FDA-UNet: Frequency-Domain Dual Attention Based on UNet for Stain Normalization

Histopathology is the diagnosis and study of tissue diseases, and staining is a crucial part of its analysis. However, existing staining methods often encounter inaccuracies, blurred boundaries, and issues such as under- or overstaining in medical image processing, impacting visualization and clinical reliability. To address these issues, we propose an improved network architecture called FDA-UNet. The core of FDA-UNet is a frequency-domain attention (FDA) module, which applies the Fourier transformation to capture frequency-domain information and then leverages both channelwise and spatial attention mechanisms operating on this frequency-domain representation. The FDA-UNet module is able to better improve foreground-background distinction, noise reduction, and feature learning, effectively mitigating staining problems. Furthermore, we introduce a novel loss function based on the LAB color space, which allows for more accurate assessment and guidance of color accuracy in stain normalization. This results in colors that are consistent with human perception and enhanced reliability. Compared with current mainstream methods, our experimental results show that FDA-UNet achieves more competitive performance on the MITOS-ATYPIS-14 contest dataset.