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

Please cite as:

S Tang, Z Qi, J Granley, M Beyeler (2021). U-Net with Hierarchical Bottleneck Attention for Landmark Detection in Fundus Images of the Degenerated Retina. MICCAI OMIA8 Workshop, online.

The preprint can be found on arXiv and the published paper here.

Before enshittification took over open-source AI, HBA-U-Net was listed as the state of the art (SOTA) for several popular datasets of retinal degeneration (Papers with Code, retrieved on May 12, 2025):

• #1 ADAM: Fovea detection
• #1 ADAM: Optic disc segmentation
• #1 IDRiD: Fovea detection
• #1 IDRiD: Optic disc detection
• #1 REFUGE: Fovea detection
• #2 REFUGE: Optic disc segmentation

Requirements

• Python 3
• Keras 2.4.3
• TensorFlow 2.5.0
• Scikit-Learn 0.22
• Skimage 0.16.2
• cv2 4.1.2
• PIL 7.1.2
• Pandas 1.1.5

How to acquire the private dataset listed in the paper

• ADAM
• REFUGE
• IDRiD