UltraGANet: a class-conditional GAN framework for breast tumor classification in ultrasound imaging

This study addresses the challenge of limited annotated data in ultrasound-based breast cancer classification by proposing a GAN-augmented deep learning framework. The aim was to evaluate whether conditional generative adversarial network (cGAN)-based augmentation could improve classification performance when integrated with a lightweight convolutional neural network (CNN), termed UltraGANet. A cGAN was developed to generate class-specific grayscale breast ultrasound images representing benign, malignant, and normal categories. These synthetic images were combined with real images from the BUSI dataset to construct an augmented training dataset. UltraGANet was evaluated using 5-fold stratified cross-validation and an independent test set comprising 146 real ultrasound images. Performance was assessed using accuracy, precision, recall, and F1-score. UltraGANet achieved a mean cross-validation accuracy of 0.914 ± 0.012 and an independent test accuracy of 0.904. On the independent test set, malignant tumor recall reached 0.92, indicating high sensitivity for malignant cases. The model showed generally consistent precision and F1-scores across the three classes. Qualitative and quantitative evaluations suggested that the generated images retained class-relevant ultrasound characteristics and moderate structural similarity to real images. The integration of cGAN-based augmentation with UltraGANet improved classification performance compared with real-only training, particularly for malignant tumor detection. The proposed lightweight framework may provide a useful approach for breast ultrasound classification in data-scarce and resource-limited settings.