Dual Attention-Based Multi-Instance Learning For Referable Diabetic Retinopathy

Classifying referable versus non-referable Diabetic Retinopathy (DR) is of key importance, in that the delayed diagnosis of DR may end up with severe vision loss. The non-referable DR is asymptomatic or mildly asymptomatic, whereas lesions (e.g., Hemorrhage, Soft exudate) appear in the referable DR (rDR) which needs immediate medical treatment. Manually annotating lesions is, however, laborious and subjective in clinical practice. In this paper, we leverage attention-based multi-instance learning (MIL) and channel-attention mechanism to propose a dual attention-based MIL to differentiate non-referable DR and referable DR. Meanwhile, the MIL technique empowers us to localize lesion areas with image-level labels. The proposed method is validated on the Eyepacs dataset achieving encouraging performance with an AUC-ROC of 0.95 and demonstrating that the contributions of lesion instances are relatively higher than those of normal regions to the final predictions.