Investigating the Use of Gan in the Classification of Brain Tumors
Michaël Jordan Ah-Shee-Tee, Rousitha Chikhuri, Zahra Mungloo-Dilmohamud · 2023
Within the healthcare industry, data assumes diverse forms, ranging from textual records to intricate medical images and even noise. Due to their subjectivity, and complexity, image interpretation can be subject to error as it is quite constrained by the significant variations between interpreters and fatigue. This research investigates the use of Generative Adversarial Networks (GANs) for the identification of brain tumors. The methodology involves training GAN networks, namely DCGAN and StyleGAN2, to generate brain tumor images. These generated images are then used to augment the training dataset for a Convolutional Neural Network (CNN) classifier. Four datasets were designed, namely the original dataset, balanced dataset with reduced number of images in some classes and datasets augmented with images from DCGAN and StyleGAN2 respectively. The objective is to evaluate the impact of GAN-generated images on the classifier's performance. Different batch sizes and training configurations were explored for the training of DCGAN and StyleGAN2, with StyleGAN2 showing more realistic and diverse images in terms of KID and FID scores. The classifier's accuracy, sensitivity, precision, and generalization capacity were evaluated for different train to test ratios and epochs. While both GANaugmented datasets improved the classifier's performance compared to the other two datasets, the dataset augmented by StyleGAN2 consistently outperformed the others.