Multi-label classification of fundus images with graph convolutional network and LightGBM

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of retinal disorders is vital to prevent irreversible vision impairment.
• Clinical settings require models that can identify multiple retinal diseases simultaneously.
• Existing models may not fully address the complexity of multi-label fundus disease detection.

Purpose of the Study:

• To develop a patient-level, multi-label fundus disease identification model.
• To enhance the accuracy of screening for various retinal illnesses.
• To address clinical needs for comprehensive retinal disease diagnosis.

Main Methods:

• Proposed a composite model integrating a backbone, hybrid graph convolution, and classifier modules.
• Utilized EfficientNet-B4 for feature extraction and LightGBM for multi-label classification.
• Employed graph convolution and self-attention to model label relationships.

Main Results:

• The MCGL-Net model achieved state-of-the-art performance on the ODIR dataset, reaching a 91.60% F1 score.
• Hybrid graph convolution improved F1 score by 2.39% with EfficientNet-B4.
• The composite model outperformed a single EfficientNet-B4 model by 5.42% in F1 score.

Conclusions:

• The proposed hybrid graph convolutional structure and composite model enhance fundus disease identification performance.
• The model effectively captures label correlations, improving multi-label classification accuracy.
• This approach offers a promising solution for clinical multi-label retinal disease screening.