Open Collections

Abstract

The success of Deep Neural Networks (DNNs) in ultrasound image analysis heavily relies on the availability of large quantity of training data. In the medical domain, however, large-scale data collection and annotation are often limited by the high cost and time demands, particularly due to the workload of clinical professionals. Furthermore, maintaining model robustness across varying imaging conditions—such as differences in ultrasound equipment and manual transducer handling—remains a critical challenge. Data augmentation has become an essential strategy to address these limitations by expanding dataset size and variability, thereby improving model generalization. Traditional augmentation techniques, such as rotation, flipping, and noise injection, have been widely adopted. More recently, the emergence of generative models, including Generative Adversarial Networks (GANs) and diffusion-based models, has introduced new possibilities for synthesizing realistic medical images for augmentation purposes. Despite their promise, there remains a lack of comprehensive evaluations comparing classical and generative augmentation strategies, particularly in breast ultrasound imaging—a domain marked by high variability in tissue density, tumor characteristics, and operator-dependent imaging quality. This thesis aims to introduce a new augmentation approach while systematically comparing the effectiveness of conventional methods and generative network-based techniques in enhancing the performance of DNNs for breast cancer classification using ultrasound images. Through extensive experiments, we assess whether the increased computational cost associated with generative augmentation is justified by measurable gains in classification accuracy and robustness. The findings of this study offer practical insights into selecting suitable augmentation strategies for medical imaging tasks and contribute to optimizing deep learning workflows for breast cancer diagnosis.