Exploring Generative Adversarial Networks (GANs) for Synthetic Medical Imaging and Improved Tumor Diagnosis

Abstract

Medical imaging is a cornerstone of tumor detection and diagnosis, yet challenges such as limited annotated datasets, high acquisition costs, and variability across modalities restrict the accuracy and generalizability of diagnostic models. This study explores the application of “Generative Adversarial Networks (GANs) for synthetic medical imaging to improve tumor diagnosis. Using three benchmark datasets BraTS (brain MRI), LIDC-IDRI (lung CT), and MIAS (mammograms) GAN variants including DCGAN, cGAN, and CycleGAN were employed to generate synthetic tumor images”. The experimental results demonstrate that GAN-generated images achieve high structural fidelity with SSIM values above 0.85 and PSNR exceeding 30 dB, comparable to real scans. When integrated into training pipelines, GAN-augmented datasets improved tumor classification accuracy by 5–6% and segmentation Dice scores by up to 0.08 compared to baseline models. Radiologist evaluations further confirmed the clinical plausibility of synthetic images, indicating their potential to supplement diagnostic training and enhance decision confidence. However, limitations such as mode collapse, image artifacts, and ethical concerns regarding synthetic data usage remain. Overall, the study highlights GANs as a powerful augmentation tool for addressing data scarcity in medical imaging and a promising pathway toward earlier and more reliable tumor diagnosis.