Comparative Evaluation of Deep Learning Architectures for Early Skin Cancer Detection Using Dermatoscopic Images

Aurel Elviolita Putri (Universitas Esa Unggul)
Agung Mulyo Widodo (Universitas Esa Unggul)
Nizirwan Anwar (Universitas Esa Unggul)
Binastya Anggara Sekti (Universitas Esa Unggul)

Journal of Technology and Informatics (JoTI)
Vol. 8 No. 2 (2026): Vol. 8 N. 2 (2026)

Early identification of skin cancer is essential for improving diagnostic reliability and patient outcomes. Although convolutional neural networks (CNNs) have demonstrated promising performance in dermoscopic image classification, comparative evaluations on imbalanced multiclass datasets remain challenging. This study compares three transfer learning-based CNN architectures—VGG16, ResNet50, and MobileNetV2—for seven-class skin lesion classification using the imbalanced HAM10000 dataset. All models were evaluated using the same dataset partitions, image resolution, and performance metrics, while architecture-specific augmentation and loss-function strategies were retained. In particular, VGG16 used MixUp, ResNet50 used CutMix, VGG16 and ResNet50 employed Focal Loss, and MobileNetV2 used label smoothing. Performance was assessed using accuracy, precision, recall, F1-score, confusion matrices, and ROC–AUC. ResNet50 achieved the best overall performance, with 85.71% accuracy and a weighted F1-score of 0.8502, ahead of VGG16 (82.52%, 0.8172) and MobileNetV2 (79.32%, 0.7739). ResNet50 also showed the highest recall for melanoma and dermatofibroma among the three tested configurations, although minority-class sensitivity remained limited. These results favor ResNet50 under the evaluated training configurations; however, differences in augmentation and loss functions prevent attributing performance gaps solely to network architecture.