Fourier Ptychographic Microscopy 기반 도메인 전환을 활용한 디지털 염색 알고리즘

Abstract

Digital staining (DS) has solved the problems of traditional histological staining (HS), which is complicated, time-consuming, and costly. However, differences in staining protocols (e.g., solution concentration, staining time, scanner used, patient variability) have not completely overcome the differences in staining color and quality. To solve this problem, we used a DS method utilizing domain transformation in unstained images. This study proposes a sin- gle-domain transformation staining method combining unsupervised and supervised learning, as well as an unsu- pervised learning-based multi-domain transformation staining method. The data were constructed using fourier ptychographic microscopy and bright-field microscopy images and optimized for training through five preprocessing steps. Afterward, we compared the proposed methods with the existing unsupervised staining method, CycleGAN. Prior to evaluation, the data were further processed through postprocessing to facilitate analysis. The visual eval- uation showed that the proposed method effectively distinguished the background compared to the control group and prevented the over-staining issue that occurs during the HS process. Quantitative evaluation using a cell nucleus detection model demonstrated that the proposed model showed improvements of more than 7.96% and 10.81% in the Dice and Jaccard indices, respectively, compared to the CycleGAN model. This indicates that the proposed model achieves digital staining that closely resembles the traditional Hematoxylin and Eosin staining method, and it sug- gests that supervised learning can be utilized even in environments without paired data, and unstained images can be used in multi-staining research.