Abstract
In ophthalmology, early fundus screening is an economic and effective way to prevent blindness caused by ophthalmic diseases. Clinically, due to the lack of medical resources, manual diagnosis is time-consuming and may delay the condition. With the development of deep learning, some researches on ophthalmic diseases have achieved good results, however, most of them are just based on one disease. During fundus screening, ophthalmologists usually give diagnoses of multi-disease on binocular fundus image, so we release a dataset with 8 diseases to meet the real medical scene, which contains 10,000 fundus images from both eyes of 5,000 patients. We did some benchmark experiments on it through some state-of-the-art deep neural networks. We found simply increasing the scale of network cannot bring good results for multi-disease classification, and a well-structured feature fusion method combines characteristics of multi-disease is needed. Through this work, we hope to advance the research of related fields.
Introduction
Most existing AI models for eye health are "specialists" trained to identify specific conditions, such as diabetic retinopathy or glaucoma, often using images from only a single eye. However, clinical reality is more complex: patients frequently present with multiple conditions simultaneously, and ophthalmologists typically examine both eyes to reach a diagnosis. The gap between academic "clean" datasets and clinical "noisy" reality creates a performance bottleneck. There is a distinct lack of public, large-scale datasets providing multi-label annotations for binocular fundus images. This paper introduces the OIA-ODIR benchmark to force models to analyze the "whole patient" rather than isolated eyes or diseases.
The OIA-ODIR Dataset
The primary contribution of this work is the Ocular Disease Intelligent Recognition (OIA-ODIR) dataset. It consists of 10,000 fundus images from 5,000 patients, collected from 487 hospitals across 26 provinces in China. As shown in the comparison below, OIA-ODIR is uniquely positioned as a multi-disease, multi-label, and publicly available dataset.
The dataset includes eight diagnostic categories: Normal (N), Diabetes (D), Glaucoma (G), Cataract (C), AMD (A), Hypertension (H), Myopia (M), and Others (O). These categories represent the most common conditions encountered in clinical screening.
Methodology
To establish a baseline, nine popular convolutional neural networks (CNNs)—including VGG-16, ResNet variants, and Inception-v4—were modified to handle binocular input. A shared-weight backbone extracts features from both eyes, which are then integrated via feature fusion. Three fusion methods were evaluated:
- SUM: Element-wise addition of features.
- PROD: Element-wise multiplication of features.
- Concat: Concatenation of feature vectors.
The fused features are passed through eight separate classifiers to enable multi-label output for each disease category.
Key Insights and Results
The benchmark experiments revealed several critical findings:
- Feature Fusion is Essential: The "SUM" method generally provided the most stable performance, while "PROD" combined with Inception-v4 achieved the highest overall score (0.7516).
- Network Architecture Matters: Surprisingly, VGG-16 outperformed several deeper ResNet models, suggesting that for multi-disease classification, raw depth is less important than effective feature extraction and fusion. Models with greater "width," like ResNeXt and Inception-v4, performed better by processing features at different scales.
- Clinical Imbalance: The dataset reflects clinical reality, where some diseases (e.g., Hypertension) are rarer than others. This class imbalance remains a significant challenge for model training.
Conclusion
The OIA-ODIR dataset provides a robust benchmark for moving ophthalmic AI toward real-world clinical utility. Future work will focus on increasing data volume, addressing class imbalance, and expanding geographic and racial diversity to ensure global applicability.
Summary updated