Assessment of Transfer Learning for H&E Virtual Staining in Veterinary Oncology

Conventional hematoxylin and eosin staining is indispensable for histopathology but is time-, labor-, and resource-intensive, environmentally unsustainable, and sensitive to operator and reagent variability. We developed a two-stage transfer-learning virtual staining (VS) pipeline for generating H&E-like images from label-free autofluorescence (AF) slides. In Stage 1, a Pix2Pix Dense U-Net was trained under paired supervision using spatially registered fluorescence images of H&E-stained sections and brightfield H&E images acquired from the same physical sections. The fluorescence inputs contained channel-dependent eosin emission and hematoxylin-associated attenuation, which provided image contrast available to the network but were not encoded as explicit loss terms. In Stage 2, the retained Pix2Pix(v2) checkpoint initialized channel-specific Pix2Pix(v3) models, which were fine-tuned on registered single-channel fluorescence–H&E pairs (DAPI, Cy3, or FITC input to the corresponding full H&E target) and then applied to native AF images from unstained sections lacking paired H&E targets. The pipeline integrates HSV masking, ORB–RANSAC registration, multi-resolution tiling, SSCOR-based stripe suppression, a fine-tuned Pix2Pix Dense U-Net, and an evaluated CycleGAN color-normalization step that was ultimately discarded because it degraded SSIM/FID and expert scores. In a four-case pilot comparison, grayscale DAPI yielded the highest overall pathology-based scores and the most coherent whole-slide VS among the tested configurations; Cy3 and FITC were prone to discontinuities and hue shifts, and multi-channel composites underperformed. In a veterinary cohort (13 soft tissue sarcomas, 12 mast cell tumors), single-observer semiquantitative pathology scoring (technical quality, H&E staining, image quality, accuracy) found that tissue-level architecture and tumor margins were preserved, but cellular and nuclear detail was limited, precluding confident diagnosis or grading. Quantitatively, improved preprocessing and model refinement lowered FID and raised SSIM versus the tested baselines. This approach reuses archived H&E slides as source-domain supervision and provides a practical route for adapting virtual staining models to label-free AF data under real-world acquisition constraints.