eSkinHealth: AI Teledermatology & Datasets

• eSkinHealth is a digital dermatology field focusing on AI-assisted teledermatology, remote skin assessments, and multimodal datasets for underrepresented populations.
• It combines clinical information systems, edge/cloud learning loops, and selfie-based mobile applications to enhance diagnostic support and longitudinal monitoring.
• Innovative computational pipelines, including transfer learning, self-supervised learning, and fairness-aware models, drive improved accuracy and bias mitigation.

Search arXiv for papers on eSkinHealth dermatology AI datasets fairness teledermatology

eSkinHealth denotes a strand of digital dermatology research concerned with remote skin assessment, longitudinal monitoring, and AI-assisted diagnostic support across consumer devices, clinical platforms, and underrepresented populations. In the cited literature, the name appears in two closely related senses: as a class of AI-assisted tele-dermatology or diagnostic support applications, and as a multimodal dermatology dataset for Neglected Tropical Diseases (NTDs) collected in West Africa. The resulting research landscape spans web-based clinical information systems, edge/cloud closed-loop learning architectures, selfie-based severity assessment, fairness-aware and privacy-preserving learning, wearable and tactile sensing, and multimodal datasets with lesion masks, captions, and clinical concepts (Alharthi et al., 2018, Chen et al., 2019, Wang et al., 26 Aug 2025).

1. System scope and principal forms

The literature presents eSkinHealth as an umbrella domain rather than a single monolithic system. A prototypical early infrastructure example is the web-based Skin Cancer Information System, designed to store patient health records, diagnoses, treatments, forms, reports, images, and pathology material in a three-tier architecture using web/application/database server components (Alharthi et al., 2018). A second line of work frames eSkinHealth as a closed-loop medical AI service: the AI-Skin framework organizes user terminals, edge nodes, a cloud platform, and remote medical sites into a continuous loop for data collection, filtering, model updating, and clinician feedback (Chen et al., 2019). A third line operationalizes eSkinHealth as direct patient-facing mobile analysis, such as selfie-based acne severity scoring deployed through a mobile application and delivered through a Python Flask API on Azure Container/Kubernetes Service (Zhao et al., 2019). A fourth line treats eSkinHealth as a data resource for global dermatology, most explicitly in the multimodal dataset collected in Côte d’Ivoire and Ghana (Wang et al., 26 Aug 2025).

Form	Core function	Representative source
Clinical information system	Record diagnosis, treatment, reports, and images	(Alharthi et al., 2018)
Closed-loop AI service	User–edge–cloud–clinician feedback loop	(Chen et al., 2019)
Selfie-based mobile application	Real-time remote acne assessment	(Zhao et al., 2019)
Multimodal dataset	NTD-focused images, masks, captions, concepts	(Wang et al., 26 Aug 2025)

This suggests that eSkinHealth is best understood as a convergence zone between teledermatology, medical AI, and digital health infrastructure. The common denominator is not a single algorithmic stack, but the attempt to support skin assessment outside conventional specialist workflows while preserving clinical structure, longitudinal traceability, and population coverage.

2. Data regimes, annotation strategies, and representational scope

A defining feature of eSkinHealth research is the shift from narrow curated corpora toward datasets with richer metadata, broader condition coverage, and clinically meaningful intermediate labels. The most direct example is the multimodal dataset "eSkinHealth," which contains 5,623 images from 1,639 cases, spans 47 skin diseases, and focuses on skin NTDs and rare conditions in West African populations. Beyond diagnosis labels and patient metadata, it includes semantic lesion masks, instance-specific visual captions, and 69-dimensional clinical concept vectors, with multimodal annotations generated through an AI-expert collaboration paradigm involving GPT-o1 for captions/concepts and SAM for lesion masking under dermatologist guidance. On a randomly selected 10% subset of image-caption-concept pairs, 84% received a rating of 3 (good) or greater on a five-level alignment scale (Wang et al., 26 Aug 2025).

Two adjacent datasets clarify why this annotation expansion matters. SkinCon contributes 3,230 images from Fitzpatrick17k and 656 images from the Diverse Dermatology Images dataset, with 48 clinical concepts, of which 22 have at least 50 images in Fitzpatrick17k; validation by additional board-certified dermatologists showed >94% inter-expert agreement (Daneshjou et al., 2023). SCIN, created by crowdsourcing dermatology images through Google Search advertisements, provides 10,408 images from 5,033 contributions collected over eight months. It reports that over 97.5% of submissions were genuine images of skin conditions, and includes dermatologist condition labels plus estimated Fitzpatrick Skin Type and Monk Skin Tone labels. Its condition mix is dominated by short-duration, allergic, infectious, or inflammatory presentations rather than the neoplasm-heavy distribution common in legacy dermatology datasets (Ward et al., 2024).

These resources change the technical basis of eSkinHealth research in three ways. First, they widen disease scope beyond melanoma and dermoscopy, especially toward common inflammatory disease and NTDs. Second, they expose models to real-world image variability and diverse skin tones. Third, they make it possible to supervise models with clinical concepts, captions, and masks rather than only class labels, which in turn supports debugging, concept bottlenecks, vision-language learning, and region-aware analysis.

3. Computational pipelines for image-based assessment

The algorithmic core of eSkinHealth systems is heterogeneous, but several recurring workflows can be identified. In AI-Skin, images and contextual data are collected at user terminals, filtered at the edge, and selectively transmitted to the cloud. The filtering stage is information-entropy based: for unlabeled image $x_i^u$, the framework computes

$$E\left(p_{x_i^u}\right) = -\sum_{j=1}^{M} p_i^{j} \log\left(p_i^{j}\right),$$

with transmission conditioned on a threshold $E_T$. The same system exposes an external algorithm load module and empirically compares LeNet-5, AlexNet, and VGG16; AlexNet is reported as both the best overall performer and sufficiently fast for real-time operation in the prototype (Chen et al., 2019).

A more specialized selfie pipeline appears in facial acne assessment. There, 4,700 selfie images were acquired and labeled by 11 internal dermatologists into five severity categories. Facial landmarks from OpenFace and fallback localization with OpenCV’s One Eye model were used to extract 2–4 facial skin patches per image. Features were produced by a fixed ResNet 152 pretrained on ImageNet, followed by a 3-layer fully connected neural network with 1024, 512, and 256 neurons, trained as a regression model. To reduce spatial sensitivity, the work introduced image rolling with

$\text{roll\_size} = \frac{X}{N}.$

On a 230-image golden test set, RMSE improved from 0.72 without image rolling to 0.482 with image rolling augmentation, compared with a mean-guessing baseline of 0.78, and the model’s correlation with consensus dermatologist ratings reached 0.755 (Zhao et al., 2019).

Limited-label settings motivate a different pipeline for eczema severity. The self-supervised framework in "Automated Measurement of Eczema Severity with Self-Supervised Learning" uses SegGPT for few-shot segmentation and DINO features from the segmented region, passed to an MLP with a 128-unit hidden layer and 0.3 dropout for 4-class severity grading. On 528 annotated “in-the-wild” eczema images, the method achieved Weighted F1: 0.67 ± 0.01, outperforming finetuned Resnet-18 (0.44 ± 0.16) and Vision Transformer (0.40 ± 0.22) (Kumar et al., 21 Apr 2025).

Dense feature localization has also been addressed directly. "Automatic Facial Skin Feature Detection for Everyone" annotated 3,755 selfie images spanning multiple ethnicities, skin tone colors, severity levels, age groups, and lighting conditions, and trained Unet++ for acne, pigmentation, and wrinkle detection. The improved model achieved IoU values of 0.2300 for acne, 0.3035 for pigmentation, and 0.1512 for wrinkles, surpassing both preliminary models and traditional baselines (Zheng et al., 2022).

Taken together, these pipelines show that eSkinHealth does not rely on a single canonical model family. Transfer learning, few-shot segmentation, self-supervised representation learning, patch-based regression, and dense segmentation all appear, with the choice driven primarily by label granularity, deployment constraints, and whether the target task is diagnostic classification, severity scoring, or localized lesion/feature quantification.

4. Fairness, bias mitigation, and interpretability

Fairness is a central technical issue in eSkinHealth because dermatology datasets and generative models are often imbalanced across skin tone and disease categories. "FairSkin" identifies a twofold bias in diffusion-based synthetic skin-disease generation: image quality for African and non-Caucasian individuals is worse, and downstream models learn critical features less effectively from minority skin tones. To address this, FairSkin introduces a three-level resampling mechanism covering training data resampling, balanced diffusion training, and downstream reweighting. Quantitatively, the FID variance among ethnicities dropped from 603.74 in vanilla DM to 327.11 in FairSkin; African skin image FID dropped from 126.22 to 114.50; Demographic parity improved from 25.04 to 9.95; ESSP improved from 3.76 to 7.78; and IS improved from 2.39±0.47 to 2.52±0.38 (Zhang et al., 2024).

Transfer learning and domain adaptation provide a second fairness pathway. In "Equitable Skin Disease Prediction Using Transfer Learning and Domain Adaptation," the DDI dataset is used as an equity-oriented benchmark, with additional adaptation on HAM10000 and DeepDerm. Among tested models, MedViT-base was the top performer. Before domain adaptation, it achieved Accuracy: 0.74 and Macro-F1: 0.60 on DDI; after adaptation on HAM10000+DDI, performance improved to Accuracy: 0.76 and Macro-F1: 0.63, while all models improved across skin tones (Dip et al., 2024).

Privacy-preserving fairness has also been studied. The federated-learning framework in "Achieving Fairness in Dermatological Disease Diagnosis through Automatic Weight Adjusting Federated Learning and Personalization" assigns higher aggregation weights to clients with higher loss,

$$w_c = \frac{e^{m \cdot \mathcal{L}_c}}{\sum_{i\in C} e^{m \cdot \mathcal{L}_i}},$$

and then applies post-FL client-specific fine-tuning. On Fitzpatrick 17k, FedAuto reduced accuracy variance to 0.00065 in the in-FL stage versus 0.00461 for FedAvg, improved overall accuracy to 0.643, and raised skin type 6 accuracy from 0.467 to 0.575. After personalization, the average accuracy gap was 0.0238, average overall accuracy 0.734, and final variance 0.00015 (Xu et al., 2022).

Interpretability is addressed most explicitly by SkinCon. Because it supplies concept-level labels such as “plaque,” “scale,” and “erosion,” it supports Concept Activation Vectors, Conceptual Counterfactual Explanations, and post-hoc Concept Bottleneck Models. In one reported experiment, a PCBM using SkinCon matched or exceeded the original black-box model’s AUC on Fitzpatrick III–IV skin, with PCBM AUC = 0.727 versus DeepDerm AUC = 0.632 (Daneshjou et al., 2023). A related but distinct strategy appears in S-SYNTH, which can systematically vary skin color, blood fraction, lesion shape, hair, and lighting, and reports that comparative segmentation trends on synthetic data follow similar trends to real dermatologic images while mitigating underrepresentation in public datasets (Kim et al., 2024).

A common misconception is that more synthetic data automatically yields equitable dermatology AI. The literature does not support that simplification. Synthetic generation can itself encode bias, requiring explicit balancing objectives, fairness metrics, and subgroup-aware evaluation.

5. Continuous monitoring, self-measurement, and nontraditional sensing

A major branch of eSkinHealth emphasizes repeated measurement rather than one-time diagnosis. In cosmetic product trials, smartphone selfies have been used to track skin colour and wrinkle changes between baseline and endpoint instrument measurements. In a study of 12 women aged 30 to 60 years, participants provided up to 3 selfies per week, averaging 8.5 images per participant, while Antera 3D measurements served as ground truth. Among normalization methods, CLAHE gave the best correlation for redness ($A^*$) with $r = 0.86$, and Antera 3D showed a significant 11.2% reduction in maximum wrinkle depth after four weeks ($p = 0.048$) (Smeaton et al., 2020).

Biophysical surrogate estimation from selfies extends this logic. "AI-driven Remote Facial Skin Hydration and TEWL Assessment from Selfie Images" collected data from 336 Chinese panelists, measured 37 facial landmarks per person, and paired patch-level images with skin hydration and TEWL values. Its Skin-Prior Adaptive Vision Transformer combines a texture-adaptive module, band-pass frequency filtering, and position adapters, while a symmetric-based contrastive regularization addresses annotation imbalance. On selfie images, the reported method achieved $R^2 = 0.264$ for TEWL and $R^2 = 0.205$ for SH, outperforming ViT-B and other baselines; for TEWL in few-shot selfie samples, MAE improved from 8.44 with ViT-B to 6.49 with the proposed method (Soh et al., 8 Sep 2025).

Instrumented sensing pushes eSkinHealth beyond RGB imaging. A distributed wireless body area network with lab-on-skin sensors has been proposed for continuous long-term monitoring, where flexible nodes measure parameters including pH level, oxygen level, temperature, bacterial load, bio-potentials, glucose levels, hydration, and heart rate, and dynamically form energy-efficient aggregation trees over abnormal skin areas (Kumar et al., 2021). At a finer spatial scale, a handheld GelSight tactile probe with a custom elastic gel and force sensing achieves a mean absolute error of 12.55 micron on wrinkle-like test objects and reports statistically significant post-moisturizer wrinkle-height reductions at the palm, wrist, and elbow in a study of 15 participants without skin disorders (Padmanabha et al., 14 Sep 2025).

These results broaden the meaning of eSkinHealth from image classification to digital phenotyping. The target variables now include not only diagnosis and lesion class, but also texture, hydration, barrier function, wrinkle depth, and spatiotemporal change.

6. Deployment models, clinical utility, and current limits

The deployment literature prioritizes low latency, portability, and operation on non-ideal images. The AI-Skin prototype reports an Edge Computation Delay (AlexNet) of ~1.2 seconds on average, an end-to-end device ↔ edge delay of around 1 second with standard deviation ~75 ms, and robustness across image resolutions through pre-filtering and adaptive compression (Chen et al., 2019). The all-inclusive smartphone application for melanoma risk assessment implements preprocessing, level-set segmentation, ABCD feature extraction, and SVM classification directly on Android, processing an image in less than a second; on a 200-image public database it reported, with SMOTE, 80% sensitivity, 90% specificity, 88% accuracy, and 0.85 AUC, and without SMOTE, 55% sensitivity, 95% specificity, 90% accuracy, and 0.75 AUC (Kalwa et al., 2022).

Real-world pediatric dermatology introduces additional constraints. The ensemble-learning system ENSEL, developed for managing children with atopic dermatitis from user-captured images, combines Mask R-CNN, YOLOv8, and EfficientNet variants with soft voting. In its second phase, it achieved Precision: 0.85, Recall: 0.98, F1-score: 0.91, and Accuracy: 0.91, with an average response time of 907 ms and most images processed in <1 s (Jeon et al., 28 Nov 2025). This suggests that clinically relevant deployment on noisy, user-generated imagery is feasible when detection, classification, and visualization are co-designed.

The literature is also explicit about limits. The acne mobile application states that it is not intended to replace clinical diagnosis but to supplement early self-assessment, tracking, and engagement (Zhao et al., 2019). The SH/TEWL work reports lower accuracy on unconstrained everyday selfies than on controlled images, and its dataset is limited to Chinese subjects (Soh et al., 8 Sep 2025). The ISIC2019 differential-diagnosis study reports that segmentation with SAM reduced accuracy rather than improving it, indicating that generic preprocessing can be counterproductive for dermatologic morphology (Anaissi et al., 1 Jan 2026). SCIN further shows that skin tone labeling itself can be subjective, with discrepancies between self-reported and expert-rated Fitzpatrick types and slight differences between Monk Skin Tone raters from different countries (Ward et al., 2024).

The likely implication is that future eSkinHealth systems will be judged less by single aggregate accuracy values than by a combination of subgroup robustness, interpretability, latency, multimodal supervision, and compatibility with clinician oversight. In that direction, the West African eSkinHealth dataset is explicitly positioned as a benchmark for disease recognition, vision-language learning, image captioning, concept bottleneck models, segmentation, test-time adaptation, and parameter-efficient fine-tuning in underrepresented dermatologic settings (Wang et al., 26 Aug 2025).