MTG-Net: Graph-Enhanced OCTA CNV Segmentation

• The paper introduces MTG-Net, a multilateral graph-based network that segments CNV and vessel structures in OCTA images using joint multi-task reasoning.
• It employs a U²-Net-style nested architecture and dual graph modules (MIGR and MRGR) to fuse region, boundary, and shape information with uncertainty-weighted loss.
• Experimental results on the CNVSeg dataset show significant improvements in Dice, IoU, recall, and precision, outperforming previous segmentation methods.

MTG-Net refers to the Multilateral Interaction-Enhanced Graph Convolutional Network developed for the segmentation of choroidal neovascularization (CNV) and associated vascular structures in optical coherence tomography angiography (OCTA) images. CNV is a primary feature of wet age-related macular degeneration (AMD) and demands precise quantification due to its clinical significance in blindness prevention. MTG-Net integrates region and vessel morphology, formulating segmentation as a joint multi-task reasoning problem under graph constraints. It addresses difficulties imposed by imaging artifacts, irregular CNV morphologies, and data scarcity by introducing joint semantic-geometric graph reasoning and providing a newly established public dataset for benchmarking OCTA-based CNV segmentation (Chen et al., 5 Aug 2025).

1. Multi-Task Region Codec Architecture

MTG-Net incorporates a U²-Net-style nested U-shaped convolutional neural network as its backbone. The encoder part consists of five levels, each capturing features at different resolutions, which are fused to form streams targeting specific sub-tasks:

• Boundary stream: $$F_{\mathrm{bou}} = R(L_2 \parallel L_3 \parallel L_4)$$
• Shape stream: $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$
• Region stream: $F_{\mathrm{reg}} = R(L_5)$,

where $L_i$ are encoder features and $R(\cdot)$ denotes a $1\times1$ convolution followed by batch normalization. Supervision for the boundary map is derived from applying the Canny operator to the region mask, and the shape map is expressed through the signed distance function of the same mask.

For vessel segmentation, the predicted region mask is multiplied elementwise with the OCTA input, and the result enters a second, identical encoder-decoder, yielding the vessel segmentation as a dependent branch.

2. Multilateral Graph-based Cross-Task Modules

MTG-Net implements two core, complementary graph modules for inter-task reasoning:

A. Multilateral Interaction Graph Reasoning (MIGR):

• Graph Projection ($G_{\text{pro}}$): Each feature map $F_i \in \mathbb{R}^{H\times W\times C}$ is projected to $F_i' \in \mathbb{R}^{(HW)\times C}$ and clustered into $K$ nodes with soft-assignment weights $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$0 based on learnable centers $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$1 and scales $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$2.
• Graph Interaction ($$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$3): Region, boundary, and shape graphs are mapped into Key, Value (region) and Query (boundary/shape) embeddings. Cross-task affinity matrices, e.g. $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$4, mediate the injection of region semantics into boundary/shape graphs.
• Intra-graph reasoning: Each graph undergoes a GCN update: $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$5.
• Graph Reprojection: Node features are mapped back and fused into spatial feature maps.

B. Multilateral Reinforcement Graph Reasoning (MRGR):

• Fuses enhanced boundary/shape features with region streams and predicts boundary/shape maps as attention weights.
• Constructs secondary k-NN graphs between region and support nodes derived from these boundary/shape attention maps, aggregates them via max-pooling, and reintegrates the outputs to reinforce spatial segmentation in the decoder path.

The iterative application of MIGR (region $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$6 boundary/shape) and MRGR (boundary/shape $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$7 region) realizes semantic and geometric duality, allowing high-order synergy across region, edge, and shape constraints (Chen et al., 5 Aug 2025).

3. Uncertainty-Weighted Loss and Training Strategy

MTG-Net employs uncertainty-weighted multi-task supervision to attenuate the influence of OCTA-specific noise and artifacts:

• Uncertainty estimation: $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$8 stochastic forward passes (MC-dropout, $$F_{\mathrm{shp}} = R(L_3 \parallel L_4 \parallel L_5)$$9) yield pixelwise mean $F_{\mathrm{reg}} = R(L_5)$0 and variance $F_{\mathrm{reg}} = R(L_5)$1.
• Uncertainty-weighted surrogate BCE loss:

$F_{\mathrm{reg}} = R(L_5)$2

• Task-specific losses: Region and vessel tasks use Dice+$F_{\mathrm{reg}} = R(L_5)$3; boundary, Dice+$F_{\mathrm{reg}} = R(L_5)$4; shape, MSE+$F_{\mathrm{reg}} = R(L_5)$5.
• Adaptive weighting: Region, boundary, shape losses are combined with weights proportional to their uncertainty; vessel loss is summed directly:

$F_{\mathrm{reg}} = R(L_5)$6

Training uses the Adam optimizer ($F_{\mathrm{reg}} = R(L_5)$7, $F_{\mathrm{reg}} = R(L_5)$8), mini-batches of 4, and extensive augmentation (Chen et al., 5 Aug 2025).

4. CNVSeg Dataset: Composition and Annotation Protocol

MTG-Net was developed alongside CNVSeg, the first public dataset tailored for CNV segmentation in OCTA, containing 224 en-face images from 184 wet-AMD cases spanning four imaging devices (Heidelberg, Optovue, Zeiss, SVision). Scan areas include $F_{\mathrm{reg}} = R(L_5)$9, $L_i$0, and $L_i$1 mm². Expert pixel-level annotation was performed on 3D scans using a multi-operator protocol: initial labeling by two ophthalmologists, arbitration by a senior, and final validation by a second senior expert. Mask generation targeted both lesion regions and vessels, with the dataset made available via GitHub, enabling reproducibility and comparative benchmarking (Chen et al., 5 Aug 2025).

5. Quantitative Outcomes and Experimental Protocol

MTG-Net was assessed on mixed-device and leave-one-device-out splits with input size $L_i$2, adopting Dice, Intersection-over-Union (IoU), recall, and precision as primary metrics. On the mixed-device test:

• Region Dice: 0.8721
• Region IoU: 0.7740
• Region Recall: 0.9068
• Region Precision: 0.8660
• Vessel Dice: 0.8812
• Vessel IoU: 0.7974
• Vessel Recall: 0.9138
• Vessel Precision: 0.8731

Single-device performance varies per scanner: for Heidelberg, region Dice is 0.8143, vessel Dice 0.8109; for Optovue, region Dice 0.9038, vessel Dice 0.9189. These results significantly surpass prior segmentation approaches, particularly in handling challenging boundaries and vessel detail (Chen et al., 5 Aug 2025).

6. Semantic and Geometric Duality in Graph Reasoning

MTG-Net systematically exploits two forms of cross-task duality:

• Semantic duality (MIGR): Region features supply supervision to boundary and shape graphs through explicit cross-graph attention, enforcing mutual semantic consistency.
• Geometric duality (MRGR): Shape and boundary inferences provide geometric guidance to region segmentation by propagating information via k-NN support graphs and max-pooled aggregations.

This iterative synergy enables the model to refine higher-order shape and edge relationships, which is essential in mitigating projection artifacts, label ambiguity, and the intrinsic complexity of CNV morphologies. The joint graph-based mechanism ensures robust delineation of both lesion core and capillary detail, advancing the state of OCTA image analysis for AMD (Chen et al., 5 Aug 2025).