Frequency-Domain Assisted Fundus Image Enhancement Network Based on the Mamba Model

Enhancing fundus images is crucial for early diagnosis and monitoring of retinal diseases. Although CNN and Transformer-based methods have made great progress, CNNs struggle with long-range dependencies, and Transformers face challenges with large model sizes and low computational efficiency. To address these issues, Selective Structured State-Space Models (SSMs) offer an efficient solution with linear complexity. However, standard Mamba models still suffer from local pixel information loss and channel redundancy, limiting their performance in fundus image enhancement. We propose MambaFE, which incorporates localized refinement and an advanced channel attention mechanism to resolve challenges such as pixel detail loss and excessive channel information. Considering that retinal structures play a vital role in fundus analysis, retaining these structures during the enhancement process is crucial. As retinal structures are predominantly located within the high-frequency components of fundus images, we introduce a frequency-guided approach to capture reliable representations of these features, ensuring improved retention of essential information during enhancement. Comprehensive experimental findings indicate that the proposed MambaFE outperforms contemporary fundus image enhancement techniques, significantly enhancing the quality of clinically low-quality and synthetically degraded fundus images, while offering robust support for real-world medical applications.