The stratum corneum (SC), the outermost layer of human skin, is a thin yet functionally critical structure that governs epidermal permeability, water retention, and protection against pathogens, allergens, and xenobiotics (Lintzeri et al., 2022). Although typically only 10–20 µm thick, the SC exerts a disproportionate influence on skin barrier function through its lipid–protein “brick-and-mortar” architecture. Even micrometre-scale deviations in SC thickness can substantially compromise barrier integrity or, conversely, modulate the penetration of therapeutic and cosmetic agents. Consequently, precise and reproducible quantification of SC thickness has emerged as a key objective in clinical dermatology, transdermal drug delivery, and personal-care science (You et al., 2023). However, accurately delineating this ultrathin, heterogeneous layer in vivo remains a major technical challenge, limiting the translation of advanced imaging modalities into routine clinical and industrial practice (Wang et al., 2024).

Optical coherence tomography (OCT) has become a leading non-invasive tool for SC assessment due to its depth-resolved imaging capability, micrometre-scale axial resolution (1–5 µm), and penetration depths approaching 1.8 mm in keratinized tissue (Lin et al., 2024). Modern spectral-domain, swept-source, and line-field confocal OCT (LC-OCT) systems can clearly distinguish the highly scattering SC from the underlying viable epidermis while preserving tissue integrity for longitudinal monitoring (Lintzeri et al., 2022). Unlike histology, OCT avoids fixation-induced shrinkage; unlike high-frequency ultrasound (Luan et al., 2023), it resolves fine epidermal layers without contrast agents; and unlike reflectance confocal microscopy (Yu et al., 2023b), it provides subsurface imaging over hundreds of micrometres at video rates. These advantages position OCT as an ideal modality for quantitative SC thickness mapping (Yang et al., 2024).

Despite these intrinsic strengths, the clinical and translational utility of OCT is constrained by a persistent bottleneck: the reliable extraction of SC boundary coordinates from raw OCT reflectivity data (Lin et al., 2024). Manual caliper-based measurements are subjective, labor-intensive, and poorly reproducible, with inter-observer variability often exceeding 15% (He et al., 2024). Histological validation, frequently treated as a reference standard, introduces tissue shrinkage artifacts of 12%–21%, undermining direct comparison with in vivo OCT measurements (Kerns et al., 2008). Automated segmentation approaches, including graph-search methods and convolutional neural network–based pipelines, have improved accuracy but remain limited by high computational demands, sensitivity to device-specific signal characteristics, and a heavy reliance on large, densely annotated training datasets (Kerns et al., 2008). These constraints hinder deployment in point-of-care settings, multicenter clinical trials, and large-scale cosmetic testing, where robustness, efficiency, and data economy are essential (Chen et al., 2025).

At the same time, the biological and clinical importance of precise SC thickness measurement continues to intensify (Gambichler et al., 2006b). Deviations from physiological SC thickness are closely associated with skin disorders (Xu et al., 2025) such as atopic dermatitis, xerosis, ichthyoses, and psoriasis, where barrier dysfunction, rather than thickness alone, correlates with disease severity and increased transepidermal water loss (Lin et al., 2024). Beyond pathology, controlled modulation of SC thickness through hydration, exfoliation, or formulation design is central to transdermal delivery and cosmetic efficacy; shifts of only 2–4 µm can alter active ingredient penetration by tens of percent. These effects are further complicated by pronounced inter-individual and anatomical variability, with SC thickness accounting for much of the epidermal thickness range observed across body sites, ages, phototypes, and environmental conditions. Detecting such subtle, early-stage changes requires measurement strategies that are both highly sensitive and biologically consistent.

To overcome these challenges, a shift is needed from purely discriminative, data-hungry segmentation pipelines toward models that explicitly integrate OCT physics with data-efficient representation learning (Tang et al., 2024). Physics-informed digital signal processing (DSP) techniques, such as speckle reduction, depth-dependent sensitivity compensation, dispersion correction, and deconvolution, can suppress imaging artifacts and linearize OCT signals before learning, reducing the burden placed on downstream models (Ozcan et al., 2007). Building on this foundation, variational autoencoders (VAEs) offer a powerful generative framework for SC analysis (Li et al., 2025). By learning continuous, regularized latent representations of skin morphology, VAEs can disentangle biologically meaningful factors, such as SC thickness and surface roughness, from noise and device-specific variability (Li K. et al., 2024). Their generative nature enforces structural coherence, reduces dependence on exhaustive pixel-level annotations, and enables intrinsic uncertainty estimation—features that are particularly valuable for clinical trust, cross-device generalization, and longitudinal monitoring.

This review synthesizes 2 decades of progress in SC thickness quantification with a particular focus on the emerging convergence of advanced DSP and VAE-based generative learning. We first outline the biological and clinical rationale for accurate SC thickness mapping and critically examine the limitations of conventional and contemporary measurement approaches. We then detail DSP strategies tailored to skin OCT and VAE architectures optimized for SC boundary delineation, benchmarking their performance against classical algorithms and discriminative deep learning models. Finally, we discuss current clinical and industrial applications, unresolved methodological challenges, and future directions, outlining a roadmap toward centimetre-scale, micrometre-accurate in vivo mapping of the human skin barrier. By integrating principles from optical physics, signal processing, and representation learning, this work aims to advance precision skin barrier assessment and support the development of personalized dermatological and cosmetic interventions.

The Early studies treated the SC reference standard as a straightforward length: place calipers on the OCT screen or excise a biopsy, embed it in paraffin, and read the distance between the corneocyte surface and the viable epidermis. Yet, each of these “conventional” routes introduces measurement error that is now well-documented (Gambichler et al., 2006a; Lintzeri et al., 2022). The manual placement of electronic calipers on individual B-scans remains the most widely used clinic-side technique because it requires no dedicated software. Unfortunately, it is both labor-intensive and operator-dependent. A recent multi-site trial that timed 1,587 manual measurements reported a mean reading time of 42 s per frame, an impractically burdensome task when hundreds of frames are generated in a 2-min sweep (Figure 1). The same study showed inter-observer coefficients of variation approaching 13% for forearm skin, even after a joint training session, largely because speckle noise masks the dermal-epidermal junction and encourages subjective placement of the basal line (Gambichler et al., 2006b). Manual protocols also require the user to down-sample three-dimensional stacks to a handful of “representative” slices, thereby forfeiting information on regional undulations and biasing group statistics, especially at acral or aged sites where papillary relief is pronounced (Jain et al., 2024).

Histology is often considered a gold standard, but its thickness numbers are themselves moving targets. From the moment a 4 mm punch is taken, contractile forces and subsequent processing steps initiate cumulative shrinkage. Optical tracking and serial caliper studies indicate that the loss in length is approximately 12%–21% and in width is 11%–23% before the slide even reaches the microscope (Kerns et al., 2008).​ Lipid extraction during dehydration collapses intercellular spaces within the SC, while paraffin infiltration hardens the sample, exaggerating the effect. Cryosectioning reduces shrinkage but sacrifices the crisp nuclear detail dermatopathologists prefer. The net result is that histological SC often reads several micrometres thinner than its in-vivo counterpart, forcing investigators to apply empirical correction factors that vary with fixation protocol, anatomic site, and patient age (Sandby-Møller et al., 2003; Kerns et al., 2008; Tran et al., 2015).

To circumvent these limitations, classical image-processing pipelines emerged in the mid-2000s. The simplest exploit the bimodal intensity profile of SD-OCT A-scans: taking the first strong reflection as the air/SC interface and a second peak as the dermal-epidermal junction. While fast, this “peak-to-peak” heuristic fails whenever the second maximum corresponds instead to a collagen bundle or papillary tip, resulting in systematic under-reads in more than 25% of scans (Gambichler et al., 2006a). Edge-based algorithms refine the approach by applying Sobel or Canny operators to each frame and then fitting polynomial splines through the strongest gradient. However, they struggle in low-contrast regions and tend to propagate single-frame errors through the spline, inflating the root-mean-square (RMS) error to 8–12 µm on healthy volar skin. The model-based methods improve robustness by incorporating anatomical priors. Active-contour “snakes” initialise near the skin surface and evolve toward energy minima that penalise curvature while rewarding high gradient magnitude. However, convergence stalls in the presence of speckle voids and requires careful tuning of elasticity parameters for each device. Graph-search formulations overcome many of these issues: they treat the B-scan as a weighted graph in which the cost of traversing a pixel is inversely proportional to its edge strength, then solve a shortest-path problem that yields globally optimal, smooth boundaries. Three-dimensional extensions that couple adjacent B-scans have reduced segmentation error by ∼20% relative to 2-D approaches and generate continuous thickness maps suitable for roughness assessment (Srivastava et al., 2018).

Despite these refinements, conventional image-processing pipelines share common weaknesses. First, most rely on handcrafted thresholds or filter kernels that were calibrated on homogeneous laboratory datasets, degrading when confronted with lower signal-to-noise ratios, darker phototypes, or atypical curvature. Second, they remain computationally heavy; a typical 3-D graph-search pass over a 512 × 1,024 × 400 volume can take tens of seconds on a CPU. Third, repeatability seldom surpasses that of a well-trained technician: even state-of-the-art convolution-edge hybrids report mean absolute errors around 10 µm and Dice overlaps near 0.83 ± 0.06 across 270 clinical OCT frames, numbers acceptable for population studies but marginal for detecting the 2–4 µm shifts that accompany early barrier impairment or cosmetic interventions (Del Amor et al., 2020).

Combined, these findings highlight a plateau in what purely manual or classic algorithmic strategies can deliver. They highlight the need for next-generation pipelines that integrate physics-informed preprocessing with learning-based inference, precisely the gap that advanced digital signal processing and variational autoencoders aim to address in the sections that follow.

The SC comprises 10–20 flattened, anucleate corneocytes embedded in a lipid matrix organised as short- and long-periodicity lamellae. The matrix, which is roughly 50% ceramides, 25% cholesterol, and 15% free fatty acids by weight, forms the only continuous diffusion pathway across the barrier. At the same time, the protein-rich corneocytes (“bricks”) provide mechanical strength (Bouwstra et al., 2023).​ Corneocytes originate in the stratum granulosum, where lamellar bodies exocytose precursor lipids and hydrolytic enzymes. Once keratinocytes enucleate, transglutaminase cross-links involucrin, loricrin, and small proline-rich proteins to form the cornified envelope; covalently bound ω-hydroxy-ceramides anchor the lipid lamellae to this scaffold (Feingold and Jiang, 2011). A downward pH gradient (≈7.0 → 4.5) activates β-glucocerebrosidase and acidic sphingomyelinase for lipid maturation, then triggers kallikrein-5/-7 to cleave corneodesmosomes during desquamation. Hydration modulates the lateral spacing of lipid bilayers, causing the SC to swell or shrink by up to 30%, which directly alters optical backscatter and, hence, OCT contrast (Table 1). Clinically, perturbations in lipid ratios or corneodesmosome turnover manifest as increased transepidermal water loss, dyschromia, or scaling disorders, such as ichthyosis (Bouwstra et al., 2023).​ Optical coherence tomography relies on low-coherence interferometry, where back-reflections from tissue microstructures interfere with a reference arm to localize scatterers with an axial resolution of Δz ≈ 0.44 λ₀ ²/Δλ. Using broadband sources centered at 840 nm (for epidermal work) or 1.3 µm (for dermal penetration) yields an axial resolution of 1–7 µm in the skin (Popescu et al., 2011).

A variational auto-encoder (VAE) links an encoder qϕ(z/x) with a decoder pθ(x/z) and is trained by maximising the evidence lower bound, which trades reconstruction fidelity against the Kullback–Leibler divergence that keeps the approximate posterior close to a simple prior. Because the reparameterization trick makes the stochastic path differentiable, the network learns latent codes that, in skin OCT, capture layer topology and vessel-scattering statistics. Random sampling in this space supports both data augmentation and principled uncertainty estimates (Wu Z. et al., 2024). Pushing the KL term with a factor β larger than one (β-VAE) forces the network to compress more aggressively, and in doing so tends to align individual latent axes with interpretable quantities such as SC thickness, surface roughness, or hydration level, trading a slight loss in structural similarity (SSIM) for far greater explainability and controllable synthesis (Aronsson, 2023). If, instead of a continuous Gaussian, the model uses a learned code-book and vector quantisation (VQ-VAE), the latents become discrete; this prevents posterior collapse and sharpens fine edges, a valuable property when millimetre-wide B-scans translate sub-pixel shifts into micrometre-scale errors.

Building on these foundations, researchers have introduced spatial-contextual and volumetric extensions that weave attention masks or 3-D convolutions through the encoder and decoder, preserving local coherence across slices; retinal-OCT anomaly detectors employing such designs exceed 0.95 AUROC with fewer than ten thousand labelled images, and volumetric-erasing tricks further exploit inter-slice continuity to lift unsupervised segmentation scores in skin OCT and MRI alike (Jebril et al., 2024). Where label scarcity is acute, conditional variants (cVAEs) append class tags, physics priors, or intermediate DSP features to both ends of the network, thereby steering reconstructions toward thickness-conditioned outputs or multi-hypothesis segmentations; probabilistic U-Net and PHISeg exemplify this approach by embedding a cVAE inside a U-Net backbone, capturing aleatoric uncertainty and furnishing confidence maps around tricky structures such as hair follicles, while newer hybrids couple Hamiltonian sampling with discriminative regularisation to sharpen boundaries in ultra-small datasets (Petersen and Kucheryavskiy, 2025).

Across all these flavours, several implementation rules consistently improve performance. First, the latent dimensionality should be large enough to encode every anatomical factor of interest, thickness, scattering slope, and curvature, yet not so large that it encourages over-fitting (Biffi et al., 2020). Second, KL-annealing or cyclical β-schedules curb early posterior collapse in the presence of high speckle noise (Huang and Yang, 2013). Third, feeding physics-informed channels, such as attenuation-corrected intensity or dispersion-compensated phase, conditions the network toward anatomically plausible solutions and can reduce the need for pixel-accurate labels by 30%–40%. Finally, depth-wise separable convolutions and latent-space clustering enable few-shot adaptation across different OCT devices while maintaining real-time throughput on portable hardware (Jin et al., 2016).

Modern dermatologic OCT pipelines rarely pass raw interferograms directly to variational autoencoders (VAEs). Instead, they rely on a sequence of physics-aware signal-processing blocks that clean, linearize, and enrich the data, allowing the network to devote its limited capacity to modeling anatomy rather than artifacts. Below, each major block is summarised (Table 2) together with its net benefit for VAE-based thickness estimation.

Speckle-noise suppression remains the single most crucial preconditioning step because multiplicative speckle not only lowers peak-signal-to-noise ratio (PSNR) by 15–20 dB but also generates false high-frequency texture that a VAE might erroneously encode as physiologic structure. Traditional adaptive filters, including enhanced Lee, Kuwahara, hybrid median, and adaptive Wiener, reduce speckle contrast by 25%–35% while preserving edge strength (Ozcan et al., 2007).​ Wavelet-domain approaches go further: the à-trous transform followed by scale-dependent thresholding removes granular noise while retaining the SC-epidermis gradient and increasing the structural similarity index (SSIM) by ≈ approximately 0.05 on test phantoms (Ozcan et al., 2007).​ Over the past 2 years, deep unsupervised methods have overtaken handcrafted filters. The Speckle Split Noise2Void (SSN2V) framework, for instance, trains an OCT-specific blind-spot network that implicitly learns the speckle statistics from paired noisy patches, delivering a ∼2.8 dB PSNR gain without clean targets (Schottenhamml et al., 2023).​ Recent optical schemes, such as aperture-phase modulation with adaptive optics, physically decorrelate speckle prior to detection and can be combined with post-hoc GAN-based multiscale denoising for an additional 1–2 dB improvement (Yu et al., 2023a; Das et al., 2024). In VAE pipelines, a denoised magnitude-only input suppresses pixel-level randomness, allowing the latent space to capture mesoscopic parameters, such as layer thickness and scattering slope, rather than fitting speckle.

Depth-dependent sensitivity roll-off compensation tackles the systematic decay of signal amplitude with optical path length that plagues spectral- and swept-source OCT. Uncorrected roll-off distorts intensity-based tissue cues, biasing thickness estimates toward shallower values. Hardware options (k-clock resampling, dual-balanced detection) help, but software compensation is now the preferred route because it adapts per scan. Calibration scans from a mirror provide a reference axial point-spread function (PSF); the inverse of this curve is then applied to each A-scan, or a polynomial/log-domain model is fitted to rescale deeper pixels. Optimized numerical k-sampling in swept-source OCT reduces roll-off to 2–3 dB over 4 mm, maintaining a 4.9 µm axial resolution across the span (Huang et al., 2024). Line-field systems integrate a similar polynomial correction to achieve a <10 dB drop over 1 mm (Chen et al., 2024). Once roll-off-normalised, the dynamic range of superficial versus deep SC becomes consistent across volumes and scanners, giving VAEs a homogeneous intensity distribution that accelerates convergence and improves cross-device generalisation.

Phase-sensitive and polarization-sensitive enhancements add entirely new information channels. Phase-resolved OCT (ϕ-OCT) registers sub-nanometre axial displacements caused by pulsatile blood flow or biomechanical waves; when mapped over time, these phase shifts reveal visco-elastic contrasts between SC and viable epidermis. Dynamic phase-sensitive optical coherence elastography, for example, tracks the speed of Rayleigh waves to grade burn severity in vivo (Liu et al., 2024). Polarization-sensitive OCT (PS-OCT) exploits tissue birefringence: the cornified envelope and ordered lipid lamellae of the SC show minimal birefringence, whereas the keratin network below exhibits measurable retardation. Full-range depth-encoded SS-PS-OCT now delivers high-sensitivity birefringence maps that clearly outline the SC boundary even when intensity contrast is weak (He et al., 2023; Wu T. et al., 2024).​ Feeding phase and retardation volumes—either as auxiliary channels or as priors in a conditional VAE, improve boundary localisation by ∼15% and permit the latent space to disentangle structural features (thickness) from polarimetric ones (birefringence).

Super-resolution and deconvolution seek to reverse the blur introduced by the coherence gate and confocal pinhole. Blind and PSF-informed Richardson–Lucy deconvolution sharpens the axial profile, recovering 15%–25% of high-frequency content and reducing the edge-spread width by ∼1.5 µm. A 2025 review catalogues emerging Bayesian and deep-unfolded deconvolution networks that explicitly model OCT noise statistics; these schemes reclaim 30 of the axial bandwidth while suppressing ringing (Abbasi et al., 2025).​ In parallel, temporal-PSF deconvolution using recurrent neural networks, initially developed for time-resolved fluorescence, is being adapted for OCT to achieve joint depth super-resolution and dispersion correction (Pandey et al., 2024).​ VAE decoders trained on deconvolved inputs no longer need to “undo” system blur. They can focus on subtle curvature cues, thereby reducing reconstruction error at the SC–epidermis junction by up to 40% on synthetic phantoms.

Time–frequency analyses provide an orthogonal approach by decomposing non-stationary backscatter patterns. The short-time Fourier transform (STFT) treats each A-scan as a spectro-temporal signal; varying window length tunes the trade-off between depth and spectral resolution, enabling selective enhancement of sparsely distributed high-k components associated with lipid lamellae (Baba, 2012).​ Continuous wavelet transforms (CWT) offer multiscale localisation; coupling CWT coefficients to a VAE gives the encoder explicit access to both fine (lamellar) and coarse (papillary) scales. Empirical mode decomposition (EMD) and its learnable derivatives iteratively strip intrinsic mode functions dominated by speckle, leaving a residue that approximates the structural signal; EMD-based denoising boosts SSIM by 0.08 over median filtering on volar skin (Myakinin et al., 2013; Velasco-Forero et al., 2022). When these spectro-temporal features are concatenated with intensity images, VAEs gain richer descriptors of layer periodicity and scattering anisotropy, yielding more geometrically faithful reconstructions and latent factors that correlate with biophysical properties (e.g., lipid order).

Deep generative learning has moved VAE-based pipelines from proof-of-concept to practical tools that rival classical graph-search and U-Net segmenters for OCT skin analysis. The key design choices, network architecture, latent-space regularisation, conditioning strategy, interaction with signal-processing blocks, and data-efficiency tactics, determine whether a model captures subtle micrometre-scale boundaries or collapses into blurry reconstructions (Figure 2).

Public resources for skin OCT are beginning to mature, yet they still lag far behind the dozens of well-curated retinal collections that dominate the field. Currently, the most widely cited open dataset is the UIUC “CNN-GS-skin” corpus, which was released with the 2024 Scientific Reports paper on rapid epidermal thickness measurement. It provides 1,575 B-scans (460 × 1,500 px) acquired with a swept-source handheld probe from five body sites in 63 healthy volunteers; both the air/SC and the SC/dermal–epidermal junction (DEJ) boundaries are traced by two independent raters and reconciled with a third in cases of ≥3 µm disagreement (Lin et al., 2024).​ Complementing the 2-D collection, the LC-OCT Healthy Epidermis volume set, published by Chauvel-Picard and colleagues, provides 30 volumetric stacks (1.2 × 1.2 × 0.5 mm, with isotropic 1 µm voxels), along with automatically derived thickness maps that have been cross-validated against histology (Cappilli et al., 2024).​ Meanwhile, several manufacturers have released demo volumes, often bundled with their proprietary analysis suites, that, although limited in number, cover pathological cases such as psoriasis plaques and early basal cell carcinomas; researchers typically augment these with small, bespoke acquisitions shared on request. Consequently, the community still relies heavily on semi-public repositories, such as GitHub links or data-use agreements negotiated with principal investigators, to achieve sample sizes suitable for deep generative training.

Because open, diverse cutaneous data remain scarce, groups have filled the gap with physical phantoms and in silico synthesis. Silicone–gelatin multilayer blocks laden with titanium-dioxide scatterers reproduce the optical attenuation and the 10–20 µm SC seen in vivo, while 3D-bioprinted constructs incorporating synthetic melanin nanoparticles span Fitzpatrick phototypes I–VI, allowing investigators to probe pigment-dependent contrast without recruiting human participants (Liu et al., 2018; Yim et al., 2023). At the numerical end of the spectrum, the open-source MCOCT Monte Carlo engine simulates A-scans under arbitrary refractive index, absorption, and anisotropy profiles. By stochastically varying layer thicknesses and scattering parameters, it produces thousands of “realistic yet label-perfect” B-scans, which are invaluable for pre-training β-VAEs before fine-tuning on scarce patient data (Erdenedalai et al., 2024).​

The value of any dataset, however, hinges on the consistency of annotation protocols. In practice, most groups adopt a two-tier pipeline: junior annotators place rough polylines around the air/SC surface and the SC/DEJ interface in tools such as ITK-SNAP; senior dermatologists then refine those curves using overlaid birefringence or phase-contrast cues when available, and a final adjudicator resolves conflicts by majority vote. In the UIUC corpus, this procedure yielded an average inter-observer standard deviation of 1.7 µm for the SC boundary and 2.3 µm for the DEJ (Lin et al., 2024).​ To streamline future efforts, recent papers recommend publishing annotation checklists that clearly outline the inclusion criteria for hair follicles, sweat ducts, and motion artifacts, along with slice-by-slice provenance, so that downstream users can quantify rater uncertainty. Alongside technical rigor, ethical, privacy, and sampling bias issues require equal attention. Unlike retinal OCT, skin images reveal body topology, tattoos, and sometimes even fingerprints, making complete anonymization impossible without aggressive cropping; institutional review board protocols therefore emphasize explicit patient consent for open release and long-term storage. Furthermore, several audits of dermatology repositories have revealed a systematic underrepresentation of Fitzpatrick phototypes IV–VI, which in turn degrades model performance on darker skin tones. A 2024 JAMA Network Open study demonstrated that crowdsourced recruitment, combined with stratified sampling, can help close this gap; however, it must be paired with bias-aware training objectives if generative VAEs are to accurately replicate lesions across all phototypes (Ward et al., 2024).​ Finally, data custodians are urged to strip or hash device identifiers to avoid inadvertent leakage of vendor IP and to publish datasheets for datasets that document imaging parameters, subject demographics, consent language, and known caveats.

Continuous advances in high-resolution OCT and the VAE-enhanced analytics described earlier are already reshaping day-to-day dermatology. To begin with, inflammatory-disease monitoring has moved beyond crude clinical scores: weekly line-field OCT (LC-OCT) scans of atopic-dermatitis lesions show that a 2–4 µm reduction in SC thickness and a parallel fall in dermal inflammatory signal precede visible improvement on EASI scores by almost a week, allowing clinicians to titrate biologics such as dupilumab with unprecedented precision (Dryżałowska et al., 2024). Similar micro-scale readouts now track psoriatic-plaque descaling and the early relapse kinetics of chronic hand eczema, replacing serial biopsies and significantly reducing patient burden. Because moisturisers, retinoids, and exfoliants target these same micrometre-level shifts, cosmetic-science groups have adopted OCT-VAE pipelines as objective endpoints in product efficacy trials. Under controlled occlusion, visible-light OCT reveals that occlusive hydration brightens and swells the SC by up to 30% within 2 hours, a change that regresses after 24 h unless an occlusive barrier ingredient is present (Revin et al., 2023).​ Manufacturers, therefore, use automated thickness maps to rank formulations, justify marketing claims, and fine-tune rinse-off times, while regulators welcome a non-invasive alternative to repeated tape stripping.

Moving from cosmetics to pharmacotherapy, transdermal drug engineers rely on the exact measurements to design microneedles and iontophoretic patches that bypass or temporarily thin the SC. Real-time OCT has demonstrated that hydrogel-microneedle arrays swell upon insertion and maintain micro-channel patency for only 15–20 min; feeding these dynamics into VAE-based latent models predicts permeation windows and optimizes patch dwell time without radio-label tracers (Wu C. et al., 2024; Omidian and Dey Chowdhury, 2025).​ Moreover, adaptive controllers that couple OCT feedback with iontophoretic power already achieve closed-loop insulin delivery in ex vivo skin, suggesting the potential for fully autonomous wearable therapies. Beyond intact skin, precision measurements have become invaluable in acute-care settings. Paediatric hand-burn teams now apply depth-resolved OCT scoring systems to determine whether to use conservative dressings or early grafting; this method reduces unnecessary excisions by a third while preserving functional outcomes (Li H. et al., 2024; Lindert et al., 2024).​ In chronic-wound clinics, longitudinal maps of neo-epidermal thickness forecast complete closure almost 2 weeks before planimetric area shrinkage reaches significance, enabling earlier discharge. Scar-revision surgeons likewise exploit SC and epidermal-thickness asymmetries to time fractional laser passes more effectively, thereby reducing postoperative hyper- or hypopigmentation.

Finally, the shrinkage of models into <100-MB, CPU-ready binaries has opened the door to genuinely portable imaging. Handheld probes now embed VAE inference on a Raspberry Pi-class board, streaming encoded latent vectors (∼1 kB per frame) to a clinician’s tablet or a cloud server for teleconsultation (Liu et al., 2025).​ Because latent traversal can visualise the algorithm’s internal notion of “thick” or “thin” skin, remote dermatologists gain interpretability and can flag suspicious scans for local follow-up. Early pilots in rural clinics demonstrate that such point-of-care workflows reduce referral delays for severe eczema by 40%, while home-monitoring studies, similar to those in ophthalmic tele-OCT, are adapting similar architectures for chronic dermatitis surveillance (Blinder et al., 2024; Dolar-Szczasny et al., 2024).​

Achieving truly marker-less, real-time thickness mapping on pocket-sized hardware remains the first major hurdle, and here the optics and the algorithms must evolve together. Recent “brief-case” and even smartphone-coupled OCT engines now weigh <1 kg and draw <10 W, yet they still offload segmentation to a laptop; shrinking DSP-preconditioned β-VAE models to sub-50 MB binaries that can run at >40 fps on ARM chipsets will eliminate that tether, provided power-aware quantisation and on-chip FFT accelerators keep latency under the 25-ms perceptual threshold (Song et al., 2021). Success would open the door to at-home eczema tracking and battlefield burn triage without fiducial markers or external calibration targets. Moving beyond one modality at a time, the next frontier is multimodal fusion. Hybrid probes that co-register cellular-resolution OCT with near-infrared Raman spectroscopy already discriminate between malignant and benign skin cells by combining micro-architecture with molecular fingerprints, while tri-modal studies couple line-field OCT to ultra-high-frequency ultrasound to extend penetration beyond the dermis and capture both scatter and acoustic impedance in a single pass (You et al., 2023; Boussingault et al., 2024). The challenge is to craft generative latent spaces that respect the physics of each signal, perhaps by training cross-modal VAEs whose shared latent manifold encodes geometry, while modality-specific branches handle optics or vibro-acoustics. Doing so could push confidence intervals below the ±2 µm clinical threshold, even in oedematous or scarred skin (You et al., 2023).

Yet richer data streams will be of little clinical value if clinicians cannot understand, trust, and legally deploy the algorithms that interpret them. Regulators are rushing: the FDA’s March 2024 AI roadmap and its draft lifecycle guidance require continuous-learning devices to document model updates, quantify performance drift, and provide human-readable rationales before market clearance. Empirically, a 2025 study of 15 dermatological AI devices (Bellemo et al., 2025) found that only 33% met FDA’s “performance drift quantification” requirement—specifically, failing to track error increases across 6 months of clinical use (e.g., a U-Net model’s MAE rose from 5.8 µm to 8.2 µm on phototype V skin due to unaccounted seasonal humidity effects). Conversely, a DSP-augmented β-VAE in the same study satisfied the requirement by integrating monthly federated fine-tuning and latent-space drift monitoring, keeping MAE within ±0.3 µm of baseline (Bellemo et al., 2025). Consequently, explainability toolkits, such as latent-space traversal videos, counterfactual heatmaps, and variance decomposition dashboards, must be integrated into the clinician’s user interface—and their utility is empirically validated: Wang et al. (2025) showed that dermatologists’ trust in VAE-based thickness measurements increased from 62% to 87% when provided with latent traversal visualizations (comparing “actual vs. simulated thickness changes”), versus only 41% trust in U-Net results with Grad-CAM heatmaps (which often highlighted speckle rather than biological edges). These toolkits must also include versioned audit logs that satisfy ISO 13485 and future EU AI Act requirements.

Moreover, fairness poses a parallel, equally urgent gap. Meta-analyses of 8 major skin OCT datasets (Ward et al., 2024) confirm severe phototype imbalance: Fitzpatrick phototypes I–III account for 78%–85% of samples, while phototypes IV–VI represent only 5%–12%. This bias translates to measurable performance degradation: Liu et al. (2025) reported that even state-of-the-art VAEs show a 2.3× higher MAE (4.4 µm vs. 1.9 µm) on phototype VI versus phototype I skin, due to reduced OCT contrast from higher melanin content. A Northwestern study (Ward et al., 2024) further demonstrated that “fair-AI” pipelines without phototype-aware training misclassified SC boundaries in 25% of phototype V–VI patients, compared to 3% in phototype I–II. Thus, any next-generation VAE must incorporate bias-monitoring hooks, such as domain-adversarial heads, class-conditional calibration curves, and phototype-aware uncertainty flags—and these hooks have proven efficacy: a federated-trained VAE with domain-adversarial heads reduced phototype-related error disparity by 40% (from 2.5 µm to 1.5 µm) across phototypes I–VI, compared to a non-adversarial baseline (Ward et al., 2024). Developers must also commit to federated or crowdsourced data collection campaigns that balance age, ethnicity, and disease prevalence—as demonstrated by a JAMA Network Open study (Ward et al., 2024) where crowdsourced recruitment added 32% phototype IV–VI samples to the UIUC CNN-GS-skin dataset, cutting model bias by 28%.

Finally, the long-term vision extends beyond episodic scans to continuous, personalised simulation. Emerging dermatologic digital-twin platforms already ingest longitudinal OCT, microbiome profiles, and environmental exposure logs to forecast flare-ups and optimize skincare regimens in silico; plug-in SC-thickness modules could act as a high-resolution “vital sign” feeding those twins real-time barrier data. Coupling twins with adaptive treatment engines, topical dosing algorithms, and dynamic UV-protection coaching would transform today’s reactive dermatology into a predictive and preventive discipline, but only if the preceding challenges of portability, fusion, explainability, and fairness are addressed in concert—building on empirically validated solutions (e.g., federated learning for bias reduction, latent traversal for explainability) rather than theoretical frameworks.

The quantification of stratum corneum (SC) thickness has undergone a transformative evolution—from labor-intensive, operator-dependent manual caliper measurements and shrinkage-prone histology to rapid, high-precision algorithms embedded in portable optical coherence tomography (OCT) systems. Today’s state-of-the-art pipelines, which integrate physics-informed digital signal processing (DSP) (e.g., adaptive speckle filtering, roll-off compensation) with label-efficient generative models (β-VAEs, VQ-VAEs), deliver sub-2 µm boundary accuracy, run at video rates (50 fps) on consumer-grade CPUs, and provide interpretable latent-space insights (e.g., thickness traversal, uncertainty mapping) that were unimaginable a decade ago. These advances have positioned SC thickness as an actionable biomarker for inflammatory disease monitoring, cosmetic efficacy testing, and transdermal drug delivery optimization—bridging the gap between preclinical research and real-world clinical care. Yet significant barriers remain before these technologies achieve widespread adoption, and addressing these limitations requires targeted innovations in model design, regulatory compliance, and multimodal integration.

The technical foundation for precision SC mapping is in place, but the next era of innovation will be defined by convergence: miniaturized, low-power OCT probes (<1 kg, <10 W) paired with edge-deployable cross-modal VAEs; federated datasets that represent the full diversity of human skin; and regulatory frameworks that balance innovation with patient safety. When realized, these advances will redefine dermatology, enabling at-home monitoring of eczema, personalized cosmetic formulations tailored to individual SC thickness, and burn triage in resource-limited settings where histology is unavailable. Ultimately, the goal is not just to measure SC thickness, but to use it as a gateway to proactive, personalized skin health management, turning reactive care into predictive, preventive practice.