An explainable deep learning framework for classification of eczema and scabies using a multiregional dermatological image dataset

Background Eczema and scabies are frequently misdiagnosed due to overlapping clinical and visual features, a challenge amplified in resource-limited settings with limited access to dermatologists. Accurate differentiation is essential, as treatment strategies differ significantly. We evaluated a deep learning framework for differentiating eczema and scabies using images aggregated from multiple regional dermatology datasets. Materials and methods A dermatologist-annotated dataset was curated from three geographically distinct sources: DermNet (Western population), PASSION MICCAI 2024 (African population), and DermaCon-IN (Indian population), comprising 2,555 training images and a single holdout validation set of 269 images. This holdout validation set was used for both model checkpoint selection and final performance reporting; no independent test set, held out from model selection, was created. Five state-of-the-art deep learning architectures, namely, ResNet-50, ConvNeXt-Tiny, EfficientNet-B0, Vision Transformer (ViT-B/16), and Data-Efficient Image Transformer (DeiT), were fine-tuned for binary classification. In training, data augmentation was used to increase image variability and avoid overfitting. Model performance was evaluated using accuracy, precision, recall, F1-score, and validation loss. Each architecture was trained in a single run without repeated seeds, confidence intervals, or formal statistical significance testing; consequently, this study was not designed or powered for statistical inference between architectures, and all comparative results should be interpreted as descriptive. Grad-CAM++ was used to interpret model decision-making. Results In this single-run, descriptive comparison using a shared protocol, the five configurations achieved accuracies between 92.94% and 96.65%, with ResNet-50 achieving an accuracy and weighted F1-score of 96.65%. Because these values were obtained on the same holdout validation set used for checkpoint selection, they are optimistically biased and do not constitute independent estimates of generalization performance. Conclusion In the current experimental setup, all five evaluated architectures showed good descriptive performance in differentiating eczema from scabies on the compiled multisource set, as a preliminary single-run observation under a shared protocol. The results, however, need to be externally validated and the evaluation conducted with stratification of skin tones before a generalization of the results can be made to other clinical populations. Incorporating explainable AI ensures clinical interpretability, enhancing trust in automated diagnostic tools. Future work should expand dataset diversity, include prospective clinical validation, and explore advanced explainability techniques to improve practical applicability.

Authors

Institutions

Publication Details

Journal
Frontiers in Medicine
Published
2026-11-17
DOI
https://doi.org/10.3389/fmed.2026.1910115
Primary Topic
Dermatology and Skin Diseases
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An explainable deep learning framework for classification of eczema and scabies using a multiregional dermatological image dataset

Dildar Hussain, Razaullah Khan, Sarwar Shah Khan, Jawad Khan et al.
3 links
Frontiers in Medicine
Dermatology and Skin Diseases
article

An explainable deep learning framework for classification of eczema and scabies using a multiregional dermatological image dataset

Dildar Hussain, Razaullah Khan, Sarwar Shah Khan, Jawad Khan, Yeong Hyeon Gu, Muneer Ahmed, Muhammad Yousaf, Muhammad Junaid, Zain Ul Aabidin
article en

Abstract

Background Eczema and scabies are frequently misdiagnosed due to overlapping clinical and visual features, a challenge amplified in resource-limited settings with limited access to dermatologists. Accurate differentiation is essential, as treatment strategies differ significantly. We evaluated a deep learning framework for differentiating eczema and scabies using images aggregated from multiple regional dermatology datasets. Materials and methods A dermatologist-annotated dataset was curated from three geographically distinct sources: DermNet (Western population), PASSION MICCAI 2024 (African population), and DermaCon-IN (Indian population), comprising 2,555 training images and a single holdout validation set of 269 images. This holdout validation set was used for both model checkpoint selection and final performance reporting; no independent test set, held out from model selection, was created. Five state-of-the-art deep learning architectures, namely, ResNet-50, ConvNeXt-Tiny, EfficientNet-B0, Vision Transformer (ViT-B/16), and Data-Efficient Image Transformer (DeiT), were fine-tuned for binary classification. In training, data augmentation was used to increase image variability and avoid overfitting. Model performance was evaluated using accuracy, precision, recall, F1-score, and validation loss. Each architecture was trained in a single run without repeated seeds, confidence intervals, or formal statistical significance testing; consequently, this study was not designed or powered for statistical inference between architectures, and all comparative results should be interpreted as descriptive. Grad-CAM++ was used to interpret model decision-making. Results In this single-run, descriptive comparison using a shared protocol, the five configurations achieved accuracies between 92.94% and 96.65%, with ResNet-50 achieving an accuracy and weighted F1-score of 96.65%. Because these values were obtained on the same holdout validation set used for checkpoint selection, they are optimistically biased and do not constitute independent estimates of generalization performance. Conclusion In the current experimental setup, all five evaluated architectures showed good descriptive performance in differentiating eczema from scabies on the compiled multisource set, as a preliminary single-run observation under a shared protocol. The results, however, need to be externally validated and the evaluation conducted with stratification of skin tones before a generalization of the results can be made to other clinical populations. Incorporating explainable AI ensures clinical interpretability, enhancing trust in automated diagnostic tools. Future work should expand dataset diversity, include prospective clinical validation, and explore advanced explainability techniques to improve practical applicability.

Frontiers in MedicineVol. 13
Gachon University (KR), Sejong University (KR), Iqra National University (PK), University of Swat (PK)
Good health and well-being
Dermatology and Skin Diseases
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon