Lecture + Exercises · Informatics / Mathematics & Natural Sciences

Computational Medical Imaging (Lecture) (CoMI)

The core module of the Computational Medical Imaging curriculum in the M.Sc. Computer Science (4 h lecture + 2 h exercises per week). From medical data and clinical evaluation to deep learning for classification, detection and segmentation, EHR and survival, graphs and foundation models, and on to explainability, uncertainty, fairness and deployment under the EU AI Act — with a Python lab notebook every week.

Current / next semester SS 2027 (to be confirmed) Thursdays, 16:00–18:00 Dates, schedule, tutors & contact →

Course structure

Every session comes with interactive slides, including in-depth material for computer scientists, and a lab notebook. All labs run on a laptop CPU with the course data pack; lab notebooks are shared with registered participants. Exercises: the weekly lab notebook, CS track, graded by automatic checks.
SessionLecture and topicsLab
Week 1L1 Why AI in medicine
clinical pressure points; promise and evidence; what goes wrong; humans and AI together
you, a model, and the two of you together (blood cells, chest X-ray)
Week 2L2 Medical data
modalities, planes, DICOM; cleaning and splitting; annotation and agreement; reconstruction (sinogram and FBP, k-space, compressed sensing, hallucinations)
voxels, tables, leaks, labels, and how an image is made (CT pack)
Week 3L3 Machine learning foundations
optimisation; unsupervised and supervised learning; classic models; bias–variance
ML foundations on blood cells
Week 4L4 Evaluation with clinical metrics
confusion matrix; ROC and PR; prevalence and operating points; confidence intervals
clinical metrics on chest X-rays
Week 5L5 Deep learning for classification and detection
perceptron to CNN; architectures and transfer; detection, IoU, mAP; class imbalance
filters, small CNNs, rare findings and boxes
Week 6L6 Beyond images
EHR and risk scores; time series and early warning; survival analysis; temporal leakage; RL for treatment; disease progression (RECIST, volume-doubling time, mixed-effects and event-based models)
time to event and progression
Week 7L7 Segmentation
classic methods; U-Net and nnU-Net; Dice, HD95, losses; few labels and SAM; registration and tracking (rigid to deformable, mutual information, lesion and cell tracking)
outline every organ, then find where it moved (CT pack)
Week 8L8 Graphs and geometric deep learning
graphs in medicine; Laplacian and spectral methods; message passing and GNNs; graph foundation models
relations, not rows
Week 9L9 Foundation models and multimodal data
self-supervision; multimodal fusion and contrastive learning; vision–language models (CLIP); clinical LLMs and agents
foundation models and multimodal data
Week 10L10 Interpretability and explainability
why XAI; taxonomy; attribution methods; evaluating explanations; explaining GNNs and VLMs
explanations you can test
Week 11L11 Uncertainty quantification
aleatoric and epistemic; MC dropout and ensembles; calibration; refer or decide; conformal prediction; uncertainty for graphs and LLMs
calibration, answer sets and shift
Week 12L12 Bias, fairness and causality
causal diagrams; dataset shift; fairness metrics; mitigation; bias in graphs and foundation models
subgroups, shift and causes (incl. a skin-tone audit of a public dermatology dataset)
Week 13L13 Trustworthy AI in practice
ethics and the EU AI Act; evidence and reporting; monitoring and drift; federated learning and privacy; carbon footprint
rules, federation, monitoring, footprint

Toolbox

Self-study in week 0, before the first lecture.

  • T1 Vectors, matrices and gradients: scalars to tensors; dot products and linear maps; distances, Euclidean to geodesic; convexity; gradient, Jacobian, Hessian
  • T2 Probability and statistics: z-scores and outliers; distributions, Bernoulli to Rician; CLT and confidence intervals; Bayes and Bayesian updating; MLE, MAP, entropy, KL
  • T3 Python toolbox: notebooks and Colab; NumPy on the course CT; pandas; scikit-learn and PyTorch cheat sheets · Lab: arrays, tables, models

Data used in the labs: MedMNIST subsets (BloodMNIST, PneumoniaMNIST, RetinaMNIST, DermaMNIST); the CoMI CT pack (40 TotalSegmentator CT subjects, published on Hugging Face under CC BY 4.0); public survival tables (gbsg, pbcseq); simulated cohorts.

Learning objectives

  • Explain where AI helps and where it fails in clinical practice, and judge the evidence behind a medical AI system (L1, L13).
  • Handle medical data: imaging modalities and DICOM, patient-level splits without leakage, annotation and inter-rater agreement, and how images are reconstructed from raw signals (L2).
  • Train and compare classic machine-learning models, and reason about optimisation and the bias–variance trade-off (L3).
  • Evaluate models with clinical metrics: ROC and precision–recall curves, operating points under real prevalence, and confidence intervals (L4).
  • Build deep networks for classification, detection and segmentation (CNNs, transfer learning, U-Net and nnU-Net), and register images (L5, L7).
  • Model data beyond images: electronic health records, time series, survival and disease progression; graphs and graph neural networks; foundation models, vision–language models and clinical LLMs (L6, L8, L9).
  • Make models trustworthy: test explanations, quantify and calibrate uncertainty (including conformal prediction), and audit bias and fairness with causal diagrams (L10–L12).
  • Take a model towards deployment: the EU AI Act and medical-device rules, reporting guidelines, monitoring for drift, federated learning and privacy, and carbon footprint (L13).
  • Implement all of the above in Python, in weekly lab notebooks with automatic checks.

Target audience

Students of the M.Sc. Computer Science at the University of Bonn (optional module, 3rd semester).

Prerequisites

  • Programming in Python (NumPy; first experience with PyTorch helps).
  • Linear algebra, calculus and probability at Bachelor level. The toolbox (T1–T3: maths, statistics, Python) is a refresher to work through before week 1.
  • No medical background needed: the clinical context is taught in the course.
  • Recommended: MA-INF 2312 Image Acquisition and Analysis in Neuroscience (or equivalent knowledge of medical imaging modalities) and/or MA-INF 2222 Visual Data Analysis (or equivalent knowledge of statistical data analysis).

Evaluation

  • Graded written exam.
  • Admission to the exam requires successful participation in the exercises.

Semesters

SemesterCourseFormatAudienceLecturers & tutors
SS 2027 (to be confirmed)Computational Medical Imaging (Lecture) (CoMI)Lecture + ExercisesInformatics / Mathematics & Natural Sciences