Uncertainty Estimation for Neural Cellular Automata in Medical Imaging
A new study on arXiv (2605.26726) proposes resilience, a method to estimate prediction uncertainty in neural cellular automata (NCA) for medical image segmentation without modifying the architecture or retraining. The approach views NCA as a dynamical system where convergent attractors indicate confident predictions. Resilience probes the stability of final predictions under small perturbations of the automaton state. Predictions returning to the same solution are deemed confident; those changing substantially are flagged as uncertain. The method is evaluated using selective prediction metrics (ΔDice@90 and AURC) and ranking measures.
Key facts
- arXiv paper 2605.26726
- Neural cellular automata (NCA) are lightweight alternatives to encoder-decoder segmentation networks
- Resilience measures prediction uncertainty without modifying architecture or retraining
- NCA viewed as dynamical system with convergent attractors for confident predictions
- Resilience probes stability under small perturbations of automaton state
- Evaluation uses ΔDice@90 and AURC metrics
- Method flags uncertain predictions that change substantially under perturbation
Entities
Institutions
- arXiv