For a summary of the talk, please see below:
Every successful AI model depends on trustworthy ground truth. In medical imaging, however, that ground truth often does not exist. This talk explores how we engineer the biological and physical ground truth needed for trustworthy AI using high-resolution bone imaging. First, controlled experiments in which bone composition is systematically altered allow us to determine what radiomic features actually measure, transforming abstract image patterns into biologically meaningful biomarkers. Second, physics-based motion simulation creates paired clean and motion-corrupted scans that cannot be acquired from patients but are essential for training and validating deep learning models for motion correction. These projects address different challenges, but they share a central principle: trustworthy AI begins with trustworthy data. Rather than accepting the limitations of available clinical datasets, we build the evidence needed for AI models that are accurate, interpretable, and capable of advancing both skeletal biology and patient care.