Comparative Study of Data Augmentation Approaches for Improving Medical Image Classification

In recent years, data augmentation has advanced to the point where it no longer relies on traditional photometric or geometric image processing techniques, such as rotation, scale, and filtering. Instead, it has turned to deep learning-based approaches like variational autoencoders (VAEs) and generative adversarial networks (GANs). This shift has made data augmentation a crucial discipline for improving artificial intelligence models, especially in fields like medicine where collecting labeled data is difficult and expensive. This study focuses on using data augmentation approaches to enhance convolutional neural networks (CNNs) classification accuracy. These approaches include geometric modifications, color space transformations, and generative techniques, and they are employed to augment the BraTS20 (Brain Tumor Segmentation 2020) dataset. After comparing the accuracy of CNNs with and without the augmented dataset, we found that DCGANs and AttnGAN achieved the highest accuracy among all techniques. They generated images that closely resembled those in BraTS20 compared to other methods.