Abstract
Ophthalmic diagnosis is increasingly reliant on deep learning, yet most existing models treat eyes as independent entities, ignoring the critical inter-ocular correlations used by clinicians. We present DMS-Net, a novel Siamese network architecture designed for binocular fundus image classification. By explicitly modeling both symmetrical and asymmetrical pathological correlations between a patient's left and right eyes, DMS-Net achieves state-of-the-art diagnostic accuracy. Evaluated on the ODIR-5K dataset, the framework demonstrates superior performance in identifying multiple ocular diseases, highlighting the clinical value of binocular analysis.
Introduction
In clinical practice, ophthalmologists rarely diagnose an eye in isolation. Many ocular conditions, such as glaucoma, typically progress symmetrically, while others, like diabetic retinopathy, may present with disparate lesion severities between eyes. Despite this, most deep learning models are "monocular," processing each eye as an independent data point. This approach often fails to resolve "boundary ambiguity"—where subtle lesions blend into healthy tissue—and misses the diagnostic "prior" provided by the contralateral eye.
DMS-Net addresses these limitations by treating binocular fundus images as a paired, dual-modal input. By leveraging the inherent symmetry and asymmetry of ocular pathologies, the network mimics human diagnostic reasoning to improve classification robustness and accuracy.
Methodology: The DMS-Net Architecture
The core of DMS-Net is a Dual-Modal Multi-Scale Siamese Network designed to extract, align, and fuse features from both eyes simultaneously.
Siamese Backbone
The model utilizes a weight-sharing ResNet-152 backbone to process left and right eye images in parallel. This ensures that the network learns a consistent semantic representation of ocular structures across both inputs.
OmniPool Spatial Integrator Module (OSIM)
To account for the varying scales of retinal lesions—from tiny microaneurysms to large-scale hemorrhages—the OSIM employs multi-scale adaptive pooling (2x2 and 4x4) alongside global pooling. This allows the model to integrate fine-grained local details with coarse-grained global context.
Calibrated Analogous Semantic Fusion Module (CASFM)
The CASFM utilizes bidirectional attention mechanisms to bridge the semantic gap between the two eyes. It helps the model align features that are common to both eyes, facilitating a more holistic representation of the patient's ocular health.
Dual-Synergy Fusion: CCAM & CIAM
The most innovative aspect of DMS-Net is its dual-pathway fusion strategy:
- Cross-Modal Contrastive Alignment Module (CCAM): Focuses on identifying and leveraging asymmetrical differences between the eyes.
- Cross-Modal Integrative Alignment Module (CIAM): Focuses on capturing symmetrical similarities, such as synchronous changes in the optic cup or disc.
Experiments and Results
Dataset and Preprocessing
The model was rigorously evaluated on the ODIR-5K (Ocular Disease Intelligent Recognition) dataset, comprising 5,000 patients. The dataset includes eight categories: Normal (NL), Diabetic Retinopathy (DR), Glaucoma (GL), Cataract (CT), AMD, Hypertensive Retinopathy (HR), Myopia (MY), and Others (OTH). To standardize image quality, the authors applied non-uniform illumination correction and utilized the CutMix algorithm for data augmentation.
Performance Metrics
DMS-Net achieved state-of-the-art (SOTA) results on the ODIR-5K benchmark:
- Accuracy: 82.9%
- Recall: 84.5%
- Cohen’s Kappa: 83.2%
Ablation studies confirmed the necessity of each module. Notably, removing the CASFM module resulted in the largest performance drop (a 2.5% decrease in Kappa), proving that effective cross-modal fusion is the linchpin of binocular diagnostic success.
Discussion and Conclusion
The success of DMS-Net underscores the "two-eye advantage" in ophthalmic AI. By explicitly modeling inter-ocular correlations, the network provides a more reliable diagnostic tool than traditional monocular systems. Interestingly, the study found that ResNet-152 outperformed Vision Transformers (ViT) in this context, suggesting that the inductive biases of CNNs remain highly effective for the structured spatial data of retinal imaging.
Future research will focus on improving model robustness against low-quality images from mobile diagnostic devices and extending the Siamese architecture to longitudinal analysis, comparing a patient's ocular state over time.
Summary updated