CiGeN 1.0: A Generative AI Enabled Medical Image Augmentation Pipeline for Deep Network Models

Deep Learning methods have gained a significant momentum in medical image classification tasks, including many ranges of datasets. However, these datasets often suffer from class imbalance, where the minority class (e.g., diseased cases) is underrepresented. This imbalance leads to biased predictions, affecting the model's performance. Generative Adversarial Networks (GANs) offer a potential solution by generating synthetic images for the minority class, helping balance the dataset. Methodology: In this study we further evaluated our Deep Learning Model CoVaD-GAN with separate set of datasets to test the dynamics behind it. We applied a Deep Convolutional GAN (DCGAN) to generate synthetic images for underrepresented classes in the chest X-ray (CXR) and Break His histopathological dataset. Augmented synthetic images were made a part of training set, and the model's performance was evaluated before and after augmentation using VGG-16, ResNet-50 with necessary metric measures as accuracy, sensitivity, and specificity were compared pre- and post-GAN augmentation. We also evaluated the quality of GAN-generated images using Fréchet Inception Distance (FID) and manual expert reviews. The results demonstrate significant improvements in model accuracy, sensitivity, and F1-scores after applying GAN-based augmentation. The use of GANs for data augmentation enhances the ability of deep learning models to generalize, particularly for minority classes. GAN-generated images were of high quality and contributed positively to model performance. The Proposed model (CiGeN 1.0) dynamically generalizes the other medical dataset introduced to it and derived a fair resulting diagnosis.