.png){kind=icon}
.png){kind=icon}
Return
Adaptive MDA-Net: Degradation-Aware Retinal Fundus Image Enhancement with Frequency-Domain and Illumination Guidance for Clinical Screening
.png){kind=icon}
OA
AI
M
M
M
R
H
DOI：10.3390/diagnostics16162555.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Background/Objective: Retinal fundus image quality directly affects ophthalmic interpretation and automated analysis. Existing multi-degradation methods provide adaptive restoration but offer limited spatial degradation encoding and frequency-level supervision for severely degraded images. This study aims to improve restoration fidelity while preserving retinal structures required by downstream clinical-analysis models. Methods: Adaptive MDA-Net combines a CNN degradation encoder, coupled synthetic degradation, latent-conditioned attention, a supervised illumination-guidance branch, and a seven-term spatial- and frequency-domain objective. The model was evaluated on held-out EyeQ data and with fixed downstream models on DRIVE and REFUGE. Results: Across five runs, the model achieved 35.86 ± 0.12 dB PSNR and 0.9732 ± 0.0010 SSIM, improvements of 0.62 dB and 0.0064 over RGB-GAP. Downstream evaluation achieved a Dice score of 0.7601 on DRIVE and an mDice score of 0.7917 on REFUGE. Conclusions: Degradation-aware conditioning improves restoration fidelity and recovers part of the downstream performance lost to image degradation. Validation remains limited to image-level EyeQ splits and synthetic supervision.
Keywords:
retinal fundus image enhancement
degradation-aware deep learning
image restoration
diabetic retinopathy screening
frequency-domain loss
Journal
IF:
3.3
Papers:
1.9W
Citations:
3.6W
Organization
Cited Papers