Polycystic Ovary Syndrome Detection Utilizing SRGAN-Generated Synthetic Images and Advanced CNN Models

Polycystic Ovary Syndrome (PCOS) is a common hormonal disorder affecting women worldwide and is detected using ultrasound scans. Accurate diagnosis is crucial but often hindered by limited data availability. High-quality data is needed for building reliable models, and with advancements in artificial intelligence, data generation through synthetic images has shown promising results. This paper explores the use of Super-Resolution Generative Adversarial Networks (SRGAN) and convolutional neural networks (CNN) for PCOS detection on a limited dataset. Synthetic images from four SRGAN variants and original ultrasound scans are used to train and test CNN models, including NasNetMobile, ResNet152, and Xception. Additionally, hybrid models combining these CNNs with the CatBoost classifier are evaluated. The Xception model and its hybrid version with CatBoost achieved the best performance, up to 99% accuracy. Some models showed a slight drop when trained with synthetic data. This study concludes that SRGAN-generated images can expand datasets and support diagnosis, though performance depends on the architecture used.