Ensemble of learning transfer models with GAN for data augmentation: Application to skin cancer detection

Deep Learning models typically require large datasets to achieve high performance, which can be challenging when applied to medical data due to the difficulty in acquiring sufficient observations, particularly when the data is imbalanced between categories like pathological and normal cases. To address this, approaches like transfer learning, ensemble methods, and data augmentation through Generative Adversarial Networks (GANs) can be utilized to improve classification tasks on small datasets. In situations with imbalanced data, resampling techniques are also crucial. This research combines transfer learning with resampling methods to boost the prediction accuracy of minority class samples in small, imbalanced datasets. Additionally, techniques like data retracing and GAN-based augmentation are applied. The dataset used includes small, imbalanced images of skin cancer, aimed at classifying them as malignant or benign. The findings reveal that models using resampling techniques achieve better results, while those without resampling underperform. This underscores the benefit of resampling in enhancing prediction, particularly for the minority class. Furthermore, using GANs for data augmentation improves model performance over those that do not incorporate this technique.