Development of radiomics techniques for contrast-enhanced mammography: In silico testing with synthetic data to encompass less common cancer subtypes

Breast cancer is one of the most common cancers, and early detection is vital. Contrast-enhanced mammography (CEM) improves diagnosis by combining low- and high-energy images, but interpreting them can be challenging. Artificial intelligence and radiomics extract hidden features from images and make predictions to help radiologists. But they require large, high-quality datasets. Since collecting medical images is difficult, especially for newer methods like CEM, synthetic data has emerged as a solution. This thesis focuses on improving synthetic data generation for CEM and assessing its role in radiomics-based cancer detection. It reviews lesion modelling in virtual clinical trials (VCT), showing that models must balance realism and versatility. Novel techniques to simulate microcalcification clusters, which are hard to detect, are introduced, alongside an enhanced hybrid simulation framework for efficient dataset creation. Experiments reveal that combining synthetic and real data improves detection, especially with limited datasets. Overall, synthetic data offers opportunities, though clinical acceptance remains a challenge.