Uncertainty estimation for reliable neural network-based radiotherapy dose modelling using Monte Carlo dropout, mean variance estimation and deep ensemble☆

Description

Neural networks can plan radiation doses in a flash, but a fast wrong answer in cancer care is dangerous. Moritz Schneider, Tabea Eberhardt, Cihan Gani, Paul Fischer, Christian Baumgartner, and Daniela Thorwarth test three ways to make a dose model confess its own doubt. They trained the model on thousands of segments from MRI-guided treatment plans, then compared dropout sampling, variance estimation, and a team-of-models approach.

All three methods' doubt tracked the real error closely. Set a doubt cutoff, and the accepted plans failed at close to the target rate, while forty-four to seventy-seven percent of segments passed. The variance method was best calibrated and cheapest to run. The team says such self-doubt scores can guard clinical use.

Catalog record

Source: https://europepmc.org/articles/PMC13572023