CNVSeg: CNV Segmentation Dataset & MTG-Net

• CNVSeg is a dataset of 224 OCTA images from wet AMD patients, offering detailed pixel-level annotations for both CNV regions and intralesional vessels.
• MTG-Net employs a multi-task U-shaped network with graph convolutional reasoning, integrating boundary detection, shape regression, and region segmentation tasks.
• Empirical results demonstrate MTG-Net’s improved region and vessel Dice scores, showcasing robust performance across multiple imaging devices.

Choroidal neovascularization (CNV) is a key pathological hallmark of wet age-related macular degeneration (wet AMD) and a major cause of vision loss globally. Reliable segmentation of CNV lesions and associated vascular structures in optical coherence tomography angiography (OCTA) images is essential for clinical diagnosis, prognosis, and treatment planning. CNVSeg, as introduced in (Chen et al., 5 Aug 2025), constitutes the first publicly accessible dataset for CNV region and vessel segmentation in OCTA, alongside a specialized segmentation model (MTG-Net) that leverages multilateral graph convolutional reasoning to tackle intrinsic challenges related to irregular lesion morphology, imaging artifacts, and dataset heterogeneity.

1. CNVSeg Dataset: Construction and Features

CNVSeg comprises 224 en-face OCTA acquisitions from wet-AMD patients, distributed across three scan areas (3 × 3 mm², 6 × 6 mm², 9 × 9 mm²) and two native resolutions (304 × 304, 384 × 384 pixels). The dataset aggregates data from multiple devices including Heidelberg (128 images at 3×3 mm², 384×384 px), Optovue (40 images at 6×6 mm², 304×304 px), Zeiss (32 images across two configurations at 384×384 px), and SVision (24 images across two configurations at 384×384 px).

Pixel-level annotations for both CNV regions and intralesional vessels were performed using ITK-SNAP by two junior ophthalmologists, leveraging 3D OCTA rendering for resolving ambiguities. This was followed by a two-stage review: consensus adjudication by a senior ophthalmologist, then a global review by a second senior expert, ensuring high annotation consistency despite the absence of formal inter-annotator κ-scores.

Preprocessing includes selection of en-face “avascular” slabs (from outer plexiform layer to Bruch’s membrane), random flip augmentations (±45°), and intensity normalization to [0,1]. The standard protocol divides data into 60% train, 10% validation, and 30% test sets (mixed-device), with cross-device splits (e.g. leave-one-device-out) to quantify generalization. CNVSeg is open-access at https://github.com/jiongzhang-john/CNV.

2. MTG-Net: Multilateral Graph Convolutional Network

MTG-Net employs a multi-task nested U-shape architecture (five encoder/decoder levels), in which internal U²Net blocks with dilated convolutions provide deep multiscale context. The framework decomposes each input image into three specialized tasks:

• Boundary detection: Edge localization for delineating lesion margins.
• Shape regression: Prediction of signed distance maps to characterize lesion shape.
• Region segmentation: Binary mask estimation for CNV extent.

Task-specific feature maps are derived through channel-wise concatenation of specific backbone layers, each regularized by a 1×1 convolution and batch normalization.

Two inter-task modules, Multilateral Interaction Graph Reasoning (MIGR) and Multilateral Reinforcement Graph Reasoning (MRGR), implement higher-order message passing within the graph domain:

• MIGR involves projection of dense feature maps into a set of K graph nodes using learnable soft clustering centers and scale parameters, followed by cross-task attention (region-to-boundary and region-to-shape) using MLP-encoded key-value-query schemes and affinity-based message propagation. Graph convolution (with learnable weights) and reprojection to image space enable context fusion.
• MRGR fuses the outputs of boundary- and shape-enhanced region features with region predictions to construct supportive node graphs via kNN affinity, producing attention-weighted features that reinforce region segmentation.

The graph reasoning is iteratively applied to allow complementary optimization between region, boundary, and shape tasks.

3. Optimization and Training Protocols

MTG-Net introduces an uncertainty-weighted loss to counteract the impact of imaging noise and projection artifacts. Using Monte Carlo Dropout (Z=10 or 15 samples per input), the model estimates pixelwise prediction variance $V$, which is used to modulate the local binary cross-entropy loss:

$$\mathcal{L}_{uce} = -\frac1Z \sum_{i=1}^Z \bigl[ (1 + \mathrm{Normalize}(V_i))\,[\,X_i\log\hat X_i + (1−X_i)\log(1−\hat X_i)\,]\bigr]$$

where $X_i, \hat X_i$ denote ground-truth and predicted probabilities for pixel $i$. Higher uncertainty assigns larger loss weight, focusing learning on challenging regions.

The composite multi-task loss combines uncertainty-weighted and task-specific objectives:

• $$\mathcal{L}_{reg} = \mathcal{L}_{dice}^r + \mathcal{L}_{uce}^r$$
• $$\mathcal{L}_{bou} = \mathcal{L}_{dice}^b + \mathcal{L}_{uce}^b$$
• $$\mathcal{L}_{shp} = \mathcal{L}_{mse}^s + \mathcal{L}_{uce}^s$$
• $$\mathcal{L}_{ves} = \mathcal{L}_{dice}^v + \mathcal{L}_{uce}^v$$

Region-related losses are weighted adaptively using uncertainty estimates across tasks.

Training uses PyTorch on NVIDIA RTX 3090, Adam optimizer (lr=$10^{-4}$, weight decay=$10^{-4}$), batch size 4, for 300 epochs, with image input size 384×384. Data augmentations mirror those used in dataset preparation.

4. Empirical Performance and Ablation

MTG-Net demonstrates superior segmentation capability on CNVSeg, as summarized below:

Test Setup	Region Dice	Vessel Dice	Δ Dice (vs. U²Net)
Mixed-device (mean)	0.8721 ±0.088	0.8812 ±0.086	+2.9%
Single-dev. (Heid.)	0.8143	0.8109	--
Single-dev. (Optovue)	0.9038	0.9189	--

Ablation studies reveal stepwise improvements: adding the boundary task (+0.56% region Dice over U²Net baseline), shape task (+0.45%), uncertainty loss (+0.89%), and finally MIGR+MRGR graph modules (+1.98%). Qualitatively, MTG-Net accurately delineates irregular lesion boundaries and captures fine vessel structures under challenging conditions (noise, projection artifacts). Competing architectures tend to under-segment (missing vessels) or over-segment (artifact leakage).

5. Application Domains and Clinical Utility

MTG-Net applied to CNVSeg enables several clinically relevant analyses:

• Pre- and postoperative CNV assessment: Automated quantification of lesion area, vessel density, and avascular zone metrics exhibit statistically significant changes ($$\mathcal{L}_{uce} = -\frac1Z \sum_{i=1}^Z \bigl[ (1 + \mathrm{Normalize}(V_i))\,[\,X_i\log\hat X_i + (1−X_i)\log(1−\hat X_i)\,]\bigr]$$0, paired t-test), supporting use in monitoring disease progression.
• Therapeutic outcome monitoring: Quantitative vessel density reduction ($$\mathcal{L}_{uce} = -\frac1Z \sum_{i=1}^Z \bigl[ (1 + \mathrm{Normalize}(V_i))\,[\,X_i\log\hat X_i + (1−X_i)\log(1−\hat X_i)\,]\bigr]$$140%) post anti-VEGF therapy, with early detection of fibrotic vessel features facilitating personalized intervention planning.
• Cross-disease generalization: Application to the CVC-300 colon polyp dataset yields a region Dice of 0.8979 and IoU of 0.8489, surpassing prior segmentation methods.

6. Limitations and Prospective Improvements

MTG-Net’s nested U²Net backbone and graph reasoning modules, combined with MC-Dropout, result in substantial computational overhead (∼84 s/epoch, RTX 3090). The authors suggest knowledge distillation as a strategy for deriving lighter-weight models. Dataset diversity is currently limited, primarily drawing from the Heidelberg scanner; expanding to other devices and centers may enhance model robustness. While the uncertainty-weighted loss improves artifact tolerance, segmentation in extremely low-contrast CNV cases remains challenging. Further advances may involve integration of physics-based artifact removal pipelines.

7. Data Accessibility and Benchmarking

CNVSeg provides the community's first open-access, pixel-level annotated benchmark for CNV segmentation in OCTA. Its comprehensive cross-device structure and rigorous annotation protocol establish a de facto standard for future benchmarking and method development in retinal lesion segmentation (Chen et al., 5 Aug 2025).