Live, label free 3D virtual HE imaging of small intestinal organoids using holotomography and generative cross-modality artificial intelligence

Organoids provide physiologically relevant 3D models of tumor biology, but current imaging depends on fluorescence labeling, clearing, or destructive sectioning, limiting longitudinal analysis and perturbing native states. Holotomography (HT) enables long-term, high-resolution, label-free 3D imaging of live intestinal organoids, capturing crypt budding, mitosis, apoptosis, and drug-induced cytotoxicity. Generative frameworks that translate refractive index (RI) tomography into multiplexed fluorescence or H&E-like contrast, including RI2FL and 3D virtual H&E, have shown accurate cross-modality prediction. Here, we integrate these technologies to achieve live, label-free 3D virtual H&E visualization of growing intestinal organoids.Mouse small intestinal organoids were imaged continuously for 120 hours using low-coherence HT. 3D RI tomograms covering whole organoids were reconstructed and stitched. A 3D generative translation model-trained on paired RI-H&E data from human colon cancer tissues and extended using RI2FL's generalizable strategy-was adapted to predict 3D virtual H&E volumes directly from RI stacks, without staining, fixation, or sectioning. Resulting images were evaluated for epithelial architecture, crypt-villus organization, nuclear morphology, luminal topology, and drug-induced injury. Cisplatin-treated organoids were monitored over days.Label-free HT revealed morphogenesis including symmetry breaking, crypt budding, epithelial migration, extrusion, and luminal remodeling. Applying the model to live RI volumes produced artifact-free 3D virtual H&E with hallmark features: basophilic nuclei, eosinophilic cytoplasm, apical organization, Paneth-like granularity, and mucin-rich domains. Depth-resolved virtual H&E preserved axial continuity of crypt-villus structures. In drug-response assays, cisplatin induced early nuclear condensation, crypt collapse, and epithelial fragmentation; virtual H&E highlighted these patterns more clearly than raw RI, while HT-derived metrics showed decreases in protein density and dry mass. The model maintained stable inference across dynamic, unlabeled, live samples, extending RI2FL's cross-system generalizability to complex multicellular organoids.This work establishes the first long-term, fully label-free 3D virtual H&E imaging of live organoids, enabling multi-day, non-destructive histology with subcellular contrast. The approach quantitatively captures pharmacodynamic injury and viability and offers a scalable platform for preclinical oncology, drug screening, and organoid pathology, introducing a new paradigm of live virtual histopathology.