Comparative Analysis of Pre-Trained Deep Learning Models for Classifying Tropical Fungal Skin Infections

Abstrak

Tropical fungal skin infections, including Tinea corporis, Tinea versicolor, Tinea pedis, and Tinea nigra, are common health problems in tropical countries such as Indonesia. Although not life-threatening, these diseases can cause discomfort, reduce self-confidence, and interfere with daily activities. Conventional diagnostic methods still rely on subjective visual observation, which is often inaccurate—especially in regions with limited infrastructure and scarce access to specialists. Moreover, existing studies rarely provide a comparative evaluation of deep learning architectures for tropical fungal infections using small and diverse datasets. Therefore, this study aims to address these challenges by conducting a systematic comparative evaluation of three pre-trained models—MobileNetV3, EfficientNet-B2, and SE-ResNet101—to determine the most accurate and computationally efficient architecture for multi-class classification of tropical fungal skin diseases. In this study, a dataset of 660 clinical skin images sourced from the Kaggle repository was used, covering four tropical fungal infection classes. The dataset consisted of 165 images per class, which were divided into training, validation, and testing subsets. Experimental results demonstrated that MobileNetV3 achieved the best performance, with a validation accuracy of 95.08%, a test accuracy of 97.34%, and the shortest training time of 15 minutes, compared to EfficientNet-B2 (16 minutes) and SE-ResNet101 (22 minutes). The main contribution of this study is to provide a systematic comparative evaluation of deep learning models for the classification of tropical fungal skin infections, while recommending MobileNetV3 as the most suitable model for practical implementation of automated image-based diagnosis in primary healthcare services with limited resources.