A GAN based approach for inferring progression trajectories of costal cartilage calcification from cross-sectional...

Costal cartilage calcification (CCC) increases with age and presents differently for men and women. In individuals, however, the cross-sectional studies that show such trends do not reveal the geometric trajectories through which calcification might accumulate across a lifetime. Generative adversarial networks have the potential to reveal such trajectories from cross-sectional data by learning population trends and synthesizing individualized images at progressive levels of calcification.

Chest wall mid-surface CT images with normalized cartilage morphologies were produced for 379 subjects aged 6 to 90, and labeled by sex and calcification severity. A conditional GAN with added loss terms to favor one-way accumulation of CCC was trained using organized image batches. GAN performance was assessed by comparing the distributions of images between the training and synthetic groups.

Synthetic images generated from a common seed for a given sex and at successive calcification severity levels showed incremental and regional growth of calcification sites. CCC patterns for synthetic male and female images matched known sex-based differences, and individual CCC growth in synthetic images was consistent with previously observed population trends. These trends in the synthetic images were also quantified by structural similarity scores. Synthetic images generated from different input seeds further showed individual variance in specific regions and trajectories of CCC accumulation.

This study inferred individual progression of CCC accumulation from uncalcified to severely calcified using cross-sectional image data. This information can inform computational models of the changing chest wall biomechanics with age, and the GAN-based technique shows potential for inferring longitudinal data from population trends in other clinical areas.