Data Balancing through Data Augmentation to Improve Transfer Learning Performance for Skin Disease Prediction

Abstract

Data balancing through data augmentation is a crucial step in improving the generalization of transfer learning for skin disease prediction. Many skin diseases, such as melanoma, are rare, and thus, the number of samples in the minority class is often limited. By using data augmentation techniques, the number of samples in the minority class can be increased, resulting in a more balanced dataset and a more robust model. In this paper, we evaluated the classification performance of the Tuned EfficientNetV2L based classifier. In the first experiment, we adopt a default dataset structure for training and testing. In the second experiment, we use the data augmentation function to balance the used dataset. The experiments are performed on a collection of medical images and associated data of skin lesions that have been sourced from various sources. In order to assess the obtained outcomes, various performance evaluation metrics were employed, including accuracy, precision, and recall. The analysis of the proposed methodology revealed that the balanced dataset improves the performance of transfer learning for skin disease prediction.