Fig. 6: Comparisons of PhaseFIT against GAN methods for virtual UEA-I painting.

UEA-I stains the secretory cells, such as Paneth cells and goblet cells. Compared to the ground truth from real UEA-I staining (a), the performance of PhaseFIT (b) is better than the GAN methods (c and d). For example, the GAN methods, especially Conditional-GAN, frequently missed the UEA-I⁺ dot-like signals in the dashed boxes (e) PhaseFIT model exhibited a superior Dice score of 0.36 ± 0.028, a recall score of 0.31 ± 0.077, a SSIM of 0.41 ± 0.061, and an MSE of 0.41 ± 0.136. This was marginally higher compared to Cycle-GAN’s 0.21 ± 0.036, 0.28 ± 0.110, 0.20 ± 0.055, and 0.73 ± 0.107; Conditional-GAN’s 0.16 ± 0.057, 0.18 ± 0.116, 0.11 ± 0.042, and 0.79 ± 0.165. p < 0.001 (n = 889, paired Wilcoxon rank-sum test). 4.12e⁻¹⁴⁷ is the smallest value