Analysis of Generative Patterns in Image Generation with Autoencoders and BigGAN: Applications in Osteosarcoma
F. Cortes, F. Soriano, N. Lopez, B. Mendez, Arturo Vera, L. Leija, Rocío Ortega-Palacios · 2025
The generation of synthetic medical images offers solutions to challenges such as limited datasets and ethical constraints in healthcare research. This work evaluates two deep learning models-Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs)-for generating and reconstructing histological images of osteosarcoma. A dataset of 1144 hematoxylin and eosin (H&E)-stained images annotated as viable tumor, necrotic tumor, and non-tumor was used to train and assess both architectures under comparable conditions. The VAE achieved a steady reconstruction loss reduction from 0.15 to 0.04 over 175 epochs. Reconstructed images reached a mean Structural Similarity Index (SSIM) of 0.552 and a mean Mean Squared Error (MSE) of 0.012, effectively preserving structural features but partially losing finer textures. Meanwhile, the BigGAN was trained for 2000 epochs, achieving a peak SSIM of 0.739 and a corresponding MSE of 38.28 during early epochs, demonstrating its capacity to generate highly detailed synthetic images with significant variability. These results underscore the complementary strengths of both models: the VAE's robust latent representation ensures reliable reconstructions, while the BigGAN excels in producing high-quality synthetic images. Combining these approaches has potential to augment histopathological datasets, advancing diagnostic algorithms and computational pathology.