Fiqri Ramadhan Tambunan – Poster

Skin cancer is a serious and potentially fatal disease, but it is often difficult to detect in its early stages. Early detection through dermoscopic images offers an effective non-invasive approach, but its success is highly dependent on the quality and balance of the training data. The HAM10000 dataset, widely used in skin lesion classification research, has an imbalanced class distribution, leading to model bias. To address this, this study proposes the use of Auxiliary Classifier Generative Adversarial Networks (ACGAN) as a data augmentation method to generate realistic synthetic images, while modifying its discriminator to function as a classification model. Experiments were conducted by combining the ACGAN discriminator and a pretrained CNN architecture such as ResNet50 into a hybrid model to improve multiclass classification performance. Evaluations were conducted using metrics such as accuracy, precision, recall, macro average, confusion matrix, and GRAD-Cam visualization. The results showed that the combination of ResNet50 and the ACGAN discriminator achieved the highest accuracy of 93.61%, as well as significant improvements in other evaluation metrics compared to the baseline method. This model also demonstrated more balanced classification capabilities between classes. These findings demonstrate that ACGAN is not only effective as a data augmentation technique, but can also be integrated as part of a robust classification system for image-based skin cancer diagnosis, thus offering an efficient and precise dual-purpose approach in the medical domain.