Binary Classification Optimisation with AI-Generated Data

In the field of machine learning, obtaining sufficient and high-quality data is a persistent challenge. This report explores the innovative solution of using synthetic data generated from existing datasets to overcome this limitation. By employing synthetic data, we not only increase the quantity of available information but also maintain the integrity and essential characteristics of natural data. This methodology allows the application of conventional data augmentation techniques, ensuring a more robust and efficient learning process. The study is based on a dataset provided by the International Skin Imaging Collaboration (ISIC), consisting of 3,323 cases divided equally between melanomas and Basal Cell Carcinoma (BCC). Using Generative Adversarial Networks (GANs), specifically StyleGAN2 with transfer learning from the Flickr-Faces-HQ (FFHQ) model, synthetic images were generated, expanding the dataset fourfold to a total of 26,584 synthetic records. The quality of the synthetic images was ensured using the Frechet Inception Distance (FID) metric [5], with BCC obtaining 22.2534 and melanomas obtaining 20.4577 according to this metric. Models trained with a hybrid approach using both real and synthetic data showed improved performance metrics (F1 0.71 to 0.79), highlighting the effectiveness of this method in enhancing binary classification tasks in medical imaging. The source code for all the research, along with the generated dataset is publicly available.