Enhanced Generation of Gastrointestinal Images for Data Augmentation Using a Modified BlobGAN

The advent of artificial intelligence (AI) and deep learning has revolutionized healthcare, particularly in the realm of medical imaging, patient care, and personalized treatments. However, the emergence of deepfake technology, while offering promising opportunities, also presents unique challenges. Generative adversarial networks (GANs) can produce highly realistic synthetic medical images (deepfakes) with potential applications in tasks like segmentation and detection. This study leverages a modified BlobGAN architecture, incorporating a self-attention block, to generate enhanced gastrointestinal images. We trained and validated our model on the Kvasir dataset, a comprehensive repository of endoscopic images. Our modifications aim to address the limitations of current deep learning models by producing synthetic images of exceptional quality that closely resemble real gastrointestinal images. Results demonstrate a lowest Fréchet Inception Distance (FID) of 15.552 and a highest Inception Score (IS) of 6.951, indicating the high fidelity of our generated images. By generating synthetic data that mirrors actual medical images, this research contributes to overcoming challenges of limited data availability and privacy concerns in the field of AI for healthcare. Ultimately, our approach has the potential to expand training datasets and improve the performance of automated classification systems.