BFENet: A two-stream interaction CNN method for multi-label ophthalmic diseases classification with bilateral fundus...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of eye diseases is crucial for preventing blindness.
• Existing methods often analyze single-eye fundus images, neglecting valuable bilateral information.
• Current approaches typically focus on classifying only one ocular disease at a time.

Purpose of the Study:

• To develop a model that utilizes bilateral fundus images for patient-level diagnosis.
• To enable multi-label classification of ophthalmic diseases simultaneously.
• To introduce the Bilateral Feature Enhancement Network (BFENet) for improved accuracy.

Main Methods:

• A two-stream interactive Convolutional Neural Network (CNN) architecture is proposed.
• A feature enhancement module uses attention mechanisms to leverage inter-eye information, improving feature representation.
• A multiscale module with dilated convolutions enriches feature maps for capturing diverse disease characteristics.

Main Results:

• BFENet achieved superior performance compared to existing methods on both off-site and on-site datasets.
• Key performance metrics include Kappa (0.535 off-site, 0.513 on-site), F1-score (0.892 off-site, 0.886 on-site), and AUC (0.912 off-site, 0.903 on-site).
• The model demonstrated strong classification accuracy for multiple ophthalmic diseases.

Conclusions:

• BFENet effectively addresses the limitations of single-eye analysis and single-disease classification.
• The model's ability to integrate bilateral information enhances diagnostic capabilities.
• The proposed method is adaptable to other tasks requiring analysis of correlated image data.