Guided Context Gating: Learning to leverage salient lesions in retinal fundus images