Fred Godtliebsen  –  Professor

Robert Jenssen  –  Professor

Puneet Sharma  –  Associate Professor

Kristoffer Wickstrøm  –  Associate Professor

Michael Kampffmeyer  –  Professor

Elisabeth Wetzer  –  Associate Professor

Martin Skjelvareid  –  Associate Professor

Benjamin Ricaud  –  Associate Professor

Hyeongji Kim  –  Postdoctoral Fellow

Elin Kile  –  Doctoral Research Fellow

Christian Salomonsen  –  Doctoral Research Fellow

Rwiddhi Chakraborty  –  Researcher

Eirik Myrvoll-Nilsen  –  Postdoctoral Fellow

Dennis Adamek  –  Doctoral Research Fellow

Tianchao Li  –  Doctoral Research Fellow

Daniel Kaiser  –  Doctoral Research Fellow

Solveig Thrun  –  Doctoral Research Fellow

Johan Mylius-Kroken  –  Doctoral Research Fellow

Magnus Oterhals Størdal  –  Doctoral Research Fellow

Changkyu Choi  –  Postdoctoral Fellow

Jørgen Andreas Agersborg  –  Doctoral Research Fellow

Lars Uebbing  –  Doctoral Research Fellow

Nikita Shvetsov  –  Postdoctoral Fellow

Jan Zavadil  –  Doctoral Research Fellow

Suaiba Amina Salahuddin  –  Researcher

Jørgen Aarmo Lund

Youssef Wally  –  Doctoral Research Fellow

Durgesh Kumar Singh  –  Doctoral Research Fellow

Marit Dagny Kristine Jenssen  –  Doctoral Research Fellow

Eirik Agnalt Østmo  –  Doctoral Research Fellow

Theodor Anton Ross

Harald Lykke Joakimsen  –  Doctoral Research Fellow

Kristian Dalsbø Hindberg  –  Researcher

Stine Hansen  –  Researcher

Joel Burman  –  Data Scientist

Ashenafi Zebene Woldaregay  –  Data Scientist

Gitta Astrid Hildegard Kutyniok  –  Professor

Sebastien Francois Lefevre  –  Professor

Karl Øyvind Mikalsen  –  Associate Professor

Shujian Yu  –  Associate Professor

Samuel Kuttner  –  Associate Professor

Mark Girolami  –  Professor

Petter Bjørklund  –  Adviser

Inger Solheim  –  Head of Office

Erik Heggeli  –  Head Engineer

Graph neural networks

We have several research projects focused on graphs and graph machine learning. The main topics are: 

• theoretical concepts, connecting Math and Physics to Machine Learning,

• using graph neural net for generative AI,

• application to document analysis.

See our Graph ML Group website for more details:

https://ngmlgroup.github.io/theses.html 

Recommended prerequisites: Knowledge of machine learning, deep learning and signal/image processing. Good programming skills (Python).

Contact person: Benjamin Ricaud (benjamin.ricaud@uit.no)

Deep learning aided quantification in PET imaging

Positron emission tomography (PET) is a medical imaging technique that visualises the distribution of an injected radioactive tracer in living subjects. Medical imaging with PET plays an important role in the detection, staging, and treatment response assessment of many diseases, including cancer, neurological and cardiovascular conditions, as well as inflammation and infection.

PET is quantitative, in the sense that it allows not only visualisation but also non-invasive quantification of regional tracer uptake. Specifically, with dynamic PET imaging, it is possible to fully assess the time-dependent tracer distribution in the body. This allows quantification of biological processes, such as glucose metabolism or blood flow, by using tracer kinetic modelling. Tracer kinetic modelling, however, requires accurate determination of an arterial input-function (AIF), i.e. the tracer time-activity curve in blood.

The gold-standard AIF is obtained by measuring the time-dependent FDG radioactivity concentration in arterial blood through invasive blood sampling. This procedure is complex, time-consuming, and potentially painful with risk for complications.

The aim of this Master´s project is to develop novel methodology for arterial input function prediction, building upon state-of-the-art deep learning methods. In addition, the project will focus on developing interpretability and uncertainty methods to explain outcomes and the variability in the predictions. Both preclinical (mice) and clinical (human) PET data will be explored in the project.

Prerequisites:

• Relevant machine learning courses, for instance FYS-2021 and FYS-3033

• Programming skills in Python, preferably using PyTorch and/or TensorFlow

• Experience with medical image processing is an advantage
  

Contact person: Samuel Kuttner (samuel.kuttner@uit.no)

Hierarchical medical image segmentation with novel metric learning

Metric learning is a technique that focuses on learning similarity by mapping data into a feature space where similar data points are closer together. Traditional prototype-based approaches achieve this through directional separation of classes, which may not adequately capture the inclusion relationships inherent in hierarchical structures. This limitation is evident in medical imaging scenarios, where organs like the left and right kidneys exhibit higher semantic similarity to each other than to other organs.

Recent advancements, such as the Anomaly Detection Network (ADNet) for medical image segmentation, introduce an anomaly detection-inspired approach that effectively distinguishes between foreground and background classes using a single prototype. By treating the foreground class as typical (close to the prototype) and the background as non-typical (distant from the prototype), ADNet employs an inside-outside separation strategy that aligns well with representing inclusion relationships among classes.

Incorporating these inclusion relationships into metric learning holds significant potential for enhancing the accuracy and semantic coherence of medical image segmentation, addressing the limitations of conventional methods.

To achieve this, this project proposes a hierarchical metric learning framework inspired by anomaly detection principles. Specifically, the method leverages an inside-outside separation strategy to effectively model hierarchical class structures at multiple levels. The method’s effectiveness will be examined using hierarchical datasets, with initial experiments planned in medical imaging contexts focusing on segmentation accuracy and hierarchical performance. Depending on data suitability, this project may be experimented on non-medical segmentation or hierarchical classification tasks.

Contact persons:

Hyeongji Kim (hyeongji.kim@uit.no)

Elisabeth Wetzer (elisabeth.wetzer@uit.no)

Digital Pathology

Deep learning plays a transformative role in digital pathology by enabling automated image analysis, enhancing diagnostic accuracy, identifying complex patterns in histopathological data, accelerating workflows, supporting personalized medicine, and facilitating large-scale research through high-throughput image processing.

Potential directions for a thesis are the following:

• Unsupervised cell type identification in highly multiplexed imaging techniques such as imaging mass cytometry

• Representation learning for multimodal image registration

• Representation learning and generative methods for modality translation to improve cell detection in label-free microscopy

• Studying tissue composition and signaling pathways of cells in the tumor microenvironment

Recommended prerequisites: Knowledge of machine learning, deep learning, signal processing and image analysis. Good programming skills (Python).

Contact person: Elisabeth Wetzer (elisabeth.wetzer@uit.no)

Explainable artificial intelligence

Explainable artificial intelligence (XAI) seeks to increase the reliability and transparency of AI systems, which is a necessity for applications in safety critical domains such as healthcare. There are many potential directions for the thesis:

• Unsupervised XAI
• XAI for vision-language models
• Self-explainable models.
• Case studies in applied settings such as healthcare and remote sensing.

Recommended prerequisites: Knowledge of machine learning, deep learning and signal/image processing. Good programming skills (Python).

Contact person: Kristoffer Wickstrøm (kwi030@uit.no)

Uncertainty modeling in deep learning

Uncertainty modeling is of vital importance when applying deep learning to tasks with safety critical outcomes. Project could take many forms:

• Fault and out-of distribution detection
• Confidence estimation in regression tasks
• Unsupervised uncertainty modeling
• Case studies in applied settings such as healthcare and remote sensing

Recommended prerequisites: Knowledge of machine learning, deep learning and signal/image processing. Good programming skills (Python).

Contact person: Kristoffer Wickstrøm (kwi030@uit.no)

Exploiting Context and Dependencies

Deep Learning has excelled in object recognition and classification tasks in computer vision. A key aspect of this remarkable success is the amount of data on which these methods train. However, for example biomedical applications face the problem that the amount of training data is limited. In particular, labels and annotations are usually scarce and expensive to obtain as they require expert domain knowledge.

One way to overcome this issue is to use additional knowledge about the data at hand. This guidance can come from expert knowledge, which puts focus on specific, relevant characteristics in the images, or geometric priors which can be used to exploit the spatial relationships in the images.

Potential directions for a thesis are the following:

• Geometric Deep Learning (Equivariant Neural Networks, Non-Euclidean Embeddings, Topology Preserving Networks)

• Physics-informed Machine Learning to integrate physics modeling and data-driven learning approaches

• Learning under privileged information by e.g. exploiting additional modalities

Recommended prerequisites: Knowledge of machine learning, deep learning, signal processing and image analysis. Good programming skills (Python).

Contact person: Elisabeth Wetzer (elisabeth.wetzer@uit.no)

Monitoring Information Flow in Concept Bottleneck Models

Concept Bottleneck Models (CBM) are neural networks that learn human-defined concepts while making predictions on images. There are a variety of techniques in what concepts can be learned, how these networks are trained, and how concept interventions (or new concepts) can be taught to the network when training.

One key issue that we are interested in is the information monitoring while training a CBM. When do CBMs "learn" a concept during training? When do dominant concepts "emerge" during training? Can concept learning be "forgotten"? These are the questions that interest us in this project.

Related Work:
1. Koh, Pang Wei, et al. "Concept bottleneck models." International conference on machine learning. PMLR, 2020.

Recommended prerequisites:
1. Comfortable knowledge of Python
2. Light experience with Pytorch (no expertise assumed).
3. Completion of the deep learning course offered by UiT MLG.

Contact person: Rwiddhi Chakraborty (rwiddhi.chakraborty@uit.no)

Probabilistic Dynamic Line Rating for Energy Analytics

As electricity demand increases and variable renewable energy sources are integrated, transmission lines are operating closer to their thermal limits. To maintain security of electricity supply and grid reliability while also making the most of available transmission capacity without taking additional risk, Dynamic Line Rating provides a powerful data-driven solution. Avju’s Observer units collect real-time measurements of conductor temperature, sag, and environmental conditions. This data are essential inputs needed to compute safe power transfer limits

This thesis will build a probabilistic forecasting pipeline that supports Dynamic Line Rating by predicting the key variables influencing ampacity. These include temperature, wind speed and wind direction, humidity, pressure, and cloud cover. Accurate forecasts with calibrated uncertainty allow grid operators to safely unlock hidden capacity and respond more flexibly to both operational needs and weather variability.

Thesis goal and approach

Contact Puneet or Jonas to set up a meeting for specific information regarding the project.

Expected impact

Well-calibrated ampacity forecasts enable smarter grid operation by safely increasing power transfer without requiring new physical infrastructure. This capability is increasingly important as energy systems face growing demand for electricity and external pressure from climate variability.

Prerequisites

Machine Learning-Based Threat Detection

Cybersecurity is complex, but has major real-world implications. To effectively protect critical assets from threats, cybersecurity must be approached through reliable, systematic, and precise solutions.

This project is done together with the company Mnemonic, and will be focused on challenges closely related to the services that Mnemonic provides. Cybersecurity systems must process a huge volume of data in a continuous stream, and high precision is of utmost importance to ensure that critical assets are protected. This project will focus on how relevant features can be extracted from cybersecurity logs and how potential threats can be precisely and reliably identified based on these features. Furthermore, aspects related to informative data visualization and explainability can also become relevant.

Prerequisites

• Relevant machine learning courses, for instance FYS-2021, FYS-3012, and FYS-3033
• Programming skills in Python, preferably using PyTorch and/or TensorFlow
• Experience with graph-based learning and cybersecurity is a plus

Contact people: Kristoffer Wickstrøm (kwi030@uit.no) and Elisabeth Wetzer (elisabeth.wetzer@uit.no)