Developing a Custom Generative Adversarial Network for Synthetic Mammogram Generation

Abstract

As one of the most commonly diagnosed cancers worldwide, breast cancer poses a significant challenge in early detection and treatment. Mammographic analysis plays a crucial role in diagnosis, with malignant calcifications and malignant masses serving as key radiological markers. However, the limited availability of annotated medical imaging datasets and high inter-observer variability pose challenges in developing robust AI-driven diagnostic models. This thesis focuses on developing a custom Generative Adversarial Network (GAN) to generate synthetic mammograms that enhance the representation of malignant calcifications and masses. Unlike studies utilizing pre-existing GAN architectures, this research builds a GAN model from scratch, with a focus on optimizing performance and improving image fidelity. The generated synthetic images aim to augment existing datasets, providing additional training data for deep learning models in breast cancer classification. The study employs a dataset containing mammographic images of malignant calcifications and malignant masses, carefully curated for training and validation. To assess the quality of the generated images, performance metrics such as Fréchet Inception Distance (FID), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) are utilized. Additionally, the GAN’s performance is evaluated using the loss graph, where the discriminator and generator losses are plotted over training epochs to analyze convergence and stability. To assess the diagnostic value of the synthetic images, a Support Vector Machine (SVM) classifier is also employed. The classifier is tested with real and GAN-generated images to evaluate their robustness and impact on classification performance. By improving the performance of a custom GAN for synthetic mammogram generation, this work contributes to addressing data scarcity, enhancing AI-based breast cancer detection, and supporting the integration of generative models in medical imaging. The findings highlight the potential of GAN-based data augmentation in improving diagnostic reliability and dataset diversity in breast cancer research.