Digital Staining Algorithm for Multi-Domain Transformation of Unstained Images

Digital staining is an artificial intelligence-based technology that replaces the conventional staining process. This technology addresses the challenges associated with the traditional staining procedure, which is complex, time-consuming, and costly, as it relies heavily on manual labor. However, inconsistencies in staining color and quality, arising from variations in staining protocols caused by factors such as solution concentration, staining duration, patient-to-patient variability, and the type of scanner used, remain unresolved. To address this issue, a multi-domain digital staining method was employed. This study proposes a multi-hematoxylin and eosin (H&E) digital staining transformation method based on the phase image of fourier ptychographic microscopy (FPM) using a single model in bright-field microscopy (BM). The proposed model was trained by incorporating a mask and latent loss within the framework of the StarGAN algorithm. The model’s performance was quantitatively evaluated using a cell detection model trained on the MoNuSeg dataset. The results demonstrated improvements of 7.96% and 10.81% in Dice and Jaccard indices, respectively, compared to the CycleGAN model. These findings indicate that the proposed model achieves digital staining that closely replicates the conventional H&E staining method. Additionally, the visual assessment results revealed that the proposed method provides clearer differentiation between background and cytoplasmic boundaries. These results suggest that the proposed approach may effectively address quality degradation issues caused by the limitations of FPM data.