Comparative Evaluation of Transfer Learning Architectures for Five-Class Skin Disease Classification with Grad-CAM Explainability

Abstract

Skin disorders affect approximately 1.9 billion people worldwide and remain significantly underdiagnosed due to the global shortage of dermatologists. This study provides a thorough comparative investigation between four pre-trained deep CNN models — ResNet-50, EfficientNetB0, MobileNetV2, and Xception — for automatic classification of five important skin disorders: Acne, Actinic Keratosis, Basal Cell Carcinoma, Eczema, and Rosacea. A balanced data collection comprising 625 images was curated from publicly accessible dermatological databases and split into training, validation, and test sets (375/65/185) to allow proper early stopping and learning-rate scheduling during fine-tuning. All models used architecture-specific preprocessing, with Xception trained at its native 299×299 input resolution while the remaining three models used 224×224. Accuracy, precision, recall, F1 score, per-class F1, AUC-ROC and confusion matrices were used as evaluation criteria. Gradient-Weighted Class Activation Maps (Grad-CAM) were computed on all four models to provide visual interpretability, highlighting the image regions that most influenced each prediction. EfficientNetB0 attains the maximum accuracy (71.89%) and macro F1 score (0.72), followed closely by Xception (68.11% accuracy) and ResNet-50 (67.57% accuracy), while also achieving the highest macro-average AUC-ROC (0.926). MobileNetV2 records the lowest overall accuracy (64.86%) but remains attractive for low-resource deployment given its lightweight architecture. A dataset-size suitability analysis also sheds light on the optimal model choice depending on the amount of data available.