<![CDATA[Skin cancer is one of the diseases that requires early detection to increase the likelihood of successful treatment. The use of artificial intelligence, particularly Deep Learning, has become an effective alternative in assisting the automatic classification of skin cancer images. However, the high class imbalance and visual similarity between lesion types in skin cancer datasets remain challenges in achieving optimal classification performance. This study aims to improve the accuracy of skin cancer image classification using an Improved Convolutional Neural Network based on Transfer Learning and Hyperparameter Tuning. The dataset used is HAM10000, consisting of 10,015 dermoscopy images across seven diagnostic classes. The architecture employed is MobileNetV2 as a feature extractor combined with a custom classification head. The training process was carried out using a two-phase transfer learning strategy, namely the backbone freezing phase and the fine-tuning phase. To address class imbalance, class weighting and data augmentation were applied, while model optimization was performed using grid search over the parameters of learning rate, dense layer size, and dropout rate. Model performance was evaluated using accuracy, precision, recall, F1-score, and Area Under Curve (AUC) metrics. The results show that the proposed model achieved a test accuracy of 85.50%, a validation accuracy of 84.75%, a macro F1-score of 83.14%, and a mean AUC of 0.94. These results indicate that the combination of two-phase Transfer Learning and Hyperparameter Tuning is capable of improving the performance of MobileNetV2 in skin cancer image classification. The contribution of this research is the development of a classification model that achieves high accuracy, is computationally efficient, and is capable of handling class imbalance in the HAM10000 dataset.]]>
Copyrights © 2026
