A two-stage multi-scale attention-based network for weakly supervised cataract fundus image enhancement.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Cataract is a leading cause of vision loss, complicating retinal image analysis.
• Enhancing cataract fundus images is difficult due to limited paired data and detail loss.

Purpose of the Study:

• To propose a novel weakly supervised network (TSMSA-Net) for cataract fundus image enhancement.
• To address the scarcity of paired training data and improve the recovery of fine details.

Main Methods:

• A two-stage approach: Stage 1 uses CycleGAN for realistic paired image synthesis from unpaired data.
• Stage 2 employs a multi-scale attention network for feature extraction and detail restoration under weak supervision.

Main Results:

• TSMSA-Net effectively enhances cataract fundus images, outperforming state-of-the-art methods.
• The network demonstrates strong generalization ability on multiple datasets, even without paired images.
• Enhanced images improve downstream tasks like vessel segmentation and disease classification.

Conclusions:

• TSMSA-Net offers a robust solution for cataract fundus image enhancement.
• The method alleviates data scarcity and improves diagnostic accuracy for vision-threatening conditions.