GenAI for Synthetic Data Creation using StyleGAN2-ADA

Deeplearning Models requires large and balanced dataset but medical datasets are limited due to privacy issues, high costs and class imbalance. Conditional GANs(CGANS) and pixel-based mixup methods improves the dataset size but this kind of methods produces unrealistic images that leads to unstable training and this doesnot make them very much useful in the medical applications. This study proposes a better data augmentation and detection framework. It leverages a label-aware generative mixing approach based on StyleGAN2-ADA.In this process we train the generator using the dataset and the dataset is conditioned on the soft labels that leads to generate realistic images.The synthetic images generated will get combined with real data to train the classification model. We measured accuracy, precision, recall, F1 score, macro-F2 and false negative rate. The results are observed that this StyleGAN2-ADA augmentation improves detection performance when compared with the CGAN-based and traditional methods