🧭 Future Directions
What to use, when to use it, and what the field still has not solved.
Part 3 — Evaluation & Reliability
You learned to interrogate uncertainty estimates rather than trust them: measuring calibration with ECE and reliability diagrams, using risk–coverage curves to choose a sensible abstention threshold, and detecting out-of-distribution inputs. Together these turn an uncertainty number into evidence about whether a model can be relied upon in practice.
About this part
The final part brings the whole course together. It compares every method you have implemented across the dimensions that actually matter for deployment — computational cost, ease of training, quality of the uncertainty estimates, and the strength of any guarantees — and offers practical guidance on choosing between them for a given clinical problem.
It also looks forward honestly. Uncertainty quantification in medical imaging is an active research area with real open problems: scaling to foundation models, handling distribution shift between hospitals, and the still-unsettled question of how uncertainty should be communicated to the clinician who has to act on it.