New solutions based on the combination of data science and medical physics can help pave the way for rapid diagnostics and personalized medicine for improved patient prognosis and quality of life.


Our research focuses on developing data science approaches for medical and veterinary applications, with emphasis on diagnostics, treatment outcome prediction, biomarker identification and automatic segmentation of radiotherapy target volumes. We analyse medical images and tabular data to develop our models.

Areas of interest are feature engineering, feature selection, multivariate and multi-source data analysis, machine learning, deep learning and related software development.

News and publications

Research areas

  • Target volume definition is a central part of the workflow in radiotherapy planning, where the aim is to deliver sufficiently large radiation doses to the target, while sparing the surrounding tissue to avoid unpleasant side effects. The current gold standard for target definition is manual contouring in medical images. However, manual delineation is both labour intensive and time consuming and can be prone to uncertainties introduced by intra- and interobserver variations.

    Our goal is to develop approaches for automatic target volume segmentation using deep learning or classical machine learning. To date we have focused on automatic segmentation of head and neck tumours and malignant lymph nodes, as well as tumours of the pelvic region (rectal and anal cancers). In addition, tumor segmentation of canine head and neck tumors has been investigated in collaboration with the Faculty of Veterinary Medicine.

    Current research interests include impacts of imaging modality (CT, PET, MRI) on segmentation performance, architecture assessment, model optimization, AI interpretability and uncertainty quantification.

  • A large amount of data is accumulated during patient workup and monitoring, such as medical history, vital signs, laboratory tests, medical images, histology, and potentially genomic data. This amounts to high-dimensional datasets consisting of very many variables and images acquired using many different sources. Our aim is to identify patterns within such datasets for the purpose of finding biomarkers and predicting patient treatment outcome using machine learning and deep learning models. Such models and biomarkers can potentially be incorporated into decision support systems paving the way for personalized medicine

    Our current approaches include radiomics, design and exploration of feature selection methods as well as investigating methods for analysing multivariate and multi-source data.

  • (to be updated)

  • We develop methods and software for feature selection, feature engineering, image segmentation, as well as multivariate and multi-source data analysis.



    Our chemometrics Python package hoggorm includes classical methods for multivariate statistics such as principal component analysis, principal component regression and partial least squares regression. See our paper in the JOSS for more details on hoggorm.

    Rent logo

    RENT – Repeated Elastic net for feature selection

    RENT is our approach for feature selection from short and wide datasets which have few samples compared to the number of features per sample. The RENT Python package has recently been published in the Journal of Open Source Software (JOSS) . Feature selection methods such as RENT are particularly relevant for medical datasets, where the number of patients is often limited while the number of features characterising each patient is high.


    Deoxys – Framework for running deep-learning experiments with emphasis on cancer tumor auto-segmentation and treatment outcome prediction

    Deoyxs is our Python framework for deep learning-based segmentation developed with emphasis on target volume segmentation in medical images. In addition, Deoxys can be used for treatment outcome prediction using convolutional neural networks (CNNs) and medical images or fully connected neural networks (FCNN) and tabular data such as radiomics and/or clinical features. The framework includes modules for data reading, image pre-processing and augmentation, 2D and 3D CNN architectures, parameter options (different loss and activation functions, layer types and optimizers), and an experimental management framework for efficient and systematic model training, evaluation, comparison and logging.


    ImSkaper – Radiomics feature extraction, feature selection, and classification

    ImSkaper is a Python package for the extraction and engineering of image features. ImSkaper extracts the standard radiomics features frequently used for patient treatment outcome prediction. ImSkaper has been extended to include Local Binary Patterns on 3D images.

Group members

  • Aurora Rosvoll Grøndahl

    Aurora Rosvoll Grøndahl

    Former PhD student

    CEHEADS member from 2017 to 2023

    Yngve Mardal Moe

    Yngve Mardal Moe

    Former PhD student

    CEHEADS member from 2017 – 2020

    Turid Torheim

    Turid Torheim

    Former PhD student

    CEHEADS member from 2013 – 2016