IMPACT OF LOSS FUNCTION ON SYNTHETIC BREAST ULTRASOUND IMAGE GENERATION

Abstract

The BUSI (Breast Ultrasound Images) dataset is small and imbalanced, which limits the effective training of deep learning diagnostic models. Generative Adversarial Networks (GANs) offer a promising and increasingly popular solution for synthesizing realistic medical images to augment scarce training data and improve overall model generalization. This study investigates the impact of loss function selection in our previously published Deep Generative Adversarial Network with Wasserstein Gradient Penalty and Transfer Learning (DGAN-WP-TL). Two configurations were evaluated: one trained using Wasserstein GAN with Gradient Penalty (WGAN-GP) and another trained using Binary Cross-Entropy (BCE) loss. The experiments were conducted on the BUSI dataset with perceptual loss weights λ = 0.5, 3.0, 5.0, 7.0, and 10.0. Model performance was comprehensively assessed using Fréchet Inception Distance (FID), Kernel Inception Distance (KID), Learned Perceptual Image Patch Similarity (LPIPS), and Multi-Scale Structural Similarity Index (MS-SSIM). Results demonstrate that WGAN-GP consistently outperformed BCE across all λ values, generating images with higher fidelity, improved realism, and greater visual diversity. The superiority was most pronounced for λ = 3.0 and λ = 5.0, where WGAN-GP achieved the lowest KID and FID and the most balanced diversity–fidelity trade-off. The best-performing DGAN-WP-TL configuration (WGAN-GP, λ = 5.0) achieved KID = 0.14, FID = 179.42, LPIPS (fake–fake) = 0.49, and MS-SSIM (fake–fake) = 0.18. These results highlight the crucial role of loss function design in medical image synthesis. Overall, the study confirms that WGAN-GP provides superior image realism and variability, making it the preferred choice for high-quality, clinically relevant synthetic data generation, while BCE remains a lightweight and practical alternative for constrained computational environments.