<![CDATA[Imbalanced class distributions in skin lesion image datasets can reduce the effectiveness of multiclass classification models. This research proposes a classification model based on the EfficientNetV2-S architecture with the application of two-stage training and loss functions that emphasize learning in classes with limited data. The models were trained using on-the-fly image augmentation and evaluated to assess generalization capabilities to the test data. In the initial stage, the model is trained by freezing the backbone and only updating the classifier layer. Next, fine-tuning was carried out on part of the backbone layer to adjust the representation of features to the image characteristics of the skin lesion. Evaluation is conducted through multiple training times with different random initializations to ensure consistency of results. The test results showed that the model experienced an improvement in performance after the fine-tuning process, with an accuracy of about 88% as well as an increase in F1-score values in some classes. Overall, the results indicate that the proposed approach may help improve classification performance when dealing with imbalanced skin cancer image data.]]>
Copyrights © 2026
