See the small lesions: Frequency-guided spatial debiasing GAN for multimodal medical image fusion

Multimodal medical image fusion uses these features from different modalities to generate fused images with richer information for diagnosis. However, in medical images, the lesion area only constitutes a small portion, and feature extraction is dominated by background regions, which weakens lesion representation and leads to pixel-level unfairness. To address this issue, this paper proposes frequency-guided spatial debiasing GAN (FGSD-GAN). The proposed method introduces a frequency-guided debiasing framework with a spatial debiasing branch, a frequency-domain debiasing branch, and a debiasing loss function. The frequency-domain branch enhances structural and contour information through subband decomposition and adaptive fusion, while the spatial branch strengthens lesion-related features and suppresses redundant background information. In addition, the debiasing loss further improves the balance between lesion enhancement and structure preservation. Experiments on the CT lung window, CT mediastinal window, and Whole Brain Atlas datasets demonstrate that FGSD-GAN achieves effective and robust fusion results across different medical imaging scenarios.