Periodic Reporting for period 1 - EchoLux (Intelligent Optoacoustic Radiomics via Synergistic Integration of System Models and Medical Knowledge)
Reporting period: 2022-07-01 to 2024-12-31
Extracting clinically relevant information from medical imaging data via mathematics and data science is called radiomics. Artificial intelligence plays a major role in modern radiomics but relies heavily on large datasets for training the AI models. Solely relying on data has major disadvantages: biases in the training dataset might propagate to the AI solution, and all information about the imaging process and disease effects is only contained implicitly in the imaging data. The EchoLux project will enable interpretable, reliable, and unbiased radiomics by integrating knowledge about the imaging process and disease effects into a clinical decision support system. The use case of EchoLux is optoacoustic imaging of peripheral nerves with the potential to enable the detection of early pathological changes in nervous tissue, for example, due to peripheral neuropathy, which is a main complication of diabetes.
EchoLux puts the biological tissue in the center of attention. It does so by linking both the raw imaging data and the target disease to tissue properties. Quantitatively reconstructing tissue properties from the raw optoacoustic imaging data is called ‘quantitative optoacoustics’ and is known to be a very hard computational problem. Making major progress towards quantitative optoacoustics is the first objective of EchoLux. Furthermore, specific disease effects on the tissues that can be visualized with optoacoustic imaging will be modelled explicitly to understand how observable tissue properties change. This approach enables to identify pathological changes specific to the target disease from the observed tissue properties, which is the second objective of EchoLux. Combining the two models to detect disease effects from raw imaging data yields the final EchoLux radiomics solution.
EchoLux puts the biological tissue in the center of attention. It does so by linking both the raw imaging data and the target disease to tissue properties. Quantitatively reconstructing tissue properties from the raw optoacoustic imaging data is called ‘quantitative optoacoustics’ and is known to be a very hard computational problem. Making major progress towards quantitative optoacoustics is the first objective of EchoLux. Furthermore, specific disease effects on the tissues that can be visualized with optoacoustic imaging will be modelled explicitly to understand how observable tissue properties change. This approach enables to identify pathological changes specific to the target disease from the observed tissue properties, which is the second objective of EchoLux. Combining the two models to detect disease effects from raw imaging data yields the final EchoLux radiomics solution.
An initial success is the publication of a foundational EchoLux paper that introduces the clinical application of multispectral optoacoustic tomography for peripheral neuropathy, the use case of EchoLux, and demonstrates the ability of the technology to visualize nervous tissue in great detail in a pilot study on 12 healthy volunteers. In addition, the paper contributes fundamental methodology for the analysis of spectral optoacoustic data and provides insights into the effects of light attenuation in tissue, which is a main obstacle towards quantitative optoacoustic imaging.
EchoLux will carry out a follow-up imaging study on a sample of the general population to gather a larger dataset of real optoacoustic images of the target region that will be used for data-driven aspects of the EchoLux framework. The ethics committee of the Technical University of Munich approved the study, and the team is starting to recruit participants.
For finetuning of the model and for method validation, the team has started to build physical phantoms of the target tissues, i.e. materials that mimic the acoustic and optical properties of tissues assembled into a geometric design similar to the anatomy of the target region.
Furthermore, the interdisciplinary research team of EchoLux has made major progress towards the goals ‘quantitative optoacoustics’ and ‘inference of medical knowledge’.
Fundamental work to characterize and model the optoacoustic imaging system has been carried out. Detailed models of the optical excitation of tissue and of the ultrasound detectors in the optoacoustic imaging system have been developed and integrated into the image reconstruction procedure to allow the EchoLux framework to be aware of the specifics of the data generated by the system. Based on these models, we developed methods for solving the optical inverse problem of optoacoustic imaging probabilistically in a Bayesian framework, with a suitable regularization scheme, and based on a physics-based effective model of the optoacoustic imaging data.
The team researched medical knowledge on the effects of different neuropathies and on confounding effects. These sources of information are currently used to implement changes in specific tissues in a numerical phantom of the upper arm.
As fundamental work towards the final EchoLux framework, we revisited the variational Bayes methodology together with our collaborators at the Max Planck Institute for Astrophysics.
EchoLux will carry out a follow-up imaging study on a sample of the general population to gather a larger dataset of real optoacoustic images of the target region that will be used for data-driven aspects of the EchoLux framework. The ethics committee of the Technical University of Munich approved the study, and the team is starting to recruit participants.
For finetuning of the model and for method validation, the team has started to build physical phantoms of the target tissues, i.e. materials that mimic the acoustic and optical properties of tissues assembled into a geometric design similar to the anatomy of the target region.
Furthermore, the interdisciplinary research team of EchoLux has made major progress towards the goals ‘quantitative optoacoustics’ and ‘inference of medical knowledge’.
Fundamental work to characterize and model the optoacoustic imaging system has been carried out. Detailed models of the optical excitation of tissue and of the ultrasound detectors in the optoacoustic imaging system have been developed and integrated into the image reconstruction procedure to allow the EchoLux framework to be aware of the specifics of the data generated by the system. Based on these models, we developed methods for solving the optical inverse problem of optoacoustic imaging probabilistically in a Bayesian framework, with a suitable regularization scheme, and based on a physics-based effective model of the optoacoustic imaging data.
The team researched medical knowledge on the effects of different neuropathies and on confounding effects. These sources of information are currently used to implement changes in specific tissues in a numerical phantom of the upper arm.
As fundamental work towards the final EchoLux framework, we revisited the variational Bayes methodology together with our collaborators at the Max Planck Institute for Astrophysics.
Foundation models, i.e. AI models that are trained in an unsupervised or self-supervised manner on large data sets, are the current state-of-the-art of artificial intelligence. They have been demonstrated to outperform other methods when finetuned for downstream tasks. The EchoLux team investigates approaches to leverage the power of foundation models for the project goals. A first success in this direction is the development of a method to condition a deep generative model of the target anatomy with optoacoustic data to allow reconstruction of the segmentation of the region into different tissues and simultaneously recover the tissue specific optical properties. A proof-of-principle for this strategy is currently in preparation for publication.
Another preliminary result of particular importance is the demonstration that the integration of the interaction of light with the skin into the image reconstruction process yields a method that strongly reduces the skin color bias of optoacoustic imaging – a major advance towards equitable optoacoustics.
Another preliminary result of particular importance is the demonstration that the integration of the interaction of light with the skin into the image reconstruction process yields a method that strongly reduces the skin color bias of optoacoustic imaging – a major advance towards equitable optoacoustics.