The review, led by Santanu Misra of Sungkyunkwan University and Chiho Yoon of Pohang University of Science and Technology, together with colleagues including Chulhong Kim and Byullee Park, takes stock of deep learning-based virtual histological staining, a technique in which neural networks learn to translate label-free images of unstained tissue—or images stained with one dye—into photorealistic syntheses of stains that were never applied. The authors frame the technology as having already left the proof-of-concept stage, with demonstrations spanning quantitative phase microscopy, autofluorescence imaging, photoacoustic microscopy, optical coherence tomography, Raman and infrared spectroscopy, and even in vivo imaging of human skin. But they warn that inconsistent data handling, model design, and evaluation practices are now actively slowing its march into the clinic.

The technical core of virtual staining is a data-driven image-to-image transformation. In the label-free setting, a network is trained on pairs of images: a tissue region imaged without dyes, and the same region after chemical staining. The network learns the mapping between intrinsic optical signals—autofluorescence from cellular metabolites and structural proteins, phase shifts from refractive index variations, or endogenous absorption measured acoustically—and the characteristic color and contrast patterns of hematoxylin and eosin, Masson’s trichrome, or immunohistochemical markers. Once trained, the model can generate stain-like contrast directly from raw, unstained images, in some cases within seconds. In the stain-to-stain setting, the model instead converts one existing stain into another, allowing a laboratory to extract additional molecular or diagnostic information from a single stained section without cutting and processing more tissue.

The lineage of the field traces back to landmark demonstrations such as PhaseStain, which digitally stained label-free quantitative phase images in 2019, and virtual H&E staining of tissue autofluorescence published the same year in Nature Biomedical Engineering. Since then, the review documents an accelerating proliferation: virtual staining of biopsy-free in vivo skin, of human carotid atherosclerotic tissue, of autopsy material, of amyloid deposits via birefringence imaging, and of glioma tissue from hyperspectral images. Diffusion models, which generate images through iterative denoising, have recently joined generative adversarial networks as workhorse architectures, with pixel super-resolution virtual staining and pathology-aware Schrödinger bridge approaches pushing both fidelity and training efficiency. Transformer-based backbones have been adapted to capture the long-range tissue context that convolutional networks can miss.

The prize is substantial. Chemical staining and the turnaround time it imposes are bottlenecks in surgical pathology, particularly during operations when frozen sections must be prepared, stained, and read in minutes. Label-free virtual staining could eliminate that delay entirely: photoacoustic-based systems have already demonstrated label-free intraoperative histology of bone tissue and rapid cancer diagnosis at subcellular resolution, allowing surgeons to receive histology-grade feedback without waiting for a cryostat. Because the tissue is never chemically processed or destroyed, virtual staining also preserves material for molecular testing, enables repeated virtual stains from a single section, and opens the door to stains that are impractical or impossible to perform chemically, such as virtual multiplexed immunostaining for assessing vascular invasion in cancer.

Yet the review’s central message is cautionary. The authors find that studies vary enormously in how imaging data are acquired, how tissues are curated, how image pairs are registered and preprocessed, how networks are configured, and—most consequentially—how results are evaluated. Because deep networks learn statistical correlations rather than physical laws, a model trained on autofluorescence images from one microscope, one tissue type, or one institution may fail silently when applied elsewhere, a problem known as domain shift. The review highlights the pathological extremes of this risk: hallucination, in which a generative network invents plausible-looking structures that do not exist in the underlying tissue. If a hallucinated morphological feature changes a diagnosis, the consequences could be severe, and recent work on scalable hallucination detection frameworks underscores how seriously the field now treats this failure mode.

To address the reproducibility gap, the authors propose a standardization blueprint that spans the entire pipeline: modality-specific data construction, model design, handling of domain shift, evaluation strategy, and safety assessment. A key deliverable is a minimum reporting checklist, analogous in spirit to the CLAIM, TRIPOD+AI, CONSORT-AI, and SPIRIT-AI guidelines that transformed reporting standards in medical imaging and clinical AI. The checklist would require researchers to disclose their datasets, imaging protocols, preprocessing steps, training configurations, and evaluation settings in a consistent format, enabling fair cross-study comparison and reproducible benchmarking. Without such disclosure, the authors argue, claims that one virtual staining system outperforms another are essentially unverifiable.

Evaluation itself receives pointed criticism. Common image-similarity metrics such as peak signal-to-noise ratio and structural similarity index, along with perceptual measures derived from deep features and distributional metrics like FID and MMD, reward statistical closeness to real stained images but do not guarantee that diagnostic content is preserved. A virtually stained image can score well on every pixel-level metric while subtly distorting nuclear morphology or inventing mitotic figures. The review calls for pathology-aware evaluation metrics, built around diagnostically relevant structures, and for expert reader studies in which pathologists assess whether virtual slides support the same interpretations as their chemical counterparts—an approach already tested in clinical-grade validation of an autofluorescence virtual staining system for prostate cancer.

The question of clinical translation is where the review is most deliberately sobering. The authors situate virtual staining within real pathology workflows, complete with whole-slide imaging, digital pathology infrastructure, and regulatory oversight, and conclude that full replacement of chemical staining is not yet routine—and should not be presented as imminent. Regulatory frameworks for AI-based diagnostic tools remain in flux, and the evidence base, while growing rapidly, still contains gaps in multicenter validation, long-term performance monitoring, and clear accountability when a virtual slide and a chemically stained slide disagree. The authors emphasize that near-term adoption is most realistic in well-defined niches: intraoperative consultation, rapid assessment where tissue is scarce, research settings requiring multiplexed stains, and adjunctive second reads rather than autonomous diagnosis.

That measured framing distinguishes the review from much of the celebratory literature. The field’s trajectory is undeniable: what began as a laboratory curiosity a decade ago now spans organ systems, imaging modalities, and stain types, with foundation models for computational pathology processing more than a hundred clinical-grade tasks. But the authors’ blueprint makes clear that the next phase of progress will be won not by bigger networks or flashier generative architectures, but by the unglamorous work of consistent reporting, rigorous benchmarking, hallucination safeguards, and regulatory engagement. If the community adopts these standards, the vision that animates the field—histology-grade images of living, unstained tissue, produced in seconds at the bedside or in the operating room—moves from a compelling demonstration to a defensible clinical tool. Until then, the dyes stay in the dish, and the burden of proof stays with the algorithms.