It's time to think about your future and your thesis project. Here, we list some tips on research groups and companies with interesting thesis project proposals.
Are you looking for students for your research group or company and want to be included on this list? Contact Linda Bresäter, collaboration coordinator at the Department of Computing Science at Umeå University.
Developing an educative videogame that train players’ spatial skills
Project description:
Spatial cognition studies have shown that there is a strong link between success in Science, Technology, Engineering and Math (STEM) disciplines and spatial abilities (Newcombe, 2010). Spatial thinking (Hegarty, 2010)involves thinking about how shapes build together a puzzle (e.g. jigsaw or tangram), how to fold a cardboad to build a box, how to stack cubes to build as stable tower, etc.
Research studies by Verdine et al. (2014) showed that students with low spatial abilities under-perform in STEM tasks and then they avoid STEM disciplines when selecting college. Children in families with low socio-economic status are also disadvantaged in spatial ability development (Wai et al., 2009). Lippa et al. (2010) also reported a gender gap in spatial ability.
Afortunately, the research studies by Sorby (2009) showed that spatial skills can be developed through practice: students that attended an engineering graphics gateway course at university to improve their ability to visualize in three dimensions, improved also their success and retention significantly, particularly female students. Thus, spatial thinking can be taught using visual and kinetic interactions offered by new digital technologies (Highfield and Mulligan, 2007) and research has demonstrated that video game training enhances cognitive control (Spence and Feng, 2010) specially when aging (Anguera et al., 2013).
Aim of the project:
The aim of this project is to design and develop a videogame with the aim of training any of the spatial skills mentioned by Newcome (2010, p.30, “Tests of Spatial Thinking”). This includes designing a story, and an avatar and levels for training, but also implementing the corresponding procedures for gathering gameplay data and store that in a database (e.g. Google Firebase) for posterior analysis.
A user study to test participants’ enjoyment and improvement while playing the game is a also relevant aspect of this project.
Ideal candidates should have a strong background in computer graphics, Unity, C#, Android programming, and databases (e.g. Google Firebase). This project can be carried out individually or in groups of 2 or 3 students.
Supervisor:
This activity will be supervised by Zoe Falomir. In this context, we have developed before three videogames for training spatial skills: “Paper Folding Game”, “Blocks in the Woods” and “Treasure Hunt” (see https://spatialreasoninggames.weebly.com).
If you think this sounds interesting, please contact zoe.falomir@umu.se for more details!
References
Anguera, J. A., Boccanfuso, J., Rintoul, J. L., Al-Hashimi, O., Faraji, F., Janowich, J., Kong, E., Larraburo, Y., Rolle, C., Johnston, E., and Gazzaley, A. (2013). Video game training enhances cognitive control in older adults. Nature, 501(7465):97–101.
Hegarty, M. (2010). Chapter 7 - components of spatial intelligence. Psychology of Learning and Motivation, vol 52, pp 265–297.
Highfield, K. and Mulligan, J. (2007). The role of dynamic interactive technological tools in preschoolers’ mathematicalpatterning. In Watson, J. and Beswick, K., editors, Proc. of the 30th annual conference of the Mathematics EducationResearch Group of Australasia, volume 1, pages 372–381.
Lippa, R. A., Collaer, M. L., and Peters, M. (2010). Sex differences in mental rotation and line angle judgments are positively associated with gender equality and economic development across 53 nations. Archives of Sexual Behavior,39:990–997.
Newcombe, N. (2010). Picture this: Increasing math and science learning by improving spatial thinking. American Educator,34(2):29–35.
Sorby, S. A. (2009). Educational research in developing 3D spatial skills for engineering students. Int. J. of Science Education, 31(3):459–480.
Spence, I. and Feng, J. (2010). Video games and spatial cognition. Review of General Psychology, 14(2): 92.
Verdine, B. N., Golinkoff, R. M., Hirsh-Pasek, K., and Newcombe, N. S. (2014). Finding the missing piece: Blocks, puzzles, and shapes fuel school readiness. Trends in Neuroscience and Education, 3(1):7–13.
Wai, J., Lubinksi, D., and Benbow, C. P. (2009). Spatial ability for STEM domains: Aligning over 50 years of cumulativepsychological knowledge solidifies its importance. Journal of Educational Psychology, 101(4): 817–835.
TRAIN Lab, Our mission is to develop new AI models that boost the transition to 5P medicine in a wide range of medical tasks. 8 project proposals
Proposal 1. Evaluating and Extending X-GeM: An In-House Multimodal Foundation Model for Medical Data Generation
At TRAIN Lab (https://sites.google.com/view/trainlab/home-page) we aim to advance artificial intelligence in medicine, with an emphasis on medical image analysis. Our mission is to develop new AI models that boost the transition to 5P medicine in a wide range of medical tasks.
To this goal, our research focuses on two main areas. On the one side, we build generative approaches that address the need for more data in medicine and enhance the quality of available data. On the other side, we develop multimodal learning methods that extract knowledge from different data modalities, ensuring fairness and transparency.
Background
The adoption of Artificial Intelligence in medical imaging holds great promise but remains hindered by challenges such as data scarcity, privacy constraints, and the need for robust multimodal data integration. Generative modelling has emerged as a promising solution, enabling synthetic data generation for augmentation and anonymization purposes. To address these challenges, we developed XGeM, a 6.77-billion-parameter multimodal generative model capable of synthesizing chest X-rays and radiological reports jointly. Unlike existing unimodal or unidirectional approaches, XGeM enables any-to-any
synthesis across modalities via a shared latent space and Multi-Prompt Training strategy. Initial validation on the MIMIC-CXR dataset demonstrated state-of-the-art performance, with positive results in both benchmark metrics and expert radiologist evaluations. While these results are promising, testing and adapting XGeM beyond MIMIC-CXR is crucial, to ensure clinical robustness and generalizability. To this end, this thesis will focus on validating and adapting XGeM on external datasets, covering different data modalities and anatomical districts.
Aim of the project
The aim of this thesis is to evaluate and adapt XGeM on external medical datasets covering diverse data modalities and anatomical districts, with a focus on testing generalization, robustness, and adaptability to new clinical contexts. Specific objectives include:
1) Benchmarking XGeM against external datasets of chest X-rays and radiological reports.
2) Investigating domain adaptation strategies (fine-tuning, prompt adaptation, data harmonization).
3) Assessing the realism and clinical alignment of generated data under distribution shifts.
Work description
The project will consist of the following tasks:
1. Dataset Preparation: select and preprocess medical datasets.
2. Model evaluation on external data: Test baseline performance of XGeM on unseen datasets, analysing limitations and performance gaps compared to baseline performance.
3. Domain adaptation: Implement fine-tuning strategies for adapting XGeM to unseen data distributions (prompt adaptation, contrastive re-alignment, or lightweight transfer learning).
4. Quantitative and qualitative validation: Benchmark model outputs with standardized evaluation metrics (e.g., FID, BLEU, clinical report metrics). Conduct expert-based assessments where feasible to evaluate clinical realism.
5. Documentation and Reporting: Document methodology, adaptation strategies, and results in a final report.
Supervisor(s):
This activity will be supervised by Paolo Soda and Francesco Di Feola. For further information, please contact paolo.soda@umu.se and francesco.feola@umu.se
Proposal 2. Embedding Time in Generative Models for Temporal Medical Image Synthesis
Generative Artificial Intelligence (AI) has emerged as a powerful tool for medical imaging, with applications in image synthesis, translation between modalities, and data augmentation. Beyond single time-point imaging, an exciting frontier is the ability to generate longitudinal sequences of medical images that capture disease progression or treatment response. Such models could support prognosis, reduce data scarcity, and provide new tools for simulation. Recently, diffusion models and other generative frameworks have shown remarkable ability to synthesize realistic and clinically meaningful images. Extending these methods to temporally-aware modeling, capturing changes between consecutive scans, offers the potential to simulate disease evolution across multiple time points. However, research in this area is still in its early stages.
Aim of the project
The aim of this thesis is to embed temporal information into generative AI models for longitudinal medical imaging. By integrating time as an explicit component of the modeling process, the project seeks to generate realistic follow-up scans that reflect disease progression and treatment response. The focus will be on developing, adapting, and evaluating temporally-aware generative models (e.g., diffusion models, conditional GANs) to improve predictive imaging and broaden applications such as prognosis support, data augmentation, and virtual disease progression modeling.
Work description
The thesis will involve the following main tasks:
1. Literature Review: survey state-of-the-art approaches in generative AI for medical imaging, with emphasis on temporal and longitudinal modeling.
2. Data Preparation: Select and preprocess longitudinal imaging datasets (e.g., CT, MRI) suitable for temporal generative modeling.
3. Model Development: Implement or adapt a generative model that explicitly incorporates temporal information (e.g., via latent space conditioning, time embeddings, or contrastive alignment).
4. Evaluation: Quantitatively evaluate generated images using technical metrics (PSNR, SSIM, FID, task-based metrics), Qualitatively assess realism and clinical plausibility with medical expert input.
5. Documentation and Reporting: Document methodology, experiments, and findings in a structured thesis report.
Supervisor(s):
This activity will be supervised by Paolo Soda and Francesco Di Feola. For further information, please contact paolo.soda@umu.se and francesco.feola@umu.se
Proposal 3. Foundation Model Embeddings and XAI for Interactive Longitudinal Patient Trajectory Analysis
Longitudinal biomedical imaging, such as repeated CT or MRI scans, provides key insights into how diseases progress under treatment. Comparing these temporal trajectories across patients is challenging due to high data dimensionality and heterogeneous evolution patterns. Foundation models trained on large-scale biomedical datasets can extract embeddings that capture relevant spatial and temporal features, enabling clustering of patients with similar trajectories. To make these analyses clinically meaningful, an interactive Human-Machine Interface (HMI) is required. Such a system should allow clinicians to visualize trajectories, identify subgroups of patients with similar evolution, and apply explainable AI (XAI) methods to highlight the features that drive differences between groups.
Aim of the project
This project aims to develop an interactive HMI for longitudinal biomedical imaging analysis. The objectives are:
1. Extract embeddings from foundation models applied to longitudinal sequences.
2. Cluster patients according to temporal disease evolution.
3. Visualize trajectories and subgroup similarities through an interactive interface.
4. Integrate XAI to explain which imaging features or temporal patterns distinguish different patient groups.
The resulting system will combine representation learning, clustering, and explainability into a user-friendly tool to support clinical interpretation of longitudinal data.
Work description
1. Literature Review: Survey methods for longitudinal imaging, foundation models, clustering, and XAI.
2. Dataset Preparation: Preprocess longitudinal biomedical imaging datasets.
3. Embedding Extraction: Generate embeddings from foundation models at multiple time points.
4. Trajectory Clustering: Group patients based on embedding similarity across time.
5. HMI Development: Implement an interface for trajectory visualization, subgroup exploration, and
interactive comparisons.
6. XAI Integration: Apply explainability methods to identify the features and time points that explain
differences between subgroups.
7. Evaluation: Test the system on real datasets and refine it based on clinical feedback.
8. Reporting: Document methods, design, and results in the thesis report.
Supervisor(s):
This activity will be supervised by Paolo Soda and Filippo Ruffini. For further information, please contact paolo.soda@umu.se and filippo.ruffini@umu.se
Proposal 4. Test-Time-Adaptation for medical decision support systems
Supervised learning has long struggled with generalizing under distribution shifts, where the training and test data come from different distributions. Even minor differences between these datasets can cause state-of-the-art models to underperform. This issue is particularly prominent in the medical field, where models trained on specific types of medical images or patient demographics often fail when applied to different populations or imaging conditions. This mismatch leads to a significant drop in performance, raising serious concerns for clinical applications where reliable generalization is critical.
Aim of the project
This project seeks to explore Test-Time-Adaptation (TTA), a technique designed to address the challenge of distribution shifts by allowing pre-trained models to adapt to new, unlabeled data from the target domain before making predictions.
Unlike traditional approaches, TTA accesses the test data during the test phase, enabling dynamic adaptation to the target distribution. The goal is to move beyond the static, fixed decision boundary typically used during testing. Instead, the project aims to blur the line between training and testing, aiming for continuous learning and adaptation even after deployment. This paradigm shift holds promise for improved model performance and robustness, particularly in domains like healthcare where data variability is high.
Work description
This project focuses on developing and evaluating TTA techniques to improve model performance under distribution shifts.
The core objective is to enable pre-trained models, to adapt dynamically to new, unseen test data from different distributions without requiring additional labelled data.
Key tasks are:
1. Literature Review: research existing approaches to distribution shift and test-time adaptation, particularly in medical imaging.
2. Algorithm Development: develop and implement TTA methods that enable existing pre-trained backbones to adjust to test-time data.
3. Model training and testing: train models on source domain data and evaluate their baseline performance on shifted target domains (for this task we plan to use public datasets such as MedMnist).
4. Performance Evaluation: i) conduct quantitative and qualitative assessments to measure how well the TTA-augmented models generalize to new data; ii) compare against traditional supervised learning approaches to highlight the improvements in generalization.
Reporting and Documentation: document all stages of the project, including methods, results, and insights gained, producing a final report.
Supervisor(s):
This activity will be supervised by Paolo Soda and Francesco Di Feola. For further information, please contact paolo.soda@umu.se and francesco.feola@umu.se
Proposal 5. GenAI unlocks whole-body virtual scanning
Image-to-image translation is a technique that transforms images from one domain to another, offering significant potential in the medical field for creating a Digital Twin—a virtual replica of a patient. This approach aims to generate accurate images across multiple modalities, thereby reducing the need for multiple scans and minimizing radiation exposure. In medical imaging, the goal is to develop a virtual scanner capable of producing faithful, multi-modal images, helping diagnostic processes and improving patient safety. Current research in medical image-to-image translation has primarily focused on translating between specific modalities, such as MRI-to-CT or PET-to-CT, with promising results. However, most of these efforts are limited to specific anatomical regions, such as the head or chest, and do not address the more complex task of whole-body image translation. Developing methods for whole-body image translation introduces challenges due to the reduced amount of available data and higher computational cost, thus making model training more difficult and resource-intensive. Nonetheless, investigating whole-body image-to-image translation represents a significant step forward in creating a fully functional virtual scanner, advancing the capabilities of medical diagnostics and reducing patient risk.
Aim of the project
This project aims to develop a whole-body virtual scanner through image-to-image translation techniques based on Generative Adversarial Networks or diffusion models. By focusing on full-body medical image translation across multiple modalities, the project seeks to create an accurate and efficient Digital Twin of a patient, reducing the need for multiple scans, minimizing radiation exposure, and improving diagnostic precision.
Work description
This project will focus on developing a whole-body virtual scanner using advanced image-to-image translation techniques.
The key objectives and tasks for this project include:
1. Literature review: research existing approaches in medical image-to-image translation, focusing on full-body imaging and cross-modality applications, starting from a selection of already reviewed papers. This task will inform the project's direction and guide the development of the next tasks.
Data Preprocessing: This task involves cleaning, segmenting, and aligning the data to ensure consistency across modalities. We plan to utilize a publicly available whole-body PET-CT dataset, comprising approximately 1,000 patients. Preprocessing will be facilitated by ongoing research and existing scripts associated with this dataset.
Model Development: Implement the image-to-image translation framework, leveraging existing architectures like diffusion models, and adapting them to handle the challenges of full-body medical imaging, as identified in task 1.
Algorithm Optimization: Fine-tune the image translation algorithm to enhance accuracy in cross-modality image generation while maintaining anatomical detail. Novel techniques will be explored to reduce computational complexity and improve efficiency during training.
Validation and Testing: Assess the virtual scanner's performance using held-out data from the original dataset (Task 2) through both quantitative metrics and qualitative evaluation. For the qualitative assessment, we may involve radiologists collaborating with our team to provide expert visual inspections.
Reporting and Documentation: document all the stages of the project, including methods, results, and insights gained, producing a final report.
Supervisor(s):
This activity will be supervised by Paolo Soda and Francesco Di Feola. For further information, please contact paolo.soda@umu.se and francesco.feola@umu.se
Proposal 6. Physics-Informed Neural Networks for Tumor Response Modeling: An Interactive Visualization Framework for Radiotherapy Monitoring
Radiotherapy is a cornerstone in the treatment of many tumors, aiming to reduce tumor volume and slow disease progression.
Longitudinal medical imaging enables clinicians to track these changes over time, yet deriving quantitative insights about tumor response remains challenging. Physics-Informed Neural Networks (PINNs) represent a new paradigm in computational oncology, integrating imaging data with partial differential equations that describe tumor dynamics under radiation. PINNs can infer hidden biological and physical parameters, such as tumor decay rates or radiation response factors, thus enhancing interpretability beyond purely data-driven models. To make this approach usable in clinical practice, a Human-Machine Interface (HMI) is needed. Such an interface should not only visualize longitudinal images but also embed the outputs of the
PINN models, including estimated tumor decay parameters, predictive tumor trajectories, and interactive graphs. This integration will provide clinicians with a unified environment where tumor screening and model-driven predictions converge, supporting treatment monitoring and personalized decision-making.
Aim of the project
This project aims to develop an interactive Human-Machine Interface (HMI) that integrates Physics-Informed Neural Networks (PINNs) with longitudinal medical imaging data of tumors under radiotherapy. The objectives are to implement a PINN framework capable of estimating tumor decay factors and predicting tumor evolution under radiation; to design an HMI that embeds both medical images visualization and PINN model predictions, offering clinicians a single platform to analyze imaging data and computational insights.
By merging physical models with interactive visualization, the project will enhance the interpretability of tumor monitoring, foster clinical adoption of PINNs, and support personalized oncology.
Work description
The project will involve the following tasks:
1. Literature Review and requirements definition: Analyze current approaches in PINNs for biomedical
applications, tumor growth modeling, and visualization tools in radiotherapy.
2. Model Development: Implement a PINN framework that integrates imaging-based tumor measurements with differential equations describing radiation-induced tumor decay. Preprocess longitudinal medical images datasets and align them with PINN model inputs.
3. HMI Design and Development: Design an intuitive interface that enables: visualization of tumor images at different time points; interactive exploration of tumor volume evolution; graphical outputs of PINN-inferred decay factors and treatment response parameters.
4. Reporting and Documentation: Document methods, implementation, and results, producing a final report that includes both technical contributions and clinical insights
Supervisor(s):
This activity will be supervised by Paolo Soda and Giulia Romoli. For further information, please contact paolo.soda@umu.se and giulia.romoli@umu.se
Proposal 7. Multi-Agent Overall Survival Prediction on In-house Multicentric Lung Cancer Data
Computed Tomography (CT) is central to assessing tumor burden and predicting overall survival in lung cancer, yet multicentric datasets exhibit substantial acquisition variability, arising from differences in scanner manufacturer, reconstruction kernel, slice thickness, and noise characteristics that introduces non-biological variability. Conventional solutions typically apply image-level harmonization, often using GAN-based style transfer, but such approaches risk oversmoothing clinically relevant details and cannot fully accommodate the diversity of acquisition domains.
In this project, we shift the focus from explicit harmonization to downstream robustness by adopting a multi-agent learning framework. Multiple specialized survival-prediction models are trained to handle distinct acquisition subspaces, while a coordinating module dynamically selects or blends experts based on metadata or learned domain representations. This strategy treats harmonization as a representation-level challenge rather than an image transformation task, enabling more reliable and center-agnostic survival prediction across heterogeneous CT data.
Aim of the project
To develop a multi-agent survival prediction framework that learns to manage acquisition variability in multicentric CT data without explicit image harmonization. The system will: 1) Train domain-specialized survival-prediction agents; 2) Learn a coordinator/gating module for dynamic expert selection; 3) Evaluate whether multi-agent cooperation improves cross-center generalization and prognostic accuracy.
Work description
The project will involve the following tasks:
Literature Review: Survey domain generalization strategies, multi-agent frameworks, and deep learning methods for survival prediction. Identify limitations in current approaches to handling CT acquisition heterogeneity.
Dataset Analysis and Preprocessing: Characterize an in-house multicentric CT dataset, extract and cluster acquisition metadata (e.g., kernel, manufacturer), perform standard preprocessing steps, and establish center-aware training, validation, and test splits.
Multi-Agent Model Development: Implement multiple domain-specialized survival prediction models and design a coordinator module for hard or soft expert routing based on metadata or learned latent representations.
Training Strategy: Train experts and coordinator jointly using survival-specific objectives (e.g., C-index optimization), apply regularization to prevent expert collapse, and encourage diversity among specialized models.
Validation and Evaluation: Assess prognostic performance, robustness to unseen acquisition domains, and model calibration. Conduct ablation studies comparing single-model, multi-agent, and metadata-agnostic variants, complemented by qualitative analyses of representation invariance.
Reporting and Documentation: Document the full methodology, experimental results, and implications for clinical deployment, producing the final thesis report.
Supervisor(s):
This activity will be supervised by Paolo Soda and Francesco Di Feola. For further information, please contact paolo.soda@umu.se and francesco.feola@umu.se
Proposal 8. Surrogate Text Embedding Learning for Self-Conditioning Cross-Modal Generation
Diffusion models conditioned on natural language have become the dominant paradigm for high-quality controllable generation. However, many data modalities used in scientific and medical domains—such as radiology images, whole-slide histopathology, tabular patient data, or sensor signals—do not possess inherent textual descriptions. This lack of linguistic information prevents these modalities from taking advantage of text-conditioned diffusion models, limiting their generative potential and hindering applications such as data augmentation, counterfactual simulation, and modality synthesis. In this project, we propose to repurpose the CLIP architecture to learn surrogate textual representations from modalities that do not naturally contain text. The method trains an auxiliary encoder that maps a non-textual input into the text-embedding space of a frozen CLIP text encoder. These learned surrogate embeddings mimic linguistic conditioning and can be directly fed into text-to-X diffusion models to guide generation of the same or other modalities. By treating the missing text as a representation-learning problem rather than relying on manually written prompts or ad-hoc heuristics, this strategy enables controllable diffusion-based generation even in domains where no text is available.
Aim of the Project
To develop a CLIP-based framework that produces surrogate textual embeddings from non-textual data and uses them to condition diffusion models. Specifically, the system will:
Train a modality-specific encoder that projects inputs into the CLIP text-embedding space;
Enforce semantic alignment through contrastive, consistency, and CLIP-guided objectives;
Integrate the learned surrogate text representations into a text-conditioned diffusion model to enable controllable generation, reconstruction, and counterfactual synthesis.
Work Description
The project will involve the following tasks:
Literature Review:
Survey CLIP-based multimodal alignment, text-conditioning in diffusion models, and techniques for cross-modal representation learning. Identify gaps in current approaches when text is absent
Dataset Preparation:
Analyze the chosen dataset, define preprocessing and augmentation steps, and establish training/validation splits suitable for contrastive and generative objectives
Model Development:
Implement a modality-specific encoder with a projection head mapping inputs to the CLIP text-embedding space. Explore minimal, efficient architectures tailored to the data modality
Training Strategy:
Train the system using a hybrid objective combining contrastive alignment, view-consistency, optional CLIP-guided regularization.
Diffusion Integration:
Use the learned surrogate text embeddings to condition a diffusion model and evaluate reconstruction fidelity, controllability, and generative consistency
Evaluation and Reporting:
Assess embedding quality, generative accuracy, and robustness. Conduct focused ablations and document the results in the final thesis report.
Supervisor(s): This activity will be supervised by Paolo Soda and Filippo Ruffini. For further information, please contact paolo.soda@umu.se and filippo.ruffini@umu.se .
Federated Learning is Not So Mysterious or Different; it is a new AI Frontier (30 ECTS).
Federated Learning (FL) is a way to train a shared machine-learning model across many devices or organizations without moving their raw data off-device or off-site. Clients compute local updates that a coordinator securely aggregates, enabling learning under privacy, data-governance, and bandwidth constraints. Position FL as often seen as mysterious (non-IID challenges, privacy, communication bottlenecks), yet these are not unique to FL. Many phenomena (convergence slowdown, personalization trade-offs, communication inefficiency) can be explained with classical distributed optimization and statistical learning frameworks. Here, we argue that FL is not mysterious or fundamentally new in its optimization/statistical challenges. Its deployment context (trust, regulation, decentralization) makes it unique.
Motivation: Federated Learning is portrayed as a “new paradigm,” but most of its challenges overlap with long-standing distributed optimization and privacy-aware ML. Hence, FL can be treated, appreciated, and evaluated as a deployment architecture.
Contribution: Show that Federated Learning phenomena can be understood using known theories (convex optimization, game theory, coding theory) and compared to centralized learning approaches.
Please contact: Feras M.Awaysheh feras.awaysheh@umu.se for more information
Open Set Learning and Anomaly Detection in Vision Tasks (15/30 ECTS)
We are announcing projects focused on open set learning, anomaly detection, and out-of-distribution detection. These areas are crucial for developing robust and reliable algorithms, particularly in applications where the ability to handle unexpected or novel inputs is essential. By leveraging deep learning (DL) and machine learning (ML) techniques, this project aims to advance the state-of-the-art in vision tasks, with potential applications in medical imaging and robotics.
For instance, consider a robotics scenario where an autonomous robot navigates through a dynamic environment. Traditional models might struggle with objects that were not present in the training data, such as new furniture or unexpected debris. This project will explore methods to identify and handle such cases, ensuring safety and efficiency when faced with unforeseen obstacles.
We are looking for motivated students who have previously taken courses in DL and computer vision, and as a result, have experience with TensorFlow and/or PyTorch. Interest in embodied AI (when AI has a "body" and as a result interacts with the environment, the way robots do) is a bonus. Ideal candidates should have a strong background in these areas and a keen interest in applying ML to real-world problems.
More info: Polina Kurtser polina.kurtser@umu.se
Path Planning and Object Manipulation with Industrial Robotic Arms (30 ECTS)
We are announcing a new MSc project focused on path planning and object manipulation using industrial robotic arms. This project will leverage reinforcement learning (RL) to learn sequences of tasks, and potentially incorporate large language models (LLMs) and foundation models to break down large tasks into smaller, manageable tasks.
For example, consider a scenario where the goal is to stack cubes. The command to “stack cubes” can be broken down into smaller tasks such as identifying the cubes, planning the path to pick up each cube, and placing them in the correct order. Each of these smaller tasks can be assigned a score based on an RL cost function, which evaluates the efficiency and effectiveness of the task execution.
We are looking for motivated students with the following skills and interests - experience with ROS (Robot Operating System) for both practical and simulation environments, comfortable with Python/C++, previous coursework in robotics and a strong interest in the field are considered a bonus.
Ideal candidates should have a solid background in these areas and a keen interest in applying advanced AI techniques to real-world industrial applications. If you are passionate about robotics and eager to contribute to innovative solutions in industrial automation, we encourage you to apply.
More info: Polina Kurtser polina.kurtser@umu.se
Multimodal Machine Learning, Reasoning, and Compositional Generalization (15/30 ECTS)
Are you interested in developing AI systems that can reason effectively across multiple modalities, such as images and text? While current multimodal models like CLIP, LLaVa, and GPT-4v demonstrate impressive performance on a range of tasks, they also suffer from reliability issues. Compositional generalization, or the ability to combine known concepts in new ways, is one such challenge. For example, a model should understand that the color of an object often does not determine its size. Many existing models struggle with this type of reasoning and instead rely on statistical patterns in the training data rather than a true compositional understanding of the world.
This project aims to develop multimodal models (e.g., by combining language and visual information) that can reason more effectively and generalize better across different tasks. You will explore methods to improve compositional generalization and/or create better frameworks to evaluate these capabilities. Examples of tasks include questions about image content and visual mathematical reasoning, where possible methods to explore include neuro-symbolic approaches that combine neural networks with frameworks for logical reasoning.
Whether you are interested in implementing new methods, evaluating existing models in new ways, or approaching this from an interdisciplinary perspective (such as cognitive science), this project offers a broad area to explore. If any of these topics interest you, please contact Adam Dahlgren Lindström at dali@cs.umu.se for more information.
Visualization and processing of atmosphere data (15/30 ECTS)
Researchers within space physics and geophysics are facing new challenges when three-dimensional data from radar measurements and photogrammetry of atmosphere and ionosphere become available with emerging technologies. Within the research project VisA, methods and tools for visualizing volumetric authentic data are developed to make the data accessible for researchers and the general public. The huge amount of data that is processed (both in time and space) entail requirements on new and effective algorithms as well as new ways to present the results. To visualize space phenomena, like Northern lights, we are using the open source planetarium software OpenSpace and for targeting the researchers we develop custom made tools.
Examples of Bachelor’s and Master’s Thesis projects are:
Within OpenSpace we have a close collaboration with Visualiseringscenter C in Norrköping. Therefore, a Master’s thesis can be conducted as a joint project with them that may include longer or shorter stays in Norrköping.
For more information
If you think this sounds interesting, please contact Stefan Johansson, stefanj@cs.umu.se.
Machine learning model for medical imaging in ovarian cancer
Machine learning has become very important in medical sciences as a means to automate for instance diagnosis and prognosis, and to personalise treatment. Machine learning can be used in medical imaging analysis and has the potential to function as a clinical decision support tool in cancer diagnostics. In collaboration with the Department of Radiation sciences, Diagnostic radiology, and Umea University Hospital, Radiology and Nuclear medicine department, we plan to set up a machine learning model for standardised classification of suspected ovarian cancer in magnetic resonance tomography (MR). We are therefore looking for Master's thesis students who are interested in taking part in the development of such a model based on locally annotated images. The method will subsequently be validated on a prospective image material from a study where the inclusion has started and will be ongoing between 2023-2027.
For more information
If this sounds like something for you, please contact Tommy Löfstedt via e-mail: tommy@cs.umu.se for more information, on this and other potential projects.
HPC TTN contraction for machine learning real-time applications (30 ECTS)
In a machine learning context, the prediction of a Tree Tensor Network (TTN) learner is performed via a contraction of the full network with a (tensorised) sample of the dataset. This contraction can be carried out in many different ways and allows parallelism and different exploits such as:
Having an optimal strategy for fast predictions is of the uttermost importance in every ML scenarios but especially in real-time deployment of ML models. Some of the most prominent applications field range from: object-detection in autonomous-driving, on-line data processing in manufacturing, big data finance analysis, etc.
More info:
https://hpac.cs.umu.se/~pauldj/tmp/MS_thesis-TTNContraction.pdf .
Contact Paolo Bientinesi for more info pauldj@cs.umu.se
HPC TTN Riemann optimisation
In a machine learning context, an interesting paradigm for Tree Tensor Network (TTN) models training is the Riemann optimisation (see https://doi.org/10.1088/1367-2630/ac0b02). With this technique, derived from differential geometry, one can optimise the complete network at once allowing direct implementation of different gradients and stochastic gradients methods. Moreover the numbers of trainable parameters can be dynamically adjusted and thus giving a better training convergence and resilience to overfitting. Technically the Riemann optimisation involves a multiple stacking operations between tensors (thus data manipulation) that are not well supported by many tensor libraries.
For more info
https://hpac.cs.umu.se/~pauldj/tmp/MS_thesis-TTNRiemann.pdf
Contact Paolo Bientinesi for more info pauldj@cs.umu.se
Ericsson Internship/Master Thesis: Automatic segmentation for semantic SLAM dataset
Motivation
The Efficient understanding of both geometric and semantic characteristics of the environment is needed so that robots or XR devices can appropriately perform requested tasks and understand the local physical environment. Simultaneous localization and mapping (SLAM) algorithms allow the XR device to simultaneously geometrically map the environment and localize itself within the environment, while object detection (OD) models can be used to semantically understand what those features in the environment represent or mean.
For a long time, SLAM and OD have been taken as separate processes or tasks and research has been carried out on separate datasets. There are many datasets with object class and location labels and ground truth in the form of bounding boxes or segmentation masks. While there are many other datasets with maps and devices that pose ground truth. But very few datasets, if any include both, and the ones that do are mostly synthetic datasets.
At Ericsson, we are investigating how to perform joint detection and SLAM in a way that such processes benefit from each other in an efficient and integrated way. For that, we need a joint dataset that allows us to evaluate the performance of SLAM and object detection or segmentation algorithms together. However, semantically segmenting SLAM datasets can be very time-consuming and a fast and efficient acceleration via automatic annotation tools is needed.
Internship Assignment
The proposed project is an iterative process that will evolve and take shape during the internship work. It is expected to have many iterations and trial and error tests in the pipeline. The project will be carried out during the period (Sep-Dec) 2022. We are looking for an intern who can help us accelerate the semantic annotation of existing visual-inertial SLAM datasets in an iterative way, reducing manual annotation efforts. The internship can be converted into a master's thesis if necessary.
The scope of the internship includes but is not limited to:
a) Become acquainted to ML and SLAM frameworks
b) Become acquainted to object detection or segmentation frameworks using RGB images.
c) Carry out a literature survey on prior works related to automatic annotation processes
d) Develop python/OpenCV scripts and/or ROS wrapper/libraries for a proof of concept
Qualifications
We are looking for a self-motivated and creative student with the following qualifications:
Supervisor(s):
The internship coordinator and supervisor is Héctor Caltenco, who will be able to help with administrative matters, purchase of material and general internship-related issues, as well as take care of project matters and research-related questions.
Further Information:
If you have any questions about this position, please contact Nithesh Chandher Karthikeyan, Research Engineer at Department of Computing Science, Umeå University.
Thesis work in Software Engineering or Computer Security (15/30 ECTS)
In the Software Engineering and Security group, we aim to improve the quality of software by identifying and removing weaknesses during its whole life cycle – from the design and implementation phase to the deployment phase. We develop tools and approaches to analyze software, test software, understand the software and protect software from bugs and vulnerabilities. Our research has reald-world impact: Covid tests may leak personal data
Our research focuses on topics in software engineering and computer security such as:
Do you plan to do your bachelor/master thesis in software engineering or computer security?
Please, contact Prof. Alexandre Bartel: alexandre.bartel@cs.umu.se or 072 208 73 56.
Are you a fast learner and like to work with Robotics/AI (30 ECTS)
Intelligent Robotics Research Group develops formal and non-formal methods for intelligent robot behavior, such as visual perception or natural language usage. We are interested in algorithm development as well as interaction and design studies. Are you a fast learner, like to work in a team, have excellent communication skills, and would like to work in Robotics/AI?
More information
Please contact Thomas Hällström, Professor at the Department of Computing Science via e-mail for a discussion about a suitable thesis project (30 ECTS).
Machine learning has become very important in medical sciences as a means to automate for instance diagnosis and prognosis, and to personalise medicine and treatments. Machine learning is for instance used in medical imaging to automate parts of the radiation treatment planning after a cancer diagnosis.
We are curious about what exactly is learned about particular patients when building deep convolutional neural networks for medical imaging applications. Since models trained on patient data are commonly and freely shared online, there are potential privacy issues involved that are currently not fully understood. We are therefore looking for Master's thesis students who are interested in projects in this area.
For more information
If this sounds like something for you, please contact Tommy Löfstedt via e-mail: tommy@cs.umu.se for more information, on this and other potential projects.
AI classically focuses on optimizing raw performance criteria (e.g., optimizing the productivity of a factory), often disregarding the impact on human variables (e.g., burned out factory workers), due to being more difficult to quantify. As a response, the field of human-aware AI seeks to develop AI technologies that can adapt their decisions to human factors. This thesis proposal is intended to provide you with an opportunity to join this emerging trend.
Methods: I have a broad technological background and can offer to supervise you along an array of methods, including LLM & NLP, ML, RL, automated planning, cognitive modelling, multi-agent systems, logics, and argumentation. Please reach out to me if there is another method you wish to use.
Specific topics: I have a set of thesis topics, each of which align a practical application, a researcher from another discipline to whom your results will matter, an opportunity streamlining your thesis into a concrete scientific contribution.
The topics evolve from year to year, but they include for example:
Tailoring: the methods and topics can be adapted and recombined. While not mandatory, I make offerings best enable a follow up into a research track (from transversal skills, to paper-writing support to connections with possible recruiters for a PhD). Within this programme, many of my students stepped up their thesis into a paper, including in top-ranked venues, and had opportunities to apply for a summer job and/or PhD by the end of their thesis.
Supervisor: I am an experienced supervisor of 50+ students, with a particular interest in the design of AI systems with high social impact. I have a particular expertise on the integration of interdisciplinary human factors aspects through e.g. my co-direction of TAIGA, Umeå’s center for transdisciplinary AI and my research project on anxiety-sensitive AI.
Contact Loïs Vanhée
AI models of human deliberation (15/30 ECTS)
The foundational mission of AI was to use computational methods in order to help better understand the phenomenon that is intelligence. Whereas this endeavor got somewhat eclipsed by the rise of high-profit applications of AI, the long-term growth of our understanding of the many ways artificial forms of intelligence can be designed is only possible through a focused research in this direction. This line of projects is dedicated to the development of foundational computational models of how humans deliberate, seeking inspiration from psychological and social literature. Applications involve both models seeking to behave like humans and models capable of seizing the psychological or social context of involved humans.
Methods: The primary activities for the human-like agents involve the investigation of psychology literature and the development of innovative decision models, primarily along e.g. agent-based modelling and social simulation, towards replicating deliberative patterns as observed in humans.
Naturally, other AI technologies are to be put in place depending on the context, including LLM & NLP, ML, RL, automated planning, cognitive modelling, multi-agent systems, logics, and argumentation.
Specific topics: I have a set of thesis topics, each of which align a practical application, a researcher from another discipline to whom your results will matter, an opportunity streamlining your thesis into a concrete scientific contribution.
The topics evolve from year to year, but they include for example:
Tailoring: the methods and topics can be adapted and recombined. While not mandatory, I make offerings best enable a follow up into a research track (from transversal skills, to paper-writing support to connections with possible recruiters for a PhD). Within this programme, many of my students stepped up their thesis into a paper, including in top-ranked venues, and had opportunities to apply for a summer job and/or PhD by the end of their thesis.
Supervisor: I am an experienced supervisor of 50+ students, with a particular interest in the design of AI systems with high social impact. I have a particular expertise on the integration of interdisciplinary human factors aspects through e.g. my co-direction of TAIGA, Umeå’s center for transdisciplinary AI, my organization of the Special Interest Group on Modelling Human Decisions, and my research project on anxiety-sensitive AI
Contact Loïs Vanhée
Deep Data Mining Group has been working on multimodal heterogeneous data analytics and data federation by applying the techniques of text mining, information retrieval, natural language processing, machine learning, and differential privacy. The main research topic are data driven and application-oriented such as entity-based social network analysis (e.g., sentiment, emotions, political view etc.), personal privacy analysis, and ontology-based knowledge graph construction etc. Additionally, the group has several connections to industry and academia who are interested to co-supervise master thesis projects, where we can conduct some interesting AI related interdisciplinary research (e.g., Covid pandemic event tracking and inference, medical screening, energy consumption and anomaly prediction, genetics&healthcare etc.). We welcome an open discussion regarding your thesis plan.
For more information
If you are interested, feel free to contact Lili Jiang, associate professor at the Department of Computing Science via e-mail: lili.jiang@cs.umu.se about a suitable master thesis project.
Autonomous Distributed Systems Lab: The group's main focus is new cloud/edge technologies with enhanced performance enabled by AI for the data processing required to support future hyper-distributed applications. Based on such the cloud/edge, we research data markets and data economy; therefore, we are currently looking for Master's students who want to do their thesis project (30hp) in our group. More specifically, we have several openings in:
a) Data markets allowing for data assets to be discoverable, efficiently and fairly priced, and shared/traded in a compliant way.
b) Scalable and AI-powered data management for data usability in different contexts, covering data provenance, data quality management, improving data interoperability, and metadata management.
For more information
Please contact Antonio Seo, eunil.seo@cs.umu.se, to discuss/define a thesis project.
When training deep neural networks in machine learning, the standard optimisation method is some variant of stochastic gradient descent. An hypothesis about why deep neural networks can generalise as well as they do, despite them typically having many more parameters than training data, is the utilisation of stochastic gradient descent. The hypothesis says that when using stochastic gradient descent, wide/flat minima are preferred over sharp/narrow ones, that stochastic gradient descent has the ability to "shake itself out" of sharp minima.
The goal of this project is to investigate different ways that stochastic gradient descent can work as a form of regulariser for the model, how it can avoid overfitting the training data. There are several different ways this can be tested, and the project would amount to select one or a few approaches, implement, and test them by comparing theoretical and practical results.
For more information
If this sounds like something for you, please contact Tommy Löfstedt, associate professor at the Department of Computing Science, through e-mail: tommy@cs.umu.se for more information on this, and other potential projects.
High-performance kernels for tensor computations (30 ECTS)
In a recent survey of software for tensor computations, we identified an enormous case of replication of efforts. Indeed, despite the importance and widespread use of tensors, a library of building blocks (i.e., the tensor counterpart of BLAS) is still missing. With this project we make a concrete proposal for a set of high-performance kernels to be used across disciplines in which computations involve tensors. Such disciplines range from machine learning and data science all the way to computational chemistry and computational mechanics.
The survey mentioned above: "The landscape of software for tensor computations".
For more information
Contact Paolo Bientinesi, pauldj@cs.umu.se, 0720 856 180 for more information.
MatchC: A C/C++ library for Pattern Matching
Mathematica (Wolfram) arguably offers the most powerful pattern matching capability of any programming language. A C/C++ counterpart, even if more limited in features and functionality, would be an extremely useful building block for a myriad of applications. This project aims at developing a high-performance pattern matching library, aiming for associative and/or commutative operators, with constraints, as well as single and multiple matches (akin to ./+/* in reg exps), and shared-memory parallelism.
In previous work, we developed MatchPy, a Python library for pattern matching. This constitutes the natural reference and starting point for this project. See also "Efficient Pattern Matching in Python".
For more information
Contact Paolo Bientinesi, pauldj@cs.umu.se, 0720 856 180 for more information.
Cybersecurity + Distributed/Tiny/Machine Learning + Computer Systems (15 ECTS/30 ECTS)
Autonomous Distributed Systems Lab - This wing of the Lab focuses on how to build adaptive, resilient, and responsible systems that operate in complex and real-world environments by investigating emerging and cutting-edge problems that remain to be addressed across domains, such as cloud-edge continuum, serverless computing, Internet of Things (IoT), and societal applications. Our research strives to find solutions that are low-cost, high-performance, reliable and unconditionally secure.
The primary goals of our research are intersections of cyber analytics and machine learning, but the areas are:
For more information
If you are ready or interested in writing your thesis (15/30 ECTS) in any of the above areas, then please don't hesitate to write a single email to Monowar Bhuyan, Assistant Professor, at e-mail: monowar@cs.umu.se.
Passion for music and programming? Do your thesis project together with Toontrack Music
Toontrack Music is a leading developer of software for music creation. Our customers range from home users to renowned producers, songwriters, and artists who use our sounds and tools daily to create music. The foundation of our business idea is that making music with the help of a computer should be easy. This has permeated our operations since the company was founded in 1999 with a vision to create the world’s best virtual drummer. Since then, we have transformed the market for computer-assisted music production.
Today, we are the largest in the industry within the drum segment and are rapidly advancing in several other areas. We employ over 45 people and have offices in Umeå and Chicago, with the entire world as our market. Our most well-known brands are EZdrummer, Superior Drummer, EZbass, EZmix, and EZkeys.
We are now looking for students who want to do their thesis project in programming with us. The exact focus of the thesis will be agreed upon together, but it will be related to computer-assisted music technology. The most important thing for us is that, in addition to being skilled in programming (C++), you are passionate about music and play one or more instruments yourself or are proficient in studio technology. Ideally, you are already familiar with using Digital Audio Workstations and plugins.
Toontrack Music is located in Umeå. Read more at: https://www.toontrack.com/
Email your CV with a short description of yourself to: Erik Phersson at erik@toontrack.com
Sandvik Coromant thesis project – Enhancing Machining Time Estimation through AI and Digital Prototyping
Are you passionate about engineering, AI, and digital innovation? Join Sandvik Coromant and help revolutionize how machining times are calculated – a key factor in efficient manufacturing. This is your chance to work with cutting-edge technologies and contribute to real-world industrial solutions! Sandvik Coromant is a global leader in tools and solutions for the metalworking industry. We drive innovation through advanced R&D and collaborate with top players in automotive, aerospace, and energy sectors. Our digital machining services, like CoroPlus® Tool Guide, Tool Library, and Tool Path, help customers optimize performance using the latest technologies.
Background and scope of the project
Accurate machining time estimation is essential for planning, cost control, and production efficiency. Traditional time study methods are often manual, slow, and not detailed enough. Existing solutions provide partial support but frequently lack flexibility and standardization. To address these issues, this thesis investigates the requirements for more accurate and efficient time study calculations. This includes gathering end-user input, analyzing current practices and existing solutions, and identifying their shortcomings. A prototype based on the proposed solution will be developed to validate the findings. The scope of the project is limited to defining and testing algorithms for selected operation types, developing a prototype to demonstrate their feasibility, and summarizing findings, challenges, and recommended solution directions. The results will be presented in both written and oral form.
Purpose
The purpose is to investigate and develop analytical and AI-supported methods for improving the accuracy and efficiency of machining time estimation in manufacturing. The thesis aims to define the necessary data requirements, design appropriate algorithms, and create a prototype to validate the proposed approach.
Read more by following this link
For more information about the thesis, please contact:
Thesis at Quicke, a part of JOST World
Are you interested in using innovative technology to contribute to the future of agriculture and transport? Then a thesis at Quicke, part of JOST World, will be something for you.
Quicke products are leveraging sustainability, productivity and safety, helping to secure local production of affordable food - globally. Let’s push the limits even further!
Software & Electronics
• Artificial Intelligence - Predictive machine maintenance based on Edge Machine Learning
• Vision-based object classification (YOLO model) training based on synthetic data modelling wear and tear
• Localization of objects based on monocular vision – method selection and evaluation in reality
• Deterministic testing framework for virtual CAN simulation for a distributed system with multiple nodes
• Design electronics (incl embedded software) for a CAN based keypad incl RGB LED indicator for each button
• Design electronics (incl embedded software) for a CAN based I/O module with high ingression protection
Hydraulics
• Optimization of passive loader suspension by simulation – model design, tuning and validation
• Active load suspension – conceptual design, prototyping and evaluation (hydraulics + controls)
• Modelling of tractor front loader hydraulics in and model evaluation in co-simulation
Mechanical Design
• Concept generation and design of electronics enclosure for Harsh environments using Injection Molding
Physics, Simulation and Controls
• “Extended kinematics-compensation” of closed-loop front-end loader (2-link arm) control
• Self-learning continuous position control of front-end loader, similar to “crane tip control”
• Design and prototyping of “adaptive-gain human machine interface” using model-based design
• Visualization of a tractor front loader HIL (hardware in the loop) simulation using a 3D game engine with integration of real HMI (human machine interface)
For more information please contact:
For Mechanical Design: Aron Lidgren aron.lidgren@jost-world.com
For all other topics Jakob Andrén: jakob.andren@jost-world.com
Join Knightech Group for your thesis work! Gain hands-on experience, work on real projects, and develop your skills in a supportive and innovative environment! 3 topics
At Knightec Group, you will be part of an entrepreneurial culture driven by innovation and collaboration. We aim to be the ultimate meeting place for professionals who want to work at the forefront of product and digital service development.
See the full version and application details via the links below:
AI-based and simulation-driven path planning for autonomous surface vessels
Unmanned surface vessels must navigate safely and predictably in inshore/coastal waters where encounters, crossing traffic, and narrow channels are common. To avoid collisions, the International Regulations for Preventing Collisions at Sea (COLREGs)—the maritime “rules of the road”—are applied; however, translating these rules into robust decision strategies under varying traffic and sensor conditions is challenging.
Machine Learning for Segmentation and Volume Analysis of 3D Point Clouds
Laser scanning of industrial environments produces very large point clouds, often with tens of millions of points. In a grinding mill, the raw data captures not only the ore mass but also walls, equipment, personnel, and parts of the surroundings, making the input noisy and complex.
Real-time detection of wear debris in hydraulic motors
Hydraulic motors generate metallic wear debris that must be captured to ensure reliable operation and long service life. Today, daily checks of the magnetic plug are done by eye, which is subjective and hard to compare over time. The goal is to replace this with a measurable, repeatable approach that estimates debris amount, particle size, and the rate at which debris forms—without stopping operation.
Clear Street – Modernizing the Brokerage Industry. 9 topics, deadline 10 oct
Clear Street is an innovative financial technology company on a mission to modernize the entire brokerage infrastructure. By building a new cloud-native platform from the ground up, Clear Street delivers a more efficient, secure, and user-friendly solution for capital markets. The company offers services in clearing, trading, and prime brokerage for institutional and professional clients. Clear Street is committed to creating a more accessible and transparent market through technology-driven innovation and customer-focused service.
Clearstreet offers degree project in the following areas, presented in a short version below and full version at this link. Also open for own ideas and/or suggestions for similar topics.
How do I apply?
1. What? CV and grades
2. Where? student_swe@clearstreet.io
3. When? Deadline is October 10th
Measuring the Impact of Microfrontend Architecture in Financial Web Portals
Study how microfrontends affect performance, scalability, and reliability in data-heavy financial apps. Focus on responsiveness, deployment independence, and user experience across modular interfaces.
A Comparative Analysis of Real-Time Communication Protocols for Modern Web Applications
Compare WebSockets, SSE, and WebTransport in terms of latency, throughput, and integration complexity. Deliver a data-driven guide for selecting real-time protocols in financial platforms.
Calibration of Monte-Carlo Models for Pricing Convertible Bonds
Develop a multi-factor Hull–White model for CB pricing. Calibrate using market data (swaps, CDS) and assess how each stochastic factor influences valuation.
Solving for Volatility in Low-Liquidity Markets
Create volatility surfaces for illiquid assets using historical data analysis and proxy modeling. Define criteria for effective proxies and apply statistical techniques to improve pricing accuracy.
Predictive Failure Propagation in Cloud-Native Distributed Systems
Use dependency graphs and ML to forecast error propagation in microservices. Explore automated remediation strategies to reduce downtime and improve fault analysis.
Just In Time UI for Trading
Design context-aware, minimal trading interfaces generated dynamically by AI. Focus on usability and responsiveness in time-critical trading scenarios.
ORM Frameworks: A Comparative Study of Hibernate and Native SQL
Evaluate Hibernate vs. native SQL in terms of performance, abstraction, and workload adaptability. Identify when ORM is beneficial and when direct SQL is preferable.
Resilient Stream Processing: Handling Kafka Failover, Offset Management, and Eventual Consistency in Modern Distributed Systems
Investigate methods to ensure failover safety, offset accuracy, and consistency in Kafka-based streaming systems. Address challenges in rebalancing and late event handling.
Log Filtering and Cost Optimization for Large-Scale Observability Pipelines
Design a host-level agent for log filtering, deduplication, and routing. Aim to reduce observability costs while preserving diagnostic value and signal quality.
Tietoevry, Validating the Fairness of Large Language Models in Recruitment
Background and Motivation
Large Language Models (LLMs) are increasingly used in recruitment systems for résumé screening, candidate ranking, and interview support. While these models can improve efficiency, they also raise concerns about algorithmic bias and fairness.
Unintended biases in training data, model design, or prompt instructions can result in discriminatory outcomes against certain demographic groups. However, “fairness” in recruitment is not purely statistical—it also reflects human values and organizational norms.
To develop trustworthy AI tools, fairness must be defined not only mathematically but also contextually, based on how HR professionals interpret fair decision-making in recruitment. This thesis aims to combine these perspectives to validate and improve fairness in LLM-assisted hiring processes.
Research Questions
1. Conceptual: How do HR professionals define and interpret the concept of fairness in recruitment and selection, and how is this knowledge transferred to AI tools?
2. Analytical: To what extent do large language models exhibit bias when evaluating or ranking job candidates?
3. Methodological: Which fairness metrics best capture relevant biases in LLM-based recruitment tasks?
4. Practical: Which mitigation strategies (e.g., prompt engineering, data balancing, or post-processing) are most effective in improving fairness without reducing performance?
Objectives
• Explore how HR professionals conceptualize fairness in recruitment decisions.
• Develop a framework for validating fairness in LLMs used in candidate selection.
• Measure bias using counterfactual testing (e.g., name or gender swaps).
• Evaluate mitigation strategies and propose practical design recommendations for fair AI use in recruitment.
Methodology
1. Qualitative phase: Conduct 4–6 semi-structured interviews with HR professionals to understand how they define fairness and identify subtle forms of bias in recruitment. Thematic analysis will be used to extract core fairness dimensions.
2. Quantitative phase: Use synthetic or anonymized résumé datasets to test LLM behavior. Generate counterfactual examples that vary protected attributes while keeping qualifications constant.
3. Evaluation: Measure disparities across groups using metrics such as demographic parity and equalized odds with toolkits like Fairlearn or AIF360.
4. Mitigation: Experiment with prompt adjustments and post-processing to improve fairness, guided by the qualitative findings.
5. Ethical Considerations: Follow GDPR and EU AI Act principles. Only anonymized or synthetic data will be used.
Expected Outcomes
• A conceptual model of fairness in recruitment grounded in HR expertise.
• An empirical assessment of bias in LLM-based candidate evaluation.
• A reproducible audit framework for validating fairness in recruitment AI.
• Recommendations for developers and HR practitioners to ensure ethically aligned AI deployment.
Timeline (6 months)
Monthly milestones
1 Literature review, ethics approval, interview planning
2 Conduct HR interviews, thematic analysis
3 Dataset preparation and counterfactual design
4 LLM experiments and fairness evaluation
5 Mitigation experiments and integration of qualitative findings
6 Analysis, writing, and final report
Keywords
Fairness • Bias • Large Language Models • Recruitment • HR Ethics • EU AI Act
Tietoevry, From Spec to Patch: LLM Task Generation + Copilot Execution
Evaluate whether LLMs can automatically decompose a feature spec into executable tasks and whether a code copilot can carry out those tasks to implement the feature (given the files that will be affected).
Short summary: This project builds and evaluates an end-to-end pipeline: a planner LLM generates a task decomposition (issues, TODOs, tests, acceptance criteria) from a feature document plus the relevant files; an executor LLM (code copilot) attempts to implement each task by producing code patches/PRs. The thesis compares LLM-generated plans to human plans, measures how many planner tasks the copilot can implement correctly (test-passing, reviewed), and analyzes failure modes, task granularity, and grounding strategies. The work emphasizes reproducible experiments across multiple repos and an empirical analysis of success factors.
Primary research question:
· How well can LLMs generate actionable task decompositions from a feature document + affected files compared to human engineers?
Secondary research questions:
· What fraction of LLM-generated tasks can a code copilot implement into correct, test-passing commits without human intervention?
· Which task granularity (coarse vs. fine) maximizes automated execution success?
· How much does providing affected files (vs. feature doc only) improve task relevance and execution?
· Which feature types (UI, API, tests, refactors) are most amenable to end-to-end automation?
· What are common failure modes (hallucination, missing edge cases) and practical mitigations?
Contact:
Tobias Sundqvist
Tobias.sundqvist@tietoevry.com
Tietoevry, Agentic RAG-Enhanced IDE Extension for Feature Implementation
Build an IDE extension that uses Retrieval-Augmented Generation (RAG) to turn feature specs into an adaptive checklist, monitor a developer’s branch in real time, and use LLM agents to guide them to complete the feature with fewer mistakes and faster.
Short summary: Many developers lose time chasing missing requirements, unclear specs, and inconsistent workflows when implementing new features. This project builds a branch-aware IDE extension that (1) ingests feature documents and project artifacts as a knowledge base, (2) generates an adaptive checklist and actionable guidance using a RAG agent, and (3) monitors commits/diffs to provide targeted, timely reminders and remediation steps. The thesis examines both the technical design (KB construction, retrieval strategy, branch monitoring) and the human impact (productivity, quality, and trust). The work includes implementation, a controlled evaluation, and analysis of failure modes and acceptance.
Primary research question:
· How does a RAG-based IDE extension affect developer productivity when implementing a specified feature?
Secondary research questions:
· Does an adaptive checklist reduce missed requirements and defects compared to static docs?
· How accurate and actionable are RAG-generated recommendations for feature work?
· How does branch-aware monitoring influence developers’ trust and adoption of the assistant?
· What KB structure (tests, specs, examples) maximizes recommendation quality?
· What are privacy/security trade-offs of local vs. cloud RAG for private codebases?
Contact:
Tobias Sundqvist
Tobias.sundqvist@tietoevry.com
AI-Driven Forecasting of Fuel Consumption in Mechanized Harvesting Based on the Site Conditions (30 ECTS)
Background and Purpose
The forestry sector remains heavily reliant on fossil fuels for machinery operation, posing a significant challenge to achieving climate-neutral wood-supply. One of the main obstacles to transitioning toward renewable or low-emission fuel alternatives is the lack of a decentralized and responsive energy supply infrastructure for fuel alternatives. Effective planning for such systems requires accurate knowledge of site-specific fuel consumption, which in turn depends on a range of external and operational factors—including terrain slope, soil moisture, surface roughness, weather conditions, and seasonal variability.
This thesis project aims to harness the power of deep learning and geospatial data integration to predict energy usage and emissions in forestry operations. By developing a model that can generalize across harvesting sites with varying conditions, this work will contribute to more informed planning and benchmarking of sustainable forest management practices, particularly in regions like Northern Sweden where challenging terrain is common.
The degree project is to be done in English. We recommend this project as a 30 credit Master’s thesis, given the complexity and technical depth of the task.
Your Task
You will develop and train a deep learning model that predicts energy consumption (and potentially emissions) based on a set of geospatial and operational parameters collected from real-world forest harvesting sites.
Key aspects of your task include:
Data preprocessing and feature engineering from existing datasets (e.g., slope, roughness, stoniness, soil water level, machine type, and weather).
Model training and validation using deep learning frameworks (e.g., PyTorch, with potential integration of TorchGeo for handling geospatial datasets).
Evaluation of model performance on unseen harvesting plots and analysis of generalizability.
Contribute to an internal benchmarking tool to assess machine performance under diverse field conditions.
Your Skillset and Interests
You should have:
A strong interest in AI, deep learning, or geospatial data analysis.
Programming experience, preferably in Python with knowledge of deep learning libraries such as PyTorch or TensorFlow.
Familiarity with or willingness to learn about forestry operations and terrain modeling.
Curiosity to work at the interface of machine learning, sustainability, and field-level forestry data.
The ability to work independently and handle real-world, possibly incomplete datasets.
What We Offer
A relevant research topic that supports the energy transition of the Swedish forestry sector.
Possibility of contributing to the scientific research of the department as the results of the degree project will be incorporated into a scientific publication, where the student can of course become a co-author
For more information about the project, contact Justin Herdegen.
E-mail: justin.herdegen@slu.se Telefon: +46722392528
Simulation model for Hydraulic system of the press (30 ECTS)
Ursviken Technology was founded in 1885, which gives the company more than 135 years of experience in machine manufacturing and genuine know-how to create the most advanced solutions for its customers. The company today has customers in over 100 countries with a strong interest in always putting the customer in focus. Ursviken manufactures high-tech press brakes with up to 10,000 tons of press force and is continuously working to create new innovations and solutions. In 2021, they proudly delivered the world's longest press brake to the USA, a 100% customized machine with countless innovations. Ursviken Technology's turnover is approximately 200 million SEK with an export share of 75%. The company has approximately 70 employees and is in Ursviken outside Skellefteå. Today Ursviken Technology is a part of Finnish Vaski Group Oy, which acquired Ursviken Technology in January 2025. Vaski Group Oy is a fast-growing company, which manufactures other types of machines with the brands Pivatic and Vaski. Vaski Group turnover will be about 50M€ in 2025 and it has more than 230 employees.
About the job
In your work as a Master thesis worker at Ursviken Technology, you will primarily create the simulation model for Hydraulic system of the press. The press is used for the bending process. The closed-loop position and force control including resistance of the formed material will be a part of the simulation. Based on this simulation model, it is possible to study various system variables, such as position, force, pressure, temperature and volume flow as a function of time. The simulation model is used for machine design, controllability and software verification purposes.
Your profile
We are looking for someone who has studies in automation, production automation or mechanics. The student program background can be mechanical, electrical, mechatronics or production automation engineering. It is also necessary to have a basic knowledge of Matlab with Simulink software and closed-loop control systems. We expect that you have both spoken and written Swedish and English skills.
We offer
Ursviken is an exciting and developing company that designs and manufactures its own products. The design department consists of 16 designers who all play an important and central role in the organization. We are a medium-sized company with short decision-making paths that safeguard the development and success of our employees.
About Skellefteå
Skellefteå is a growing city with a long tradition of entrepreneurship and industrial development. Skellefteå has a broad range of career opportunities in engineering and innovation sector, a great range of options for life outside of work, openness to change, creative ideas and new people. This combination has made Skellefteå an increasingly strong alternative for people and families in all phases of life. Read more about the location at www.skelleftea.se
Application
If you have any questions, please contact Oskar Vestermark at oskar.vestermark@ursviken.com
Once you have shown your interest, we will match your application with open position and we will contact you via email or phone for feedback.
We look forward to receiving your application!
Digital Twin platform for software verification with Visual Components (30 ECTS)
Ursviken Technology was founded in 1885, which gives the company more than 135 years of experience in machine manufacturing and genuine know-how to create the most advanced solutions for its customers. The company today has customers in over 100 countries with a strong interest in always putting the customer in focus. Ursviken manufactures high-tech press brakes with up to 10,000 tons of press force and is continuously working to create new innovations and solutions. In 2021, they proudly delivered the world's longest press brake to the USA, a 100% customized machine with countless innovations. Ursviken Technology's turnover is approximately 200 million SEK with an export share of 75%. The company has approximately 70 employees and is in Ursviken outside Skellefteå. Today Ursviken Technology is a part of Finnish Vaski Group Oy, which acquired Ursviken Technology in January 2025. Vaski Group Oy is a fast-growing company, which manufactures other types of machines with the brands Pivatic and Vaski. Vaski Group turnover will be about 50M€ in 2025 and it has more than 230 employees.
About the job
In your work as a Master thesis worker at Ursviken Technology, you will primarily create Digital Twin platform for software verification with Visual Components software. This simulation model should also include the connection for the PLC via OPC-UA interface, which allows the model to operate via PLC. The simulation model is used for software verifications, bottleneck analysis and the optimization of the production systems.
Your profile
We are looking for someone who has studies in automation, production automation or mechanics. The student program background can be mechanical, mechatronics or production automation engineering. It is also necessary to have a basic knowledge of Visual Components software or similar, and 3D CAD software (CREO or SolidWorks) would be beneficial. We expect that you have both spoken and written Swedish and English skills.
We offer
Ursviken is an exciting and developing company that designs and manufactures its own products. The design department consists of around 20 designers who all play an important and central role in the organization. We are a medium-sized company with short decision-making paths that safeguard the development and success of our employees.
About Skellefteå
Skellefteå is a growing city with a long tradition of entrepreneurship and industrial development. Skellefteå has a broad range of career opportunities in engineering and innovation sector, a great range of options for life outside of work, openness to change, creative ideas and new people. This combination has made Skellefteå an increasingly strong alternative for people and families in all phases of life. Read more about the location at www.skelleftea.se
Application
If you have any questions, please contact Oskar Vestermark at oskar.vestermark@ursviken.com
Once you have shown your interest, we will match your application with open position and we will contact you via email or phone for feedback.
We look forward to receiving your application!
Angular measurement system. Press brake application (30 ECTS)
Ursviken Technology was founded in 1885, which gives the company more than 135 years of experience in machine manufacturing and genuine know-how to create the most advanced solutions for its customers. The company today has customers in over 100 countries with a strong interest in always putting the customer in focus. Ursviken manufactures high-tech press brakes with up to 10,000 tons of press force and is continuously working to create new innovations and solutions. In 2021, they proudly delivered the world's longest press brake to the USA, a 100% customized machine with countless innovations. Ursviken Technology's turnover is approximately 200 million SEK with an export share of 75%. The company has approximately 70 employees and is in Ursviken outside Skellefteå. Today Ursviken Technology is a part of Finnish Vaski Group Oy, which acquired Ursviken Technology in January 2025. Vaski Group Oy is a fast-growing company, which manufactures other types of machines with the brands Pivatic and Vaski. Vaski Group turnover will be about 50M€ in 2025 and it has more than 230 employees.
About the job
In your work as a Master thesis worker at Ursviken Technology, you will work with an angular measurement system to create feedback values for angle compensation all over the bending length. The measurement system is implemented for the bend part during the bending process and measured values are used for the compensation to reach high quality bends for different materials, thicknesses and bending lengths.
https://www.youtube.com/watch?v=AyEn1-xcB1A
Your profile
We are looking for someone who has studies in automation, software or sensor technology. The student program background can be electrical, mechatronics or software engineering. It is also necessary to have basic knowledge of laser measurement technology or machine vision systems. All Experience regarding signal or data processing, simulation and programming would be beneficial. We expect that you have both spoken and written Swedish and English skills.
We offer
Ursviken is an exciting and developing company that designs and manufactures its own products. The design department consists of 16 designers who all play an important and central role in the organization. We are a medium-sized company with short decision-making paths that safeguard the development and success of our employees.
About Skellefteå
Skellefteå is a growing city with a long tradition of entrepreneurship and industrial development. Skellefteå has a broad range of career opportunities in engineering and innovation sector, a great range of options for life outside of work, openness to change, creative ideas and new people. This combination has made Skellefteå an increasingly strong alternative for people and families in all phases of life. Read more about the location at www.skelleftea.se
Application
If you have any questions about the Thesis work, please contact Oskar Vestermark at oskar.vestermark@ursviken.com
Once you have shown your interest, we will match your application with open position, and we will contact you via email or phone for feedback.
We look forward to receiving your application!
Anyfin looking for students interested in Fintech
Anyfin is a Fintech scale-up with BIG ambitions. What are those, you ask? Our answer: We are passionate about changing the financial industry. High interest rates, unfair terms and hidden fees have to become a thing of the past. As soon as possible.
Do you have an exciting thesis idea and want to make an impact? At Anyfin, we’re always on the lookout for curious and ambitious students who want to explore innovative topics that align with our mission. If you have a strong academic background and an idea you’re passionate about, we’d love to hear from you!
We are looking for students starting their Master’s Thesis in August 2025. We’re interested in students pursuing:
- MSc in Industrial Engineering and Management, Physics,
Mathematics, or similar fields.
- Master's programs specializing in Applied Mathematics, Machine
Learning, Mathematical Statistics, Financial Mathematics, or
related areas.
We’re always excited to hear fresh perspectives and original thesis proposals. To give you a sense of the kind of topics that could be a great fit at Anyfin, here are a few examples we find particularly interesting:
Modeling Customer Risk Through Transactional Behaviour.
Build statistical or machine learning models to uncover patterns in bank transaction data that indicate financial risk, enabling smarter credit assessments.
Dynamic Pricing.
Build a pricing model to optimize customer lifetime value.
Advanced, Explainable Credit Risk Modeling.
Use advanced machine learning to predict customer default risk, combining strong predictive power with explainability for transparent underwriting.
We review applications on an ongoing basis and aim to provide feedback as soon as possible. Make sure to check out our website to apply.
Further development of an Android application that diagnoses hearing deficits (15/30 ECTS)
Do you want to contribute to an exciting project that will help people test their hearing at home without any need to visit the hearing clinic?
We have developed a clinically-validated Android application that has the potential to replace the standard clinical diagnostics. Your main role would be to assure that the app’s user understands how the test is done and also come up with innovative ways to report the results to both the user and the medical personnel.
Who are you? We are looking for a motivated student who has good knowledge of object-oriented programming (specially in JAVA language) with some knowledge/interest in interaction design (ID). It is meritorious if you have prior experience with Android application prototyping.
What will you gain from this? You will be developing a mobile application that is potentially going to be introduced by Region Västerbotten to decrease the waiting time for people who need to visit an audiologist for assessing their hearing.
The project is done in collaboration with Uminova and other companies.
If interested, please contact amin.saremi@umu.se
Development of Video Analysis Model for Mill Control in the Mining Industry ABB+Boliden (30 ECTS)
Are you interested in working with advanced technology and contributing to efficiency in the mining industry? ABB is now looking for a driven student for a unique thesis work that aims to develop a video analysis model to facilitate the assessment of overfilling in a mill. The work will include a close collaboration with our partner Boliden, where you as a student will have the opportunity to broaden your network in the mining business and the chance to test your work in a real process environment.
The work involves evaluating the method and training a video analysis model with the goal of optimizing the feed to a mill to maximize production without risking overfilling. Currently, there is no way to control overfilling, leading to conservative operations and unused capacity.
From January to June, Location: ABB Umeå Nygatan 25
Contact: Jenni Sandgren, application engineer, jenni.sandgren@se.abb.com
Elastisys is looking for individuals interested in cloud technology
Elastisys is a pioneer in cloud services, helping companies develop critical software faster—without compromising on security or regulatory compliance in applications or data storage. We develop and provide products and services for robust, high-performance, and cost-efficient cloud applications within DevOps and Kubernetes, and we actively contribute to the open source community.
The company was founded in 2011 by researchers from the Department of Computing Science and the UMIT Research Lab at Umeå University. We are an international company with offices in Umeå and Lund. Our team includes both junior engineers—many of whom have written their theses with us—and senior engineers, several of whom hold PhDs in cloud technology.
We are looking for individuals who are interested in cloud technology and curious to explore this field together with us. At Elastisys, you’ll have the opportunity to grow and work with passionate people who care about you.
Apply by sending your CV and a short personal letter to Student@elastisys.com, including:
We have several relevant thesis topics available, but we also welcome your own ideas for your master’s thesis—these can be developed in dialogue with us.
If you have any questions or thoughts, feel free to contact us at Student@elastisys.com.
Unveiling Duplication Patterns in Trouble Reports - Leveraging Machine Learning for Efficient Issue Tracking
In large systems such as the Radio Access Network, a continuous stream of tickets is generated whenever an issue occurs.
Many of these tickets may stem from the same root cause, resulting in numerous duplicate tickets that require analysis.
The objective of this thesis is to develop a machine learning (ML) algorithm that can learn from past tickets and determine whether a new ticket is likely a duplicate of a previous one.
Throughout this thesis, you will collaborate closely with other AI/ML experts to devise a solution aimed at enhancing the current ticket service tool
For more informtion contact Tobias Sundqvist tobias.sundqvist@tietoevry.com
+46 70 323 64 15
Join our project to help omics research (30 ECTS)
Join our project to help omics research. We are building a comprehensive genome knowledge database to unlock the potential of genomic, proteomic and metabolomic data and enable discoveries to optimize our cell line development platform.
Our project objectives include integrating diverse genomic data sources, internal and publicly available; building a database; creating a robust update management. Students will gain practical database management experience, expand their knowledge, and benefit from industry networking opportunities.
The ideal start date is September 2023 but the project timeline is flexible to accommodate students' academic commitments. The timeline compromises a 10-12 week course work followed by a thesis course.
Contact Shanti Pijeaud for more information shanti.pijeaud@sartorius.com
Benefits and limitations of moving to Microsoft Orleans from a monolithic system grid computing system (30 ECTS)
There are many challenges with developing complex, low-latency distributed systems: scalability, concurrency, reliability, performance and fault tolerance. The virtual actor model claims to help with all these challenges.
The goal of this project is to investigate the benefits and limitations of moving from a monolithic system with features of grid computing to a system using Microsoft Orleans.
A proof of concept will be implemented based on the real-life challenges of a high-throughput, complex distributed system with millions of updates every minute, and thousands of clients worldwide.
If this sounds interesting, please contact Simon Renström, founder of RebelBetting, at simon@rebelbetting.com.
Image processing and deep learning on tomography data of phosphorus-rich ash
The transition to a sustainable society requires efficient use of materials and energy, where recycling nutrients is an important part. Phosphorus is an important nutrient that today is only recycled in small amounts. This could be improved by resource recovery from residual streams such as ash or biochar formed during the thermal conversion of societal waste streams. The overall aim of our research is to improve knowledge of the ash's chemical and physical properties such as microstructure, porosity, and surface area and to study how these can affect the availability of nutrients that can be used as fertiliser.
The project therefore aims to develop tools for analysing tomography data with so-called deep learning. Computed tomography provides a 3D image of a sample and with the help of advanced image processing, detailed characteristics of the sample can be obtained. For this, deep machine learning can play an important role in streamlining and improving performance in data evaluation. The goal of the project is to develop interactive segmentation methods, and evaluate their performance, by comparing to manual segmentation and applicability to different samples.
For more information
If this sounds interesting, contact Anna Strangberg, from the Department of Applied Physics and Electronics, through e-mail: anna.strandberg@umu.se for more information.
Use of machine learning methods to identify patterns of seedling response to drought treatment with help of time-lapse images
The goal of the proposed project is to apply ML tools to learn predictive patterns of response to drought in Scots pine seedlings exposed to drought conditions with use of time-lapse images. Machine learning (ML) methods such as deep learning, elastic net and lasso have been tested to train models for organisms and object recognition. The focus within this project is to train ML models to identify patterns of Scots pine seedling change in responses to drought. Model training will be done on images obtained by time-lapse imaging during the process of drought treatment.
For more information
If this sounds interesting, contact Rosario Garcia Gil, from the Department of Plant Physiology, through e-mail: m.rosario.garcia@slu.se for more information.
Training machine learning methods to interpret genetic effects on wood properties
Development of cost-efficient DNA sequencing technology offers the possibility to generate genome data for thousands of samples, which opens the possibility to apply machine learning methods to predict relationships between an organism’s genotype and phenotype. A recent study on yeast, rice, and weat showed that almost all standard machine learning methods performed better than methods from classical statistical genomics. The goal of the proposed project is to develop an efficient machine learning method to learn predictive relationships between the exome’s genome data (genetic DNA information) and spruce growth and wood properties.
For more information
If this sounds interesting, contact Rosario Garcia Gil, from the Department of Plant Physiology, through e-mail: m.rosario.garcia@slu.se for more information.
Do you want to work interdisciplinary with medical technology at Region Västerbotten? (15or 30 ECTS)
Our mission is to conduct research and development in close collaboration with medical researchers. Our expertise in biomedical engineering provides hospital personnel with technical support in different research and development projects. Thesis has always been an important part of the activities of the dept. This uses the student’s theoretical knowledge in practice in a hospital setting and at the same time the student gets to know a future employer. The student helps us with tasks and to find new technical solutions to the challenges of our research.
Interested?
Please contact Helena Grip, Hospital engineer, assoc. prof. in biomedical engineering and adjunct university lecturer. Research and development within medical technology and movement analysis, email: helena.grip@regionvasterbotten.se or telephone: 090-785 40 29
Evity is a startup in Umeå that develops an HR tool for small and medium-sized companies. We use automation and form tools to create KPIs and processes in the tool. Now, we are looking for a student who wants to work in .net within Azure where we can find a collaboration to develop new features or improve existing ones. Size of project 15 or 30 ECTS.
We are an international team as our co-founder and CTO is from New Zealand, where he also works from. You can read more about us on www.evity.hr
Interested?
Please contact CEO, Tommy Eriksson, telephone: 070-696 70 22 or email: tommy@evity.hr
NONPI Medical develops and implements software tools for medical image analysis. The proposed project focuses on deep learning methods for medical imaging, e.g. tumor segmentation, super-resolution, or distortion correction.
There are many different frameworks available for implementing deep learning models, which is good for development and research, but also divides the field and published code into fractions, depending on the particular framework that was used. The result is that the field is divided, and the results are difficult to reproduce between framework users. An open source standard for machine learning models has been established (ONNX), however this format is yet to be commonly deployed, and especially so in medical imaging research and development.
To help spread developed models, and the ONNX format, we are currently looking for:
(a) master's thesis student(s) (30 p) who want to establish an open source collection of current state-of-the-art deep learning models and methods for medical imaging, mainly challenge winners, state-of-the-art models, and models from well-cited research papers. The project involves collecting and categorising pre-trained models, converting them from a diverse set of formats to the ONNX format, and testing them on medical image data in MICE Toolkit, a software for medical imaging developed by NONPI Medical AB.
For more information
If you are interested in learning about the state-of-the-art deep learning models and unifying the research community, please contact attila@nonpimedical.com
NONPI Medical AB develops and implements software tools for medical image analysis. The proposed project focuses on developing tools to design, alter, and train deep learning models and full machine learning pipelines for medical image processing, such as tumour segmentation, super resolution, or distortion correction.
There is much interest within for instance radiation therapy to use deep learning models for automation, but many in such fields lack the background to make substantial changes in existing models, and to develop their own specific models programmatically. With this background, NONPI Medical AB has developed a graphical tool to define the architecture for deep learning models. The existing tool is built on top of Python, which adds an extra layer of complexity since the rest of the program is written in C#, and also has problems whenever the underlying deep learning framework is updated.
The goal of this project it to evaluate different approaches to build deep learning models modularly in C#, and to find efficient graph algorithms to execute different steps of the machine learning pipeline on available hardware resources.
The goal of these developments is to aid the adaptation of existing deep learning models, and to make it possible for more researchers and medical practitioners to develop their own and be able to change or improve existing models and the entire maching learning pipeline. NONPI Medical AB is therefore looking for (a) master's thesis student(s) (30 hp) that wants to develop such tools for use within medical image analysis.
For more information
If you are interested in learning more about the project or how we use state-of-the-art deep learning models, please contact: attila@nonpimedical.com
AI to understand large scale biological data
MIMS has with the help of Johan Henrikssons research group established a new lab to produce single-cell data. With this method it is possible measure thousands (even millions) of cells, one at a time. The enormous datasets enable us to take a new approach to understanding biology, where we use AI to break down the complex measurements into knowledge.
This comes with computational challenges, from "how can we use AI to find a compressed data representation?" to "how do we efficiently handle several terabytes of sparse data?". Like the particle accelerator at CERN, biology now requires us to solve advanced computer scientific problems before we can go back to the biological questions.
We welcome thesis workers on all levels, from those who want fairly isolated problems to those who dare holistic approaches. It is a good thing if you want to learn some biology. Python is currently our standard language but we also use R, Java and C/C++ when the situation calls for it.
For more information
Please contact Johan Henriksson, researcher at MIMS. See also our website at www.henlab.org
