EchoLux
Intelligent Optoacoustic Radiomics
Synergistic Integration of System Models and Medical Knowledge
an ERC Starting Grant project
Radiomics – the extraction of medical information from imaging data via mathematics and data science – is on the verge of becoming a main player in clinical medicine. However, the current radiomics workflow lags behind the state-of-the-art in explainable artificial intelligence. We will integrate the whole imaging workflow – from imaging hardware to clinical interpretation – into an intelligent software environment, thereby realizing the transition from black box machine learning to intelligent radiomics using novel frameworks like information field theory.
The clinical use case is imaging of peripheral nerves with optoacoustic tomography, which can visualize nervous tissue in unprecedented detail. The project, thus, has the potential to enable early detection of pathological changes in peripheral neuropathy, e.g., in conjunction with diabetes.
About EchoLux
Quantitative optoacoustic imaging via probabilistic modelling
An open problem in optoacoustic tomography is the reliable quantification of tissue constituents from spectral data acquired at multiple wavelengths of the illuminating light. The reason for this shortcoming is the non-linear and ill-posed nature of this problem. To overcome these problems, we combine powerful solvers for non-linear problems with suitable informative priors to regularize the problem by building on the novel information field theory framework. This permits to utilize prior knowledge about the composition of specific tissues and their anatomical arrangement and provides uncertainty quantification, which is especially important in the medical context for reliable clinical decision making.
Modelling disease effects on tissue
Classic radiomics approaches try to predict disease labels from imaging features in a black box manner. To transition to explainable radiomics, we propose to model the known effects of disease on the tissue expressions that can be captured with the imaging modality. This approach allows us to integrate medical knowledge explicitly into the radiomics workflow. Detection of disease effects in the imaging data, thus, can be formulated as an inverse problem (similar to image reconstruction), which allows to utilize the rich mathematical toolbox in this field. We will focus on the clinical use case of peripheral neuropathy, where early detection of pathological changes in vascularization and morphology is potentially possible with optoacoustic imaging.
From sinograms to disease
Finally, the computational challenge is to achieve transparent reasoning through the whole workflow – from sinogram data to quantitative optoacoustic images to tissue features to disease effects. Recently, deep learning architectures have successfully be used to implement fast (differentiable) building blocks that allow to realize such complex reasoning tasks. We will build on this development to enable a fully interpretable and transparent reasoning framework for clinical optoacoustic tomography.
Scientists in Echolux
Postdoctoral fellow - Inverse Problems in Optoacoustics
Physician scientist - Clinical Epidemiology
Ph.D. Student - Probabilistic Reconstruction and System Characterization
Ph.D. Student - Probabilistic Reconstruction and Fluence Modeling
Ph.D. Student - Tissue Characterization and Data Analysis in Optoaoustics
Contact
Head of project