Confidence-aware multi-modality learning for eye disease screening.

Area of Science:

• Ophthalmology
• Computer Vision
• Machine Learning

Background:

• Multi-modal ophthalmic image classification is crucial for diagnosing eye diseases.
• Current methods prioritize accuracy but often overlook prediction confidence and robustness.

Purpose of the Study:

• To propose a novel multi-modality evidential fusion pipeline for eye disease screening.
• To enhance the confidence, robustness, and reliability of diagnostic predictions.

Main Methods:

• Utilized normal inverse gamma prior distributions for uni-modality uncertainty quantification.
• Employed a mixture of Student's t distributions for confidence-aware multi-modality fusion.
• Introduced a confidence-aware multi-modality ranking regularization term.

Main Results:

• The proposed model demonstrated superior robustness in scenarios with Gaussian noise and missing modalities.
• Achieved strong generalization capabilities on out-of-distribution data.
• Outperformed existing methods in reliability and accuracy.

Conclusions:

• The novel evidential fusion pipeline offers a promising solution for multimodal eye disease screening.
• The approach effectively integrates multi-modal data, enhancing diagnostic performance and reliability.