Research map / Graph learning / Knowledge and temporal graphs

SIGraph: Saliency Image-Graph Network for Retinal Disease Classification in Fundus Image

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Graph reasoning for visual recognition · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SIGraph, which combines saliency-based lesion perception with a graph of lesion pixels (Manhattan four-quadrant graph construction and mixed graph pooling) to classify retinal diseases from fundus images.

Problem

Retinal disease classification from fundus images often ignores the spatial distribution patterns of scattered subtle lesions, which clinicians use in diagnosis.

Approach

Identify latent lesion pixels via saliency, build a sparse fully connected lesion graph using minimum Manhattan distance in four quadrants (guaranteeing global connectivity), and summarize it with a mixed edge-cut plus node-cluster graph pooling to capture global lesion distribution.

Key results

• SIGraph reaches Kappa 93.2 on EyePACS and F1 98.0 on RFMiD, with at least 1.5% Kappa gain over the strongest Transformer baselines.
• Adding the image-graph block to SatFormer-B raises EyePACS Kappa from 90.8 to 92.7 and RFMiD F1 from 95.8 to 98.0.
• Manhattan-4 graph construction outperforms 4-NN (Kappa 92.7 vs 90.6), showing that guaranteeing connectivity matters.

Measured effects reported in the paper

Surprising findings

• Adding the image-graph block to Swin gave a smaller gain than for other transformers, which the authors attribute to window-based attention limiting global perception.

Limitations

• Private dataset is not released, limiting reproducibility.
• Compute cost (104 GFLOPs) is higher than several baselines such as CrossFormer-L (16.1 GFLOPs) and Swin-B (47 GFLOPs).

Experimental setup

Datasets: EyePACS RFMiD private multi-disease fundus dataset

Benchmarks: EyePACS RFMiD private fundus dataset

Metrics: accuracy AUC F1-score Kappa AUROC AUC-PR

Baselines: Zoom-in-Net GREEN-ResNet50 GRADING ViT-B/16 Swin-B PVTv2-B5 CrossFormer-L MIL-VT SatFormer-S SatFormer-B RETFound

Builds on: SatFormer Zoom-in-Net GRADING DiffPool CutPool Graph Attention Networks

Closest papers by meaning

• HACDR-Net: Heterogeneous-Aware Convolutional Network for Diabetic Retinopathy Multi-Lesion Segmentation (AAAI 2024 · similarity 0.86)
• Pathological Evidence Exploration in Deep Retinal Image Diagnosis (AAAI 2019 · similarity 0.85)
• WDT-MD: Wavelet Diffusion Transformers for Microaneurysm Detection in Fundus Images (AAAI 2026 · similarity 0.85)
• FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.84)
• Like an Ophthalmologist: Dynamic Selection Driven Multi-View Learning for Diabetic Retinopathy Grading (AAAI 2025 · similarity 0.83)
• Deep Hierarchies and Invariant Disease-Indicative Feature Learning for Computer Aided Diagnosis of Multiple Fundus Diseases (AAAI 2025 · similarity 0.83)