After following this course, the student is able to
- implement and train deep neural networks for medical image reconstruction, segmentation, and registration.
- integrate classical mathematical image analysis and deep learning techniques.
- quantitatively evaluate methods for medical image reconstruction, segmentation, and registration.
- apply methods for interpretability, explainability, and uncertainty quantification to deep neural networks.
- provide unsupervised, semi-supervised and active learning solutions for working with limited or low-quality real-world medical data.
In recent years, the automated analysis of medical images like computed tomography (CT), ultrasound, and magnetic resonance imaging (MRI) has been revolutionized by deep learning. This umbrella term covers a wide range of machine learning methods that optimize|
artificial neural networks to perform tasks such as image reconstruction, segmentation, or registration. It is expected that deep learning will significantly impact image-driven medical specialties like radiology, radiotherapy, pathology, and dermatology. The advent of deep learning builds on decades of research in mathematical medical image analysis, combined with strong influences from computer vision and machine learning. However, genuinely successful deep learning in medical image analysis also requires domain knowledge about the clinical problem visualized, the physics underlying image formation, and the mathematics governing image reconstruction.
This course equips students with an understanding of the relationship between key concepts in this rapidly developing field and practical skills to address commonly occurring medical image analysis problems. Main topics include:
- The medical imaging pipeline and common image analysis problems
- Convolutional neural networks on images and manifolds
- Deep learning for image reconstruction, segmentation, and registration
- Mathematical image analysis and its relation to deep learning
- Quantitative evaluation of medical image analysis problems
- Interpretability, explainability, and uncertainty estimation in deep learning models
- Unsupervised, semi-supervised learning, and active learning on real-world data
- Approaches to working with multi-modal imaging and clinical data
Written exam (70%) and final project report + presentation (30%)
Assumed previous knowledge
- Knowledge of:
- linear algebra (e.g. module 1 AM 202001325, module 3 CS 202001205, module 3 BME 202001203, module 4 EE 202001209, module 4 TM)
- calculus (e.g. module 2 AM 202001223, module 2 CS 202001197, module 2 BME 202001195, module 1 TM)
- probability theory (e.g. module 4 AM 202001344, module 4 CS 202001233, module 8 EE 202001235)
- Basic programming experience in Matlab or (preferably) Python.
- Knowledge of machine learning. E.g., courses Machine Learning I (201600070) and Deep Learning - From Theory to Practice (201800177)
- Knowledge of inverse problems in imaging (201900209)
|Master Applied Mathematics|
|Master Biomedical Engineering|
|Master Technical Medicine|
|Master Electrical Engineering|
|Master Business Information Technology||Required materials|
|Chapters from: Handbook of Medical Image Computing and Computer Assisted Intervention, ISBN 9780128161760. The book is available via UT subscription.|
|Chapters from: Deep Learning - Goodfellow et al. ISBN 9780262035613. Freely available online.|
|Chapters from: Image Processing, Analysis, and Machine Vision - Sonka et. al. ISBN 9781133593690.|
|Written Exam, Final Project|