Label free virtual HE from 3D holotomography enables reliable nuclear and microanatomical readouts in thick cancer tissues

Conventional H&E staining is destructive, 2D-limited, and incompatible with longitudinal or scarce-sample studies. Holotomography (HT) enables label-free 3D refractive index (RI) imaging of thick tissue, but pixel-accurate supervision for virtual staining is unattainable due to sectioning-induced deformation. Building on our recent demonstration of 3D virtual H&E for colon and gastric cancer tissues up to 50 µm thick, we developed a framework that learns hematoxylin- and eosin-like contrast directly from HT volumes using weakly aligned chemical H&E supervision. The method produces stable, interpretable nuclear and stromal features, enabling nuclei-centric quantitative readouts in a fully label-free workflow.FFPE breast cancer tissues were imaged with 3D HT to obtain quantitative RI volumes at subcellular resolution. A diffusion-based image-to-image model was trained with weak HT-H&E alignment and a nucleus-aware auxiliary loss applied only during training. Inference requires HT alone, maintaining a consumable-free, non-destructive pipeline. Fidelity was evaluated by (i) nuclear morphology agreement, (ii) patch- and field-level realism, (iii) smooth z-slice consistency, and (iv) standardized nuclei quantification. Cross-tissue generalizability was assessed using previously validated colon and gastric datasets spanning 10-50 µm thickness.The virtual H&E preserved basophilic nuclear contrast and eosinophilic stromal features, accurately reconstructing glandular, stromal, and muscular boundaries even in thick sections. Nuclear contours remained continuous across axial planes without flicker artifacts. Whole-field nuclear counts showed 97.8% concordance with chemical H&E, and nuclear area/shape metrics showed no significant deviation. Appearance error was notably lower than prior virtual-staining approaches, and expert reviewers confirmed faithful chromatin texture, nucleoli visibility, and nuclear crowding patterns essential for cancer grading. The method remained robust across tissue thicknesses (4-50 µm), organ types, and institutional sources.Virtual H&E generated directly from label-free HT enables high-fidelity, nuclei-driven readouts without staining, sectioning artifacts, or tissue loss. This supports rapid, consumable-free tissue assessment and reproducible quantification of nuclei-centric biomarkers. Its compatibility with thick tissue volumes positions the technology for drug-response profiling, mechanism-of-action studies, tumor microenvironment analysis, and high-content phenotypic screening. Ongoing multi-site validation and development of 3D diffusion models for whole-slide inference aim to establish label-free 3D virtual histopathology as a scalable alternative to conventional H&E for research and translational oncology.