U-Net with Hierarchical Bottleneck Attention for Landmark Detection in Fundus Images of the Degenerated Retina

Fundus photography has routinely been used to document the presence and severity of retinal degenerative diseases such as age-related macular degeneration (AMD), glaucoma, diabetic retinopathy (DR) in clinical practice, for which fovea optic disc (OD) are important landmarks. However, occurrence lesions, drusen, other abnormalities during severely complicates automatic landmark detection segmentation. Here we propose HBA-U-Net: a U-Net backbone enriched with hierarchical bottleneck attention. The network consists novel attention block that combines refines self-attention, channel attention, relative-position highlight may be OD segmentation degenerated retina. HBA-U-Net achieved state-of-the-art results on across datasets eye conditions (ADAM: Euclidean Distance (ED) 25.4 pixels, REFUGE: 32.5 IDRiD: 32.1 pixels), AMD Dice Coefficient (DC) 0.947), DR (IDRiD: ED 20.5 pixels). Our suggest well suited variety diseases.