290 Background: Haematoxylin and eosin (H&E) staining is central to routine histopathological diagnosis, but limited automation and rising case volumes can delay slide availability. Virtual staining, where deep learning models generate H&E-like images directly from greyscale inputs, has emerged as a potential way to shorten turnaround time. In this study, we apply generative models for large-scale virtual H&E synthesis and evaluate which training checkpoint yields the most robust staining performance for clinical integration. Methods: A lung cancer tissue microarray (TMA) comprising 25 cores was imaged with differential interference contrast (DIC)-like microscopy both before and after H&E staining on the Axioscan 7 (Zeiss, Germany). 20 cores were used for model training, 1 for validation, and 4 were held out for testing. Whole-slide images were tiled into 256×256-pixel patches and cross-modal registration was performed to ensure accurate pixel-level alignment. Two Pix2Pix generative adversarial network (GAN) models were trained on DIC-like images at different staining checkpoints: Model A used post-H&E stained DIC-like images, while Model B used pre-H&E stained DIC-like images. Both models employed identical architectures (U-Net generator, 70×70 PatchGAN discriminator), training strategies and image preprocessing workflows. Quantitative evaluation included tile-wise nuclei count correlation and structural similarity index (SSIM) compared with ground-truth H&E to assess morphological fidelity. The best-performing model based on validation metrics was selected for testing. Deep features from generator layers were visualized with UMAP and clustered using K-means to assess feature divergence. Results: Model B demonstrated superior structural fidelity (SSIM: Model A = 0.8631, Model B = 0.9077) and was found to preserve sharper nuclear boundaries and more realistic staining contrast upon visual inspection. UMAP evaluation revealed distinct separation between models, indicating capture of different structural representations despite using the same tissue sections. Conclusions: Virtual staining of pre-H&E DIC-like images generates realistic, high-fidelity images. This staining checkpoint aligns with existing clinical workflows and may facilitate future clinical translation. To assess clinical readiness, we have launched a global validation study involving more than 50 pathologists across the world to determine whether H&E 3.0 is comparable to conventional H&E. This ongoing validation, together with continued dataset expansion and scanner-agnostic optimization, will support the integration of virtual H&E into routine digital pathology workflows.
Li et al. (Tue,) studied this question.