Image Synthesis for Sperm Dataset Augmentation using WGAN-GP

Abstract

This research explores the efficacy and limitations of applying a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to generate synthetic human sperm microscopy images for data augmentation. We assessed the WGAN-GP's performance on a complex, heterogeneous dataset where images contained multiple object types. Despite achieving stable training convergence, the model's output quality was suboptimal, as evidenced by a high Fréchet Inception Distance (FID) score of 134 and qualitative signs of partial mode collapse. The generator struggled to capture the complete morphological diversity of the sperm cells. A second experiment using a dataset pre-sorted into distinct classes (Normal, Abnormal, Non-Sperm) yielded a marked improvement. This approach led to substantially lower FID scores (59.19, 74.92, and 83.56) and exhibited more robust training dynamics. Our findings underscore a critical conclusion: the success of WGAN-GP in this domain is fundamentally tied to the simplicity of the data distribution. We recommend that future efforts leverage class-conditioned models, simplified data structures, and refined generator architectures to achieve high-precision augmentation for medical imaging tasks.