DermaAI: Deep Learning-Based Skin Disease Classification and Detection

Abstract

Skin diseases represent a vast proportion of primary care visits, and their prompt and correct diagnosis may necessitate expertise on the side of a specialist. Artificial intelligence has potential to help screening skin diseases through analyzing skin lesion images. In this paper, the author will discuss the implementation of DermaAI, a multi-class skin disease detection system based on integrated deep-learning. We have prepared a source labeled dataset of dermatological pictures (cz/dermatitis, psoriasis, fungi, acne, rosacea, and vitiligo) and trained an EfficientNetB0 convolutional neural network (CNN) with transfer learning. This model was trained in two stages: one with training the newly-added layers only, and second fine-tuning the entire network, and the methods include data augmentation, class re-sampling, and learning rate adjustment using the form of a callback. It was implemented through an offline inference application using TensorFlow Lite and a Flask web interface and Android application. DermaAI has high accuracy (around 92 percent on a held-out test set) on classifying the target conditions. An error analysis (with confusion matrix) indicates that the error rate is strong on all classes with most confusion made among clinically similar categories. The system gives probabilities per class to help with the estimation of confidence. These findings indicate that DermaAI may be used as an effective dermatological tool, particularly in resource-constrained environments with a low number of specialists available.

Introduction

Skin diseases are highly prevalent, accounting for a large share of primary care visits, yet accurate diagnosis often requires dermatologists who are scarce in rural and under-resourced regions. This leads to delayed treatment and misdiagnosis. To address this, the study proposes DermaAI, an AI-based system that uses deep learning to classify multiple skin diseases from images.

DermaAI is built using an EfficientNetB0 CNN with transfer learning, trained on a multi-class dataset of skin conditions such as eczema, psoriasis, fungal infections, acne, rosacea, and others. The system optionally includes lesion segmentation and outputs disease predictions along with confidence scores. It is deployed both as a Flask-based web application and an Android mobile app (TensorFlow Lite), enabling online and offline use.

The system pipeline includes image preprocessing, CNN-based classification, Softmax probability estimation, and result visualization. To handle class imbalance, techniques like data augmentation, oversampling, and weighted loss were used. Training used transfer learning in two stages (freezing and fine-tuning) with Adam optimization and early stopping.

DermaAI achieved strong performance, with about 94% validation accuracy and 92% test accuracy, showing reliable classification across most disease categories. Errors mainly occurred between visually similar conditions (e.g., psoriasis vs. eczema), which reflects real-world diagnostic challenges. Overall, the system demonstrates that an explainable, mobile-accessible deep learning model can support early and accessible skin disease detection, especially in areas with limited dermatological care.

Conclusion

We have introduced the DermaAI, a complete deep learning framework of skin diseases identification, which includes a high-performance CNN (EfficientNetB0) along with easy-to-access deployment. With transfer learning and trained on a curated multi-class dataset, our model attained high accuracy (~92) in the classification of the common dermatoses. The implementation on the system is done in two programs (both web and mobile) hence wide applicability. We believe that, based on our assessment, the DermaAI will be able to deliver accurate and timely disease predictions and probability, which can be used to assist with early screening and triage. Although it is not intended to replace clinical judgment, this tool has shown how AI can contribute to the dermatological care, particularly in underserved areas. The architecture of DermaAI (the integration of the segmentation, the confidence scoring, and the delivery on cross-platform) covers most of the gaps evident in other solutions. Our work will in the future validate the system in the clinical setting and keep on the further improvements of its performance and scope. We hope that DermaAI can become a successful example of AI-based healthcare support, which will allow making dermatology available to anyone on the tip of the hat in the hallmark of the healthcare inquiry.

References

[1] “The Use of Artificial Intelligence for Skin Disease Diagnosis in Primary Care Settings: A Systematic Review.” [Online]. Available:https://pubmed.ncbi.nlm.nih.gov/38921305/ [2] “AI-Skin: Skin Disease Recognition Based on Self-Learning and Wide Data Collection Through a Closed-Loop Framework.” [Online]. Available: https://www.sciencedirect.com/science/article/abs/pii/S1566253519300867 [3] “AI-Based Localization and Classification of Skin Disease With Erythema.” Scientific Reports. [Online]. Available: https://www.nature.com/articles/ s41598-021-84593-z [4] “Building Powerful Image Classification Models Using Very Little Data.” [Online]. Available: https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html [5] O. Karunarathne, “A Deep Dive Into Oversampling and Undersampling for Class Imbalance in Image Classification.” [Online]. Available: https://medium.com/@okeshakarunarathne/a-deep-dive-into-oversampling-and-undersampling-for-class-imbalance-in-image-classification-2e031d8e2b00 [6] “Image Classification via Fine-Tuning With EfficientNet.” [Online]. Available:https://keras.io/examples/vision/image_classification_efficientnet_ fine_tuning/ [7] “Transfer Learning and Fine-Tuning.” TensorFlow Core. [Online]. Available: https://www.tensorflow.org/tutorials/images/transfer_learning [8] “ReduceLROnPlateau.” Keras API Documentation. [Online]. Available: https://keras.io/api/callbacks/reduce_lr_on_plateau/ [9] “EarlyStopping.” Keras API Documentation. [Online]. Available: https://keras.io/api/callbacks/early_stopping/ [10] N. Ding et al., “Differential Diagnosis of Eczema and Psoriasis Using Routine Clinical Data and Machine Learning: Development of a Web-Based Tool in a Multicenter Outpatient Cohort,” Frontiers in Medicine, vol. 12, 2025.

Copyright

Copyright © 2026 Suyash Modi, Yogendra Kumar Maurya, Manoj Kumar Dixit. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.