PTB has increased its expertise in AI in medicine 

Artificial intelligence (AI) methods are enjoying much success – even in the medical sector – as they promise to be of great benefit for patients and offer an enormous potential for the economy. However, AI is also creating new challenges. As a National Metrology Institute, the Physikalisch-Technische Bundesanstalt (PTB) is Germany's highest authority with a legal mandate when it comes to evaluating measurement procedures and PTB is rising to these new challenges. It has expanded its interdisciplinary team which is developing new procedures for the standardized evaluation of AI methods in various fields by 10 early stage researchers. The thematic focus is on explainability, robustness and uncertainty of AI systems in the fields of imaging procedures as well as on radiation and laboratory medicine. 

Due to rapidly advancing digital transformation, methods of artificial intelligence are increasingly finding their way into all areas of medicine and are also presenting PTB with new challenges. The complexity of correlations, the progress made in measurement technology, the great benefit for patients and the enormous potential for the economy will rapidly increase AI applications in the health sector. AI techniques could be the solution to many challenges: For example, imaging procedures generate large volumes of data which are unmanageable for physicians. According to the German Cancer Research Center (Deutsches Krebsforschungszentrum, DKFZ) in Heidelberg, a total of 675 exabytes of image information were generated by radiologists in 2019, which corresponds to about 13.5 trillion tomographic images – but only about 7 % of the data has been used in medicine. At the same time, the number of registered practice-based radiologists decreased by 2.2 % in Germany from 2018 to 2019, whereas the total number of practice-based physicians increased. This is where AI can be of help in the future: Tasks for which physicians would require years of experience can be processed by artificial intelligence within a few seconds. However, there is to date no generally recognized anchor of trust in quality infrastructure.

PTB is intensively dealing with the subject of AI within the scope of its legal mandate, given, for example, by the German Medical Devices Act. PTB is furthermore developing modern methods for determining the image quality of X-ray images, e.g. in mammography. The aim is to use objective parameters for the image quality and dose to be able to develop an objective target value and optimize these radiological diagnoses – with a view to achieving the best possible image quality at low doses. In magnetic resonance tomography, robust methods for image reconstruction are being developed. Here, machine learning allows, for instance, the image reconstruction of functional MR images to be accelerated in such a way that examinations of moving tumors are also possible. Another example is the world's largest freely accessible ECG database published by PTB, which provides machine-readable findings and assessments by cardiologists. The dataset can be used as a reference for methods on the market that have often been trained with proprietary data so far. In this way, greater transparency is achieved.

The research program "Metrology for AI in Medicine" (M4AIM) is interdisciplinary and covers AI applications in different fields of medicine. Depending on their special fields, the new staff members are employed at one of PTB's two sites, Braunschweig or Berlin, and work both with experienced PTB colleagues and with experts from the academic AI research and from cooperating university clinics and hospitals. However, even if the concrete tasks vary, the focus is on the team character as well as on the three major pillars of metrology for AI applications. The first pillar is comprehending the reasons for an AI result. The second pillar is about determining the uncertainty of AI methods, which is the metrological approach to the accuracy of the corresponding measurement methods. The third pillar stands for the generalization and robustness of AI methods. Here, the aim is to develop parameters which can be standardized to evaluate robustness against noisy, unknown or harmful input data. This is the general perspective for the entire team.

Further information on the research program "Metrology for AI in Medicine" (M4AIM), e.g. descriptions of the single projects, can be found on its website.

Relevant PTB news (sorry, some of them are in German only):

Training machine learning algorithms for assessing ECGs

Großer maschinenlesbarer EKG-Datensatz für Entwicklung und Test maschineller Lernverfahren veröffentlicht

Maschinelles Lernen zur Qualitätssicherung in der Mammografie

When the MRI scanner does the analyses on its own

Model-based observer and machine learning for high-quality images

Matters close to the heart: artificial intelligence

Selected publications 

• L. Hoffmann, C. Elster, Deep neural networks for computational optical form measurements, Journal of Sensors and Sensor Systems, 2020. https://doi.org/10.5194/jsss-9-301-2020
• T. Kretz, K.-R. Müller, T. Schäffter and C. Elster, Mammography Image Quality Assurance Using Deep Learning, IEEE Transactions on Biomedical Engineering, 2020. https://doi.org/10.1109/TBME.2020.2983539
• J. Martin and C. Elster, Inspecting adversarial examples using the Fisher information, Neurocomputing, 2020. https://arxiv.org/abs/1909.05527
• J. Martin and C. Elster, Detecting unusual input to neural networks, Applied Intelligence, 2020. https://doi.org/10.1007/s10489-020-01925-8
• E. Montagnon et al., Deep learning workflow in radiology: a primer, Insights into Imaging, 2020. https://doi.org/10.1186/s13244-019-0832-5

• M. Anton, W. J. H. Veldkamp, I. Hernandez-Giron and C. Elster, RDI - a regression detectability index for quality assurance in: x-ray imaging. Physics in Medicine & Biology, 65(8), 085017, 2020. [https://doi.org/10.1088/1361-6560/ab7b2e]
• T. Kretz, M. Anton, T. Schaeffter, C. Elster, Determination of contrast-detail curves in mammography image quality assessment by a parametric model observer. Physica Medica (2019), https://doi.org/10.1016/j.ejmp.2019.05.008
• N. Strodthoff, P. Wagner, T. Schaeffter, W. Samek, Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL. IEEE Journal of Biomedical and Health Informatics. 2020, early Access Article, https://doi.org/10.1109/JBHI.2020.3022989
• P. Wagner, N. Strodthoff, R.-D. Bousseljot, D. Kreiseler, F. I. Lunze, W. Samek, T. Schaeffter: PTB-XL, a large publicly available electrocardiography dataset. Scientific Data 7, 154 (2020) https://doi.org/10.1038/s41597-020-0495-6
• C. Cai, A. Hashemi, M. Diwakar, S. Haufe, K. Sekihara, S. S. Nagarajan, Robust estimation of noise for electromagnetic brain imaging with the champagne algorithm, NeuroImage, 2021. https://doi.org/10.1016/j.neuroimage.2020.117411
• A. Hashemi, C. Cai, G. Kutyniok, K.-R. Müller, S. S. Nagarajan, S. Haufe, Unification of Sparse Bayesian Learning Algorithms for Electromagnetic Brain Imaging with the Majorization Minimization Framework, bioRxiv, 2020. https://doi.org/10.1101/2020.08.10.243774