WAFMGAN: A frequency-aware CycleGAN for structure-consistent virtual staining of transmission electron microscopy...

Conventional transmission electron microscopy (TEM) staining improves image contrast but requires additional sample preparation, consumes reagents and tissue, and may introduce deformation, local mismatch, or tissue loss. When strictly paired stained and unstained sections are difficult to obtain, unpaired virtual staining provides a practical route for improving the readability of unstained TEM images while retaining correspondence with source-image structural cues. This paper presents WAFMGAN, a frequency-aware CycleGAN framework that combines Wavelet Asymmetric Frequency Mamba (WAFM) blocks for low-frequency-dominant global staining transfer and long-range structural coordination with a Frequency-Adaptive Gated Refinement (FAGR) block for later-stage boundary and texture refinement. On the mouse renal TEM test set, WAFMGAN achieved an FID of 14.8064 and a KID( $\times 100$ ) of 0.5074 $\pm$ 0.0417, representing reductions of 23.0% and 47.3% relative to CycleGAN. At the original-micrograph-pair level, WAFMGAN achieved SSIM, MS-SSIM, CSS, and Gradient-SSIM values of 0.9796, 0.9874, 0.9825, and 0.9889, respectively, and all four structure-related metrics were significantly higher than those of every evaluated baseline after Holm-Bonferroni correction. Output-level frequency analysis further showed lower wavelet-domain discrepancy than CycleGAN on all four tissue sets and lower high-frequency Fourier discrepancy on the renal, heart, and liver sets. These results demonstrate that WAFMGAN improves stained-domain distribution and frequency matching while maintaining strong source-image structural consistency. WAFMGAN therefore provides a useful frequency-aware approach for improving the readability of unstained TEM images while maintaining source-image structural cues.