The Detection of Skin Dermatology Disease with an improved Activation Function in Deep Learning

Abstract

Introduction: Skin dermatological diseases are among the most prevalent health concerns worldwide, affecting individuals across various age groups and demographics. Traditional diagnostic methods—such as visual inspection and manual assessment—are often subjective, time-consuming, and prone to inconsistencies due to variability in the examiner’s expertise. These limitations call for more objective and automated approaches to enhance the accuracy and efficiency of dermatological diagnostics

Objectives: This research aims to develop a stable and effective classifier for dermatological diseases using deep learning. The specific objectives include reshaping the dataset, addressing class imbalance using SMOTE, comparing the performance of various deep learning models, and identifying the most accurate model for skin disease classification.

Methods: A balanced dataset consisting of eight classes of skin diseases was used for the study. Data preprocessing techniques were applied, including the use of Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance. Three deep learning models—CNN, VGG16, and EfficientNet-B2—were trained and evaluated for performance. Model fine-tuning was carried out to optimize classification accuracy.

Results: Among the models tested, EfficientNet-B2 achieved the highest accuracy of 84%, demonstrating its superior architectural efficiency for the classification task. The findings also highlighted the importance of data preprocessing and model fine-tuning in achieving robust diagnostic performance.

Conclusions: This study demonstrates that deep learning models, particularly EfficientNet-B2, can significantly improve the accuracy and efficiency of dermatological disease classification. The results suggest that proper data handling and model optimization are crucial in overcoming the limitations of traditional diagnostic methods. Future work will explore the integration of attention mechanisms with the Inception model to further enhance diagnostic capabilities.