{
  "abstract": "Background Haematoxylin and eosin (H&E) staining remains the clinical gold standard for visualising tissue structures essential for diagnosis. Hospitals with high caseloads and lacking in-house automated H&E staining may have prolonged turnaround times for clinical diagnosis. Virtual staining, where trained deep learning models generate various stains from unlabelled digital pathology images, could be employed to overcome this. 1–4 Here, we used generative models for high throughput generation of virtual H&E images from greyscale images and investigated the possible ideal staining checkpoint for implementation of virtual H&E staining.Methods As described in figure 1, one lung cancer tissue microarray (TMA) slide containing 25 patient cores underwent deparaffinization and differential interference contrast (DIC)-like imaging before and after H&E staining using Axioscan 7 (Zeiss, Germany). Data splitting was performed: 20 cores for training, 1 for validation, and 4 held-out for testing. Images were tiled at 256×256-pixels and registered to ensure precise spatial alignment between modalities.Two Pix2Pix generative adversarial network (GAN) models5 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.Model evaluation involved standardized nuclei segmentation on virtual and real H&E images. Predicted nuclei counts were compared with ground truth on a tile-by-tile basis using Pearson correlation coefficient. The best-performing model (based on the validation set correlation coefficient) was selected for subsequent testing and feature analysis (figure 2A). Deep features were extracted from intermediate layers of the Pix2Pix generator and visualised using Uniform Manifold Approximation and Projection (UMAP), followed by K-means clustering to assess feature differences.Results Both models produced biologically plausible virtual H&E images. On the held-out test set, Model A had marginally higher Pearson correlation (r=0.828) than Model B (r=0.812). However, visual inspection demonstrated Model B preserved clearer nuclear boundaries and more realistic staining contrast, especially in tumour regions ( figure 2B). UMAP of extracted deep features suggested distinct separation, indicating that models captured different structural features despite identical tissue sections being used (figure 2C-D). This representational divergence highlights that the staining checkpoint significantly influences the key structural features identified by the models.Conclusions Both models generated accurate virtual H&E images. Future work will expand datasets, optimise preprocessing, and explore scanner-agnostic virtual staining. These steps aim to enhance model robustness, improving integration into clinical workflows, particularly in resource-limited settings without conventional staining capabilities.Acknowledgements We thank our institutions for supporting this work.References Azam AB, Wee F, Vayrynen JP, Yim WWY, Xue YZ, Chua BL, et al. Training immunophenotyping deep learning models with the same-section ground truth cell label derivation Methods improves virtual staining accuracy. Front Immunol. 2024;15:1404640.Jabbar MADA, Yim WWY, Wee F, Chong LY, Yeong J. 1250 H&E-based cell prediction multi-classification models to capture morphologically distinct subpopulations of CD8+ T cells. J Immunother Cancer. 2024;12.Jackson CR, Sriharan A, Vaickus LJ. A machine learning algorithm for simulating immunohistochemistry: development of SOX10 virtual IHC and evaluation on primarily melanocytic neoplasms. Mod Pathol. 2020;33(9):1638–1648.Rivenson Y, Liu T, Wei Z, Zhang Y, de Haan K, Ozcan A. PhaseStain: the digital staining of label-free quantitative phase microscopy images using deep learning. Light Sci Appl. 2019;8:23.Isola P, Zhu JY, Zhou T, Efros AA. Image-to-image translation with conditional adversarial networks. Proc IEEE Conf Comput Vis Pattern Recognit. 2017;1125–1134.Ethics Approval This study was approved by the Agency for Science, Technology and Research; IRB number 2021-188.Abstract 1106 Figure 1Overall workflow. (A) Preparation of stained (post-H&E) and unstained (pre-H&E) DIC-like images for Model A & B; (B) Registration and 256×256 patching; (C) Pix2Pix GAN model structure; (D) Feature extraction and UMAP visualizationAbstract 1106 Figure 2Quantitative and representational comparisons between Model A and Model B. (A) Top 5 GAN checkpoints by Pearson correlation on validation set. (B) Ground truth vs. virtual H&E images. (C) UMAP + k-means of extracted features. (D) Joint UMAP showing feature differences",
  "authors": [
    {
      "affiliations": [
        "Agency for Science, Technology and Research (A*STAR), Singapore, Singapore"
      ],
      "name": "Ruisi Li"
    },
    {
      "affiliations": [
        "IMCB/A*STAR, Singapore, Singapore"
      ],
      "name": "Felicia Wee"
    },
    {
      "affiliations": [
        "Agency for Science, Technology and Research (A*STAR), Singapore, Singapore"
      ],
      "name": "Craig Joseph"
    },
    {
      "affiliations": [
        "Agency for Science, Technology and Research (A*STAR), Singapore, Singapore"
      ],
      "name": "Rachel EA Fincham"
    },
    {
      "affiliations": [
        "Yong Loo Lin School of Medicine, National University of Singapore, Singapore, Singapore"
      ],
      "name": "Siyu Zhang"
    },
    {
      "affiliations": [
        "Agency for Science, Technology and Research (A*STAR), Singapore, Singapore"
      ],
      "name": "Timothy O Wang"
    },
    {
      "affiliations": [
        "National University of Singapore, Singapore, Singapore"
      ],
      "name": "Jiangfeng Ye"
    },
    {
      "affiliations": [
        "Singapore General Hospital, Singapore, Singapore"
      ],
      "name": "Joe Poh Sheng Yeong"
    }
  ],
  "title": "1106 H&E 3.0: Pix2Pix-based virtual haematoxylin and eosin staining from greyscale images with feature comparison",
  "uid": "22993bd2-203b-5a94-910e-ebfe5ce1759b"
}