1. INTRODUCTION

Comprehensive spatial insights from tissue and organs at the “mesoscopic” centimeter scale are of growing interest. Fluorescence imaging of large intact samples is a key enabler in protein expression studies [1] and organ-level spatial transcriptomics [2]. Confocal [3], multiphoton microscopy [4], and light-sheet microscopy with tissue clearing [5] are important techniques today but are costly and complex commercial offerings, usually available as shared infrastructure in privileged scientific organizations. An uncomplicated and economical microscope for rapid cellular imaging of thick, large samples has been a long-standing challenge in surgical pathology [6] and is emerging in life science with stringent requirements on quantitative imaging, molecular specificity, and compatibility with laboratory assays [7]. Specifically for tissue, a fast analog to conventional hematoxylin and eosin (H&E) histology is needed for thick or intact samples [8]. Label-free microscopy of biological samples based on optical coherence tomography [9] or quantitative phase imaging [10] lacked molecular multi-color contrast, which hampered image interpretation and multiplexed readouts.

The use of deep ultraviolet (DUV) illumination for fluorescence excitation is a recent promising approach for thick-sample microscopy without lasers or scanning. DUV light penetrates unsectioned tissue superficially, providing optical sectioning, and can simultaneously excite multiple fluorophores using a single wavelength [11,12]. Also known as microscopy with ultraviolet surface excitation (MUSE), this technique’s simplicity has been previously highlighted for its potential in intra-operative imaging [13,14], although demonstrations in life science have been limited [15,16]. We propose fiber optic microscopy with UV surface excitation (FUSE), featuring a new fiber optic illumination strategy [Fig. 1(a)] that enables high-quality imaging up to ${50 \times}$, 0.55 NA [Fig. 1(c)], surpassing the capabilities of prior MUSE designs. Leveraging precise angular control of the fiber optic illuminators, near-horizontal (up to ${\sim}{87^ \circ}$ from vertical in air, ${\sim}{46^ \circ}$ in tissue) illumination minimized the effective penetration depth and enhanced optical sectioning for high magnification [Fig. 1(b)] without the need for an immersion medium in an inverted setup suitable for thick tissue samples and rugged deployments. The fiber optic delivery also enabled oblique illumination for low-cost objective lenses with conventionally small working distances (down to ${\sim}1\;{\rm mm}$), more efficient heat management of light sources at a distance, and notably, micro-scale 3D topography based on sequential omnidirectional illumination and photometric stereo reconstruction [17] [Fig. 1(f)]. Automation with motors enabled efficient large-field imaging [Fig. 1(d), Supplement 1, Table 1], and illumination was synchronized strictly to camera capture for minimization of DUV exposure to tissue samples. Virtual H&E recoloring [18,19] can be performed to suit standard histopathology appearance for ease of clinical interpretation [Fig. 1(e)].

 

Fig. 1. Optical and computational capabilities of fiber optic microscopy with ultraviolet surface excitation (FUSE). (a) DUV illumination leveraged fiber optics for enhanced axial sectioning. (b) Images of ${\gt}1\;{\rm mm} $ thick formalin-fixed mouse liver slice show improved optical sectioning with increasing illumination angle at ${50 \times}$. (c) Multi-scale (${4 \times}$, 0.10 NA; ${10 \times}$, 0.25 NA; ${50 \times}$, 0.55 NA) imaging of ${\gt}1\;{\rm mm} $ thick fresh mouse kidney slice shows nuclei-dense renal corpuscle. (d) Stitched image depicting hand-cut cross-section of fresh rat heart, highlighting capability to quickly (${\lt}1.5 \min $) capture large (${8}\;{\rm mm} {\times} {8}\;{\rm mm}$) regions of interest with facile (${\lt}2 \min $) sample preparation. (e) Virtual H&E recoloring applied to fluorescence histological images to suit pathological workflows. (f) Switching illumination between multiple optical fibers produced 3D fluorescence microtopography of mouse liver bile duct for textural imaging usable in advanced histological techniques. (g) Conditional normalizing flow enhanced low magnification (${4 \times}$) images through learned statistical relationship between high-resolution detail and coarser structural elements.

Download Full Size | PDF

Generative visual deep learning has greatly influenced the biomedical AI landscape following the impressive performance of generative adversarial networks (GANs) [20,21] in image reconstruction and processing. The advent of diffusion [22] expanded the applicability of generative AI, now encompassing areas such as medical imaging [23] and computational microscopy [24,25]. Normalizing flows [26–29], a class of generative models less common in biomedical research, are popular in the physical sciences [30,31] for their strong statistical basis. Unlike typical generative models that learn using approximated objectives (VAEs [32], diffusion models [33], and GANs [34]) that prioritize generative variation, normalizing flows directly map complex data distributions to familiar priors, such as the Gaussian or uniform distributions. This property makes normalizing flows the ideal model class to infer model uncertainty, a frequently sought-after capability for dependable biomedical AI applications [35]. We built FUSE-Flow, a conditional normalizing flow model that learns the statistical relationship between high-resolution detail and coarser structural elements [Fig. 1(g)], bringing variational inference, computational super-resolution and thick sample microscopy together for the first time.

2. RESULTS

A. Fiber Optic DUV Illumination

Previous MUSE microscope designs had LEDs pointed directly at the sample, some with additional light guides or optics [14,36]. While this approach delivers maximal optical power to the sample in principle, there is very limited control of the oblique angle of illumination due to practical considerations. Previous studies found that illumination angle and axial resolution could be enhanced by the use of an immersion medium [13,37], but there has been little else reported on the potential for enhancing MUSE capabilities with more degrees of directional control. Our setup used multi-mode fibers coupled to LEDs and positioned at the sample with customized miniature optomechanics. The flexibility and scalability of fiber optic illumination enabled fibers to be precisely adjusted in the polar axis (angle from horizontal), and multiple fiber sources to be positioned surrounding the sample at different azimuthal angles in the horizontal plane. On top of logistical advantages provided by the compactness and arbitrary length of optic fibers such as the use of generic short-working-distance (down to ${\sim}1.5\;{\rm mm} $) objectives at ${4 \times}$ and ${10 \times}$, fiber optic illumination further enabled two transformative advances.

First, illuminating at a near-horizontal polar angle in air produced optimal axial sectioning to the extent of enabling ${50 \times}$ magnification imaging. The use of higher-magnification objectives enhances lateral/axial resolution and reduces the depth of field (DOF), but the effective axial resolution (related to both the penetration depth of the DUV light in MUSE and the objective DOF) also has a strong effect on image quality from a thick scattering sample. Ideally, the objective DOF should match or exceed the UV penetration depth to obtain sharp images [13,14,37], but this becomes more challenging at higher optical magnification or numerical aperture. Our experiments showed that oblique illumination at ${\sim}{60^ \circ}$ in air produced blurred images at ${50 \times}$ magnification. Only when the illumination angle was optimized could cellular images of sub-nuclear resolution be obtained, albeit still with limited DOF on the order of a few microns as indicated by the size of nuclei in focus. We found that the enhanced angular control afforded by the positioning of the illumination fibers enabled sharp images at moderate NA (0.25) with inverted air objectives, as well as up to 0.55 NA with further optional enhancement by focus stacking [14,37] while noting that image quality and effective DOF showed some dependence on tissue type (Supplement 1, Fig. S7). It would be extremely inconvenient, if not impossible, to achieve such an illumination angle with direct illumination from a bulky LED even with focusing. Combining angular control of fiber illumination with an immersion medium may further reduce UV penetration depth and improve image quality, but further study is needed (Supplement 1, Fig. S6). Second, optical fibers enabled sequential switching of single illumination directions to generate a set of images that could be used to produce a 3D reconstruction of the sample’s microscopic surface [Fig. 1(f); Fig. S1]. Obliquely illuminated images, while two dimensional, often showed a quasi-3D effect on uneven surfaces due to texture and shadowing [12]. We used the photometric stereo algorithm [17] to estimate depth from a single fixed field-of-view using fluorescence 2D images excited from multiple azimuthal directions, highlighting microstructural features and minute textural differences indiscernible from thin-sectioned microscopy.

B. Image Enhancement with FUSE-Flow

Accurate visualization of nuclei is crucial for numerous pathological applications, particularly in cancer assessment. Greater magnifications are ideal for such applications as low-magnification objectives often produce images with noise and poor color contrast, hindering effective nuclear characterization. However, higher-magnification objectives introduce complex technical challenges such as challenges with illumination, aberrations, and diminished DOF. Typical solutions necessitate added hardware complexity, which precludes affordability and simple usage.

FUSE-Flow, along with a preprocessing step, effectively resolved nuclear boundaries with high contrast to adjacent cytoplasm, mirroring that observed in the ${10 \times}$ references [Figs. 2(a) and 2(b)], while retaining ${4 \times}$ structural features such as the position and general shape of nuclei, macro tissue architectures, illumination uniformity [Fig. 2(c)], and consistent focal clarity [Fig. 2(d)]. Ideally, we aim to deduce the conditional standard error of predictions as a metric of uncertainty to represent the confidence we should have in the predictions. Given that the exact parametric nature of the conditional data distribution learned by the normalizing flow model is unknown, direct inference using parametric families would be inaccurate. We employed Monte Carlo simulations to calculate pixel-wise standard deviation as an empirical estimation for local standard error [Fig. 2(e)]. Regions with a five-sigma ($p \lt 3{e^{- 7}}$, commonly used in the physical sciences to denote highly unlikely events) level of uncertainty highlight the aleatoric uncertainty inherent in this ill-posed challenge—low-resolution and noisy inputs lack the comprehensive information needed for definitive output generation. FUSE-Flow performance was compared to a CycleGAN [38] and conventional image enhancement tools (Supplement 1, Fig. S11).

 

Fig. 2. Image enhancement with FUSE-Flow. (a) Performance overview on fluorescent histological images (held-out) of fresh mouse kidney slice. FUSE-Flow performed domain alignment of input ${4 \times}$ images to reference images in color and detail while preserving input’s coarser features like nuclei positioning and tissue texture. (b) FUSE-Flow enhanced nuclear margin sharpness and increased contrast between nuclei and cytoplasm. (c) ${10 \times}$ images display clear bias in upper right corner due to non-uniform illumination. Bias was absent in model-enhanced images as evidenced by intensity maps correctly corresponding to tissue features. (d) FUSE-Flow outputs show no out-of-focus areas, typically seen in higher-magnification images due to tissue regions falling outside objective depth-of-field. (e) Multiple samples ($n = 64$) drawn from learned posterior distribution could estimate conditional standard error to identify regions with highly aleatoric uncertainty. $\sigma = 5$ (or ${\rm p} {\text -} {\rm value} = 3{e^{- 7}}$) is highlighted.

Download Full Size | PDF

C. Fluorescence Histology of Thick Tissue

UV-excited fluorescence has a unique capacity to deliver nuclear-level functional insights from thick tissue samples with minimal complexity. It involves straightforward sample preparation, imposes no laser requirements, and is compatible with RGB cameras, which readily support multiplexed imaging and can be processed with standard image analysis and spectral unmixing techniques. We found DUV excitation well-suited for a wide range of fluorescence assays, further facilitating multiplexed imaging with the color camera [Fig. 3(a)].

 

Fig. 3. Fluorescence histology of thick fresh tissue. (a) Range of stains that fluoresced upon DUV excitation illustrated with images at ${10 \times}$ magnification. Nuclear (Hoechst 333422, SYTO 9, PI) and cytoplasmic stains (Rhodamine B, Eosin Y) served as analogs to H&E, revealing the inner surface of a blood vessel from fresh mouse kidney (Hoechst, Rhodamine) and surface of fresh rat liver (Hoechst, Eosin). LIVE/DEAD staining provided viability readouts of fresh renal tissue (SYTO9, PI). Immunofluorescence staining provided organ-level protein expression in fixed mouse uterus (Fibronectin, Alexa Fluor 488; Laminin, Alexa Fluor 594, AF: autofluorescence) (b) Multi-scale imaging of Hoechst and Rhodamine stained fresh murine organs with H&E comparisons. Developing secondary follicle of the ovary was staged by the diameter of the follicle. Stratum functionalis (white arrowheads and dashed line) was useful for staging of estrus cycle.

Download Full Size | PDF

Prior studies had shown close correspondence between fluorescence histology and H&E histology [3,12]. We broadly showcased this capability on our platform across the renal, digestive, and female reproductive systems [Fig. 3(b)]. Fresh samples, which are important in clinical applications and biological viability studies, were our primary focus although we noted that fixed samples could achieve clearer preservation of cellular details and an even closer correspondence to histology. Tissue preparation protocols for DUV fluorescence histology did not appear to affect any downstream H&E processing, and the fluorescence images showed excellent nuclear visualization with a high level of structural correspondence to the matching H&E images.

Large field-of-view coverage at cellular resolution enabled quantitative measurements at the organ scale that could correlate with physiology or phenotype, such as the size of a follicle indicating maturation and endometrial thickness as a marker of estrous stage. Quantitative FUSE analysis enabled the measurement of vascular structure dimensions and functional assessment of the uterus. Oblique illumination provided a shadow effect and a quasi-3D appearance, a unique perspective of the organ revealing the inner surface of the endometrium. The layer of simple columnar epithelial cells lining the endometrium corresponded well to histology. In rat kidneys, renal corpuscles, including the Bowman’s capsule and glomerular vasculature, could be distinguished, potentially revealing insights into kidney physiology and injury. High-quality en face views of the inner stomach surface showed gastric pits. While the visualization of intact organs is limited to the surface, alternate imaging planes may be developed either by manual cutting or microtome. The cross-sectional structure of the gastric mucosa was studied via a “Swiss roll” preparation of tissue cut, rolled up and mounted on its side.

D. Large-Field Viability Analysis for Renal Toxicity from Antibiotic Insult

Vancomycin remains a standard treatment of methicillin-resistant Staphylococcus aureus (MRSA) [39] despite being linked to nephrotoxicity and acute kidney injury for decades [40,41]. It is known to affect the proximal convoluted tubules of the kidney nephron due to the drug’s mechanism of action. We hypothesized that cell viability measurements at the organ scale could be a surrogate quantitative metric for toxic insults to renal function.

The SYTO 9 and propidium iodide (PI) viability assay based on membrane permeability visualized the effects of increasing vancomycin dosage, including standard treatment dosage of 5 mg/mL [42,43] and PBS control, at ${4 \times}$ magnification. FUSE imaging displayed diverse structures, revealing the kidney’s innate anatomical variability [Figs. 4(a) and 4(b)]. Functional components like the renal calyx, corpuscles, and tubules were identifiable due to their unique nuclear density and response to vancomycin. Pixel-wise comparison of the green and red color channels respectively segmented live and dead cells, independent of brightness variation biases [Fig. 4(c)]. Significant differences ($p \lt 0.0001$) in viability between treatments and control groups in all regions [Fig. 4(d)] highlight clear effects of vancomycin-induced toxicity. The medulla and cortex showed similar viability patterns, suggesting comparable effects of vancomycin. In contrast, the corpuscle and tubule displayed unique viability profiles. Treated corpuscles exhibited drastically reduced viability while even untreated corpuscles demonstrated varied viabilities, suggesting an inherent propensity towards cell death. We observed a marked difference ($p \lt 0.05$) in viability between corpuscles and the surrounding cortex tissue across treatments [Fig. 4(e)], which might be due to prolonged warm ischemia of the kidney tissue after harvest, where histone localization in glomerular cells may have contributed to the lower corpuscle viability compared to the cortex [44]. This contrast was absent in tubules, despite known vancomycin-induced nephrotoxicity, and may be observable only from in vivo clearance. Although the clinical exposure of vancomycin to kidneys occurs during drug clearance in renal excretion and differs from that of our ex vivo kidney treatment, investigating the effects of antibiotic treatment on specific functional regions within tissue could still offer meaningful insight into the drug’s effects, while also relevant to more sophisticated in vivo tissue models.

 

Fig. 4. Ex vivo vancomycin antibiotic treatment on kidney tissue viability. (a) Kidney slices from treated and control groups show effects of vancomycin on various organ regions at ${4 \times}$. (b) Structure and stain intensity of renal calyx stand out from surrounding inner medulla. Renal corpuscles (marked by $^\ast$) were more susceptible to vancomycin toxicity. Decreased viability extended to proximal convoluted tubule (marked by †) of kidney nephron. (c) Multiple segmentation strategies were employed to isolate specific regions of interest for downstream comparative assessment. Cells were deemed as live or dead depending on dominant image color channel. (d) Decline in viability highlighting toxicity of vancomycin. Distributional comparisons of regional viability estimates show statistically significant ($p \lt 0.0001$) differences between treatment and control groups regardless of region. (e) Higher corpuscle viability compared to surrounding cortical tissue across treatment groups, indicating consistent corpuscle susceptibility to external effects. Error bars denote confidence interval based on t-statistic calculated at 95% confidence level.

Download Full Size | PDF

E. Differential Uterine Protein Expression Quantification with Immunofluorescence

The human female reproductive system displays reduced fertility and hormonal dysfunctions [45] with aging. While the mechanisms remain elusive, research indicates a correlation between extracellular matrix (ECM) protein composition and aging [46,47]. Fibronectin is a multi-functional protein that is important for ECM assembly, embryogenesis, cell adhesion, and cell growth [48,49]. Understanding the changes in protein composition that occur with aging could provide insights into reproductive health [50] and potentially help prevent the onset of diseases.

Immunofluorescence quantification with FUSE at ${10 \times}$ magnification assessed differences in uterine fibronectin expression between mice at the onset (6 weeks) and end (32–36 weeks) of their reproductive life span (Fig. S4b). Expression level assessed antibody binding concentration, which indicates protein abundance, whereas expression coverage captures protein localization. Significant differences ($p \lt 0.05$) in both expression levels and coverage across all regions indicated a decline in fibronectin expression with age [Figs. 5(a) and 5(b)], suggesting a potential impact on uterine functionality. The higher protein expression in the myometrium compared to the endometrium for both age groups ($p \lt 0.05$) may be linked to the need for myometrial fibronectin for uterine stretch during pregnancy. Our study did not consider the estrous cycle of the mice, which may lead to different expression patterns [51]. The quantification of protein expression may be sensitive to biases in illumination or other optical artifacts. Other protein analysis methods (e.g., western blot or mass spectrometry) may be required to substantiate any claims in specific regional expression differences.

 

Fig. 5. Senescence on uterine extracellular matrix fibronectin expression and hepatic polyploidization in young (6 weeks) and aged (32–36 weeks) mice. (a) Increased fluorescence in young mouse uterus slices indicates higher fibronectin expression, correlating to well-understood changes in uterine architecture and reproductive health. (b) Image-based statistical assessments ($n = 25$) reveal significant ($p \lt 0.05$) variations in fibronectin expression between the two groups. Expression level estimated fibronectin concentration from the signal intensity and proportion of tissue coverage reflected its distribution and localization in tissue. No differential expression was observed between endometrium and myometrium, indicating a consistent decline in fibronectin with uterine aging. (c) Multimodal distribution of hepatocyte nuclear area with peaks corresponding to occurrence of hepatic polyploidization (diploids, 2 n; tetraploids, 4 n). Rightward translation in distribution of aged population indicating a general increase in DNA contents of nuclei regardless of polyploidization. Right skew of aged group suggests increased polyploidization occurrence, with increased presence of higher-order polyploids in place of diploids. (d) Fluorescence imaging (${50 \times}$, 0.55 NA, air immersion objective) of diploid and polyploid hepatocyte nuclei alongside higher-powered brightfield H&E images (${60 \times}$, 1.40 NA, oil immersion objective). Similarly sized hepatocyte nuclei (not exact correspondence to FUSE images) from same tissue sample reveal close correspondence in sub-nuclear structures.

Download Full Size | PDF

F. Sub-nuclear Characterization of Hepatic Polyploidization

Polyploidization (duplication of chromosomes) is a unique feature of mature hepatocytes. It increases the DNA content (diploid, 2 n; tetraploid, 4 n; octoploid, 8 n; and so on) and has been linked to cellular senescence and stress (such as oxidative injury, surgical trauma, metabolic overload, and exposure to toxic insults). Open questions include the significance of hepatocyte polyploidization phases observed in normal development and aging, the role of its alteration in various pathological conditions, and the causal effect between rejuvenation of senescent hepatocytes and polyploidy reversal [52]. The aged liver may have higher numbers of senescent cells due to decreased regenerative capacity, and may exhibit higher polyploidization levels in these mature hepatocytes. Obtaining nuclear-level information (${50 \times}$) with FUSE could help with understanding this process and its significance during postnatal development and with senescence.

Nuclear size distributions of hepatocytes sampled from liver slices from young (6 weeks) and aged (32–36 weeks) mice revealed three distinct trends [Fig. 5(c)]. First, the observed multimodal distributions aligned with recognized hepatic polyploidy phases [53]. Second, a right-shifted distribution of the aged population indicated a general increase in the DNA content of nuclei in aged mice. This is consistent with prior studies that found escalating hepatocyte DNA content with age, even after accounting for polyploidization [54]. Finally, the aged group displayed a notable right skew, with its 2 n frequency peak matching the 4 n peak. In contrast, the young group’s 2 n frequency peak is approximately twice the height of its 4 n peak, implying increased polyploidization and proliferation of higher-order polyploids in the aged group. The hepatocyte cell cycle phase could be inferred from the sub-nuclear DNA content, as shown with H&E comparisons [Fig. 5(d)]. Currently, obtaining such sub-nuclear insights requires thinly sectioned liver, single-cell cultures, fine needle aspiration, or complex and costly systems to optically section and provide sufficiently high axial resolution. While our findings are well supported by prior work, it should be noted that we did not account for any specific liver lobule and location from which the samples were harvested. Hepatocytes located at the origin (portal tract) may be younger than cells near the hepatic vein [54], and this relationship could affect the DNA content and nuclear area distributions within this study.

3. DISCUSSION

DUV-excited fluorescence microscopy, also known as MUSE, has been celebrated for its simplicity and effectiveness in concept—the combination of oblique DUV illumination and widefield detection. However, prior published illustrations of real-world implementations [14,15,36,37,55,56] showed inevitably large and complex setups, with the notable exception of Pocket MUSE, which is generally incompatible with large surgical or mesoscale samples [16]. Also notable in the MUSE literature was the challenge of imaging at high optical magnifications or numerical aperture (${\gt}40{ \times}$, ${\gt}0.3 \;{\rm NA}$), a mainstay of biomedical inspection and unusually difficult in MUSE due to a mismatch of objective DOF and the relatively larger (tens of microns) penetration depth of DUV illumination [13,14,37], as well as the limited working distance of conventional high-magnification objectives. Our FUSE concept of fiber optic-delivered DUV illumination was not only space-efficient but also enabled several technical innovations that significantly expanded the capabilities of MUSE. Light delivery via a set of small diameter and easily micro-positioned fiber tips greatly relaxed the logistics of oblique illumination, to the extent of enabling the economical use of no-brand, short working distance objective lenses normally for standard widefield inspection. Fiber illumination at near-horizontal angles in air (${\sim}{46^ \circ}$ in tissue) was also found to greatly enhance DUV optical sectioning for high-power microscopy at ${50 \times}$ 0.55 NA using an air objective in an inverted configuration suited for thick tissue imaging, with further DOF enhancement possible by focus stacking (Supplement 1, Fig. 7).

Researchers have appreciated the phenomenon of meaningful shadowing produced by oblique DUV illumination, which generated a pseudo-3D effect from coarsely cut samples and organ surfaces [12]. While multi-directional illumination would already improve uniformity, we further exploited directional control to produce 3D surface micro-topography using photometric stereo [17]. Fluorescence is typically assumed to be isotropic [57] and has been previously studied as a feasible modality for photometric stereo reconstruction [58]. Although we did not exploit micro-topography for a comprehensive study beyond technical validation, we believe the technique could have applications in rapidly evaluating the morphology of 3D models as they gradually form microstructure or texture, which may reflect cellular heterogeneity and could have future relevance as a malignancy marker in histopathological evaluation [59], or even highlighting paths of low resistance for cell migration [60].

Generative AI has electrified the microscopy field, with extremely high-fidelity image reconstructions for resolution enhancement or modality transfer enabled by adversarial models [61]. However, there are several issues inherent to the adversarial training strategy of classical GANs [62], such as the lack of a defined convergence state, vanishing gradients [63], and mode collapse [34]. These issues have tempered the enthusiasm for GANs to be more broadly used in biomedical microscopy, where most granular image observations may be used to prove a hypothesis or make a diagnosis. The acclaim for visual generative AI could be attributed in part to the rise of variational models, which include diffusion, variational autoencoders, and flow. We chose flow because it is a classic technique underused in bioimaging for the direct statistical modeling of the data distribution, and not necessarily preferred over diffusion, which learns a denoising process and has similar properties. In demonstrating variational image generation in use cases where mode collapse or hallucination would be catastrophic, we argue that this class of models is critical for principled yet high-quality generative imaging in biomedicine and microscopy. FUSE-Flow reconstruction of ${10 \times}$ data from ${4 \times}$ magnification has great practical value since ${10 \times}$ is the predominant setting used in MUSE, but it also provides a codebase and scalable framework for more ambitious efforts in computational super-resolution.

Although multiple studies have already shown traditional MUSE to generate exquisite images from unsectioned tissue potentially for intra-operative histopathology, our work also featured a selection of murine tissue examples loosely themed around relatively large organs. This was partly to demonstrate our unique optical design to be competitive with the state of the art, and also to present a few new perspectives based on large-field mapping to motivate the serious consideration of FUSE as a tool for quantitative tissue biology. While we acknowledge that fixation is ideal for visualization, we presented several fresh and unfixed biological preparations to envision a translational step towards DUV excitation for live samples and in vivo models, both important use cases in biological microscopy. Despite the well-known mutagenic effects of DUV light, the very brief sub-second exposures needed for high-quality imaging may be tolerable, and motivate dosimetry studies. Given concurrent advances in DUV light sources, fiber optics, and microscopy, the potential of minimally invasive ultraviolet endoscopy [64] for real-time cellular interrogation of living organs could be due for a revisit.

The centerpiece of our multi-disciplinary effort is a curated set of hypothesis-driven biological investigations that are modestly scoped to corroborate known science, namely, antibiotic toxicity, ECM expression, and polyploidy while demonstrating a novel quantitative approach using thick-sliced tissue sampled from a relatively large set of intact organ samples. While we emphasize the ability to infer histological and biological insights in a large region of interest without the need for thin sectioning, being constrained to surface-level information can be restricting for 3D biological samples; deeper information beyond surface texture is largely unobtainable. The time-sensitive nature of tissue scale viability precluded the ability to perform orthogonal validation with standard photometric assays. FUSE presents a new opportunity for biologists to rapidly interrogate coarsely sliced tissue at the cellular level from numerous candidates that would normally be logistically overwhelming to work through with standard confocal microscopy or cleared light-sheet microscopy. This high-throughput approach could work as a screening modality much like a tissue microarray at “mesoscale,” for which the most promising candidates can be selected for more thorough standard investigation. As the science of tissue omics accelerates and the lines between clinical pathology and tissue biology continue to blur, we humbly anticipate FUSE and FUSE-Flow to represent important inroads made towards a fast platform technology for multiplexed, high-throughput tissue insights.

4. METHODS

A. Microscope Design and Components

The deep ultraviolet (DUV) fiber optic fluorescence microscope platform was based on an epi-illuminated inverted microscope design. The microscope sat on a $30 \times 30\,{\rm cm}$ optical breadboard, comprising the illumination, optical, sample holder, and mechanical and electrical components (Fig. S2a).

Illumination components. DUV illumination at a peak wavelength of 277 nm was produced by DUV light-emitting diodes (LEDs) (ILR-OV01-O275-LS030-SC201.; Osram) powered by a constant 700 mA current supply (RECOM) and cooled with a heat sink. Magnetic butt-coupling of multi-mode optical fibers (57-070; Edmund Optics), with an outer diameter of 710 µm, to the LEDs was achieved with a custom 3D-printed magnetic fiber coupler (Fig. S2b). The magnetic coupler allowed hot swapping of the illumination source and the optical fibers. It takes approximately 2 min to switch between objectives. Each objective is paired with fibers at preset angles and is swapped as a set. The butt-coupled optical fibers had a measured output power of around 6.29 mW and a coupling efficiency of roughly 5%. The DUV-emitting end of the fiber optic was controlled via a 3D-printed custom optomechanical fiber holder jig that provided four degrees of freedom, allowing fine control and angle adjustments of the fiber for uniform sample illumination. An extreme oblique incident angle of illumination (${\sim}{87^ \circ}$) was used in the ${50 \times}$ objective setup and (${\sim}{85^ \circ}$) in the ${10 \times}$ objective setup. The incident angle of illumination was (${\sim}{60^ \circ}$) for the ${4 \times}$ objective setup (Fig. S2c). With the ${4 \times}$, 0.1 NA objective, sharp images could be obtained without requiring the fiber optic to be positioned at a near-horizontal angle. The smaller incident angle (${\sim}{60^ \circ}$) of illumination was preferred for ${4 \times}$ as it provided stronger illumination while still maintaining sharp images. The areas of illumination for the ${10 \times}$ and the ${4 \times}$ setups were measured to be approximately $4.88\;{{\rm mm}^2}$ and $8.58\;{{\rm mm}^2}$, respectively. The intensities of the ${10 \times}$ and the ${4 \times}$ setups were calculated to be approximately $1.29\;{\rm mW}/{{\rm mm}^2}$ and $0.733\;{\rm mW}/{{\rm mm}^2}$, respectively. The area of illumination and intensity of the ${50 \times}$ setup was estimated to be similar to the ${10 \times}$ setup.

Optical components. Digital images were taken using a color camera (BFS-U3-120S4C-CS; Teledyne FLIR) magnified by either ${50 \times}$ finite conjugate telecentric objective (375-052; Mitutoyo), or ${10 \times}$ or ${4 \times}$ finite conjugate DIN objectives (43-907, 67-706; Edmund Optics). The working distances for the ${50 \times}$; 0.55 NA, ${10 \times}$; 0.25 NA, and ${4 \times}$; 0.10 NA objectives were 13.00 mm, 1.50 mm, and 15.25 mm, respectively. A custom 3D-printed objective holder was non-mechanically coupled to a tube assembly (54-868; Edmund Optics), enabling hot swapping between the ${50 \times}$, ${10 \times}$, and ${4 \times}$ objective setups (Fig. S2c). The objective holder included a filter slot that enabled suppression of specific wavelengths, which could minimize unwanted signals, including autofluorescence from samples. The tube assembly connected the objective holder with the color camera. The effective resolution was measured by conducting the knife-edge resolution test across the sharp edge of a blade. The 10%–90% intensity transition, which corresponds to the Rayleigh resolution, was measured at nine different points and averaged to obtain the half-pitch spatial resolution. The half-pitch spatial resolutions were measured to be 3.60 µm, 1.11 µm, and 0.311 µm for the ${4 \times}$, ${10 \times}$, and ${50 \times}$ objective setups, respectively. The effective field-of-view was determined to be $0.748 \times 0.561\;{\rm mm}$ (${10 \times}$) and $1.85 \times 1.39\;{\rm mm}$ (${4 \times}$) by imaging a checkerboard calibration target. For ${50 \times}$, the effective field-of-view was calculated to be $0.150 \times 0.112\;{\rm mm}$.

Sample holder, and mechanical and electrical components. Histology cassettes were modified to include a 150 µm-thick UV-transmissible quartz window to contain tissue. A custom-machined aluminum sample holder supported by four optical posts was used to mount the histology cassettes. Standard histology cassettes (C33118001PH, Uni-Sci) were modified by cutting a hole approximately $11 \times 20\;{\rm mm}$ before applying liquid epoxy to adhere a quartz coverslip (JGS2, Latech) to the outer surface covering the hole. The quartz coverslips acted as a flat and UV-transmissible surface for samples to be placed on for imaging. X-axis and Y-axis movement, at sub-micrometer precision, of the sample relative to the camera was enabled by mounting the sample holder on a stack of two manual linear translation stages. Z-axis movement was enabled by a manual vertical translation stage. Two 3D-printed stepper motor holders and stepper motor couplers were designed to co-axially couple the stepper motors to the micrometer knob of the stage. A microcontroller controlled the stepper motor by either parsing a set of sequential instructions allowing for automated stage movement or inputs from an analog stick for manual stage movement of the microscope in the X and Y directions. The flashing of the LEDs was controlled by the microcontroller during image acquisition, synchronized to the camera’s exposure time, to reduce unwanted DUV exposure to samples during stage movement of the automated imaging process.

B. FUSE-Flow Development

Architecture. Our conditional normalizing flow was developed with inspiration from existing open-source implementations (Fig. S3a). SRFlow [29], a recent demonstration of normalizing flow for single-image super-resolution, formed the foundational structure for our model. We adopted advanced techniques from Flow++ [28], such as variational dequantization and the application of gated residual networks [65,66], to enhance the model’s performance. Moreover, our design includes a custom-made adaptive U-Net tailored for arbitrary input sizes, ensuring the extraction of features with appropriate dimensions to feed the primary normalizing flow. We also assessed the integration self-attention mechanisms, squeeze-and-excitation [67] and convolutional block attention module [68], tailored for convolutional neural networks. Ablation studies, performed on the CelebA [69] dataset, are detailed in the Supplement 1 (Figs. S9, S10).

Training strategy. The training of the model for our unique fresh tissue preparations and general-purpose application presented a multi-faceted challenge. This complexity prevented the use of common training strategies for image super-resolution such as supervision [70] (using perfectly registered high- and low-magnification images as targets and inputs, respectively) and self-supervision [71] (where the inputs are modified low-resolution versions of the high-resolution targets). Obtaining perfectly registered targets and inputs was challenging due to the inherent dynamic nature of fresh tissue, largely credited to the presence of moisture that can induce real-time fluidic shifts, altering tissue morphology and positions. Conversely, non-supervised augmentation of the target to match the input domain is not viable, given the difficulty in precisely emulating the photometric distortions specific to each optical and experimental setup. These variations arise from factors such as differing chromatic characteristics of magnification objectives, fluorophore photobleaching, sample degradation, human error in preparation, and fluctuations in environmental conditions. We developed a unique semi-supervised strategy for the training of FUSE-Flow tailored to our distinctive fresh tissue imaging application. This approach was assessed on a common benchmark image dataset CelebA [69] (Fig. S9) as well as fluorescence histological images of sliced fresh mouse kidneys and acquired using FUSE. The model was trained using ${10 \times}$ magnification images as targets, while the inputs were augmented versions of these targets, with supervised adjustment to match the style (color and resolution) of the corresponding loosely registered ${4 \times}$ magnification images. Subsequently, we evaluated it on a held-out (not part of the training dataset) kidney slice imaged at ${4 \times}$ magnification, with corresponding ${10 \times}$ images as references for comparison.

Data. Fresh kidneys from a sacrificed mouse were immediately prepared and imaged post-harvesting. One kidney was sliced with a vibratome longitudinally and the other transversely, yielding 30 slices in total. From the extensive set of over 4,000 images from large-field imaging on both ${4 \times}$ and ${10 \times}$ magnification, a small subset of 36 pairs of high-quality images (in focus and lacking imaging artifacts) that displayed important structural features was selected. A comprehensive data processing pipeline (Fig. S3b) generated input and target patches from the 36 image pairs for model training. Fast non-local means denoising [72] despeckled the original images, enabling the model to concentrate on structural variations without noise-induced biases. The high-resolution images from the ${10 \times}$ set were simulated as low resolution through imputation downsampling and blurring. Histogram matching aligned the color profile of these simulated images with the corresponding ${4 \times}$ images. The final processing step involved splitting the images into $96 \times 96\;{\rm px}$ and $24 \times 24\;{\rm px}$ patches (target and input, respectively) for training, with selective minor blurring to mitigate any artifacts introduced from the previous step. The training utilized a set of 120,000 patch pairs.

Training. The model was trained on an NVIDIA RTX A6000 GPU (48 GB memory) over 48 h, spanning 32 epochs with a batch size of 4. We set the learning rate at $1.0{e^{- 4}}$, decaying by a multiple of 0.9 each epoch. The posterior was mapped to a standard normal distribution (Gaussian with a standard deviation of 1.0). The adaptive U-Net underwent pre-training with binary cross-entropy as the loss function, and once trained, it was frozen during the normalizing flow training. During training, horizontal and vertical flips were applied with a probability of $p = 0.5$.

Prediction. FUSE-Flow took 13.5 min to generate a single image of size ${0.555 \times 0.74}\;{\rm mm}$ (${3000 \times 4000}\;{\rm pixels}$, 10,980 patches) with a batch size of 512. Multiple samples of each patch can be drawn to improve uncertainty estimation accuracy and reduce patch artifacts. This would multiply memory consumption, reducing batch size, and consequently increase the time taken for prediction. Generating four samples of variational output dropped batch size to 128 and required 54 min.

C. Image Acquisition and Large-Field Mosaicking

Imaging settings and processing. Color images were captured with exposure ranging from 0.1 to 3 s. The images were saved in PNG file format. Image processing was done with the open-source ImageJ (Fiji) [73] and limited to brightness and contrast, color balance, and despeckling to remove hot pixel artifacts from the color camera. No background correction, subtraction, or deconvolution was performed on the images.

Automated image acquisition. Image acquisition was facilitated by an Arduino, Python code on a laptop attached to the microscope, and the Teledyne FLIR Spinnaker SDK. Manual image acquisition was done using an analog stick with the Arduino for motor control and the SpinView application from the Spinnaker SDK for the viewing and saving of images. Automated image acquisition for large mosaic images was coordinated by our custom Python software. It received and processed user input, then controlled the camera through the Python interface from the Spinnaker SDK and the stage by sending condensed instructions to the Arduino.

Large-field mosaicking. Large-field mosaicking was enabled by the Python software integrated with the microscope and the freely available Microsoft Image Composite Editor. Images were automatically and sequentially captured with an adjustable overlap (usually set at 10%/15% margins) and stitched together using Image Composite Editor to achieve gigapixel-sized mosaic images comprising hundreds of individual images. The sample stage was precise enough for images to be manually focused prior to imaging, with the Python software subsequently automatically controlling the stepper motors’ movement, flashing of the DUV LEDs, and the camera capture of images, without additional refocusing required, over the entire sample area.

D. 3D Microtopography with Photometric Stereo

Sequential illumination and image acquisition. To obtain the micro-topographical information over a region of interest on a sample, each of the four fiber optics was coupled to an independently controlled DUV LED. A reference image of the region of interest was then captured with illumination from all the fiber optics. Four sequential images with directional illumination from each fiber optics were obtained while maintaining the same field-of-view over the region of interest. The exposure of these four sequential images was increased (up to 3 s exposure) compared with the reference image to compensate for the dimmer illumination coming from a single fiber optic.

Image processing. Despeckling was performed on all images using ImageJ (Fiji) to remove hot pixel artifacts as the bright pixels, which would affect the subsequently obtained surface normals map. Following this, the four sequential images were histogram-matched with the reference image to obtain the different direction-of-illumination images with the same relative brightness. Histogram matching was done to force a more uniform brightness between the four images given the slight variations in illumination intensity in the four optical fibers.

Photometric stereo for surface normals and depth map. With the four histogram-matched different direction-of-illumination images, four representative vectors for each direction of fiber optic illumination were then inputted into our implementation of the photometric stereo algorithm (vectorized and ported to Python, based on an open-source MATLAB implementation [74]). The four representative vectors reflected the 90° rotation about the image’s center to match our setup’s fiber orientation. The photometric stereo algorithm produced the surface normals and depth map of the region of interest (Fig. S1). It was noted that changing the input vector angle of illumination (obliqueness of illumination angle) did not significantly affect the obtained depth map as the depth map displayed the normalized depth rather than true depth. The depth map was rendered in 3D, with either the H&E recolored image or the despeckled reference image used as the surface texture in PyVista, to produce an interactive 3D micro-topographical reconstruction of the sample region of interest.

E. Histological Studies

For tissue samples, female NCr nude mice were euthanized at the humane or experimental endpoint of a separate study in accordance with the approved A*STAR Institutional Animal Care and Use Committee (IACUC) protocol. Mice were dissected and routine post-study tissue collection was performed. Alternatively, whole mice and rat carcasses or specific harvested organs were purchased (InVivos). Organs were harvested and washed in ${1 \times}$ phosphate buffered saline (PBS) (BUF-2040-10X4L, 1st BASE), and sliced either using a vibratome (VF500-0Z, Precisionary Instruments) or manually using a scalpel depending on the use case. For vibratome slicing, the organ was embedded in 2% Type I-B agarose (Sigma Aldrich). In brief, the sample was glued onto the plunger of the specimen tube, and liquid agarose was pipetted in excess to cover the tissue before the agarose was solidified using a chilling block. The specimen tube containing the sample was loaded onto the vibratome and sliced into 400 µm to 1 mm-thick slices at various oscillation and advance speeds depending on the organ type. The surrounding agarose was delicately removed using forceps before staining. Hoechst 33342 (62249, Thermofisher Scientific) for DNA staining and Rhodamine B (A5102, TCI) counterstain were prepared at various concentrations for optimal staining of different tissue types. Tissue samples were stained for 30 s to 1 min, before being washed in ${1 \times}$ PBS twice, for 30 s each. Samples were blotted dry with Kimwipes and placed on modified histology cassettes for imaging (Fig. S4a). 0.1 mg/mL Hoechst with 2 mg/mL Rhodamine B was identified as an effective fluorescent concoction for general staining over a wide range of fresh and fixed murine tissue including organs like the heart, liver, stomach, and kidney, as well as the female reproductive tract. 0.1 mg/mL Hoechst with 4 mg/mL Rhodamine B was also used for higher contrast of structural features such as the tubules of fresh kidneys. Organs that featured a high level of nuclear concentration such as the ovaries and gastric surface of the stomach were better highlighted with a lower level of nuclear stain ratio of 0.1 mg/mL Hoechst with 1 mg/mL Rhodamine B and 0.05 mg/mL Hoechst with 2 mg/mL Rhodamine B, respectively. Fixed mouse livers in Fig. 1(b) were sliced with a vibratome into 1 mm slices. Fresh mouse kidney, fresh rat liver, fresh mouse kidney, and fixed mouse uterus in Fig. 3(a) were sliced with a vibratome. Fresh mouse ovary and fresh mouse stomach (gastric mucosa and gastric pits) in Fig. 3(b) were hand-cut. Fresh mouse uterus and fresh mouse kidney in Fig. 3(b) were sliced with a vibratome.

F. Renal Viability Studies

Procedure. We harvested 12 kidneys from six mice and classified them into three treatment groups: control, low dose of vancomycin (5 mg/mL), and high dose of vancomycin (15 mg/mL). The fresh kidneys were sliced coronally with a vibratome into 400 µm slices to obtain five mid-section slices per kidney, totaling 60 slices. Samples in the treatment group were treated with vancomycin (vancomycin hydrochloride, 1709007, US Pharmacopeia) for 1 h. After treatment, kidney slices were washed in PBS before incubating in 5:1 SYTO 9 and PI nucleic acid stain (7.5 µL for 15 min with shaking) (LIVE/DEAD BacLight Viability Kit, L7012, Invitrogen). Samples were blotted dry with Kimwipes and placed on modified histology cassettes for imaging (Fig. S4b).

Viability quantification. Our general strategy for image-based viability quantification involves two stages: segmentation followed by pixel-wise relative green and red comparison [Fig. 4(c)]. Specific segmentation strategies differ for each use case. Color comparisons begin with first splitting the color image into its red, blue, and green (RGB) channels. A map of pixels where the green channel has a higher intensity than the red channel representing the live regions of the tissue is then created. This is a convenient strategy as the viability assay of choice stains in green (SYTO 9) and red (PI) at similar intensities. Since PI is only permeable to dead cells and has a stronger affinity for nucleic acids [75], the point where the red signal is greater than the green represents a significant uptake of PI and thus differentiates dead cells from live ones. This strategy is also robust to any local and global variation in brightness that tends to happen during the imaging process.

Single cells were segmented using Otsu’s method of thresholding before pixel-wise relative color comparison. Kidneys were segmented from their background using constant thresholding. Kidney regions of interest like the cortex, medulla, corpuscles, and proximal convoluted tubules were cropped using various strategies. The cortex and medullas were manually cropped using ImageJ (Fiji) for 12 kidney slices per group with visible cortical and medullary regions. All discernible longitudinal proximal convoluted tubules were manually cropped from each kidney sample due to their non-uniform shape. Corpuscles were detected using template matching and cropped to a fixed size, generating over 4,777 crops. The Welch Two Sample t-test was used to determine the statistically significant differences between treatment groups [Figs. 4(d) and 4(e)]. Statistical tests and plots were done using R programming language and Tidyverse [76].

G. Uterine Senescence Studies

Procedure. We harvested five uteruses each from young (6 weeks) and aged (32–36 weeks) mice. Mouse uteruses were sliced with a vibratome transversely into 400 µm slices. Mouse uterus cross-section tissue was fixed for 30 min in 4% formaldehyde (VWR) before blocking for 30 min in 10% Normal Goat Serum (50062Z, Life Technologies). Thereafter, primary antibody 1/50 Anti-Fibronectin (ab2413, Abcam) was added to the uterus slices and incubated for 16 h. The slices were washed twice in ${1 \times}$ PBS before the secondary antibody Alexa Fluor 488 (A11034, Life Technologies) was applied and left to incubate for 5 h. After incubation, samples were washed twice in ${1 \times}$ PBS prior to imaging. A total of 25 young uterus and 25 aged uterus slices were imaged at ${10 \times}$ magnification (Fig. S4c).

Quantification and analysis. Large-field mosaicking of the uterus images was enabled by Python code in post-acquisition processing. Gigapixel-sized mosaicked images were obtained. The image background and bright tissue borders in the perimetrium were removed by manual cropping using ImageJ (Fiji). Uterus images were further cropped to obtain endometrium and myometrium areas for downstream protein expression quantification. Threshold value ($\alpha$ value) was ascertained for each individual tissue. $\alpha$ value was manually determined through analysis of pixels with prominent green intensity in its fluorescence signal. The selected $\alpha$ value returns a RGB image with adjusted blue intensity, and subsequently the blue ${\rm B} \gt {\rm G}$ and ${\rm G} \gt {\rm B}$ masks. The green mask was verified to be representative of the signal region. $\alpha$ values ranging from (0.3 to 0.7) were applied to select for the green signal with Alexa Fluor 488 tagged secondary antibodies binding, correlating to fibronectin protein expression in uterine tissue. Expression level was determined by the mean average green intensity of the image, with intensity values between 0–255. Expression coverage represented the expression area ratio of the green intensity over the tissue area. The Welch Two Sample t-test was used to determine the statistically significant differences between treatment groups [Fig. 5(c)]. Statistical tests and plots were done using R programming language.

H. Hepatic Senescence Studies

Procedure. For the young group, livers were harvested from three 6-week-old mice, and for the aged group, livers were harvested from three 32–36-week-old mice. Livers were fixed overnight in 4% formaldehyde (VWR) before being transferred to ${1 \times}$ PBS. The samples were sliced 1 mm thick with a vibratome. The fixed liver slices were stained for 6 min in Hoechst 33342 (0.4 mg/mL) and Rhodamine B (2 mg/mL) mix. The stained liver slices were placed on the modified histology cassettes and gently compressed with biopsy pads. After imaging, the stained liver slices were processed through the usual H&E workflow and imaged on a brightfield microscope (Nikon ECLIPSE Ni-E, DS-Ri 2) with a high-powered PlanApo ${60 \times}$, 1.40 NA oil immersion objective.

Image analysis. 10 images were obtained from each liver using the ${50 \times}$ objective. From the 60 images (30 images from young mice, 30 from aged mice), hepatocyte nuclei were manually cropped using the elliptical tool in ImageJ (Fiji) [73]. 421 and 207 hepatocyte nuclei were cropped from the images of the young and aged mice livers, respectively. The area of the hepatocyte nuclei was measured in ImageJ (Fiji). The hepatocyte nuclei area was plotted against the normalized frequency to obtain the frequency distribution of hepatocyte nucleus sizes for both the young and aged groups of mice.