascular structure, and downstream measurements.Source: §I · p. 2, B. Retinal Vessel Segmentation · p. 2
I. INTRODUCTION
- WOIVES provides 206 eyes from 152 participants with soft vessel annotations, and PG-Mamba adds two polar scan orders and bottleneck gating for UWF segmentation.The paper evaluates seven baselines using multiple metric types and reports improvements across five of six primary metrics, with lower annotation-derived errors for VD, FD, and VLD.Source: §I · p. 2
C. State Space Models in Medical Imaging
The paper places PG-Mamba within work on visual state-space scan design and soft-label evaluation, then describes WOIVES acquisition. Its dataset uses widefield swept-source OCTA alongside other ophthalmic images and clinical metadata.
C. State Space Models in Medical Imaging
- Visual state-space models serialize 2D features into sequences, making scan order consequential; PG-Mamba adds two polar-coordinate scans to existing directional scans and uses bottleneck gating.The paper situates this design among work exploring scan configurations and adaptive feature modeling.Source: C. State Space Models in Medical Imaging · p. 2
D. Soft Labels and Probability-Level Evaluation
- Soft labels retain continuous annotation information, while OCTA evaluation can combine overlap, probability-sensitive, topology, caliber, and downstream-measurement analyses.The paper motivates this broader evaluation because relationships among soft agreement, vessel caliber, topology, and vascular measurements have received less attention.Source: D. Soft Labels and Probability-Level… · p. 3
III. WOIVES DATASET
- WOIVES is a dataset construction effort covering acquisition, image processing, annotation, and dataset statistics.The supplied dataset-section passage introduces these components without further detail.Source: §III · p. 3
A. Data Acquisition and Eligibility Criteria
- Imaging used a swept-source OCTA system with a maximum 24×20 mm FOV, acquiring superficial, deep, and full-retina OCTA plus three choroidal images per eye.Participants were classified as emmetropic, myopic, or highly myopic using spherical-equivalent thresholds.Source: A. Data Acquisition and Eligibility… · p. 3
B. Data Processing
The dataset workflow quality-controls and standardizes acquired images, then creates expert-refined soft vessel annotations. WOIVES separates its full imaging cohort from a clinically characterized subset and offers a substantially larger field than comparison datasets.
B. Data Processing
- Images with severe artifacts, defocus, motion blur, or exposure errors were excluded, while retained images were desensitized, resized, normalized, and paired with de-identified clinical metadata.The processing stage aims to ensure uniform dimensions and standardized pixel-intensity distributions.Source: B. Data Processing · p. 3
C. Annotation Pipeline
- Five junior ophthalmologists used SAM-OCTA-assisted prompts to create masks that were averaged into soft probability maps, then two senior experts refined missed vessels, false positives, and boundaries.The final maps retain uncertain boundary zones rather than forcing a binary label, allowing training on graded confidence.Source: C. Annotation Pipeline · pp. 3–4
- WOIVES retains soft probability labels because peripheral vessel boundaries can be ambiguous, preserving annotator disagreement that binary thresholding would discard.The paper links continuous labels to soft-label training and finer-grained boundary estimation for downstream measurements.Source: C. Annotation Pipeline · pp. 3–4
D. Dataset Statistics and Comparison
- The imaging cohort contains 206 eyes from 152 participants, while 177 eyes from 130 participants have complete clinical and biometric metadata; its 480 mm2 FOV is 13.3× OCTA-500 and 3.3× DRAC.The full cohort supports segmentation and vascular-measurement analyses; the smaller subset provides additional ophthalmic metadata.Source: D. Dataset Statistics and Comparison · p. 4
IV. METHOD
The method section introduces PG-Mamba as a Mamba-based U-shaped network for UWF OCTA vessel segmentation. Its preliminaries explain how state-space models process sequences and why feature scan order matters.
IV. METHOD
- PG-Mamba is presented as a Mamba-based U-shaped network, with the method section covering its architecture, proposed modules, and training objective.The supplied overview does not specify those components further.Source: §IV · p. 4
A. Preliminaries: State Space Models
- State-space models map an input sequence to an output through a latent state, and discrete computation uses a recurrence.For visual features, Mamba processes serialized sequences; scan order determines how spatial locations are arranged for state propagation.Source: A. Preliminaries: State Space Models · pp. 4–5
B. Overall Architecture
PG-Mamba uses a U-shaped encoder–decoder for single-channel UWF OCTA segmentation and modifies visual state-space scanning with two polar-coordinate orders.
B. Overall Architecture
- A U-shaped encoder–decoder maps single-channel OCTA tiles to full-resolution vessel-probability maps through four Polar Block stages, skip connections, and bottleneck gating.The encoder and decoder use mirrored hierarchies, with patch merging and patch expansion for down- and upsampling.Source: B. Overall Architecture · p. 5, Figure 2 · p. 6
C. Polar Block and Polar-Scan SS2D
- Polar-Scan SS2D adds Radial Out and Radial In to horizontal, vertical, main-diagonal, and anti-diagonal scans, ordering features jointly by radius and angle.The opposite radial-term signs create opposite dominant radial orderings.Source: C. Polar Block and Polar-Scan SS2D · p. 5
- Six scan routes are restored to the original spatial layout and directly summed; the Polar Block replaces the VSS block's selective scan with this operation.The architecture depicts the Polar Block with six scan orders, including both proposed polar scans.Source: C. Polar Block and Polar-Scan SS2D · pp. 5–6, Figure 2 · p. 6
D. Dynamic FOV Gating
Dynamic FOV Gating modulates bottleneck features using a content-adaptive soft mask and a fixed radial constraint.
D. Dynamic FOV Gating
- Dynamic FOV Gating applies spatial modulation at the encoder–decoder bottleneck using normalized polar coordinates constructed from bottleneck features.The polar grid is concatenated with the features to generate the adaptive mask.Source: D. Dynamic FOV Gating · p. 6
- The module combines the soft and hard masks to produce gated bottleneck features.The soft mask is generated from the concatenated polar grid and bottleneck features; the hard mask supplies fixed radial support.Source: D. Dynamic FOV Gating · p. 6
- A content-adaptive soft mask provides feature-dependent weighting, while a fixed radius-thresholded hard mask imposes a radial constraint on bottleneck features.Both masks are spatial and shared across channels.Source: D. Dynamic FOV Gating · p. 6
E. Loss Function
The loss combines soft Dice and mean squared error to supervise the soft probability vessel annotations.
E. Loss Function
- The training objective combines soft Dice loss with mean squared error because WOIVES annotations are retained as soft probability maps.All compared methods use the same loss function.Source: E. Loss Function · pp. 6–7
- The soft Dice term promotes spatial overlap between predicted vessel probabilities and continuous annotation values.The annotation value for each pixel is represented by g_i ∈ [0, 1].Source: E. Loss Function · p. 7
- Mean squared error provides pixel-wise supervision for the continuous annotations.The compared methods are trained with the same composite loss.Source: E. Loss Function · p. 7
V. EXPERIMENTS
The experiments use subject-level five-fold cross-validation on the WOIVES cohort, with both eyes from each participant kept in the same fold.
1) Dataset and Cross-Validation:
- Subject-level five-fold GroupKFold assigns both eyes from each participant to the same fold and keeps participants disjoint between training and test sets.A subject-disjoint validation split is held out from the training portion within each fold.Source: 1) Dataset and Cross-Validation · p. 7
- Reported values are mean±standard deviation across the five test folds unless otherwise stated.The cohort comprises 206 eyes from 152 participants, using superficial-retina OCTA images.Source: 1) Dataset and Cross-Validation · p. 7
- Eye-level paired analyses pool scores across test folds (n = 206), with a participant-level sensitivity analysis averaging paired differences within participants.The sensitivity analysis is provided in the supplementary material.Source: 1) Dataset and Cross-Validation · p. 7
2) Evaluation Metrics:
PG-Mamba leads seven comparators on five of six segmentation metrics, while H2Former has the lowest Brier score. Controlled ablations show progressive gains from adding polar scans and then Dynamic FOV Gating.
2) Evaluation Metrics:
- The evaluation combines thresholded overlap, soft-label agreement, pixel-wise error, and vessel-topology measures; binary metrics use a fixed threshold of 0.5.Soft Dice uses continuous predictions and labels, while clDice evaluates agreement between skeletonized vessel structures.Source: 2) Evaluation Metrics · p. 7
3) Implementation Details:
- Training used 512×512 tiles, AdamW, cosine-annealed learning rates, and subject-disjoint validation for early stopping and model selection.All models used the same composite loss and a fixed random seed.Source: 3) Implementation Details · p. 7
4) Compared Methods:
- The seven retrained comparators span CNN, Transformer, and Mamba architectures and use the same single-channel input and composite loss.They include U-Net, UNet++, R2U-Net, Swin-UNet, H2Former, VM-UNet, and AC-MambaSeg.Source: 4) Compared Methods · p. 7
B. Comparative Segmentation Performance
- PG-Mamba leads on Dice (89.81%), IoU (82.20%), soft Dice (85.99%), clDice (91.97%), and MAE (0.0257); H2Former leads on Brier score (0.0146).PG-Mamba and AC-MambaSeg share the second-lowest Brier score (0.0156).Source: B. Comparative Segmentation Performance · p. 7, Table II · p. 8
- PG-Mamba’s gains over VM-UNet are modest: mean Dice improves by 0.44 percentage points and clDice by 0.38 percentage points, with eye-level differences significant for all six metrics.The authors caution that eye-level tests do not fully account for bilateral correlation.Source: B. Comparative Segmentation Performance · p. 7
C. Ablation Study
- Adding radial-outward and radial-inward scans improves Dice from 89.43% to 89.67% and clDice from 91.54% to 91.73%, while reducing Brier score from 0.0162 to 0.0158.The Dice improvement is significant in the eye-level paired analysis (p = 1.3 × 10^-7).Source: C. Ablation Study · pp. 7–8
- Adding Dynamic FOV Gating further improves Dice from 89.67% to 89.81% and clDice from 91.73% to 91.97%, while reducing Brier score from 0.0158 to 0.0156.Eye-level paired differences are significant for Dice (p = 1.2 × 10^-5) and clDice (p = 1.1 × 10^-7).Source: C. Ablation Study · p. 8
D. Vascular Measurement Analysis
PG-Mamba has the lowest median absolute errors for vessel density, fractal dimension, and vessel length density, but not for normalized connected-component count. The analysis compares prediction-derived measurements with annotation-derived references and cautions that this is not independent validation of clinical measurement accuracy.
D. Vascular Measurement Analysis
- The analysis derives VD, FD, normalized VLD, and normalized NCC from binarized predictions and annotation-derived references, using a fixed 0.5 threshold.NCC counts 8-connected components per 10^5 FOV pixels; VLD counts skeletonized vessel pixels per 10^3 FOV pixels.Source: D. Vascular Measurement Analysis · p. 8, Figure 3 · p. 9
- 9.2508 is PG-Mamba’s median absolute error for NCC, higher than U-Net’s 7.9610 and VM-UNet’s 7.7960 but lower than R2U-Net’s 10.0523.PG-Mamba has significantly lower NCC error than R2U-Net but significantly higher error than U-Net and VM-UNet.Source: D. Vascular Measurement Analysis · p. 8, Figure 3 · p. 9
- 0.0081 for VD, 0.0109 for FD, and 1.3414 for VLD are PG-Mamba’s lowest median absolute errors among the compared models.For each of these measures, paired error differences versus U-Net, R2U-Net, and VM-UNet remain significant after Holm correction (pHolm < 0.05).Source: D. Vascular Measurement Analysis · p. 8, Figure 3 · p. 9
- Agreement with annotation-derived measurements is not independent validation of clinical measurement accuracy.NCC is especially sensitive to small disconnected regions and local breaks, unlike measures reflecting vessel area, geometric complexity, or skeleton extent.Source: D. Vascular Measurement Analysis · p. 9
VI. DISCUSSION
WOIVES provides a large-field, biometry-linked UWF SS-OCTA resource, while PG-Mamba combines complementary polar scans with bottleneck gating. The reported segmentation and vascular-measurement gains are accompanied by metric-specific and dataset-scope qualifications.
VI. DISCUSSION
- WOIVES combines 24×20 mm2 UWF SS-OCTA vessel annotations with ocular biometry across emmetropia, myopia, and high myopia, filling a public-resource gap for large-field analysis.Its characteristics may also support future investigation of myopia-related vascular changes and multimodal ophthalmic analysis.Source: §VI · p. 9
- Soft-label evaluation exposed metric-dependent rankings: H2Former had the lowest Brier score, while AC-MambaSeg had PG-Mamba-comparable Brier but lower soft Dice and higher MAE.Higher recall in the small local-thickness region did not necessarily correspond to better overall segmentation, supporting complementary overlap, probability-sensitive, and topology-aware metrics.Source: §VI · p. 9
- PG-Mamba's measurement advantage was not uniform: its NCC error was lower than R2U-Net's but higher than U-Net's and VM-UNet's.VD reflects vessel area, FD geometric complexity, VLD skeleton extent, and NCC connected-component structure.Source: §VI · p. 9
- WOIVES is single-center and single-device, has more myopic and highly myopic than emmetropic eyes, and currently annotates only superficial retinal vessels.Its soft probability maps represent multi-annotator consensus, not calibrated biological probabilities.Source: §VI · p. 9
- PG-Mamba uses fixed polar scans and bottleneck-only Dynamic FOV Gating; vascular measurements assess agreement with expert-annotation-derived references, not independent clinical validity.The authors identify adaptive scan orders and multi-scale spatial modulation as future directions.Source: §VI · p. 10
VII. CONCLUSION
- Across five of six primary metrics, PG-Mamba modestly and consistently improved over strong CNN, Transformer, and Mamba baselines; it also achieved the lowest annotation-derived errors for VD, FD, and VLD.The median absolute-error differences for VD, FD, and VLD remained significant after Holm correction.Source: §VI · p. 9, §VII · p. 10