Uncertainty Quantification in Medical Imaging Analysis
Teaching models to know what they don't know — from clinical motivation to working PyTorch code.
The aim
This course is about a single, practical problem: how do you build a medical imaging model that knows when it might be wrong? Most machine learning teaching stops at accuracy, but a model deployed in a hospital needs to do more than be right most of the time — it needs to signal the cases where a clinician should look more carefully. We start from the clinical reasoning behind that idea, then work through the main techniques for quantifying uncertainty, implementing each one in PyTorch on real chest X-ray data, and finish by evaluating the uncertainty those methods produce — because an uncertainty estimate is only useful once you can show it is trustworthy. By the end you should be able to take an existing model, attach a well-founded uncertainty estimate to its predictions, check honestly whether that estimate can be trusted, and decide which method fits the problem in front of you.
Prerequisites
Python
Machine learning fundamentals
Deep learning & PyTorch
Medical imaging data
What you'll learn
Course outline
Clinical motivation, the two failure modes of AI (overconfidence and underconfidence), the two kinds of uncertainty, and how clinicians already reason probabilistically.
The methods themselves — Bayesian neural networks, variational inference, MC Dropout, Deep Ensembles, Evidential Deep Learning and Conformal Prediction. Each concept session is followed by a PyTorch implementation on real chest X-ray data.
Having produced uncertainty estimates, we ask whether they can be trusted: calibration, risk–coverage trade-offs, and detecting inputs that fall outside the training distribution.
Pulling the pieces together: how to choose a method for a real clinical problem, what is still unsolved, and where the field is going.
Get started
Begin at the beginning, or jump to any session from the sidebar — each one stands on its own.
Start with Part 1 →