Progressive Growing of Generative Adversarial Networks (PGGAN) Approach to Synthesize Medical Images
Vishal Raner, Amit D. Joshi, Suraj T. Sawant, P. S. Tamizharasan · 2025
The area of medical imaging diagnostic capability has been greatly enhanced by recent advancements in deep learning and high-performance computing. These developments provided a platform for researchers to follow their own goals to provide crucial illness diagnostic and therapeutic support. There have been many different GAN architectures proposed and investigated ranging from the most straightforward DCGAN to the most intricate style-based GANs. Consistent training is still difficult despite GAN's excellent results. In this work, the Progressive Growing of Generative Adversarial Network (PGGAN)-based method is devised to synthesize medical images. PGGAN allows to produce images of previously unknown quality by significantly speeding and stabilizing the training process. The PGGAN model is trained on three different medical imaging datasets, namely: brain MRI, CT Scan, and retina fundus images. The key idea during PGGAN training is to start at low resolution and keep on introducing new layers gradually. Each new layer introduced doubles the output size. This process keeps on repeating until we achieve the desired resolution. The model training issues of conventional GAN models have been effectively addressed by the PGGAN methodology. This work generated images of brain MRI, CT scan, and retina fundus with sizes of 2,048 × 2,048, 1,024 × 1,024, and 1,024 × 1,024, respectively. The RESNET-50 model trained for multi-class classification achieved accuracy of 87.52%, 91.47%, and 90.78% for fundus-, MRI-, and CT scan-generated images, respectively.