This course is about deep learning for biomedical images: CT and MRI volumes, mammograms, ultrasound, microscopy and whole-slide images of tissue. The networks are the ones you know from computer vision. What changes is everything around them. The input is a physical measurement, not a photograph. Labels are scarce, expensive and uncertain. Images from a new hospital may not resemble the training data. And a prediction feeds a decision about a patient, so a good score on a test set does not show that a model helps anyone.
The path from a patient/sample to a clinical/biological decision, with the module in this course that covers each step.
Have a look at the following pages, which are relevant throughout the course:
Foundations Review
Prerequisite knowledge the course builds on, from supervised learning, neural networks, CNNs and transformers, to probability.
Notation
The mathematical symbols used throughout the course notes.
Clinical and biological terms
The biological and medical vocabulary that later pages may take for granted.
What does each module cover?
- Biomedical Image Acquisition and Data Modalities: what each imaging modality measures, how its signal becomes a digital image, and what an image file stores besides the pixels.
- Data scarcity, domain gap, transfer learning: training when labeled data are scarce, and test data come from other devices or hospitals.
- Segmentation in biomedical image analysis: from how labels are made to architectures, losses and post-processing.
- Statistical analysis & Evaluation metrics: how to measure performance without falling into traps, and how to tell a real difference between models from noise.
- Trustworthy AI: what a model needs beyond accuracy before it can be used in a clinic: robustness, fairness, uncertainty, explanations, and how it can fit in the clinical workflow.
- Multi-modal & Multi-view learning: models that combine several modalities, or several views of the same patient.
- Generative AI in Medical Imaging: generative models and their uses in biomedical image analysis.
How should you use these notes?