Improving skin diseases prediction through data balancing via classes weighting and transfer learning

Skin disease prediction using artificial intelligence has shown great potential in improving early diagnosis and treatment outcomes. However, the presence of class imbalance within skin disease datasets poses a significant challenge for accurate prediction, particularly for rare diseases. This study proposes a novel approach to address class imbalance through data balancing using classes weighting, coupled with transfer learning techniques, to enhance the performance of skin disease prediction models. Two experiments were conducted using a tuned EfficientNetV2L based classifier. In the first experiment, a default dataset structure was utilized for training and testing. The second experiment involved employing classes weighting approach to balance the dataset. The effectiveness of the proposed approach is evaluated using the ISIC 2018 dataset, which comprises a diverse collection of skin lesion images. By assigning appropriate weights to different classes based on their prevalence, the proposed method aims to balance the representation of rare disease classes. To evaluate the performance of the proposed methodology, several performance evaluation metrics, including accuracy, precision, and recall, were employed. These findings revealed that the balanced dataset achieved enhanced generalization, mitigating the biases associated with class imbalance. As a result, the efficacy of artificial intelligence models is enhanced.