Exploring the Impact of Data Augmentation : Traditional Methods vs. Deep Learning Approaches

Data augmentation enhances the performance and resilience of machine learning and the deep learning models, particularly for scarce or imbalanced datasets. Conventional methods like geometric transformations, noise addition, and intensity shifting are computationally efficient but lack the ability to introduce diverse, high-level variations needed for complex tasks. Deep learning-based augmentation, including Generative Adversarial Networks (GANs); Neural Style Transfer (NST) and Variational Autoencoders (VAEs); leverages neural networks to generate realistic and context-rich data variations. These advanced techniques improve model generalization and performance across various applications such as medical imaging, natural language processing (NLP), and computer vision.This paper compares traditional and the deep learning-based augmentation methods. A key novelty of this work lies in integrating GAN-based augmentation with domain-specific lung cancer datasets and benchmarking its performance against traditional techniques using standardized metrics like AUC-ROC and F1 Score to determine the optimal augmentation strategies based on dataset characteristics and application requirements.