Speaker
Description
To predict radiotherapy patient outcomes under clinically realistic right-censoring, we
model longitudinal gross tumor volume (GTV) dynamics using a geometric adversarial
learning framework rather than predefined mathematical growth equations. In this setting, patient trajectories are observed only up to a given follow-up time, after which
tumor evolution continues but remains unobserved. Instead of assuming a specific
functional form (e.g., exponential, logistic, or Gompertz growth), our approach (GAN-
GAF framework) first transforms observed GTV trajectories into Gramian Angular Fields
(GAFs), two-dimensional representations that encode pairwise temporal correlations
across the entire observation window, and then learns to complete these correlation
manifolds using adversarial training (GAN). Thus, right-censoring is implemented by
withholding the terminal portion of each patient’s trajectory, with the generator inferring
plausible continuations by restoring missing correlation structure in GAF space. This
enables uncertainty-aware reconstruction of post-observation tumor dynamics that
preserves amplitude, trend, and long-range temporal dependencies reflective of
individual treatment response heterogeneity, without requiring mechanistic assumptions
or parameter fitting. The GAN-GAF framework operates directly on real-time clinical
data, enforces global geometric consistency, and provides a data-driven alternative to
classical mathematical modeling for inference under right-censored radiotherapy follow-
up.
