A Frequency-Aware Self-supervised Learning for Ultra-Wide-Field Image Enhancement

Weicheng Liao, Zan Chen, Jianyang Xie, Yalin Zheng, Yuhui Ma, Yitian Zhao

Introduction

Ultra-Wide-Field (UWF) imaging has revolutionized ophthalmology by providing a 200° view of the retina, significantly broader than the 45° offered by traditional fundus cameras. However, this expansive field of view introduces technical challenges: peripheral blurring and uneven illumination caused by the laser-scanning process. These artifacts can obscure critical pathological features like microaneurysms or fine vessels, hindering both manual clinical assessment and automated diagnostic tools.

Existing enhancement methods often struggle with UWF data due to two primary factors: the lack of high-quality ground truth pairs for supervised learning and the unique color properties of laser-scanned images, which do not adhere to standard "gray world" assumptions. This paper proposes a self-supervised, frequency-aware framework to address these specific limitations.

Proposed Methodology

The authors introduce a two-stage pipeline designed to decouple and address blurring and illumination separately.

1. FREquency-Decoupled Deblurring (FRED)

The first stage utilizes Average Pooling Separation (APS) to decompose the image into high-frequency (fine details) and low-frequency (structural) components. A key innovation is the Asymmetric Channel Integration (ACI) unit, which processes these streams in parallel. By combining global context via channel-wise attention with local detail preservation, FRED restores clarity without losing pathological markers.

2. Retinex-guided Illumination Compensation (RICE)

The second stage corrects lighting based on Retinex theory. To prevent the color distortion common in medical image enhancement, the Color Preservation Unit (CPU) employs a 2D Discrete Wavelet Transform (DWT). This allows the model to estimate and adjust illumination primarily within low-frequency sub-bands, shielding high-frequency color information from unwanted artifacts.

The proposed architecture featuring FRED and RICE stages, including detailed views of the ACI and CPU modules.

The entire framework is trained in a self-supervised manner. For deblurring, the model learns from synthetic training pairs created by applying random blur to high-quality images, bypassing the need for manual labeling.

Experimental Results

The framework was evaluated on a dedicated dataset of 834 UWF images and a disease grading dataset of 902 images categorized by Diabetic Retinopathy (DR) severity.

Quantitative and Clinical Performance

The proposed method outperformed six state-of-the-art techniques across multiple no-reference quality metrics, including NIQE, BRISQUE, and PIQE. Most importantly, the enhancement significantly improved downstream diagnostic performance. When using a ResNet34 model for DR grading, accuracy increased from 57.13% on raw images to 69.49%—a jump of over 12%. The F1-score similarly saw a massive 15.8% improvement.

Quantitative results and ablation study showing the proposed method's superiority in image quality metrics and diagnostic accuracy.

Discussion and Conclusion

The study demonstrates that frequency decoupling is essential for UWF image enhancement, as it prevents interference between structural lighting adjustments and fine-detail restoration. While the results are promising, future work is needed to address the gap between synthetic and real-world blur and to optimize the computational efficiency of the wavelet-based pipeline for real-time clinical use. Nonetheless, this frequency-aware approach sets a new benchmark for preparing UWF images for both human and AI-driven diagnostics.

Summary updated

BibTeX
@inproceedings{liao2025frequency,
  author = {Weicheng Liao and Zan Chen and Jianyang Xie and Yalin Zheng and Yuhui Ma and Yitian Zhao},
  title = {{A Frequency-Aware Self-supervised Learning for Ultra-Wide-Field Image Enhancement}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2025},
  doi = {10.1007/978-3-032-05169-1\_1},
}
Figure