Enhancing Diagnostic Accuracy in Medical Imaging: Integrating GAN-Based Data Augmentation for Balanced Dataset Creation
Hemanth Karnati, SujayKumar Reddy M · 2024
Advancements in medical imaging have been substantially driven by deep learning technologies, particularly Convolutional Neural Networks (CNNs). A critical hurdle in this domain is the imbalance of datasets, where certain medical conditions are underrepresented, leading to potential biases in diagnostic models. This research addresses the imbalance in medical imaging datasets, specifically in chest radiography, by leveraging Generative Adversarial Networks (GANs) for data augmentation. The study utilizes the ChestXray2017 dataset, which is skewed towards pneumonia cases, resulting in a dearth of normal chest X-ray images. To counter this, Deep Convolution Generative Adversarial Networks (DCGAN) were employed to generate synthetic images of normal chest X-rays, thus aiming to balance the dataset. In this study, we conducted a comparative analysis of a Convolutional Neural Network's (CNN) performance on a chest radiography dataset, before and after augmenting it with Deep Convolution Generative Adversarial Network (DCGAN)-generated images. Initially, the CNN trained on the un-augmented dataset achieved 93% training accuracy and 87% validation accuracy. After integrating 400 synthetic normal chest X-ray images, the training accuracy slightly increased to 95%, while the validation accuracy notably improved to 89%. This enhancement in validation accuracy demonstrates the model's improved generalization capabilities due to a more balanced training dataset. Our results indicate that GAN-based data augmentation effectively addresses class imbalances in medical imaging datasets, potentially leading to more accurate and reliable diagnostic models. However, the study also underscores the need for further research into the quality and ethical implications of using synthetic images in medical diagnostics. Overall, the integration of GAN-generated images into CNN training presents a promising method for improving classification performance in medical imaging, offering a practical approach to overcome challenges associated with data scarcity and imbalance.