Deep UV Fluorescence Scanning Microscopy

• DUV-FSM is a microscopy technique that uses deep-UV (250–300 nm) excitation to achieve surface-selective, sub-micron imaging of biological tissues.
• It supports rapid block-face and direct surface scanning methods, enabling 3D histology and intraoperative cancer margin assessment.
• Integration with deep learning pipelines (e.g., ResNet and ViT) facilitates automated feature extraction and high-accuracy tissue classification.

Deep Ultraviolet Fluorescence Scanning Microscopy (DUV-FSM), also referred to as “MUSE Milling” or MUVE in the literature, is a surface-selective microscopy technique that utilizes deep-ultraviolet (UV) excitation (typically 250–300 nm) to generate visible fluorescence from endogenous or exogenous fluorophores in biological tissue. Its unique combination of deep-UV excitation, rapid block-face imaging, and compatibility with paraffin-embedded and fresh tissue workflows enables high-resolution, three-dimensional imaging at cellular and sub-cellular scales. DUV-FSM technology supports both traditional 3D histology via serial sectioning and contemporary data-driven approaches for intraoperative cancer margin assessment, with demonstrated efficacy in whole-slide image (WSI) acquisition, feature extraction, and deep learning-based tissue classification.

1. Optical Configuration and Hardware Architecture

The foundational principle of DUV-FSM is shallow penetration of deep-UV photons into tissue, confining excitation to a surface layer typically less than 10 μm thick and inducing strong visible fluorescence from biomolecules such as NADH, FAD, and collagen (Guo et al., 2019). Endogenous and exogenous labeling protocols can be used depending on the application. In block-face implementations, a FireEdge FE200 UV‐LED (centered at λ = 280 nm, with up to 300 mW output) and custom quartz optics focus the UV beam to a spot of ≈1 mm diameter. Standard objectives (Olympus Plan Fluorite 10×, NA = 0.3; Nikon 40×, NA = 0.6 for resolution testing) and a line-scan color camera (Thorlabs 1501C-GE, 1392 px × 23 Hz readout, 96 kB/s throughput, Fig. 1a) allow collection with minimal filtering, as glass optics block residual UV and pass visible emission.

For scanning surface (non-sectioned) tissue, deep-UV laser or high-pressure mercury lamps (≈280 nm) coupled with quartz objectives (e.g., 20×, NA = 0.5; lateral resolution ~300 nm) and scientific CMOS (sCMOS) or EMCCD detectors ensure high quantum efficiency in the visible range. Motorized XY stages, with step precision <1 μm, automate raster scanning and facilitate WSI mosaic generation (Afshin et al., 12 May 2025).

Sample preparation protocols differ by application: paraffin embedding with UV27 dye (0–14 wt%, infiltration at 60°C) enables block-face imaging in serial sectioning contexts; fresh tissue imaging for intraoperative analysis is label-free or utilizes rapid ex vivo stains (propidium iodide, eosin Y), and takes advantage of autofluorescence for contrast (Ghahfarokhi et al., 2024, Afshin et al., 12 May 2025).

2. Imaging Workflow, Sectioning, and Data Acquisition

Block-face DUV-FSM employs serial sectioning via a motorized microtome (e.g., Thermo Fisher HM355S, section thickness 0.5–100 μm, blade mounted at 10° cutting angle). The imaging cycle per slice involves:

1. Microtome cut,
2. Stage stabilization (∼1 s),
3. UV illumination (280 nm, ∼100 ms exposure),
4. Image acquisition and
5. Repeat for the next section.

Camera acquisition is synchronized via TTL trigger signals, and randomized cutting velocities mitigate periodic knife-chatter artifacts. Rigid-body alignment (OpenCV, parametric registration) compensates for minor lateral drift between slices (Guo et al., 2019).

In surface scanning DUV-FSM, mosaicking of contiguous tiles (400×400 px, ~0.5 μm/px) generates gigapixel-scale WSIs without need for physical sectioning. Frame-wise dark-field subtraction and flat-field correction standardize pixel intensities; foreground-masking removes regions with >80% background prior to patch-level deep learning (Afshin et al., 12 May 2025). Exposure times for DUV fluorescence images typically range from 50–200 ms.

Table 1. Imaging Parameters for Representative DUV-FSM Implementations

Parameter	Block-Face (MUVE) (Guo et al., 2019)	Surface Scanning (Afshin et al., 12 May 2025, Ghahfarokhi et al., 2024)
Excitation Wavelength	280 nm	280–285 nm
Objective	10× NA 0.3, 40× NA 0.6	20× NA 0.5; 4× NA ~0.13
Sampling/Pixel Size	0.37 μm lateral (40×)	2–3 μm/pixel (4×)
Section Thickness	0.5–100 μm (microtome)	Not applicable (direct surface scan)
Detector	Thorlabs line-scan camera	sCMOS/EMCCD
FOV (single)	~1.4 mm (10×), <200 μm (40×)	1.5 mm × 1.5 mm

3. Resolution, Optical Sectioning, and Penetration Depth

DUV-FSM achieves diffraction-limited lateral resolution, $d_{xy} \approx 0.61\,\lambda_{em}/{\rm NA}$, e.g., for λ_em ≈ 515 nm and NA = 0.6, $d_{xy} \approx 525\,\mathrm{nm}$ (Guo et al., 2019). Axial resolution is defined by optical penetration depth and, in sectioning systems, by physical mill thickness $\Delta z_{phys}$. UV27 doping of paraffin restricts optical section thickness to ~10 μm via exponential absorbance, and Monte Carlo PSF simulations demonstrate much sharper axial profiles compared to conventional confocal systems (Fig. 2, (Guo et al., 2019)). Empirical point spread function (PSF) measurements using fluorescent bead phantoms verify increased axial confinement (Fig. 4). For slice-based systems, net axial resolution approaches the microtome setting.

Penetration depth is effectively unlimited for volumetric imaging, since successive block-face milling removes occluded tissue layers. Surface scanning DUV-FSM provides shallow imaging (<10 μm) without need for tissue clearing or long working-distance objectives.

4. Data Processing, Deep Learning Pipelines, and Automated Classification

Post-acquisition, DUV-FSM datasets undergo rigid-body registration, brightness/contrast normalization, and direct stacking into 3D volumes or large-scale WSIs. For downstream automated analysis, each WSI is partitioned into non-overlapping patches (256×256 px (Ghahfarokhi et al., 2024), 400×400 px (Afshin et al., 12 May 2025)), with background areas filtered out.

Recent work applies deep learning for breast cancer classification. Feature extraction utilizes ResNet50 (ImageNet pre-trained, 2048-D embeddings per patch) (Ghahfarokhi et al., 2024), or a patch-level Vision Transformer (ViT-Base/16, 12 transformer layers, D=768, h=12) (Afshin et al., 12 May 2025). Each patch is classified via XGBoost or a transformer head, and patch-level predictions are fused into whole-surface decisions by weighted majority voting or regional saliency weighting.

Grad-CAM++ saliency maps quantify spatial importance for each patch, using pixel-wise scores computed from convolutional feature gradients:

$$L_{ij}^c = \mathrm{ReLU}\left( \sum_{k} \alpha_k^c \frac{\partial y^c}{\partial A^k_{ij}} \right)$$

Patch importance $r_j$ is the average saliency in area $A_j$, and fusion steps discard low-saliency regions (r_j < 0.30 (Afshin et al., 12 May 2025), r_i^j < 0.25 (Ghahfarokhi et al., 2024)) prior to final classification.

Dataset augmentation techniques, notably diffusion probabilistic models (DPM), convert Gaussian noise into realistic synthetic patches, balancing malignant and benign sample counts and improving model diversity. The DPM implements forward diffusion

$$q(x_{t}\mid x_{t-1},c) = \mathcal{N}\bigl( \sqrt{1-\beta_t} x_{t-1}, \beta_t I \bigr)$$

and reverse denoising with a U-Net plus ResNet blocks:

$$p_\theta(x_{t-1}\mid x_t,c) = \mathcal{N}\left( \mu_\theta(x_t,t,c), \Sigma_\theta(t) \right)$$

where

$$\mu_\theta(x_t,t,c) = \frac{1}{\sqrt{1-\beta_t}} \left( x_t - \frac{\beta_t}{\sqrt{1-\alpha_t}} \varepsilon_\theta(x_t,t,c) \right)$$

with a training objective combining denoising and variational components (Ghahfarokhi et al., 2024).

5. Performance Evaluation and Comparative Metrics

DUV-FSM imaging supports high-throughput acquisition, with ∼4 s per slice (block-face), enabling ~2000 slices in ∼2 h for 3D volumes (e.g., mouse midbrain, 389 × 241 × 2134 μm, Fig. 5 (Guo et al., 2019)). Surface scanning achieves rapid frame rates for gigapixel WSI mosaics. Performance metrics in breast cancer classification tasks include accuracy, sensitivity (recall), specificity, and ROC AUC. Table 2 provides 5-fold cross-validation results for several pipelines.

Table 2. Classification Performance Metrics

Method	Accuracy (%)	Sensitivity (%)	Specificity (%)	Reference
Affine Augmentation	93	94	76	(Ghahfarokhi et al., 2024)
ProGAN Augmentation	93	94	76	(Ghahfarokhi et al., 2024)
DPM Augmentation (ResNet + XGBoost)	97	97	93	(Ghahfarokhi et al., 2024)
ResNet-50 + XGBoost + Grad-CAM++	95	100	87.5	(Afshin et al., 12 May 2025)
PatchViT + Grad-CAM++	98.33	100	96.0	(Afshin et al., 12 May 2025)

The PatchViT + Grad-CAM++ pipeline demonstrates the highest performance (accuracy 98.33%, sensitivity 100%, specificity 96%), with inference that AUC exceeds 0.98. Augmentation via DPM yields significant improvement over standard affine and GAN-based strategies.

6. Advantages, Limitations, and Prospective Directions

DUV-FSM offers several advantages: sub-micron lateral and axial resolution, fast volumetric acquisition (<minutes), unlimited penetration depth by serial milling, direct compatibility with routine FFPE histology, and effective integration with deep learning for automated diagnosis (Guo et al., 2019, Ghahfarokhi et al., 2024, Afshin et al., 12 May 2025). Label-free imaging enables contrast without exogenous dyes, facilitating rapid intraoperative margin assessment in breast-conserving surgery. "A plausible implication is that slide-free, rapid cellular imaging could supplant traditional frozen-section analysis for margin status review."

Limitations include scanning speed (stage movement, exposure time), field-of-view constraints imposed by objective NA and UV transmission, and generalizability—current datasets (e.g., 60 WSIs (Afshin et al., 12 May 2025)) do not capture full clinical heterogeneity. Sample volume is limited only by instrument capacity for block-face imaging.

Future directions encompass hardware advancements (faster UV sources, larger-field objectives, GPU-accelerated mosaicking), adaptive scanning guided by preliminary AI predictions, extension to other resection margins (glioma, head & neck), and multi-center datasets to enhance model robustness. Real-time integration of DUV-FSM with surgical workflows and end-to-end deep learning interpretability frameworks remain active research areas.

7. Applications and Impact in Biomedical Imaging

DUV-FSM supports large-scale, high-resolution 3D tissue modeling in research and clinical histopathology. Applications demonstrated include microvascular segmentation, cell-vessel co-segmentation, and graph-theoretic analysis of tissue microarchitecture (Fig. 6 (Guo et al., 2019)). Intraoperative tissue characterization with DUV-FSM has proven efficacy for breast cancer margin assessment, minimizing false negatives and re-excision rates through robust deep learning pipelines (Ghahfarokhi et al., 2024, Afshin et al., 12 May 2025). Compatibility with standard dyes (DAPI, eosin, thionine, India ink) facilitates molecular investigations and augments contrast mechanisms. "This suggests growing utility for rapid, automated tissue evaluation in both preclinical and clinical domains, as well as a path toward AI-enabled, real-time histology."

DUV-FSM, by integrating deep-UV photophysics, high-throughput imaging instrumentation, and scalable deep learning, defines a versatile platform for quantitative tissue analysis at the cellular level. Its expanding role in digital pathology and computational histology is underscored by continuing methodological developments and performance improvements.