Continuous-Time Deep Glioma Growth Models

The ability to estimate how a tumor might evolve in the future could have tremendous clinical benefits, from improved treatment decisions better dose distribution radiation therapy. Recent work has approached glioma growth modeling problem via deep learning and variational inference, thus dynamics entirely real patient data distribution. So far, this approach was constrained predefined image acquisition intervals sequences of fixed length, which limits its applicability more realistic scenarios. We overcome these limitations by extending Neural Processes, class conditional generative models for stochastic time series, with hierarchical multi-scale representation encoding including spatio-temporal attention mechanism. result is learned model that can be conditioned on an arbitrary number observations, produce temporally consistent trajectories continuous axis. On dataset 379 patients, successfully captures both global finer-grained variations images, exhibiting superior performance compared other models.