Medical Image Analysis

Medical image analysis takes place at the intersection of image acquisition, processing, analysis and interpretation. While our main focus is on the modelling and analysis of data, this can only be done through a detailed understanding of the image data and the underlying anatomical context. Our vision is to push the frontiers of integrated image acquisition, modelling and its interpretation.
Our research is grounded in industrial, clinical and societal needs.

Below, you find the following highlights:

Magnetic resonance imaging (MRI) of the brain
Explainability, fairness and uncertainty quantification
Topology aware learning for medical imaging
Shape and appearance modelling
Geometric machine learning
Cross-disciplinary applications
Synchrotron imaging for validation and prediction models of diffusion MRI
Light-sheet Fluorescence microscopy is crosslinked with MRI

Magnetic resonance imaging of the brain

Magnetic resonance imaging (MRI) is an extremely versatile modality that allows us to acquire 3D images of the brain with many different “contrasts” leading to the collection of 4D or sometimes N-D volumes. Modeling these images means applying and developing signal/image processing, biophysical models, and mathematical procedures to report summary parameters that can reveal the presence of pathology like stroke, edema, trauma, neurodegeneration, demyelination, etc. There is a variety of MRI techniques like functional and structural MRI (relaxometry, susceptibility, etc.).

For example see:

Diffusion MRI

We have a long research history in diffusion MRI (dMRI) which, by tracking the motion of water within the brain tissue, enables us to reconstruct the 3D anatomy of the brain white matter (tractography, see the basics), identify the different neural connections between cortical regions of the brain (connectivity), and quantify the tissue composition at every voxel (microstructure imaging), e.g. quantity and size of cells, axons, myelin, etc.. When multiple MRI contrasts are acquired of the same brain we refer to multi-dimensional MRI analysis. These analyses not only are of clinical relevance but are also fundamental in basic neuroscience where they can be integrated with different other types of modalities such as functional and metabolic imaging as well as other non-MRI imaging techniques.

For example see:

Denoising and artifact correction

A key aspect of quantitative imaging is precision, for which the attenuation of noise is paramount. For MRI, we research different strategies for the study of noise characteristics and for denoising images before modeling them. Moreover, we are interested in modeling and rectifying all those sources of uncertainty and imperfections that arise along the acquisition and processing pipeline of images.

For example see:

Pizzolato et al., 2020, Adaptive phase correction of diffusion-weighted images, NeuroImage Video here

Explainability, fairness and uncertainty quantification for medical imaging

Predictive models are an integral part of medical image analysis, forming the backbone not just for image guided diagnosis and prognosis, but also for image processing tasks, such as segmentation and registration, that are widely adopted even in the clinic. The output returned from these models is used to improve diagnosis and treatment, as well as to enhance our anatomical knowledge. It is therefore crucial for these models to also come with information on their own limitations. We develop methods for quantification of uncertainty, explainability, and for detecting bias and enhancing fairness in predictive models for medical imaging.

For example see:

Topology aware learning for medical imaging

While most modern medical imaging takes a very local, pixel-focused approach to problems such as image segmentation or registration, these often lead to suboptimal performance when viewed globally, in the sense that topological constraints naturally inherent in the data are violated. This can take the form of segmented structures taking on an incorrect topology, or image registration algorithms incorrectly representing the topology of the underlying anatomy. Our research includes topology-aware deep learning models for medical image processing.

For more information, see:

Shape and appearance modelling

The statistical modelling of shape and appearance was pioneered by the active shape and appearance models. This has been and is still a very active research topic, where recent developments within deep learning has further increased the potential in these methods. We are using these methods in different fields including modelling of the human face, ear and skull for hearing aid design. Another example is shape modelling of the human cochlea and the surrounding structures for surgical planning. A major research theme is also the modelling of the human heart for risk assessment and intervention planning.

Geometric machine learning

Recently, deep learning has been applied to non-structured 3D geometries as for example, surface meshes and a new area of research named geometric deep learning (GDL) has emerged. We have been using methods from geometric deep learning to analyze facial geometries, heart anatomies and are developing methods for the analysis of the human inner ear system. One fundamental aspect of GDL is the representing of the geometries that should be analyzed. Classical representations like triangulated surfaces are not easy to adapt to a deep learning setting due to the non-deterministic sampling properties. We are exploring new implicit representations like unsigned distance fields (UDF) that are potentially better suited for deep learning.

See for example:

Cross-disciplinary applications

A highly cross-disciplinary teamwork resulted in the development of the Carbo-gel: A multimodal soft tissue marker that bridges the gap between high-resolution diagnostic imaging and therapeutic intervention. During the project, medical image analysis was used to quantify and validate the behavior, e.g. efflux of solvent, surface volume, gel degradation and image contrast.

For example see:

Paper in Science Advances, 2020

Synchrotron imaging for validation and prediction models of diffusion MRI

X-ray synchrotron imaging uses a large-scale synchrotron research facility that basically can be seen as a gigantic nano-scope to create unique 3D images of intact tissue samples. There exist different synchrotron facilities in the world and we mostly use the MAXIV in Sweden, the ESRF in France, and the Swiss Light Source (SLS) in Switzerland etc.. Each synchrotron facility has different beamlines with highly specialized experimental setups, some of which allow us to observe anatomical features with nanometer resolution. As an example, we use phase-contrast imaging to map 3D anatomical morphology and architecture of axons, their myelin, and cell bodies, as well as the effects of pathology. We develop methods to segment, analyze and quantify 3D morphology and architecture features from synchrotron images. These are used for the prediction and validation of microstructural estimates obtained from the low-image resolution MRI scans i.e., diffusion MRI.

For example see:

Light-sheet Fluorescence microscopy is crosslinked with MRI

Light-Sheet Fluorescence microscopy imaging is a microscope technique that can generate 3D images of fluorescence-labeled substances of whole rodent brains. To scan a brain with light we first need to make the brain transparent and we use the DISCO+ technique where the brain becomes hard like a glass brain. The tissue-clearing technique introduces non-linear deformations of the tissue making it hard to align with standard brain atlases such as the world-known Allan Mouse Atlas (AIBS CCFv3). With collaborators, we are creating a framework to non-linearly align the 3D light-sheet images to the Allan Mouse Atlas as well as to MRI e.g. diffusion MRI. This allows us to find a voxel-by-voxel correspondence between modalities which enables a rigorous cross-modality image correlation analysis.