Label-free whole slide virtual multi-staining using dual-excitation photon absorption remote sensing microscopy

Abstract

Histochemical staining is essential for visualizing tissue architecture but is destructive and limited by tissue availability. Virtual staining with label-free microscopy offers a non-destructive alternative, enabling multiple stains from the same section. We present a dual-excitation Photon Absorption Remote Sensing (PARS) microscopy system, the first application of 355 nm UVA alongside 266 nm UVC excitation. The 355 nm source extends PARS contrast to include red blood cells, melanin, and enhanced stromal architecture through complementary radiative and non-radiative absorption. Using the RegGAN framework, we demonstrate the first PARS virtual staining across routine (H&E) and specialized (Masson’s trichrome, PAS, Jones methenamine silver) stains in human and murine tissues. Quantitative metrics show dual-excitation improves virtual stain similarity over single-wavelength inputs. A masked evaluation by expert pathologists provides an initial, coarse assessment indicating that virtual stains achieve ratings on a visual diagnostic quality scale comparable to chemical counterparts under limited evaluation conditions. These results support dual-excitation PARS as a promising non-destructive approach for multi-stain virtual histology. Corresponding whole-slide image pairs are available at the BioImage Archive (https://doi.org/10.6019/S-BIAD2232).

Introduction

Histochemical stains are the primary method for visualizing the microanatomy of tissue and remain essential in disease diagnostics and life-science research¹. By binding to specific biochemical structures, stains impart contrast in otherwise transparent tissue sections. The stained samples are examined using light microscopy, often with brightfield whole-slide scanners to capture high-resolution, sub-cellular detail across the sample for digital viewing, navigation, and computational analysis^2,3. The most widely used stain is hematoxylin and eosin (H&E), which stains nuclei purple and cytoplasm and extracellular components pink, and is routinely used to assess tissue architecture and cellular morphology, particularly in cancer diagnosis⁴. In addition, specialized stains provide more targeted contrast, such as Masson’s trichrome for collagen analysis, Periodic acid-Schiff (PAS) for carbohydrate-rich structures, and Jones methenamine silver (JMS) stain for highlighting glomerular and tubular membranes in renal pathology^1,4.

While invaluable, histochemical staining has many drawbacks. The process is destructive, and the same section cannot always be re-stained or used for ancillary diagnostic assays⁵. Multiple stains require additional cutting and consumption of limited biopsy tissue, and adjacent cuts inevitably have misalignments or mismatched structures. Beyond tissue loss, staining protocols are complex, multi-step procedures that require trained histotechnologists and dedicated laboratory infrastructure, and stain quality can be highly variable⁶. These workflows also generate chemical waste, add cost, and, when additional stains are required, can lengthen turnaround times and delay diagnosis and treatment⁷.

To address these challenges and complement traditional staining, a range of label-free microscopes have been developed that capture endogenous tissue contrast without the need for chemical markers. These methods leverage light-matter interactions such as scattering (e.g., quantitative phase imaging⁸, QPI) and electronic absorption via radiative and non-radiative pathways (e.g., autofluorescence⁹, photoacoustic¹⁰, and photothermal microscopy¹¹), along with non-linear counterparts including multiphoton fluorescence¹², harmonic generation¹³, and stimulated Raman scattering (SRS) microscopy¹⁴. Advances in computational methods have extended the utility of label-free imaging modalities in histology through virtual staining, which uses deep learning to translate raw images into familiar brightfield or fluorescent images of standard histochemical or immunohistochemical stains⁷. This enables multiple stains to be visualized on the same section, without the time, cost, or tissue loss associated with chemical staining, while preserving samples for downstream histological or molecular analyses.

A variety of label-free modalities have been explored for deep learning-based virtual staining. Autofluorescence imaging has been the most widely explored, having shown virtual H&E and other histochemical stains^{15,16,17,18,19,20}, as well as immunohistochemical (IHC) stains²¹. Fluorescence lifetime imaging (FLIM) has similarly been applied to virtual H&E staining^22,23, while scattering-based methods such as QPI^24,25 and brightfield imaging^26,27 have been used to generate multiple virtual stains. Non-radiative absorption techniques, including ultraviolet (UV) photoacoustic-based approaches have been used for virtual H&E staining^28,29,30,31. Nonlinear techniques, including stimulated Raman scattering (SRS)^32,33 and multimodal nonlinear methods³⁴, have also demonstrated the ability to generate H&E-like virtual stains.

The performance of virtual staining depends on the endogenous contrasts captured by the label-free modality. Input that provides greater biochemical specificity or complementary contrast enables the model to better distinguish tissue components and reconstruct the features of the target stain^15,22,35. For this reason, Photon Absorption Remote Sensing (PARS) microscopy is well suited for virtual staining because it simultaneously captures radiative and non-radiative absorption signals from a single excitation event, effectively combining the contrast information typically provided by autofluorescence (radiative pathway) and photoacoustic or photothermal methods (non-radiative pathway). Uniquely, the PARS signals are intrinsically linked to the same absorption event and therefore encode the total absorption of the target and the ratio of relaxation fractions (quantum efficiency ratio, QER^36,37). For label-free histology, PARS has primarily been implemented with a deep UV 266 nm excitation source. With this wavelength, non-radiative (NR) absorption provides important nuclear contrast, while radiative (R) autofluorescence highlights cytoplasmic and extracellular matrix (ECM) structures, together producing contrast highly analogous to H&E^36,38. In this configuration, PARS has previously been used for virtual H&E staining using both paired^39,40 and unpaired training schemes^41,42, and has demonstrated diagnostic concordance in masked studies of cancerous human breast⁴² and skin tissues⁴⁰.

While previous PARS histology studies have focused exclusively on H&E emulation, this work presents the first demonstration of PARS virtual staining across multiple specialized stains, including Masson’s trichrome, PAS, and JMS, in addition to H&E. This advance is achieved with a newly developed dual-excitation PARS system, representing the first application of long-wave UVA 355 nm excitation alongside the established 266 nm UVC source. The addition of 355 nm extends PARS contrast by targeting the absorption of multiple histologically relevant chromophores, providing complementary radiative and non-radiative contrast mechanisms to 266 nm^9,10,43,44. In particular, radiative signals enable visualization of stromal architecture via collagen and elastin autofluorescence⁴³, while non-radiative absorption highlights red blood cells (RBCs) through hemoglobin absorption and reveals melanin-containing structures such as melanocytes¹⁰.

Virtual staining is demonstrated on whole-slide images (WSIs) across a set of tissues and disease contexts, including human kidney with clear cell renal cell carcinoma (ccRCC), human skin with nodular melanoma and fungal infection, and multiple murine organs such as kidney, gastrointestinal tract (GI), and brain. An improved virtual staining model, RegGAN⁴⁵, is employed for image translation. This supervised framework integrates a registration network to improve robustness to residual misalignment between paired samples and is observed to produce higher-fidelity virtual stains than Pix2Pix and CycleGAN across the tissue types and stains examined. Furthermore, results demonstrate that the dual-excitation input consistently outperforms 266 nm or 355 nm excitation alone, supported by quantitative similarity and perceptual metrics and by qualitative assessments showing improved reconstruction of histology-specific features such as nuclei, collagen, RBCs, and fungal hyphae. A masked evaluation by three expert pathologists provides an initial, coarse assessment indicating that virtual stains achieved comparable ratings on a visual diagnostic quality scale to their chemical counterparts under the limited evaluation conditions of this study, and that image origin could not be reliably distinguished. For each tissue-stain combination, corresponding chemical and virtual WSIs are publicly available at the BioImage Archive⁴⁶ (see Data Availability).

Results

Expanded Contrast and Biomolecular Specificity with Addition of 355 nm Excitation

The dual-excitation PARS microscope captures label-free contrast through complementary radiative and non-radiative mechanisms at two ultraviolet excitation wavelengths. To illustrate the distinct features captured with each wavelength, Fig. 1 overviews images from human skin, human kidney, and mouse kidney samples, and demonstrates the value of combining these contrasts in a single visualization. The system achieved lateral spatial resolutions (FWHM) of 563.8 nm for the 405 nm detection beam, 370.8 nm for the 266 nm excitation and 520.3 nm for the 355 nm excitation (see Supplemental Fig. 1), enabling clear visualization of cellular details comparable to a 40x brightfield scanner⁴⁷. Across all samples, the measured pulse energies were 163 pJ (266 nm) and 808 pJ (355 nm), with a peak optical fluence at the focal plane estimated to be 1.05 nJ/μm² and 2.64 nJ/μm², respectively, assuming a Gaussian profile using the 1/e² spot size radius. These fluence values are approximately two orders of magnitude below reported deep-UV photodamage thresholds measured in living cells ( ~ 0.5–2 μJ/μm²) under deep-UV illumination⁴⁸ and provide a useful order-of-magnitude reference for fixed FFPE tissue. Consistent with this comparison, no histochemical artifacts or observable tissue damage was detected in the samples imaged.

Fig. 1: Summary of label-free contrasts captured by 266 nm and 355 nm excitation in the PARS system, showing strong correspondence with H& E and specialized stains.

a PARS images of human skin malignancy (nodular melanoma) showing separate non-radiative (NR) and radiative (R) contrasts at 266 nm and 355 nm. The 266 nm NR highlights nuclei, while 355 nm NR emphasizes RBCs and surrounding tissue detail. Radiative signals from both wavelengths outline connective tissue structures (e.g., sweat glands), with 355 nm better emphasizing the basement membranes. b Composite PARS image and corresponding H& E-stained section from the same sample. b-ii Melanoma with strong 355 nm NR sensitivity to melanin (blue); b-iii Tumor with melanin, necrosis, and vertical growth against the epidermis. c PARS image of human kidney with ccRCC, compared to Masson’s trichrome. The yellow 355 nm R signal corresponds well to the blue collagen-rich stroma in the stain. d PARS image of healthy mouse kidney and PAS stain. The PAS stain highlights carbohydrate-rich structures in a magenta color. Similar features are visible between the PARS and PAS image including nuclei (red, 266 nm NR), RBCs (blue, 355 nm NR), reticular fibres in the blood vessel, and the basement membranes of the glomerulus and surrounding tubules (yellow/green, R signals).

Figure 1a, b presents representative images acquired from human skin tissue diagnosed with nodular melanoma. Figure 1a shows a breakdown of the contrast channels obtained at 266 nm (left) and 355 nm (right), with each wavelength producing both non-radiative and radiative signals. This yields four distinct contrast channels per scan, each reflecting different structural and biochemical properties within the tissue. Importantly, because the 266 nm and 355 nm excitations are interlaced within a single scan trajectory, these four contrasts are captured essentially simultaneously, without compromising imaging speed. The 266 nm excitation provides strong nuclear contrast through non-radiative absorption of nucleic acids, and radiative signals from connective tissues and fluorophores in the ECM, providing clear visualization of surrounding tissue structures such as sweat glands.

In contrast, 355 nm excitation reveals biologically meaningful tissue features that are less prominent or absent under 266 nm. The non-radiative contrast strongly highlights RBCs through hemoglobin absorption¹⁰ and also provides signal in surrounding tissue structures not as well emphasized with 266 nm non-radiative. The radiative emissions at 355 nm similarly highlights many fluorophores and structures of the ECM (e.g., collagen and elastin), but with different spectral weightings than 266 nm radiative. For example, basement membranes of sweat glands are more pronounced in the 355 nm radiative channel, reflecting differences in absorption and emission efficiencies.

Figure 1b shows the composite PARS image from all four channels, alongside the corresponding H&E-stained section from the same malignant human skin sample. In Figure 1b-ii, a region of melanoma was selected to demonstrate the strong 355 nm non-radiative sensitivity to melanin. The 355 nm non-radiative signal (blue), clearly outlines heavily pigmented tumour cells, corresponding to dark melanin regions in the H&E image, highlighting the diagnostic value of 355 nm excitation.

Figure 1b-iii shows a nodular tumour with characteristic morphology, including nuclear pleomorphism, necrosis, and extensive melanin deposition. The epidermis is clearly delineated, and the tumour mass appears to press against it, consistent with the vertical growth patterns typical of nodular melanoma⁴⁹. The combined 266 nm and 355 nm contrasts align closely with the H&E reference but 355 nm enhances visibility of tumour architecture and melanin accumulation.

Figure 1c shows a human kidney sample with ccRCC and compares the composite PARS image to the corresponding Masson’s trichrome-stained section. Masson’s trichrome selectively stains collagen-rich regions in blue, helping identify areas of fibrosis and stromal remodeling⁴. These same collagen rich regions are clearly visible in the radiative channels of the PARS image, particularly in the 355 nm autofluorescence response (yellow). While collagen exhibits autofluorescence under both excitation wavelengths, the 355 nm response is more prominent, while the 266 nm radiative signal appears more uniform across the tissue.

Figure 1d shows a healthy mouse kidney sample imaged with PARS and compared to the corresponding PAS-stained section. The PAS stain highlights carbohydrate-rich structures, such as polysaccharides (e.g., glycogen), glycoproteins, and mucosubstances. It is commonly used in kidney pathology to delineate the basement membranes of glomerular capillary loops and renal tubules, which appear magenta⁵⁰. In the PARS image, both radiative channels show the tubule structures, and the capillary loops of the glomerulus correspond to the pink-stained regions in the PAS image. Nuclei in the glomerulus are shown in red (266 nm, non-radiative), and RBCs in the capillaries are visible in blue (355 nm non-radiative). A blood vessel on the left is outlined in yellow and green from both radiative signals and aligns well with the pink reticular fibers outlined in the PAS image.

Overall, the addition of 355 nm excitation expands the structural and biomolecular contrast available in the PARS system, complementing the contrast provided by 266 nm. In many cases, the two wavelengths highlight common tissue features, doing so with different intensity distributions, offering an alternate visualization of shared structures. Together, the combined contrasts correspond well with the patterns seen in both routine H&E and more specialized stains, suggesting their potential utility as input for label-free multi-stain virtual histology.

Virtual Staining Across Multiple Histochemical Stains

The complementary radiative and non-radiative contrasts captured by the 266 nm and 355 nm excitation wavelengths provide diverse and diagnostically relevant structural and biochemical information within the tissue, forming a strong input foundation for virtual staining. Building on the previous section, the following demonstrates the system’s ability to replicate a range of histochemical stains across both healthy and diseased samples, using the RegGAN-based staining framework described in Deep Learning Based Virtual Staining and the full dual-excitation input.

In evaluating the virtual staining performance of the models, three key aspects of the experimental design should be mentioned. First, all testing results were generated from WSIs that were excluded from training and validation. However, as summarized in Table 4, many test slides represent serial sections cut from the same tissue blocks used in training, while others originate from unseen tissue blocks or patients. As such, the reported results primarily assess generalization across adjacent tissue sections and, in some cases, across different tissue blocks, rather than strict patient-level generalization. Even across serial sections, meaningful differences in local morphology and cellular composition arise from sectioning and slide preparation, and the evaluation therefore remains informative for assessing whether the learned mapping preserves histologic structure and contrast, including the contribution of the additional 355 nm excitation to recovering stain-specific features.

Second, because PARS imaging is label-free and operates at low pulse energies that do not visibly damage the tissue, the same slides were chemically stained after imaging. This was core to the training process but is also important for evaluation because it allows direct one-to-one comparisons with the exact chemical counterpart, which is not always feasible in virtual staining studies.

Lastly, histochemically stained WSIs exhibit a wide range of complex tissue structures across many scales, from global patterns to fine cellular details and layouts. It can be challenging to convey staining performance across these scales through isolated crops alone. For this reason, in addition to showing representative areas, WSI virtual and chemical pairs have been made publicly available at the BioImage Archive (see Data Availability). Slides are standard OME-TIFF format and can be viewed and compared with standard pathology viewing software such as QuPath

Figure 2 presents representative virtual staining results across five stain-tissue combinations. Each example includes the label-free PARS input (left), virtual stain (middle), and chemical reference (right), with a zoomed-out whole slide view and high-resolution crop to demonstrate performance across scales. The figure includes human kidney and skin tissues, primarily from disease cases such as ccRCC, melanoma, and fungal infection, along with one mouse kidney example. Standard H&E staining is shown first, followed by three specialized stains (PAS, Masson’s trichrome, and JMS) to demonstrate more specific structural and biochemical features. In addition to the disease-focused cases in Fig. 2, Supplemental Fig. 4 shows results in healthy mouse tissues, demonstrating model performance and PARS contrasts on additional organs, including brain, gastrointestinal (GI) tract, and kidney.

Fig. 2: Virtual staining results using the dual excitation PARS input and the RegGAN framework across healthy and diseased tissue cases.

Examples span multiple tissue types and include both routine H& E and specialized histochemical stains. For each row: (left) label-free PARS input image, (middle) virtual stain, and (right) chemically stained reference. a H& E stain of human kidney with ccRCC, showing tumor nests with distinct cell borders and clear cytoplasm, well-defined fibrous septa with a rich capillary network, and regions of necrosis. b H& E-stained human skin with nodular melanoma, capturing melanin-laden tumor cells and nuclear pleomorphism within the tumor nodule. c PAS stain of human skin with fungal infection, revealing PAS-positive fungal hyphae in the stratum corneum. d Masson’s trichrome stain of ccRCC kidney, displaying the collagen-rich stroma (bottom right), low-collagen tumor nests (upper left), and a clear tumor-stroma interface. e JMS stain of healthy mouse kidney, showing silver-positive (black) tubular basement membranes and coloring of the Nuclear Fast Red counterstain (pink).

The first two rows of Fig. 2 present H&E-stained sections of human kidney and skin. In both cases, the virtual outputs closely resemble their chemical counterpart, properly staining nuclear structures in purple and cytoplastic content and surrounding connective tissue pink. In the malignant kidney tissue (Fig. 2a), the virtual stain effectively captures the ccRCC tumor architecture, including nests of cells with distinct cell borders and abundant clear cytoplasm (high power area)⁵⁰. The presence of necrosis is clearly visible in the PARS input and is clearly preserved in the virtual stain. This finding is associated with disease aggressiveness. In addition, the thin fibrous septa that separate the tumor nests are clearly visualized in the virtual stain and the fine vascular structures and RBCs within the capillary lumens are well rendered. In the human skin tissue with nodular melanoma (Fig. 2b), the virtual stain highlights the melanin-laden tumor cells and architecture of the tumor nodule⁴⁹. Within the tumor, the nuclei display marked pleomorphism, and the model is able to reproduce the small nucleoli inside these cells (higher power area). The distribution and localization of melanin within the tumor visually corresponds to the chemical stain, aided by the 355 nm non-radiative absorption contrast input.

The third row, Fig. 2c, presents a case of dermatophytosis, a superficial fungal infection of the skin commonly diagnosed using the PAS stain, which highlights polysaccharides in fungal cell walls (e.g., chitin)⁵¹. In the virtual stain, PAS-positive fungal hyphae are clearly visible as thin magenta-colored filaments clustered along the stratum corneum. The distribution and localization of these fungal organisms closely resemble the chemical reference, as do their color and orientation. The virtual output also faithfully reproduces the structure and organization of the surrounding epidermis layers. Notably, the fungal elements are discernible in the label-free PARS input, providing valuable contrast to support virtual staining of these diagnostically important structures.

In the ccRCC kidney stained with Masson’s trichrome (Fig. 2d), the virtual output closely matches the chemical reference and enables visualization of the fibrosis surrounding the invasive tumor. The intricate blue and red staining in the collagen-rich area at the bottom right is effectively translated in detail in the virtual stain. In contrast, the tumor growth in the upper left has minimal collagen content, and the virtual stain is also consistent with this, providing clear visualization of the tumor-stroma interface. The higher power region further highlights the model’s sensitivity to subtle collagen cues within the tumor nests, which could aid the evaluation of tumor-driven stromal remodelling.

The last row, Fig. 2e, shows the JMS stain applied to healthy mouse kidney tissue. This stain highlights glomerular and tubular basement membranes in black and is invaluable for the assessment of non-neoplastic kidney biopsies⁵². Overall, the virtual output closely matches the chemical reference, sharply staining the basement membranes between tubules black and coloring the pink nuclear and background counterstain (Nuclear Fast Red). The higher magnification area showcases a distal tubule structure with a well-defined black border and clearly resolved epithelial nuclei, matching well with its chemical reference.

To further highlight the utility of the virtual staining models, Fig. 3 presents simultaneous virtual H&E, Masson’s trichrome, PAS, and JMS generated from a single multichannel PARS input in mouse kidney tissue. Figure 3a highlights a region containing both abnormal glomeruli and surrounding normal structures, while Fig. 3b shows a normal glomerulus for comparison. The observations in Fig. 3a were made in an aging mouse, where spontaneous, age-related glomerular alterations are not uncommon^53,54.

Fig. 3: Simultaneous virtual H& E, Masson’s trichrome, PAS, and JMS staining of mouse kidney tissue from a single PARS input.

a Region containing two abnormal glomeruli with nodular expansion of mesangial areas, along with a blood vessel and surrounding renal tubules. b Normal glomerular architecture from another mouse kidney section, shown in the same four virtual stains for comparison.

The two glomeruli in Fig. 3a show globally disrupted architecture, with disrupted and compressed capillary networks due to mesangial area expansion by pale-staining, amorphous material. Each of the virtual stains contributes information about the nature of the glomerular alteration. In the virtual H&E, the glomeruli appear pale pink. The virtual Masson’s trichrome shows the mesangial nodules to appear pale blue-gray (versus the darker blue color typical of collagen). The virtual PAS and JMS stains display reduced staining of the amorphous material, with PAS appearing pale and JMS a pink-gray color, lacking the strong magenta color of PAS or the sharply defined black staining of JMS seen in the normal glomerulus in Fig. 3b. These combined staining insights are most consistent with glomerular amyloidosis⁵⁵, a condition involving extracellular protein deposits (i.e., amyloid) within glomeruli that disrupts glomerular filtration barrier and impairs selective permeability.

The region in Fig. 3a also includes several well-preserved normal structures. A blood vessel with clearly defined smooth muscle is visible across all stains, with surrounding collagen clearly identifiable in the virtual Masson’s trichrome. Reticular fibers can also be seen within the vessel wall in the virtual JMS. The surrounding renal tubules are also properly stained and sharply outlined with the virtual PAS and JMS. Supplemental Figs. 5 and 6 provide additional examples of simultaneous virtual multi-staining in healthy mouse GI tissue and human ccRCC kidney tissue, where the multiple stains offer complementary views of tissue structure and pathology.

Together, the example demonstrates how each virtual stain contributes distinct structural and biochemical information that, when combined, allows for a more comprehensive visualization of the glomerular pathology than routine H&E or any single stain alone could provide. For example, work by Haan et al. demonstrated that virtual multi-staining, achieved through stain-to-stain transformations, can improve diagnostic accuracy in renal pathology, supporting the value of combining multiple stain types⁵⁶. In addition to the virtual stains, the label-free PARS input here offers an additional distinct visualization of tissue structures, with a unique contrast mechanism independent from chemical staining. Moreover, because the sample remains unaltered, subsequent staining with other specialized markers can be performed. For example, to definitively confirm amyloid deposition in Figure 3a, the same kidney section could be chemically stained with Congo red and imaged under polarized light microscopy⁵⁵.

Each virtual staining model in this study is trained independently for a specific histochemical stain, and cross-stain structural consistency is therefore not explicitly enforced during training. Figure 3 and Supplemental Figs. 5 and 6 present representative examples in which multiple virtual stains are generated from identical multichannel PARS inputs.

In Fig. 3, the glomerular structures are maintained across the virtual H&E, Masson’s trichrome, PAS, and JMS outputs generated from the same PARS field of view. Similarly, the blood vessel structure and surrounding renal tubules are preserved across all four stains, and the general distribution of cellular nuclei appears consistent across stains.

Supplemental Fig. 5 demonstrates similar agreement in mouse GI tissue, where smooth muscle fibres share the same directional alignment across virtual H&E, Masson’s trichrome, and PAS stains. Surrounding connective tissue structure is preserved across stains, and the overall architecture of the intestinal crypts, along with associated nuclear features, appears consistent.

Supplemental Fig. 6 provides an additional example in human ccRCC kidney tissue, where virtual H&E and Masson’s trichrome are generated from the same PARS field of view. Both low- and high-magnification views show correspondence of connective tissue and nuclear structures. In particular, the higher-magnification views highlight consistent multi-staining of the nested tumor architecture, RBCs, and alignment of connective tissue fibres and nuclear structures across the two virtual outputs.

In addition to demonstrating cases where the virtual staining closely resembles the chemical reference, it is also helpful to highlight situations where the output deviates from the ground truth to better understand model behavior. Accordingly, a qualitative assessment of some representative failure cases observed across the virtual staining results is provided in Supplemental Fig. 2. Firstly, virtual staining failures often occur in areas where there are slide artifacts or acquisition-related issues, such as tissue folds, out-of-focus areas, or debris on the sample (e.g., dust or hair). These artifacts typically generate strong radiative signals which locally dominates the PARS input and wash out other non-radiative contrasts, reducing the available structural cues, leading to weakened or missing structures or local miscoloration in the virtual stain (Supplemental Fig. 2a–e).

Additional examples are shown where the virtual staining deviates from the chemical reference. In Supplemental Fig. 2f, blue surgical margin ink appears blue in the chemical stain but is not rendered as blue in the virtual output, instead appearing eosin-like or RBC-like coloring. Notably, the margin ink is present in the raw PARS input and produces signal in the 355 nm non-radiative channel (blue), a channel that also typically highlight RBC structures in the tissue. In Supplemental Fig. 2g, fungal hyphae in PAS-stained skin are generally detected in the virtual output but appear incompletely stained or fragmented in some regions. In addition, Supplemental Fig. 2h shows an example of a hair follicle base where the hair shaft is miscolored with dark eosin-like staining that does not match the chemical reference, potentially reflecting limited representation of such structures in the training data.

In addition to these failure examples, the overall global color was assessed using L*a*b* color space histograms comparing virtual and chemical stains. Supplemental Fig. 3 shows overlaid distributions of the L* (lightness), a* (green-red axis), and b* (blue-yellow axis) channels for virtual and corresponding chemical stains, aggregated across randomly sampled image patches for each stain type (N = 5000). This analysis provides an overall, global view of color consistency and is intended to assess whether systematic color shifts are present.

Across all stains, the virtual staining outputs exhibit close overlap with the chemical reference distributions in each channel, indicating good overall agreement in luminance and chromatic components. Minor distributional differences are observed, but do not suggest widespread or systematic color or intensity bias. This consistency is likely aided by the fact that chemical staining was performed within a single facility using standardized protocols, reducing inter-slide variability during training.

Pathologist Image Assessment Results

The goal of the masked pathologist evaluation in this work was to obtain a coarse, initial assessment of the perceived diagnostic quality (i.e., overall visual adequacy for histopathologic review) of virtual stains relative to their chemical counterparts, and to examine whether pathologists could distinguish image origin under masked conditions. Toward this end, a limited set of 20 images was evaluated, consisting of 10 chemical and 10 virtual image pairs taken from exactly matched fields of view. A summary of diagnostic quality (DQ) scores and image origin classification responses is provided in Table 1, with the full set of individual pathologist responses reported in Supplemental Table 1.

Table 1 Summary of masked pathologist evaluation of chemical and virtual stains

For each pathologist, diagnostic quality scores assigned to chemical and virtual image pairs were similar, with both image types consistently rated as “good” or “excellent”. Diagnostic quality scoring is inherently subjective and was used here as a coarse measure to assess relative agreement rather than specific diagnostic performance. To quantify the agreement in diagnostic quality scores between the paired chemical and virtual images, a linear weighted Cohen’s kappa was used. Pathologist 1 showed substantial agreement (κ = 0.615), pathologist 2 showed moderate agreement (κ = 0.483), and pathologist 3 showed complete agreement (κ = 1.000).

When asked to classify image origin, pathologists were unable to reliably distinguish chemical from virtual images under masked conditions. When excluding responses marked as “uncertain”, image origin classification accuracy was 50.0% (95% CI: 27.2–72.8%) for pathologist 1 (who classified 100% of images as chemical) and 47.1% (95% CI: 23.0–72.2%) for pathologist 2, while pathologist 3 marked all responses as uncertain. When pooled across pathologists (n = 37 “yes” or “no” responses), overall origin classification accuracy was 48.6% (95% CI: 31.9–65.6%), suggesting that image origin was not reliably distinguishable in this evaluation. A substantial fraction of responses were marked as “uncertain” (38.3% of all classifications), which reflects the difficulty of confidently distinguishing chemical from virtual images under masked conditions. While uncertainty responses were excluded from the primary accuracy analysis, treating these responses as incorrect would further reduce observed classification accuracy.

Overall, within the limits of the small sample size used here, several observations can be made. Diagnostic quality scores assigned to virtual images were generally similar to those assigned to chemical stains across the images evaluated, with no obvious or systematic differences observed. In addition, under masked conditions, pathologists were not able to reliably distinguish virtual images from chemical images, with a substantial proportion of responses marked as uncertain. These observations provide supportive evidence of visual similarity between virtual and chemical images; however, future work involving larger studies with increased sample size and more structured diagnostic tasks that are specific to tissue type and stain would be required for a rigorous assessment of diagnostic equivalence and comparative performance.

Contribution of Individual Excitation Channels to Virtual Staining Performance

To evaluate the individual and combined contributions of 266 nm and 355 nm excitation contrasts, additional models were trained using only 266 nm input and only 355 nm input. These were compared to the dual excitation models presented in the previous section for all tissue and stain combinations used in this study (Table 4). Relative performance was then quantified using two metrics: (1) Multi-Scale Structural Similarity Index (MS-SSIM)⁵⁷, which captures pixel-level structural similarity across scales, with higher values indicating better similarity; and (2) Deep Image Structure and Texture Similarity (DISTS)⁵⁸, which captures perceptual similarity based on structural and textural fidelity, with lower values indicating better similarity. The overall aggregated results across all tissue and stain combinations are summarized in Table 2. For completeness, the per-dataset breakdown of these metrics is provided in Supplemental Table 2 and Supplemental Table 3.

Table 2 Summary quantitative scores (Avg. ± Std) across all datasets for different excitation wavelength input combinations

In all cases, models trained with both excitation contrasts achieved improved scores (higher for MS-SSIM and lower for DISTS) compared to those trained with either wavelength alone. These results are consistent with the idea that the complementary contrast from both wavelengths provides more biochemical specificity and structural information to the model, supporting improved virtual staining performance. In general, 266 nm models outperformed 355 nm models for most tissue-stain combinations, likely attributed to the essential nuclear contrast present in the 266 nm non-radiative channel. One notable exception was observed in the skin melanoma sample, where the 355 nm model achieved a better DISTS score, reflecting improved replication of melanoma texture from the 355 nm non-radiative channel.

While these metrics help assess overall staining quality, further analysis was done to assess how the addition of 266 nm or 355 nm affects the virtual staining of more specific histological features. Figure 4 presents three representative examples, collagen in Masson’s trichrome, nuclei in H&E, RBCs, and fungal hyphae in PAS-stained skin, with each demonstrating how dual excitation improves staining of clinically relevant features.

Fig. 4: Dual-excitation models improve staining of histological features across multiple stains.

a Masson’s trichrome, ccRCC kidney: Each region shows a collagen segmentation map (bottom left) and Collagen Proportionate Area (CPA). Violin plots of CPA distribution (N = 4,205, 256px patches) show the dual excitation model (median 0.57) closely matches the reference (0.59), versus the 266 nm only model (0.40). b H& E, mouse kidney: Dashed outlines and arrows highlight missing nuclei in the 355 nm only output. Adding 266 nm contrast improves nuclear staining. Violin plots of nuclear count distribution (N=17,821, 256px patches) show close agreement between the dual excitation model (median 42.00) and the reference (41.57), compared to the 355 nm-only model (34.95). c Improvement of RBC staining with addition of 355 nm excitation. In c-i and c-ii, RBCs are absent or miscolored in the 266 nm only model; in c-iii, false RBC-like structures appear where none exist in the reference. d PAS, skin tissue: Dual excitation improves fungal hyphae staining. Adding 355 nm results in more clear hyphae staining, whereas the 266 nm only model shows incomplete or missed structures.

Figure 4a compares virtual Masson’s trichrome staining of ccRCC kidney tissue using models trained with 266 nm only versus dual excitation input. Using QuPath digital pathology software, a pixel classifier was trained to segment collagen-positive pixels. The collagen proportionate area (CPA), defined as the fraction of tissue area classified as collagen, was computed across 256 × 256 px patches in regions with at least 5% collagen. While CPA is primarily used in liver biopsies to assess fibrosis and clinical outcomes⁵⁹, here it serves as a quantitative proxy for evaluating collagen staining accuracy. Three example regions are shown in Fig. 4a, in order of increasing CPA values. In each example, the 266 nm model underestimates collagen presence, mis-staining collagen areas in red, leading to significantly reduced CPA values. By adding the 355 nm channels, the model more closely matches the reference stain, as reflected in the segmentation maps (bottom left), which show a similar distribution of collagen. These qualitative findings are further supported in the violin plots below, which show CPA distributions across N = 4205 patches. The dual excitation model achieves an almost symmetric distribution (median 0.57 ± 0.25), closely matching the ground truth (0.59 ± 0.25). In contrast, the 266 nm model exhibits a skewed distribution (median 0.40 ± 0.25), reflecting an overall underestimation of collagen. The improvement likely results from the addition of the 355 nm radiative channel, which provides complementary autofluorescence contrast to 266 nm. While collagen fluoresces under both excitations, many surrounding structures do as well. Access to both radiative channels allow the model to better stain collagen based on learned differences in emission intensities across tissue structures. Surrounding tissue structures also exhibit non-radiative contrast at 355 nm, which may further aid in distinguishing collagen.

While quantitative metrics (CPA and nuclear counts) demonstrated the model’s performance on collagen and nuclei, the evaluation of red blood cells, fungal hyphae, and melanin remains qualitative in this study. Detailed manual annotation of these features for quantitative benchmarking was outside the scope of this work; however, the visual correspondence between the virtual and chemical stains indicates that the dual-excitation input provides essential contrast for these structures.

Figure 4b demonstrates the benefits of including 266 nm input for nuclear staining. Examples are shown in H&E mouse kidney and compare the 355 nm only model to the dual excitation output. In each case, the 355 nm model struggles to accurately reconstruct nuclei, and several nuclear structures are missed entirely and replaced by pink eosin staining. Although 355 nm lacks nuclear contrast, the model does successfully infer some nuclear structures from the surrounding tissue, but this is not always reliable. In contrast, by adding 266 nm excitation, the model consistently identifies and correctly stains nuclei. The quantitative nuclear counts reinforce this difference. Using StarDist⁶⁰ for segmentation, nuclear counts were computed for N = 17,821 patches. The dual excitation model closely matches the chemical H&E distribution (median 42.00 ± 14.50 vs. real 41.57 ± 14.23), while the 355 nm model shows a similar distribution shape but with a shifted center, reflecting the overall undercounting of nuclei (34.95 ± 12.54).

Figure 4c shows three examples of how the addition of 355 nm channels improves staining of RBCs. While the 266 nm model can often identify RBCs using structural cues, it lacks hemoglobin-specific contrast which can lead to errors. In Figure 4 examples (c-i) and (c-ii), RBCs are either missed or incorrectly stained purple as nuclei. While in Figure 4c-iii, the 266 nm model hallucinates red RBC-looking structures where none exist. In contrast, the 355 nm non-radiative channel (shown in blue) captures hemoglobin absorption, providing both positive and negative cues that help the model reliably distinguish RBCs and reduce such false positives.

Lastly, Fig. 4d shows three examples of PAS staining of fungal hyphae in skin tissue. The 266 nm model identifies some fungal structures but often underrepresents or only partially stains them. With both excitation contrasts, the model more clearly defines the fungal elements, aided by the 355 nm radiative channel (shown in yellow), which enhances their visibility and supports virtual staining. Based on this qualitative assessment, the dual excitation output more closely matches the reference PAS stain. While some deviations remain and the staining is not perfect, the addition of 355 nm contrast leads to a noticeably improved staining of fungal hyphae.

Comparative Performance of Virtual Staining Models

The RegGAN model was chosen for virtual staining because it enables supervised training while remaining robust to imperfect alignments by integrating a registration network during training. This makes it well-suited for cross-modality tasks where precise pixel-level alignment is difficult to achieve. To isolate the contribution of this registration component, RegGAN was benchmarked against Pix2Pix in an ablation comparison, with Pix2Pix serving as a paired baseline trained on conventionally registered WSI pairs but without a learnable registration network. Both RegGAN and Pix2Pix employed identical ResNet-based generator and PatchGAN discriminator backbones and were trained on the same datasets pre-registered using the Warpy pipeline⁶¹. In addition, comparisons were made against CycleGAN (unpaired) to evaluate performance in the absence of paired supervision.

Figure 5 presents qualitative and quantitative comparisons across the models, using a subset of examples that highlight each stain applied to a relevant tissue type or disease case featured in this study. Figure 5a-e show: (a) H&E-stained human kidney with ccRCC (N = 60,531), (b) H&E-stained human skin with nodular melanoma (N = 21,739), (c) PAS-stained human skin with fungal infection (N = 11,406), (d) Masson’s trichrome-stained kidney with ccRCC (N = 66,672), and (e) JMS-stained healthy mouse kidney (N = 16,473). Each row in the figure includes the 266 nm and 355 nm PARS input, the corresponding chemical stain, and outputs from the three models. In addition, violin plots showing the relative performance of each model are provided for both the DISTS (perceptual similarity) and MS-SSIM (structural similarity) metrics. The overall aggregated quantitative scores for RegGAN, Pix2Pix, and CycleGAN across all tissue-stain combinations are summarized in Table 3. Detailed per-dataset results are reported in Supplemental Tables 4 and 5.

Fig. 5: Visual and quantitative comparison of virtual staining performance for RegGAN, Pix2Pix, and CycleGAN models.

Each example shows the PARS input image, ground truth chemical stain, and outputs from all three models, along with distributions of quantitative metrics (DISTS and MS-SSIM). The comparison covers all stains used in this study and includes representative disease cases where possible: a H& E-stained human kidney with ccRCC, (b) H& E-stained human skin with nodular melanoma, (c) PAS-stained human skin with fungal infection, (d) Masson’s trichrome-stained human kidney with ccRCC, and (e) JMS-stained healthy mouse kidney.

Table 3 Summary quantitative scores (Avg. ± Std) for RegGAN, Pix2Pix, and CycleGAN models across all datasets

The Pix2Pix model provides a practical paired baseline illustrating the limitations of relying solely on conventional whole-slide image registration for supervised virtual staining. Despite being trained on WSI pairs pre-aligned using the Warpy framework, Pix2Pix consistently produces smoother and blurrier outputs than RegGAN across all stains and tissue types (Figure 5). This is evident both qualitatively in the visual comparisons and quantitatively in degraded MS-SSIM and DISTS scores (Table 3). In practice, this results in a loss of fine structural detail at higher magnifications, where diagnostic features such as mitotic figures, nuclear atypia, and subtle architectural cues are critical. Examples include reduced nuclear detail (Fig. 5b), poorly defined glomerular basement membranes in JMS-stained kidney tissue (Fig. 5e), and incomplete PAS fungal staining (Fig. 5c). While Pix2Pix captures overall stain color and coarse tissue structure, these limitations constrain its utility for high-resolution histological assessment. The comparatively sharper visual appearance and improved similarity metrics achieved by RegGAN support the value of incorporating a learnable registration mechanism to mitigate residual misalignment during training.

The CycleGAN model, trained in an unpaired fashion, produces sharper images and retains textural detail more effectively. However, its less constrained training approach limits its ability to learn more complex contrast mappings. For example, RBCs are poorly reproduced in the H&E kidney (Fig. 5a); the blue collagen stain is exaggerated and poorly localized in the trichrome example (Fig. 5d); fungal elements are entirely missed in the PAS skin (Fig. 5c); and melanin and nuclei are poorly rendered in the H&E melanoma case (Fig. 5b). In the JMS-stained kidney (Fig. 5e), CycleGAN roughly applies the correct stain color and texture but fails to accurately stain the real basement membrane structures.

In contrast, by incorporating a registration network into the training process, RegGAN benefits from paired training feedback while being robust to data misalignments. It preserves high-resolution texture and structure better than Pix2Pix, particularly evident in the accurate rendering of nuclear detail (Fig. 5b) and basement membranes (Fig. 5e), while also achieving more faithful stain translation than CycleGAN. It successfully captures collagen in the trichrome example (Fig. 5d), the PAS-stained fungal elements (Fig. 5c), and melanin in the skin example (Fig. 5b), producing the most consistent alignment with the chemical reference across all examples.

These qualitative observations are supported by the quantitative metrics. RegGAN achieves the best performance across all stain–tissue combinations under the DISTS metric, which captures perceptual similarity based on structural and textural fidelity (lower values reflect better similarity). CycleGAN generally achieves better DISTS scores than Pix2Pix, particularly in Fig. 5b, d, e, where Pix2Pix exhibits more pronounced blurring and textural loss. However, CycleGAN underperforms in specialized staining tasks such as PAS fungal staining (Fig. 5c), where it fails to capture key features that Pix2Pix better preserves.

When evaluated with MS-SSIM, which captures pixel-level structural similarity across scales (higher values reflect better similarity), RegGAN again outperforms both models. CycleGAN performs the worst in all MS-SSIM examples, reflecting its failure to accurately map certain structural features and stain distributions, despite its sharper overall appearance and textural similarity.

RegGAN demonstrated the strongest performance in virtual staining; however, the results highlight the broader utility of incorporating registration directly into a training pipeline. It also helps mitigate challenges in cross-modality learning tasks, particularly those involving mismatched resolutions or differences in acquisition, such as optical sectioning, illumination, or other imaging conditions. By addressing these discrepancies as part of the training and optimization process, registration-aware approaches enable more scalable and robust use of imperfectly paired data. Beyond virtual staining, this strategy may benefit a range of cross-modality learning tasks in biomedical imaging where alignment is a limiting factor.

Discussion

A primary limitation of the PARS configuration presented here is the total scan time compared to conventional brightfield whole-slide scanners. As mentioned in Interlaced Scanning, Signal Acquisition, and Image Reconstruction, acquisition of individual 0.5 mm × 0.5 mm sections takes  ~ 5 min, with imaging speed limited by the excitation laser pulse repetition rate (50 kHz) and overhead associated with mechanical stage scanning. At this rate, imaging a representative 1 cm² area at 40 × would take approximately 33 h, whereas modern clinical brightfield scanners scan comparable areas at 40 × conservatively in under five minutes⁶². However, conventional brightfield necessitates prior chemical staining, adding preparation time ranging from tens of minutes for H&E to multiple hours for special stains (e.g., Masson’s trichrome) or over a day for IHC assays²¹.

This difference in imaging speed reflects the point-scanned, stage-based raster acquisition used for PARS in this study, combined with the relatively low excitation laser repetition rate, rather than a fundamental limitation of the PARS absorption contrast mechanism. The system presented here was not optimized for imaging speed, as the primary objective of this work was to demonstrate the expanded label-free contrast provided by dual-excitation PARS and demonstrate label-free virtual multi-staining on held-out slides, supported by an initial assessment of diagnostic quality. As such, further increases in throughput would be required for routine whole-slide imaging applications.

Substantial improvements in imaging speed are achievable through higher-repetition-rate UV excitation and fast beam-scanning architectures. Concurrent work has already demonstrated a system utilizing a 1 MHz 266 nm excitation source coupled with a hybrid optomechanical scanning configuration⁶³. This platform achieved acquisition speeds of  ~ 130 min/cm² at 40 × equivalent resolution and  ~ 35 min/cm² at 20 × ⁶³. Importantly, the hybrid scanning architecture reported in that work is inherently scalable to higher repetition rates, and with a 5 MHz source, imaging times would drop to approximately 35 min/cm² at 40 × and 15 min/cm² at 20 × .

In addition to acquisition time, end-to-end performance is influenced by the inference time required to generate virtual stains from whole-slide PARS data. In this work, virtual staining was performed using patch-wise inference on 256 × 256 pixel tiles. With the single RTX 4070 Super GPU used in this study, inference required approximately 2.94 ms per tile. Given a tile size corresponds to 64 μm × 64 μm of tissue area, virtual staining a representative 1 cm² WSI ( ~ 1.6 gigapixels) would require  ~ 72 s per virtual stain. While this inference time is short relative to the PARS acquisition time reported here, further reductions would likely be needed for integration into time-constrained clinical workflows, particularly if multiple virtual stains are generated for an entire slide. Inference and model speed was not a primary focus of this study, and further improvements in whole-slide virtual staining throughput remain an area for future investigation.

Importantly, the dual-excitation PARS virtual staining approach presented here is not intended to replace routine brightfield H&E scanning, which remains highly optimized for throughput and well established in clinical pathology workflows. Instead, PARS may offer complementary value in scenarios where access to additional histochemical contrast from a single tissue section, or selective analysis of regions of diagnostic interest, is desired. While the current work focuses on virtual staining, the underlying PARS signals may also provide complementary morphological or compositional information that could be useful in certain research or exploratory settings.

Because PARS imaging is label-free and non-destructive, the same FFPE tissue section can be virtually stained and subsequently used for conventional chemical staining or more advanced downstream assays such as IHC. This capability may be particularly valuable in cases involving limited biopsy material, or in research workflows where virtual staining could be used to provide additional histological context prior to tissue consumption. Furthermore, virtual staining does not necessarily require inference over an entire whole-slide image. In some settings, virtual stains could be generated selectively for regions of interest identified either by a pathologist or by computational pre-screening.

In addition to the throughput and integration considerations discussed above, limitations regarding data generalization should also be noted. While tissue samples in this study were sourced from multiple institutions, all PARS imaging was performed using a single dual-excitation PARS system, and ground-truth histochemical staining was carried out within a single laboratory. As a result, the models were trained and evaluated within a consistent imaging and staining setting. Training against a consistent histochemical reference allows the learned mapping to focus on biologically meaningful contrast rather than being confounded by inter-laboratory staining differences (which virtual staining has the potential to reduce); nevertheless, specific FFPE and staining procedures and pre-analytic factors do vary across different laboratories and centers, and the present results do not establish robustness across such variations. In addition, test sets include a mixture of serial sections and slides from unseen tissue blocks or patients, and strict patient-level separation was not enforced. Accordingly, the results demonstrate the feasibility of virtual staining within the specific imaging, staining, and data-splitting contexts examined here, but do not establish robustness across independent patients, institutions, imaging systems, or staining workflows. Evaluation on larger, multi-center datasets with enforced patient-level separation remains an important direction for future work.

Finally, the quantitative evaluation in this study encompassed full-reference image quality metrics (MS-SSIM, DISTS) for model benchmarking, more targeted evaluation of specific histological features including CPA and nuclear counts, and masked pathologist scoring. However, for other diagnostically important structures highlighted in the manuscript, such as RBCs, fungal hyphae, and melanin pigment, evaluation relied on qualitative visual correspondence with chemical references. While MS-SSIM and DISTS were effective for comparing overall quantitative model performance, these measures are not tailored to specifically verify the accuracy of complex, irregular histological structures which can be difficult to quantify at a sufficient scale or in a diagnostically meaningful way. Similarly, the masked pathologist evaluation in this work was scoped to obtain a coarse, initial assessment of the perceived diagnostic quality (i.e., overall visual adequacy for histopathologic review) and distinguishability of image origin, rather than to validate structured diagnostic tasks. Consequently, qualitative comparison was emphasized for various histological features, and corresponding real and virtual whole-slide images have been made publicly available for further qualitative comparison.

In conclusion, this study introduces a dual-excitation PARS microscopy system for label-free virtual multi-stain histology, combining 266 nm and 355 nm ultraviolet excitation to capture complementary radiative and non-radiative tissue contrast. The addition of 355 nm excitation expands the range of histological features that can be visualized, including red blood cells and melanin-containing elements, and enhances depiction of stromal architecture such as collagen and elastin, while 266 nm excitation provides strong nuclear contrast and surrounding connective tissue structure.

Using the RegGAN image-translation framework, this work presents the first demonstration of PARS virtual staining across multiple specialized histochemical stains, including Masson’s trichrome, PAS, and JMS, in addition to H&E, spanning healthy and pathological human and murine tissues. Evaluation on unseen whole-slide images showed close visual correspondence between virtual and chemical stains, with improved similarity and perceptual metrics, and enhanced visualization of histology-specific features when both excitation wavelengths were used. A masked pathologist evaluation further provided an initial, coarse assessment indicating that virtual stains achieved ratings on a visual diagnostic quality scale comparable to their chemical counterparts, and that image origin could not be reliably determined under masked conditions.

Together, these results demonstrate the technical feasibility and potential utility of dual-excitation PARS virtual staining as a non-destructive approach for generating multiple histological contrasts from a single tissue section. However, the present study is limited by its retrospective design, a modest masked evaluation, single-center imaging, and the current acquisition throughput of the PARS scanning architecture. As such, dual-excitation PARS virtual staining should be viewed as a promising step toward nondestructive multi-stain histology. Future work will require larger, multi-center validation studies, expanded pathologist assessments with task-specific diagnostic endpoints, and continued advances in acquisition speed. Within these constraints, the present work establishes a technical foundation for future studies aimed at advancing label-free virtual multi-staining in digital pathology.

Methods

Dual-Excitation PARS System Architecture and Configuration

The dual-excitation PARS histology microscope system used in this study is illustrated in Fig. 6a. The system includes two excitation sources: a 400ps pulsed 266 nm UVC laser (Wedge XF 266, Bright Solutions) and a 1.5 ns pulsed 355 nm UVA laser (ULPN-355-10-1-10-M, IPG Photonics), both operating at a 50 kHz pulse repetition rate. These excitation wavelengths were selected to provide complementary contrast for virtual staining. As shown in previous works^36,38, 266 nm offers strong absorption of nucleic acids, providing important nuclear-specific non-radiative contrast, along with radiative emissions from ECM structures. The 355 nm source is similarly broadly absorbed, providing complementary autofluorescence from ECM components such as collagen and elastin⁴³, along with non-radiative signals from absorbers like hemoglobin for RBC visualization, melanin, and other pigments¹⁰.

Fig. 6: Dual-excitation PARS architecture, interlaced scanning pattern, and signal extraction.

a Overview of the PARS microscope optical system. Component labels: AC (achromatic collimator), BT (beam trap), BS (beam sampler), Col. (collimator), HS (harmonic separator), MMF (multi-mode fiber), M (mirror), NF (notch filter), OL (objective lens), PD (photodiode detector), VBE (variable beam expander), Cond. (condenser). b Interlaced excitation scan pattern. A delay generator alternately triggers 266 nm and 355 nm pulses to create interlaced excitation pulses on the sample. Pulses are inhibited during stage turnaround via a programmed stage GPIO signal. A high-speed digital acquisition (DAQ) card is triggered at each excitation pulse to collect time-series data at each event. c Signal extraction from system photodetectors. At each pulse,  ~ 500 ns segments are recorded from three photodetectors and compressed into pixel values for each contrast. This produces non-radiative and radiative images for each excitation wavelength, and a 405 nm scattering image. Non-radiative contrast is calculated as the modulation from the scattering baseline; radiative signal is defined as the peak value; and reference powers (266 nm pulse amplitude and 405 nm power) are extracted for post-acquisition power correction.

The output from the 266 nm source first passes through a CaF₂ prism (PS862, Thorlabs) to spectrally separate residual 532 nm green light, which is directed to a beam trap (BT: BT610, Thorlabs). The 266 nm beam is then 3 × expanded using a variable beam expander (3 × VBE: BE03-266, Thorlabs). The adjustable collar of the expander allows for collimation control of the 266 nm laser for axial beam alignment. The expanded beam is directed toward a Nd:YAG harmonic separator (HS₂: 37-721, Edmund Optics) where it is combined with the 355 nm source.

The output from the 355 nm source is first condensed by 2 × and then re-expanded with a 2 × variable expander (2 × VBE: BE02-355, Thorlabs), resulting in an overall magnification of 1×. As with the 266 nm, this is done so the adjustable collar of the expander can be used to fine-tune the collimation of the 355 nm laser for axial alignment purposes. The 355 nm beam is then reflected off a second Nd:YAG harmonic separator (HS₁: 37-721, Edmund Optics) and transmitted through HS₂ to align with the 266 nm path.

A continuous wave 405 nm detection source (OBIS-LS 405, Coherent) is used to probe the sample for scattering signals and non-radiative contrast. The fiber-coupled laser is first collimated (Col.: C40APC-A, Thorlabs) and directed through both HS₁ and HS₂ to co-align with the two excitation beams.

All three beams (266 nm, 355 nm, 405 nm) are then co-focused onto the sample using a 0.42NA UV objective lens (OL₁: NPAL-50-UV-YSTF, OptoSigma). The forward-propagating light, including scattered 405 nm and 355 nm and radiative emission from both UV sources, is collected using a 0.7 NA top objective lens (OL₂: 278-806-3, Mitutoyo). The collected light is coupled into a multimode optical fiber (MMF: M133L01, Thorlabs) using a high-NA achromatic coupler (AC₁: F950FC-A, Thorlabs), and then re-collimated into free space by AC₂ for downstream spectral separation.

The scattered 355 nm light is immediately sent into a beam trap using a 355 nm notch filter (NF₁: 39-387, Edmund Optics). The transmitted 405 nm beam is then separated from the radiative emissions using a 405 nm notch filter (NF₂: NF405-13, Thorlabs). The radiative emissions and non-radiative scattering modulation are each recorded on avalanche photodetectors (PD: APD130A2, Thorlabs) following condenser lenses (Cond.: ACL25416U-A, Thorlabs). An additional avalanche photodetector is used in the forward path to record the 266 nm pulse amplitude and 405 nm power level; the 355 nm source was not recorded given its lower pulse variability. These beams are sampled using beam samplers (BS1: BSF10-UV and BS2: BSF10-A, Thorlabs) for post-acquisition power correction.

To estimate the effective spot sizes of the detection and excitation beams at the focal plane, edge spread functions (ESF) were collected, sampled densely at 50 nm/pixel, across sharp edge features (see Supplemental Fig. 1). From this, the Full Width Half Maximum (FWHM) of the resulting line spread functions (LSF) were calculated, and the beam waist (1/e² width) was used to estimate peak fluence based on the pulse energies used during imaging.

Interlaced Scanning, Signal Acquisition, and Image Reconstruction

An overview of the scanning process, signal acquisition, and image reconstruction pipeline is illustrated in Fig. 6b, c. A programmable delay generator (DG645, Stanford Research Systems) is used to alternately trigger the 266 nm and 355 nm lasers, each at 50 kHz, producing spatially interlaced excitation spots along the scan trajectory. By interlacing the pulses rather than performing two sequential scans, the dual-excitation system acquires radiative and non-radiative contrast from both excitation sources essentially simultaneously, without compromising imaging speed.

During a scan, the mechanical stages move the samples in an S-shaped raster pattern. In the linear velocity regions, the stages move at 12.5 mm/s, spacing the alternating 266 nm and 355  nm pulses by 125 nm, and maintaining a 250 nm step between successive pulses of the same wavelength.

During the turnaround (non-linear) regions, both excitation lasers are disabled to avoid high-density UV pulses and potential photobleaching. A TTL high signal from the stage is used to inhibit the delay generator, effectively gating off the laser triggers during these periods.

At each excitation pulse, a high-speed digitizer (Gage CSE1442, 200 MS/s) captures a  ~ 500 ns time segment from each system photodetector, as shown in Figure 6c. In total, three photodetectors (PDs) are used to collect signals that produce a total of five distinct PARS contrasts in a single pass: 405 nm scattering, 266 nm and 355 nm non-radiative (PD1), and 266 nm and 355 nm radiative, along with 266 nm and 405 nm input reference powers (PD3).

PD1 captures the forward-scattered 405 nm signal and non-radiative response. The scattering contrast is calculated as the average pre-excitation scattering intensity. These scattering images are extracted during the autofocus phase of the whole-slide scan to ensure high-SNR acquisition. Non-radiative contrast is computed from the transient modulation of the scattering signal relative to a predicted baseline following excitation. As described in⁶⁴, the scattering baseline is estimated using a first-order linear fit of the scattering signal before and after excitation. The non-radiative signal is then calculated as the percentage deviation from this predicted baseline. This approach has been shown to improve SNR⁶⁴, particularly at structural edges where scattering intensity varies rapidly, by reducing error from motion-induced baseline shifts.

PD2 captures the radiative emission signals from both excitation wavelengths. The radiative contrast is extracted by taking the peak signal amplitude within the acquisition window. PD3 measures the 266 nm and 405 nm input reference powers. The average scattering input power is estimated from the pre-excitation signal level, while the 266 nm pulse energy is calculated as the difference between the peak signal amplitude and the scattering power reference.

Following the extraction of pixel intensities for each excitation pulse, the 266 nm and 355 nm pixel values are deinterlaced and reconstructed into separate image sections by mapping each pixel onto a Cartesian grid based on their stage positions. Each reconstructed section represents a 0.5 mm × 0.5 mm field of view. Using the workflow detailed in³⁸, whole slide imaging is performed by autofocusing across the sample using the scattering signal and capturing a series of overlapping sections. The total imaging speed is primarily limited by the 50 kHz pulse repetition rate of the excitation sources used in this work. At a pixel spacing of 250 nm, this yields a theoretical imaging capacity of approximately 5.3 min/mm². However, the effective acquisition time includes overhead from the mechanical dynamics of the scanning stages (acceleration/deceleration during turnarounds) and autofocusing routines. Therefore, capturing the 0.5 mm × 0.5 mm sections takes  ~ 5 min when considering stage constraints and autofocusing times³⁸. While the interlaced nature of the 266 nm and 355 nm pulses allows for simultaneous dual-contrast acquisition without extending scan time, the current throughput is restricted by the pulse repetition rate and stage mechanics. Finally, these captured sections are stitched and blended to form a WSI of each contrast.

Deep Learning Based Virtual Staining

The process used to prepare the training data for the virtual staining model and to perform WSI inference is illustrated in Fig. 7. Following PARS imaging, the same unstained tissue sections were subsequently stained and brightfield scanned. Due to the differences in the imaging modalities and tissue deformation during the staining, spatial misalignments between the PARS and stained images were then corrected using the Warpy registration pipeline⁶¹. Warpy applies a combination of rigid and non-rigid transformations to closely align the two WSI pairs, a necessary starting point for the model training. All virtual staining models evaluated in this study (RegGAN, Pix2Pix, and CycleGAN) were trained and evaluated using the same conventionally registered WSI pairs generated via the Warpy pipeline.

Fig. 7: Workflow for preparing training data and performing whole slide virtual staining.

(1) PARS and stained WSIs are aligned using the Warpy registration framework. (2) Tissue borders and artifacts are annotated in QuPath, and the slides are tiled into artifact-free patches. (3) These patches are used to train the virtual staining model. (4) Virtual staining is applied to full PARS WSIs using overlapping patch-wise inference, followed by linear blending and tissue masking to produce the final virtual stained slide. The example shown depicts mouse H& E gastrointestinal tissue.

After registration, QuPath was used to annotate the tissue borders and label artifacts across the tissue. These artifacts include tissue tears, cover-slip bubbles, folds, and out-of-focus areas in the stained images, as well as paraffin obstructions and dust artifacts in the PARS images. Annotations were exported in GeoJSON format and processed by a custom Python script, which excluded the artifact regions and tiled the WSIs into smaller, co-registered patches. Tiling was required to divide the high-resolution slides into manageable chunks for training.

The resulting datasets were used to train the virtual staining model, described in the following section. Once trained, the model was applied to the entire PARS WSIs using a patch-based inference with 40% overlap between adjacent tiles. Overlapping patch regions were linearly blended to ensure visual continuity, resulting in a virtual WSI resembling a conventional stained slide. Following inference, a tissue mask was applied to remove any background artifacts. The mask was generated by thresholding the 355 nm radiative contrast channel from the PARS image, and the background regions were smoothly faded to white.

Virtual staining employs image-to-image translation models, most of which follow one of two training paradigms: supervised approaches like Pix2Pix and variants^65,66, and unsupervised methods such as CycleGAN⁶⁷. Both approaches have been widely used in virtual staining, with unpaired methods more common due to general scarcity and difficulty in obtaining perfectly registered training data⁶⁸.

Supervised models like Pix2Pix are effective when training pairs are precisely aligned, however, achieving such alignment is particularly difficult in high-resolution, cross-modality tasks like PARS to brightfield translation. Even with registration pipelines like Warpy, small residual misalignments exist due to differences in the optical systems, imaging conditions, and local tissue deformation during staining. These inconsistencies introduce label “noise” that can result in blurred or smeared output quality. Pix2Pix has previously been applied to PARS data for H&E emulation^39,40, demonstrating strong performance and high concordance under carefully registered conditions; however, at the cost of labor-intensive preprocessing and time-consuming annotation to identify and exclude poorly aligned regions, which also reduced the effective dataset size.

CycleGAN-based models avoid the alignment issue by eliminating the need for paired training data. However, the lack of pixel-level supervision results in weaker constraints on both contrast translation and structural preservation. CycleGAN has previously been applied to loosely paired PARS data for virtual H&E staining⁴¹, successfully avoiding the blurring observed with Pix2Pix under misalignment. While this worked well for PARS to H&E, where the mapping between modalities was more direct, the stains explored in this work involve more complex or subtle structural and contrast relationships, making a supervised approach better suited to learn these mappings.

For these reasons, the virtual staining model used in this work is based on RegGAN⁴⁵, which offers a more effective balance between the paired and unpaired training regimes. RegGAN treats misalignments in the target image as a form of noisy labelling and incorporates a learnable registration network into the training process. This network adaptively estimates deformation fields that align the generated image to the target domain. As such, the model retains the benefits of paired supervision while remaining robust to residual misalignment, making it well suited for high-resolution virtual staining across imaging modalities.

The RegGAN training scheme is illustrated in Fig. 8a. The generator network (G) receives a multichannel PARS input patch (x) and produces a virtual stained output (G(x)). This output, along with its corresponding misaligned real stain image ($\widetilde{y}$), is concatenated and passed through the registration network (R). The registration network predicts a deformation field (T), which is applied to G(x)) to generate a registered virtual output $G\widetilde{(x)}$ that is more closely aligned with the real stain. An L1 correction loss is computed between the registered virtual output and the real stain image to encourage structural and visual similarity in the generated output.

Fig. 8: Overview of the RegGAN-based virtual staining framework and network architectures.

a Schematic of the RegGAN training workflow. The generator network G receives a multichannel PARS input x and produces a virtual stain G(x). This is concatenated with the misaligned real stain $\widetilde{y}$ and passed through the registration network R, which predicts a deformation field T that is applied to align the virtual stain. A correction loss is computed between the registered output and the real stain, while a smoothness loss is applied to T to enforce spatial smoothness. An adversarial discriminator D guides G via a least squares loss. b Architecture of the discriminator D, based on a 70 × 70 PatchGAN, which outputs a patch-wise classification map for real/fake discrimination. c Architecture of the generator G, a ResNet-based encoder-decoder that processes four-channel 266 nm and 355 nm contrast images and outputs an RGB virtual stain. d Architecture of the registration network R, a ResUNet-style model that predicts a 2D deformation field from a six-channel concatenation of the real and virtual stain. IN = Instance Normalization; ReLU = Rectified Linear Unit.

In parallel, the discriminator (D) learns to distinguish between real and virtual images, while the generator works to fool the discriminator by producing increasingly realistic virtual stains. To improve training stability, a least squares adversarial loss is used, per the LSGAN framework⁶⁹. Additionally, a smoothness loss is applied to the predicted deformation field, encouraging spatial smoothness by penalizing large spatial gradients.

Figure 8b–d detail the convolutional architectures of the discriminator (D), generator (G), and registration (R) networks, respectively. Full layer configurations, including convolutional kernel sizes, strides, and feature dimensions, are provided within the figure. The discriminator network, shown in Figure 8b, follows the 70 × 70 PatchGAN architecture which outputs a patch-wise classification map, where each value reflects the likelihood of a local region being real or generated. The localized feedback helps the generator in producing realistic fine-scale texture and structural detail.

The generator network, illustrated in Figure 8c, adopts a ResNet-based encoder-decoder structure, commonly used in image-to-image translation tasks⁶⁷. The inputs to the generator are the concatenated non-radiative and radiative contrasts from both the 266 nm and 355 nm wavelengths. The network consists of an encoder with two downsampling layers, a bottleneck with nine residual blocks, and a decoder that upsamples the features and outputs a three-channel RGB virtual stain image.

The registration network, shown in Figure 8d, is a ResUNet-style architecture from⁷⁰. It receives a six-channel input formed by concatenating the virtual and real stain images. The architecture consists of seven downsampling levels, a central bottleneck with three residual blocks, and a symmetric decoder with bilinear upsampling and skip connections from the encoder. The final output is a two-channel deformation field representing the predicted x and y displacements.

An overview of the tissue types, stains, and dataset sizes used in this study is provided in Table 4. The datasets include mouse gastrointestinal (GI), kidney, and brain tissues; human kidney samples with clear cell renal cell carcinoma (ccRCC); human skin with fungal infection; and human skin with nodular melanoma. The histochemical stains considered include hematoxylin and eosin (H&E), Masson’s trichrome, periodic acid-Schiff (PAS), and Jones’ methenamine silver (JMS). Separate virtual staining models were trained independently for each tissue-stain combination listed in Table 4. Each model maps the multichannel PARS input to a single target histochemical stain.

Table 4 Summary of patient, block, and slide distribution across training, validation, and testing partitions

Table 4 also summarizes the patient-, block-, and slide-level composition of the datasets used for training, validation, and testing across all tissue and stain combinations. The tiles from the training and validation WSIs were pooled and randomly split into 80% training and 20% validation sets, without slide-level separation. In contrast, test slides were held out entirely at the slide level and were not used during training or validation. As shown in the table, test sets include a mixture of serial (adjacent) tissue sections and slides originating from unseen tissue blocks or patients. Accordingly, while test slides were excluded from training and validation, strict patient- or block-level separation was not enforced across all datasets. For each WSI, a global mean and standard deviation were computed for each contrast channel, and each PARS input patch was normalized by subtracting this mean and dividing by the standard deviation.

All models were trained using the Adam optimizer with betas (0.5, 0.999). The generator and registration networks used a learning rate of 2 × 10⁻⁴, while the discriminator used a lower learning rate of 2 × 10⁻⁵. The generator was set to update four times for every update of the discriminator and registration networks. A batch size of 1 and a fixed patch size of 256 × 256 pixels was used. All LeakyReLU activations had a negative slope of 0.2. The smooth loss, correction loss, and adversarial loss were weighted by 10, 20, and 1, respectively. Data augmentation included random horizontal and vertical flips, along with 90° rotations.

Training was conducted for up to 70 epochs or until validation losses plateaued. The discriminator was updated using a history buffer containing the 50 most recent generated images, rather than relying solely on the latest outputs, as suggested in⁷¹. All models were implemented in Python 3.11.9 using PyTorch 2.4.1+cu118, and training was performed on a single NVIDIA RTX 4070 Super GPU using mixed-precision computation.

In addition to the primary models with all contrasts, ablation experiments were conducted to evaluate how different subsets of PARS contrast channels affect virtual staining performance. Separate models were trained using only 266 nm radiative and non-radiative inputs, as well as only 355 nm radiative and non-radiative inputs. All training parameters remained consistent with the primary models. Some contrast configurations had more limited input information, and the generator update frequency was initially increased during the first  ~ 10 epochs to improve early training stability, after which the standard 4:1 generator to discriminator and registration network update ratio was resumed.

Sample Preparation

Tissue samples were sectioned at  ~ 5μm thickness onto microscope slides from formalin-fixed paraffin-embedded (FFPE) blocks prepared using standard histological methods. Freshly resected tissues were fixed in formalin within 20 min of excision and stored for 24 to 48 h. Samples were then dehydrated in ethanol, cleared with xylene, and embedded in liquid paraffin wax at 60 °C. Slides were baked at a low temperature (50–55 °C) to smooth the paraffin surface for PARS imaging.

After PARS imaging, tissues were stained with one of the following: hematoxylin and eosin (H&E), Masson’s trichrome, PAS, or JMS, and scanned at 40 × magnification using a brightfield microscope (Leica Aperio AT2, Robarts Research Institute).

The ccRCC human kidney samples were provided by the Ontario Tumour Bank. Human skin tissues (nodular melanoma and fungal infection) were provided by collaborators at the Cross Cancer Institute (Edmonton, Alberta, Canada). Murine tissues were resected in-house and sent to the Robarts Research Institute (University of Western Ontario) for FFPE processing.

The mouse tissue used in this study was obtained from animals housed and maintained at the University of Waterloo Central Animal Facility. All procedures were conducted under an approved institutional animal care protocol (University of Waterloo Animal Care Committee Protocol ID #44703), in accordance with relevant guidelines and regulations. Human samples were from anonymous patient donors and no information regarding patient identity was provided to the researchers. Patient consent was waived by the ethics committee because these archival tissues were no longer required for patient diagnostics. Samples were collected under protocols approved by the Research Ethics Board of Alberta (Protocol ID: HREBA.CC-18- 0277) and the University of Waterloo Health Research Ethics Committee (Photoacoustic Remote Sensing (PARS) Microscopy of Surgical Resection, Needle Biopsy, and Pathology Specimens; Protocol ID: 40275). All human tissue experiments were conducted in accordance with the government of Canada guidelines and regulations, such as “Ethical Conduct for Research Involving Humans (TCPS 2)".

Masked Pathologist Evaluation

To obtain an initial assessment of diagnostic quality and visual similarity between chemical and virtual stains, a masked image evaluation was performed with three board-certified pathologists. A total of 20 tissue images were evaluated, consisting of 10 chemically stained images and 10 corresponding virtual stains acquired from exactly matching fields of view of the same tissue sample.

Each image was presented at both low-magnification (overview) and high-magnification (cellular detail). The survey images included all stain types used in this study and incorporated both healthy and disease cases across multiple tissue types. The image order was randomized, rotated, and flipped to minimize recognition of paired samples, and pathologists were masked to image origin.

Pathologists were asked to provide a score for the diagnostic quality of each image on a three-point scale (1 = poor, 2 = good, 3 = excellent). They were also asked to identify whether the image was chemically stained or not, with the options “yes,” “no,” or “uncertain.” Responses were collected independently for all participants. The agreement in diagnostic quality between paired chemical and virtual images was quantified using linear weighted Cohen’s kappa, and image origin classification accuracy was evaluated using binomial confidence intervals.