HEEPidermis: Convergent Epidermal Sensing

• HEEPidermis is a convergent research theme that spans imaging pipelines, bioinstrumentation, and histological workflows for epidermal analysis.
• In the imaging context, it leverages deep learning models like VGG16 and ResNet18 with manual annotations to estimate epidermal thickness with high accuracy.
• The mixed-signal SoC version enables programmable bioimpedance measurements and on-chip processing, offering a versatile platform for epidermal sensor integration.

HEEPidermis is a name used in recent arXiv literature for more than one epidermis-centered research construct. In one usage, it denotes a non-invasive imaging pipeline for estimating epidermal thickness from skin images by combining a VGG16-based heat-map generator, manual Roboflow annotations, a ResNet18 classifier, and pixel-to-millimeter conversion (GR et al., 2024). In another, it denotes a mixed-signal System-on-Chip for biological impedance recording that integrates dual current DACs, dual VCO-based ADCs, a RISC-V CPU, and event-based sub-sampling (Sapriza et al., 3 Sep 2025). The term also appears in relation to reconstructed human epidermis workflows for H-DAB-stained histology, where the EpidermaQuant pipeline is recommended for integration into HEEpidermis workflows (Zamojski et al., 2024). The current literature therefore uses the label across measurement, analysis, and instrumentation settings rather than for a single canonical method.

1. Terminological scope and research contexts

The name appears in at least three technically distinct but adjacent contexts.

Usage of the term	Core technical content	Source
HEEPidermis pipeline	VGG16 heat maps, Roboflow masks, ResNet18, epidermal thickness estimation	(GR et al., 2024)
HEEPidermis SoC	2× 8-bit iDACs, 2× VCO-based ADCs, RISC-V CPU, dLC event sub-sampler	(Sapriza et al., 3 Sep 2025)
HEEPidermis workflows	Integration context for EpidermaQuant in reconstructed human epidermis analysis	(Zamojski et al., 2024)

In the imaging usage, the focus is wounded-skin layer recognition and thickness quantification. In the histological usage, the focus is unsupervised detection and quantification of DAB-based immunohistochemical staining in reconstructed human epidermis. In the hardware usage, the focus is programmable BioZ acquisition with on-chip feature extraction and closed-loop capability. This distribution of meanings suggests that HEEPidermis functions as a systems-level label spanning epidermal phenotyping, epidermal image analysis, and epidermal bioinstrumentation rather than as a narrowly delimited algorithmic object.

2. Non-invasive epidermal thickness estimation pipeline

The HEEPidermis imaging pipeline described in the detailed summary of Devakumar et al. operates on approximately 200 labeled images of skin distributed across five clinical classes: scar1, scar2, wound, healthy_men, and healthy_women (GR et al., 2024). A pretrained VGG16, used off-the-shelf, generates a single-channel heat map $H(x,y)$ by removing the final fully connected layers and projecting the last convolutional feature map through a $1\times1$ convolution followed by ReLU. These heat-map images are imported into Roboflow and manually annotated with three masks corresponding to the stratum corneum, epidermis, and dermis. The resulting masks are combined into a 3-channel input

$$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$

Classification is then performed with a standard 18-layer residual network with 4 residual blocks, global-average pooling, an 84-unit fully connected layer, and Softmax output for five classes. The forward model is given as $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$, with categorical cross-entropy

$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$

Training uses Adam with $\beta_1=0.9$, $\beta_2=0.999$, initial learning rate $\eta_0=10^{-3}$, and early stopping with patience 10 epochs (GR et al., 2024).

Thickness estimation proceeds columnwise. If $p_{\mathrm{epi}}(x)$ denotes the number of epidermis pixels in column $x$ and $1\times1$0 the known scale factor in mm/pixel, then

$1\times1$1

and the average thickness is

$1\times1$2

Over $1\times1$3 test images, the reported mean and standard deviation are $1\times1$4 mm and $1\times1$5 mm, with a 95% confidence interval of $1\times1$6 mm under approximate normality (GR et al., 2024).

Performance is highly learning-rate dependent. On the held-out 20% split, the epidermis-class results for ResNet18 peak at $1\times1$7 with 96.1% accuracy, 0.97 precision, 0.95 recall, and 0.96 F1-score. The same detailed summary reports an overall accuracy of 97.67%, while the abstract reports that both EfficientNet and ResNet18 attained accuracy rates of almost 95.35%, and that each reached 95.35% at learning rate 0.0001 (GR et al., 2024). This makes the pipeline noteworthy less for architectural novelty than for the specific sequencing of interpretability-oriented preprocessing, manual layer localization, supervised classification, and direct geometric readout.

3. Relation to reconstructed human epidermis histology

A separate but closely related computational pipeline is EpidermaQuant, which targets unsupervised detection and quantification of epidermal differentiation markers on H-DAB-stained images of reconstructed human epidermis (Zamojski et al., 2024). The dataset contains 598 unannotated JPEG images of size 1936×1460 px, acquired at 40× magnification with a ZEISS AXIOCAM 503, showing formalin-fixed, paraffin-embedded cross-sections of in vitro reconstructed human epidermis (HaCaT-based RHE). The marker distribution is explicitly given as 189 filaggrin images, 131 keratin 10 images, 96 Ki67 images, and 182 HSPA2 images, with positive and negative-control counts specified for each marker except Ki67, which has 96 positive and 0 negative images (Zamojski et al., 2024).

The pipeline consists of five main stages: color normalization, color deconvolution, morphological background detection, automatic image rotation, and clustering-based marker detection. The most effective combination is Reinhard’s normalization, the Ruifrok and Johnston color deconvolution method, the proposed image rotation method based on boundary distribution of image intensity, and k-means clustering using DAB stain intensity (Zamojski et al., 2024). In Reinhard normalization, the per-channel transformation is

$1\times1$8

while Ruifrok–Johnston deconvolution is based on optical density,

$1\times1$9

followed by concentration recovery through $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$0.

Background segmentation uses Otsu-thresholded hematoxylin, a disk structuring element of radius 15 px, opening, closing, hole filling, selection of the largest connected component, and a second round of opening/closing for edge refinement. Automatic rotation searches $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$1 in $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$2 steps and selects the angle that maximizes the rowwise sum of mask pixels after rotation (Zamojski et al., 2024). DAB-negative images are excluded using the Average Proportion

$$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$3

ROC analysis on the FLG dataset yields $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$4, and images with $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$5 are discarded. Negative controls were excluded automatically above 99% specificity at $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$6 (Zamojski et al., 2024).

The final readout is the percentage of DAB-stained tissue per image. Median values for positive samples are 4.1% for FLG, 2.5% for Ki67, 25.3% for HSPA2, and 39.5% for K10 (Zamojski et al., 2024). No pixel-wise ground truth is provided, so classical supervised metrics such as accuracy, precision, recall, and IoU against a manual gold standard were not reported. The significance of this work for HEEPidermis is operational: it supplies an unsupervised histological quantification layer for reconstructed epidermis, and the authors explicitly recommend its incorporation into HEEPidermis workflows.

4. HEEPidermis as a BioZ System-on-Chip

In the hardware literature, HEEPidermis is explicitly the title of a mixed-signal SoC for tissue impedance measurement (Sapriza et al., 3 Sep 2025). The architecture is composed of three major subsystems: a CVE32E20 RISC-V CPU with instruction and data SRAMs, dual 8-bit arbitrary-signal current DACs, and dual VCO-based ADCs. Data movement is handled by two independent DMA engines, while a digital Level-Crossing block implements event-based sub-sampling. The digital back-end is based on the open-source X-HEEP platform, and both the digital back-end and behavioral models of the analog front-end are open-source (Sapriza et al., 3 Sep 2025).

The current-generation path uses a binary-weighted current mirror with $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$7 nA and an output range of 0–10 µA in 40 nA steps. The ideal output relation is

$$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$8

Post-layout linearity at the worst FF corner at 50 °C is reported as DNL $$I_{\mathrm{in}}(x,y,c) = [H(x,y)\cdot M_{\mathrm{stratum}},\; H(x,y)\cdot M_{\mathrm{epi}},\; H(x,y)\cdot M_{\mathrm{dermis}}].$$9 LSB and INL $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$0 LSB. Arbitrary waveform generation is supported via periodic DMA updates up to 200 kHz, limited by iDAC settling (Sapriza et al., 3 Sep 2025).

The sensing path uses 31-stage CMOS ring-oscillator ADCs with approximate voltage-to-frequency characteristic

$f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$1

with $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$2 kHz at 408 mV, $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$3 kHz at 800 mV, and $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$4 kHz/mV. Digitization uses a non-reset 26-bit counter, with one tap per stage yielding 6 bits of phase information that can be added for improved resolution. Effective resolution is stated as up to 12 bits over DC–10 kHz bandwidths, and the reported SQNR is 72 dB for a 1 kHz sine and 65 dB for a 10 kHz sine (Sapriza et al., 3 Sep 2025).

On-chip computation supports resistive-mode extraction using $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$5, spectroscopy mode with two phase-shifted sines, threshold-based compression through the dLC block, and further run-length or packet encoding on the CPU. The compression ratio is defined as

$f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$6

where $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$7 counts samples satisfying the level-crossing condition. Reported post-layout operating points include 64 µW total power for GSR measurement at 0.28 µA DC and 2 sps, and 72 µW for impedance spectroscopy at 5.12 µA peak and 20 sps (Sapriza et al., 3 Sep 2025). Total die area is 0.81 mm², with analog front-end plus reference and LDO occupying approximately 17% of core area and the digital back-end approximately 83%.

Within the supplied comparison excerpt, HEEPidermis supports 2× DC–200 kHz at 40 nA–10 µA arbitrary waveform injection, 2× 12-bit 0.2 mHz–10 kHz VCO-ADC acquisition, on-chip RISC-V processing, and event-based decimation at 72 µW acquisition. The comparison entry for MAX30001G is 1× DC–131 kHz at 55 nA–96 µA square-wave drive, 1× 20-bit $f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$8 ADC at 25–64 Hz, and 158 µW (Sapriza et al., 3 Sep 2025). In that sense, HEEPidermis is positioned as a programmable research platform rather than as a fixed-purpose impedance front-end.

5. Mechanical and electronic context around epidermal systems

The broader epidermal research environment in which the term is used includes explicit modeling of epidermal mechanics. Kobayashi et al. present a particle-based model of self-replicating cells on a deformable dermis–basement-membrane substrate and show that the model reproduces dermal papilla formation and preferential stem-cell localization on papilla tips (Kobayashi et al., 2018). The model distinguishes dermal particles, membrane particles, stem cells, and transit-amplifying cells; assumes overdamped dynamics; and defines the total potential as

$f_\theta(I_{\mathrm{in}})=\hat y\in\mathbb{R}^5$9

A coarse-grained one-dimensional envelope equation,

$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$0

is used to explain why differential adhesion, rather than buckling alone, drives both outward papilla formation and stem-cell tip preference (Kobayashi et al., 2018). The reported steady growth of $$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$1, $$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$2, and $$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$3 to positive values of approximately 0.2–0.5 supports this interpretation.

A second contextual strand is epidermal electronics. “Skin Controlled Electronic and Neuromorphic Tattoos” shows that the hypodermis can serve as an electrolyte gate for 2D-material devices, with an electrical double-layer capacitance of approximately 1.4 µF/cm² and mobility extraction from

$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$4

Reported skin- or gel-gated mobilities are up to 6500 cm²·V$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$5·s$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$6 for graphene transistor tattoos, approximately 30 cm²·V$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$7·s$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$8 for MoS$$L(\theta) = - \sum_{i=1}^5 y_i \cdot \log(\hat y_i).$$9, and approximately 1 cm²·V$\beta_1=0.9$0·s$\beta_1=0.9$1 for PtSe$\beta_1=0.9$2 (Kireev et al., 2024). With Nafion 211 beneath graphene tattoos, the devices exhibit neuromorphic behavior, including ramp tests with 20 positive pulses at $\beta_1=0.9$3 µA and $\beta_1=0.9$4 ms, write-retention experiments with 50 write pulses at 1 µA and 100 ms width, endurance beyond 40 alternating-polarity pulse cycles, and per-write energy of approximately 5 nJ (Kireev et al., 2024).

These two lines of work are not themselves the HEEPidermis SoC or the HEEPidermis imaging pipeline. They nonetheless define a relevant systems context: one provides a mechanistic account of epidermal–dermal geometry and stem-cell patterning, while the other provides a skin-mediated interface model for epidermal electronics. The term “HEEPidermis applications” is explicitly used in the tattoo summary to denote wearable neuromorphic processors, health monitoring and closed-loop therapy, and prosthetic or BCI control (Kireev et al., 2024).

6. Limitations, misconceptions, and likely convergence points

A common misconception would be to treat HEEPidermis as a single standardized platform. The supplied literature does not support that reading. Instead, the name spans at least a wound-imaging pipeline, a BioZ SoC, and a workflow context for reconstructed epidermis histology (GR et al., 2024, Sapriza et al., 3 Sep 2025, Zamojski et al., 2024).

The image-analysis branches have explicit limitations. EpidermaQuant reports occasional false positives when DAB-colored medium or brown artifacts join tissue, states that DAB intensity is only semi-quantitative because there is no strict quantitative stain–antigen stoichiometry, and notes that the absence of pixel-level ground truth limits performance benchmarking (Zamojski et al., 2024). The wound-thickness pipeline is non-invasive, but its operation depends on manual Roboflow annotations, retraining across learning rates, and uncertainty arising from inter-annotator variability, classification errors, and imaging noise (GR et al., 2024). These constraints mean that the most reliable interpretation of current outputs is as experimentally useful quantitative support rather than as definitive histometric truth.

The SoC branch is open and programmable, but its reported key numbers are post-layout figures and its analog front-end is exposed to the reader through behavioral models rather than through a general-purpose software stack alone (Sapriza et al., 3 Sep 2025). The architecture is therefore best understood as an instrumentation substrate for rapid system-level simulations, repurposing, and closed-loop experimentation rather than as a complete end-user product.

A plausible implication is that the three branches can be combined into a multiscale epidermal measurement stack. The histology pipeline supplies ex vivo marker quantification in reconstructed human epidermis, the wound-imaging pipeline supplies non-invasive geometric estimates of epidermal thickness, the SoC supplies low-power embedded BioZ acquisition and on-chip control, the mechanics model supplies a formal account of epidermal–dermal morphogenesis, and skin-controlled tattoos supply a route to conformal interfaces and neuromorphic signal transduction (Zamojski et al., 2024, GR et al., 2024, Sapriza et al., 3 Sep 2025, Kobayashi et al., 2018, Kireev et al., 2024). This suggests that “HEEPidermis” is evolving less as a single named artifact than as a convergent research theme centered on epidermal sensing, quantification, and closed-loop interaction.