Cross-modality transfer learning with knowledge infusion for diabetic retinopathy grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Ultra-wide-field (UWF) fundus photography offers a wider view for diagnosing eye diseases like diabetic retinopathy (DR).
• Training AI models for DR detection in UWF images is challenging due to limited labeled data and the need for enhanced model performance.
• Existing models lack prior knowledge crucial for effective learning in this domain.

Purpose of the Study:

• To develop a robust model for diabetic retinopathy (DR) severity assessment using unsupervised lesion-aware domain adaptation in ultra-wide-field (UWF) images.
• To transfer knowledge from large-scale, annotated color fundus image datasets to the UWF domain via unsupervised domain adaptation.
• To improve the utility of computer-aided diagnosis for DR in UWF imaging.

Main Methods:

• Implemented unsupervised lesion-aware domain adaptation for DR grading in UWF images.
• Integrated an adversarial lesion map generator to leverage detailed annotations from public datasets.
• Incorporated auxiliary lesion information, mimicking clinical DR evaluation methods.

Main Results:

• Achieved an accuracy (ACC) of 68.18% and a precision (pre) of 67.43% among six representative DR grading methods.
• Conducted quantitative and qualitative evaluations to assess the proposed method's performance.
• Performed ablation studies to validate the contribution of each component within the proposed method.

Conclusions:

• The developed method enhances the accuracy of diabetic retinopathy (DR) grading using UWF images.
• The approach improves the interpretability of diagnostic results, offering a reliable grading scheme for clinicians.
• Successfully addresses challenges of limited labeled UWF data through knowledge transfer and lesion-aware adaptation.