CoMI-L — Winter 2026/27 | Albarqouni Lab

Lab: Computational Medical Imaging (CoMI)

Lecturers: Prof. Dr. Shadi Albarqouni
Tutors:
Module: MA-INF 2230
Credits: 4 SWS, 9 CP (270h workload)
Programme: M.Sc. Computer Science
Mode: Optional (2. or 3. Semester)
Language: English
Format: Lab (Praktikum), groups of max. 3 students
Frequency: Every year

Description

The Computational Medical Imaging (CoMI) Lab is a project-based course in which students carry out a semester-long research project in computational medical imaging. Working in small groups, students identify a well-defined problem, collect or are provided a suitable dataset, implement baseline and novel approaches, and evaluate their results both quantitatively and qualitatively.

Possible project topics (chosen or proposed by the instructors) include:

Anatomical Structure Segmentation (e.g., multi-atlas, CNN-based)
Deformable Registration (e.g., classic demons, learning-based registration)
Lesion/Anomaly Detection (e.g., out-of-distribution detection in X-ray)
Radiomics & Quantitative Biomarker Extraction (e.g., texture features, deep radiomics)
Uncertainty Estimation in Deep Learning (e.g., MC-Dropout, Bayesian networks)
Causality & Bias Analysis in Medical Datasets (e.g., domain shift, fairness)

Learning Goals

Technical Skills

Acquire a deep understanding of a selected problem in computational medical imaging: from problem identification to data processing, algorithm design, implementation, and evaluation.
Implement, test, and validate state-of-the-art image-analysis algorithms (e.g., segmentation, registration, reconstruction, classification) on real medical data.
Develop proficiency in using libraries and toolkits commonly used in medical imaging (e.g., ITK, SimpleITK, PyTorch for deep learning, nibabel/dipy).
Learn to document software systematically (API documentation, code comments, unit tests) and to generate reproducible experiments (version control, containerization).

Soft Skills

Collaborate effectively in small groups over the entire semester, dividing tasks (e.g., data curation, algorithmic implementation, evaluation).
Prepare clear, concise technical documentation and design rationales.
Present project progress and final results in oral form (midterm demo, final demo).
Critically assess one’s own results and position them within the state-of-the-art in computational medical imaging.

🗓 Project (Winter 2026/27)

Who Does the Model See? Bias of Ultrasound Foundation Models across Countries and Income Settings in Fetal Ultrasound

Ultrasound is the imaging modality of antenatal care worldwide, and in low- and middle-income countries scans are increasingly done by non-specialists on low-cost devices. Ultrasound foundation models promise label-efficient AI support exactly there — but they are pretrained almost entirely on data from high-income hospitals. Students build a harmonised benchmark from 12 openly licensed fetal-ultrasound releases (13 countries, four World Bank income groups) and test whether frozen foundation models work equally well for fetuses scanned in Malawi, Uganda, Sierra Leone, Ghana, Egypt, Algeria, Brazil and China as in Spain, the Netherlands, Romania and the UK.

Bias metrics: worst-group performance, between-group gaps and calibration, with patient-level confidence intervals.
Clinical highlight: a measurement study — do errors in the cephalic index, abdominal circumference and cardiothoracic ratio, and so the share of fetuses flagged as abnormal, differ by site, device and income group?
Models: FetalCLIP, USFM, SonoNet-64 (USF-MAE secondary).
Team: 3 students; all share data harmonisation and the final synthesis, each leads one experimental work package. Students implement their own toolkit (fusbias) against a provided specification and acceptance tests.

WP	Topic	Lead
1	Data harmonisation, de-duplication, patient-level splits and quality control	all
2	Zero-shot and training-free standard-plane recognition	S1
3	Frozen-feature probing and bias audit	S2
4	Clinical measurements: cephalic index, abdominal circumference, cardiothoracic ratio	S3
5	Representation analysis, synthesis and report	all

Learning Format

Students work in groups of up to 3 on a semester-long project. Each group will:

Identify a well-defined problem and collect or be provided a suitable dataset (publicly available or hospital data under ethics approval).
Perform a literature review to select baseline methods.
Implement baseline and at least one novel or improved approach.
Evaluate quantitatively (e.g., Dice, Hausdorff, MSE, classification accuracy) and qualitatively (expert review where possible).
Document the software (README, API docs) and provide a reproducible environment (Docker/Singularity).
Present midterm progress (oral demonstration + short written summary) and final results (oral + poster or full report).

Assessment

Graded Exam:
Oral presentation (final demo + Q&A, 20 minutes per group). Written report (12–16 pages, including background, methods, results, discussion).

Ungraded Coursework (required for admission to the exam):
Project work; attendance in course sessions in accordance with the exam regulations of 2023, § 12(6).

Recommended Prerequisites

At least one of the following:

MA-INF 2222 – Visual Data Analysis
MA-INF 2312 – Image Acquisition and Analysis in Neuroscience
MA-INF 2220 – Lab Visualization and Medical Image Analysis (or equivalent)

A solid background in Python programming (and/or C++), linear algebra, calculus, probability theory, and numerical algorithms is required. Familiarity with deep learning frameworks is highly recommended.

References and Resources

Relevant Journals:

Nature Machine Intelligence
Medical Image Analysis (MedIA)
IEEE Transactions on Medical Imaging (TMI)
Radiology: AI, European Radiology, IEEE Journal of Biomedical and Health Informatics

Relevant Conferences: