You've reached the end
Nineteen sessions ago you started with a simple, uncomfortable observation: a model that is always confident is a model you cannot safely put in front of a patient. Since then you have worked through the clinical reasoning behind uncertainty, the main methods for measuring it, and the tools for checking whether those measurements can actually be trusted.
That is genuinely a lot of ground. Bayesian neural networks, variational inference, MC Dropout, Deep Ensembles, Evidential Deep Learning and Conformal Prediction are each substantial topics in their own right, and you have not just read about them — you have implemented them on real chest X-ray data and seen where each one helps and where it falls short.
Take the habit with you rather than any single method. Ask what your model does not know, insist on evidence that its confidence is calibrated, and design for the cases it should hand back to a human. That instinct is what makes clinical AI trustworthy, and it will outlast whichever technique happens to be state of the art.
What you learned
Foundations
Clinical motivation, the cost of confident errors, aleatoric versus epistemic uncertainty, and the softmax illusion.
Core UQ Methods
Bayesian methods (variational inference, MC Dropout), Deep Ensembles, Evidential Deep Learning, and Conformal Prediction.
Evaluation & Reliability
Calibration and reliability diagrams, risk–coverage analysis, and out-of-distribution detection.
Future Directions
Choosing a method for a real clinical problem, open challenges, and where uncertainty quantification goes next.
About the authors
Medical Student, Medical Imaging AI Researcher
Take it further
Every notebook, figure and the source of this site live in the public repository. Issues and pull requests are welcome, and a star helps the course reach the people who need it. The accompanying paper describes the course and evaluates it as a knowledge source for large language models.