无监督掩码循环对抗网络实现细胞虚拟染色

Key Points

• The aim is to develop a framework for virtual staining that minimizes reliance on paired data while improving image quality.
• Proposed MVS-CycleGAN framework utilizing masked self-supervision for image generation without paired data.
• Implemented a random masking reconstruction task to capture both global shape and local texture.
• Conducted experiments on three types of cell datasets to evaluate performance.
• MVS-CycleGAN outperforms traditional methods, achieving FSIM scores of 0.784 for BJ-5ta cell membrane and 0.565 for cell nucleus, with overall improvements observed.
• Scores for HEK293T are 0.854 for membrane and 0.830 for nucleus, while Neuromast achieves 0.657 and 0.740, respectively.
• Downstream segmentation experiments confirm effectiveness in quantitative analysis.