A Controlled Comparative Study of EfficientNetB0, ResNet50, and DenseNet121 for Multi-Class Skin Disease Classification

Skin disease classification from images is difficult because different disease types share visual characteristics, and image appearance can be altered by image quality and acquisition conditions. A controlled comparison of EfficientNetB0, ResNet50, and DenseNet121 was conducted under a common experimental configuration. The Skin Disease Dataset was obtained from Kaggle, consisting of 4,357 images of skin diseases divided into five classes: Acne, Eczema, Psoriasis, Tinea, and Vitiligo. Common pre-processing, training-only data augmentation, class weighting, and two-phase transfer learning from ImageNet-pretrained models were applied to all models. The accuracy, precision, recall, F1 score, confusion matrices, and results for three training seeds were used for the evaluation. Classical machine-learning baselines, including a support vector machine (SVM) and a random forest, were also evaluated. Results: ResNet50 achieved the highest mean accuracy (80.5% ± 0.9%) and macro-F1 (81.8% ± 0.9%). DenseNet121 achieved 74.6% ± 0.5% accuracy and 76.4% ± 0.5% macro-F1, while EfficientNetB0 achieved 72.2% ± 2.8% accuracy and 73.8% ± 2.5% macro-F1. Random Forest achieved 57.8% accuracy and 58.3% macro-F1, while SVM achieved 53.8% accuracy and 53.9% macro-F1. All deep-learning models appeared to classify images of Acne and Vitiligo more successfully, whereas images of Eczema, Psoriasis, and Tinea were more frequently misclassified. ResNet50 appears to have achieved the best overall classification accuracy under the experimental conditions. The results, however, are only specific to the chosen evaluation setup and dataset. More generalizability testing is required to determine if observed performance applies across populations, skin tones, and image-acquisition conditions in larger, more diverse, independent datasets.