Joint DR Lesion Analysis and Grading: External Validation of Multi-Task Deep Learning

A multi-task AI model for diabetic retinopathy grading achieved kappa 0.858 on internal test, 0.821 on external validation, with referable-DR sensitivity 0.928. Cross-task attention was key.

• The model jointly segments and detects lesions while grading DR, providing lesion-level evidence for grading.
• Removing cross-task attention led to the largest drop in grading performance; single-task training performed worse.
• External segmentation performance (Dice 0.489) was based on only 81 images, limiting generalizability claims.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.