Other Projects

Due to Europe's ageing society, there has been a dramatic increase in the occurrence of osteoporosis (OP) and related diseases. Sufferers have an impaired quality of life, and there is a considerable cost to society associated with the consequent loss of productivity and injuries. The current understanding of this disease needs to be revolutionized, but study has been hampered by a lack of means to properly characterize bone structure, remodeling dynamics and vascular activity. This project, 4D nanoSCOPE, will develop tools and techniques to permit time-resolved imaging and characterization of bone in three spatial dimensions (both in vitro and in vivo), thereby permitting monitoring of bone remodeling and revolutionizing the understanding of bone morphology and its function.

To advance the field, in vivo high-resolution studies of living bone are essential, but existing techniques are not capable of this. By combining state-of-the art image processing software with innovative 'precision learning' software methods to compensate for artefacts (due e.g. to the subject breathing or twitching), and innovative X-ray microscope hardware which together will greatly speed up image acquisition (aim is a factor of 100), the project will enable in vivo X-ray microscopy studies of small animals (mice) for the first time. The time series of three-dimensional X-ray images will be complemented by correlative microscopy and spectroscopy techniques (with new software) to thoroughly characterize (serial) bone sections ex vivo.

The resulting three-dimensional datasets combining structure, chemical composition, transport velocities and local strength will be used by the PIs and international collaborators to study the dynamics of bone microstructure. This will be the first time that this has been possible in living creatures, enabling an assessment of the effects on bone of age, hormones, inflammation and treatment.

Introduction: Antitrust laws (also known ascompetition laws) are designed to encourage strong competition and are designedto protect consumers from predatory commercial practices. The primary goals ofantitrust law are to ensure the functioning of the markets and to ensure faircompetition. A prominent example of an antitrust violation is illegal pricefixing. By definition, it is an agreement between competitors that fixes pricesor other competitive conditions, and thus violates the principle of the pricingmechanism through free market forces. A typical feature of illegal price fixingis verifiable communication (written or verbal) between human marketparticipants. However, in the age of artificial intelligence and e-commerce,the definition and the detection of this illegal practice faces new challengesas collusive behaviors that violate antitrust laws, such as the pricingmechanism, can be partially or fully automated [1]. Furthermore, thecommunications between market participants can be both overt and covert. Finally,market participants can be artificial agents which might affected by perverseinstantiation [2]. In other words, new technological possibilities areavailable to disguise illegal pricing policies and business practices.

Recent research, mainly from theeconomic and jurisprudence point of view, concludes the intensive applicationof AI algorithms in E-commerce will increase the extend of known forms ofanticompetitive behaviors [3][4]. However, the questions regarding whether andto which extent collusive behaviors will emerge by AI itself (which is anunknown form of anticompetitive behaviors) are rarely understood. Feasibilitystudies and comprehensive analysis comprising the implementation of AI methods andvalidation of the derived hypothesis has not been conducted so far. Therefore,the main goals of this project are to investigate the possibilities of collusivebehaviors stimulated and/or emerged by AI algorithms on digital marketplace andderive consequences on the antitrust law as well as competition policies. Tothe best of our knowledge, this is the first time that a research project inthe field of Antitrust and AI (AAI) is focusing on the mathematical andalgorithmic perspective of the question to which extend the utilization of AImethods is facilitating the collusive behaviors in the era of digital economy.

Objectives: In order to validate the hypothesesthat AI algorithms is able to develop and communicate collusive behaviors ondigital marketplaces both in overt and covert fashion, comprehensive emulatorsof online marketplaces in different setups will be implemented.  Furthermore, different communication channels(both overt and covert) of digital marketplaces will be discovered and understood,as it is highly relevant to the detection of collusive practices. Finally, differentonline trading scenarios utilizing AI algorithms will be established and theimpact on antitrust law and competition polices will be derived. In total, the mainaspects in the intended DFG-application can be summarized as follows:

1.      Asthe research topic belongs to a highly interdisciplinary field, a comprehensiveliterature review is necessary to define the problem space of the research andis of great importance to conduct the subsequent experiments successfully.Therefore, a comprehensive literature review on the aspects of antitrust law, gametheory, artificial intelligence and cyber security will be conducted.

2.      Firststep of the implementation is the holistic emulation of the digitalmarketplace. The market emulator should have the capability to emulate the digitalmarket following various rules (e.g., Cournot vs. collusive competition) indifferent size (i.e., with different amount of market participants). Moreover,state-of-the-art algorithms for dynamic pricing should be replicated andintegrated into the market emulator as well.

3.      Afurther aspect of this project is the communication mechanism in the era of E-commerceand AI. The know form of collusions mostly utilize overt communications.However, covert communication channels (i.e., communication channels that are notoriginally designed for the communication purpose, therefore hardly to bedetected [5][6]) poses further vulnerabilities of online marketplaces. The mechanismsand capacities of covert channels facilitating the collusive behaviors (e.g.,illegal price fixing) as should be investigated with the implemented marketemulator.

4.      Finally,artificial agents for price definition of different products should be proposedand implemented following different competition models as well as marketcomplexities, aiming at understanding the central research questions of thisresearch project, i.e., capabilities and conditions of emerging collusivebehaviors of artificial agents by themselves. This particular step can beachieved by using reinforcement learning techniques. Technical opportunitiesand challenges for the discrimination of collusive and non-collusive behaviorsthat are potentially emerged by the artificial agents should be explored aswell.

The entire project will besupervised by experts from three disciplines. Prof. Jochen Hoffmann (chair of Private Business Law) will support this research project with his knowledges and expertiseon antitrust law, Prof. Felix Feilling (chair of Cyber Security) will advise onthe aspects that are related to covert communication and cyber security, andProf. Andreas Maier (pattern recognition lab) will mentor this project from theAI point of view.

[1] KünstlicheIntelligenz als Marktteilnehmer – Technische Möglichkeiten, Maier A., Bayer S.,Mohr Verlag. Submitted, unpublished.

[2] BostromN. Superintelligence: Paths, Dangers, Strategies. Minds & Machines 25, Seite 285–289 (2015).

[3] Petit, N. Antitrust andartificial intelligence: A research agenda. In: Journal of European CompetitionLaw and Practice. Vol. 8, Issue 6, pp. 361–362. Oxford University Press.(2017)

[4] Beneke, F., Mackenrodt, M.,Remedies for algorithmic tacit collusion, Journal of Antitrust Enforcement,Volume 9, Issue 1, Pages 152–176 (2021).

[5]Hans-Georg Eßer, Felix C. Freiling. Kapazitätsmessung eines verdecktenZeitkanals über HTTP, Univ. Mannheim,Technischer Bericht TR-2005-10, November 2005. (2005)

[6] Freiling F.C., Schinzel S.Detecting Hidden Storage Side Channel Vulnerabilities in NetworkedApplications. In: Camenisch J., Fischer-Hübner S., Murayama Y., Portmann A.,Rieder C. (eds) Future Challenges in Security and Privacy for Academia andIndustry. SEC 2011. IFIP Advances in Information and Communication Technology,vol 354. Springer, Berlin, Heidelberg. (2011)

Modern diagnostic CT systems comprise a variety of measures to keep patient dose at a minimum. Of particular importance is the tube current modulation (TCM) technique that automatically adapts the tube current for each projection in a way to minimize the tube current time product (mAs-product) for a given image quality. This can also be regarded as maximizing the image quality for a given mAs-product. TCM performs a modulation depending on the angular position of the x-ray tube and depending on the z-position of the scan. Measures related to TCM are the automated choice of the mean tube current and of the optimal tube voltage. These three dose reduction methods are also known under the term automatic exposure control (AEC).As of today, however, the AEC does not minimize the actual patient dose and thereby the actual patient risk. It rather minimizes surrogates thereof. The surrogate of TCM is the mAs-product. The surrogate used to automatically select the tube voltage is the CTDI value or the dose length product (DLP). A direct minimization of the weighted summed organ dose values and thereby the patient risk is currently not practicable due to a) the very high computation times of the Monte Carlo dose calculation algorithms and b) due to the lack of a reliable segmentation of the radiation sensitive and risk-relevant organs.We therefore plan to use artificial neural networks to solve the above-mentioned problems and to realize a new AEC which is capable of directly minimizing patient dose and risk instead of using surrogate parameters. A first neural net will convert the patient topogram(s) into a coarse estimate of the CT volume. In cases where only a single topogram is available information of the table height will be used for this estimation. A second neural net will segment the relevant organs. We can here partially use prior work of a previous DFG project (KA 1678/20, LE 2763/2, MA 4898/5). A third network will use further scan parameters (table increment, pitch value, rotation time, collimation, tube voltage, …) to compute the expected dose distribution per projection. This, together with the segmentation of the organs, shall be used to compute the effective dose (or risk) or the patient per projection. A minimization algorithm will then find the optimal tube current curve that minimizes patient risk at a given image quality or that maximizes image quality at a given patient risk.To evaluate our deep AEC algorithm diagnostic CT data will be collected. The data will be retrospectively converted to the desired tube current curve by adding noise to the rawdata followed by another reconstruction. Experienced radiologists will then perform a blinded study where they read images produced without AEC, with the conventional AEC as of today, and with our new deep AEC algorithm.

Zur Bewertung von Sprechstörungen von Patienten mit Lippen-Kiefer-Gaumenspalten fehlen bisher objektive, validierte und einfache Verfahren. Im klinischen Alltag werden Lautbildungsstörungen bisher üblicherweise durch eine subjektive, auditive Bewertung erfasst. Diese ist für die klinische und v.a. wissenschaftliche Nutzung nur bedingt geeignet. Die automatische Sprachanalyse, wie sie für Spracherkennungssysteme genutzt wird, hat sich bereits bei Stimmstörungen als objektive Methode der globalen Bewertung erwiesen, nämlich zur Quantifizierung der Verständlichkeit. Dies ließ sich in Vorarbeiten auch auf Sprachaufnahmen von Kindern mit Lippen-Kiefer-Gaumenspalten übertragen. In dem vorliegenden Projekt wird ein Verfahren zur automatischen Unterscheidung und Quantifizierung verschiedener typischer Lautbildungsstörung wie Hypernasalität, Verlagerung der Artikulation und Veränderung der Artikulationsspannung bei Kindern und Jugendlichen mit Lippen-Kiefer-Gaumenspalten entwickelt und validiert. Dies stellt die Basis für die Ermittlung ihres Einflusses auf die Verständlichkeit sowie zur Erfassung der Ergebnisqualität verschiedener therapeutischer Konzepte dar.

Tomographic reconstruction is the enabling technology for a wide range of transmission-based 3D imaging modalities, most notably X-ray Computed Tomography (CT). The first CT scanners built in the 70ies used parallel geometries. In order to speed up acquisition, the systems soon moved to fan-beam geometries and a much faster rotation speed. Today´s CT systems rotate four times per second and use a cone-beam geometry. This is fast enough to cover even complex organ motion such as the beating heart. However, there exists a large class of specialized CT systems that are not able to perform such fast scans. Systems using a flat-panel detector, as they are employed in C-arm angiography systems, on-board imaging systems in radiation therapy, or mobile C-arm systems, face mechanical challenges as they were mainly built to perform 2D imaging. About 15 years ago flat-detector scanners have been enabled to acquire three dimensional data. 3D imaging on these systems, however, is challenging due to their slower acquisition speed between five seconds and one minute and a small field-of-view (FOV) with a diameter of 25 to 40 cm. These drawbacks are related to the scanners´ design as highly specialized modalities. In contrast to other disadvantages of flat panel detectors like increased X-ray scattering and limited dynamic range, they will not be remedied by hardware evolution in the foreseeable future, e.g. faster motion is impossible because of the risk of collisions in the operation room. As a result, Flat-Detector Computed Tomography (FDCT) will continue to be more susceptible to artifacts in the reconstructed image due to motion and truncation. The goal of this project is to extend existing data consistency conditions, which can be practically used for FDCT to remedy intrinsic weaknesses of FDCT imaging, most importantly motion and truncation. Our goal is the practical applicability on clinical data. Thus, the new algorithms will be tested on physical phantom and patient data acquired on real FDCT scanners of our project partners. Our long-term vision behind this project is to find a concise and complete formulation for all redundancies within FDCT projection data for general trajectories and fully exploit them in the reconstruction process. Redundancies are inherent to every FDCT scan done in today´s clinical practice, but they are ignored entirely as a source of information. Data consistency conditions do not require any additional effort during the acquisitions and only little prior knowledge such as the object extent or even no assumptions about the underlying object, unlike for example total variation regularization in iterative reconstruction. Hence, they rely solely on information which is naturally present in the data.

Quite a number of portraits of Martin Luther - known as media star of 16th century – can be found in today’s museums and libraries. However, how many portraits indeed exist and which of those are contemporary or actually date from a later period? So far unlike his works, however, the variety of contemporary portraits (painting and print) is neither completely collected nor critically analyzed. Thus, a joint project of the FAU, the Germanisches Nationalmuseum (GNM) in Nuremberg and the Technology Arts Sciences (TH Köln) was initiated. Goal of the interdisciplinary project covering art history, art technology, reformation history and computer science is the creation of a critical catalogue of Luther portraits (1519-1530). In particular, the issues of authenticity, dating of artworks and its historical usage context as well as the potential existence of serial production processes will be investigated.

Breast cancer is one of the leading causes of death in the field of oncology in Germany. For the successful care and treatment of patients with breast cancer, a high level of information for those affected is essential in order to achieve a high level of compliance with the established structures and therapies. On the one hand, the digitalisation of medicine offers the opportunity to develop new technologies that increase the efficiency of medical care. On the other hand, it can also strengthen patient compliance by improving information and patient integration through electronic health applications. Thus, a reduction in mortality and an improvement in quality of life can be achieved. Within the framework of this project, digital health programmes are going to be created that support and complement health care. The project aims to provide better and faster access to new diagnostic and therapeutic procedures in mainstream oncology care, to implement eHealth models for more efficient and effective cancer care, and to improve capacity for patients in oncologcal therapy in times of crisis (such as the SARS-CoV-2 pandemic). The Chair of Health Management is conducting the health economic evaluation and analysing the extent to which digitalisation can contribute to a reduction in the costs of treatment and care as well as to an improvement in the quality of life of breast cancer patients.

In the course of massive digitization, a large part of libraries’ archived documents are currently being converted into electronic formats. However, the digitization is also reaching its limits. Scanning robots cannot digitize documents whose condition due to natural ageing or external influences prohibit a conventional, optical based processing. Our own preliminary work has shown that the three imaging techniques X-ray Computed Tomography, Phase Contrast X-ray Computed Tomography and Terahertz Imaging are suitable for providing non-invasive insights into such documents, allow the acquisition of digital imaging information and are capable to re-enable an efficient automated process to digitalize cultural heritage documents.This research project is the first to develop a concrete digitization strategy or method for such documents. This structured evaluation will be based on a quality value that allows statements to be made about the expected result of digitization with one of the mentioned modalities for certain historical materials. From this, the most suitable imaging procedure can be determined. Based on these findings, a guideline for the digitization of fragile documents will be developed to predict the quality, feasibility and possible damage before a scan. In addition, algorithms will be developed that virtually process the generated data and make it readable for the human eye. Three concrete goals will be pursued to carry out the research project. By evaluating the modalities for selected historical materials, the most appropriate procedure for a specific document should then be identified. At the end of the project, a guide will be made available and the possibilities of each modality will be demonstrated by specifying material combinations and relevant parameters. It will be possible to test the variation of the recording parameters and to display exemplary results using the generated database. This also makes it possible to calculate a quality value. The basis for such a guide is the evaluation of the three modalities for relevant materials. For this purpose, realistic test specimen are produced. Both the scan quality and resolution as well as possible damage to the document must be considered. The guide will then be used to identify the most suitable procedure for a specific document. This statement is based on the mentioned quality value, which will also be used to predict the optimal digitization modality and the quality for an unknown document.The evaluation of several modalities as well as the development of algorithms are to be seen as central challenges of the research project. It will be possible to store endangered holdings in a digital format without destroying their structure through manual intervention. In the second funding phase, a multi-modal solution should be investigated in which disadvantages and limitations of individual modalities will be compensated by combining several modalities.

From Micro To Macro: Multiscale Multimodal Data Analysis for Breast Cancer Research

The dark-field signal of an X-ray phase-contrast system captures the small-angle scattering of microscopic structures. It is acquired using Talbot-Lau interferometer, and a conventional X-ray source and a conventional X-ray detector. Interestingly, the measured intensity of the dark-field signal depends on the orientation of the microstructure. Using algorithms from tomographic image reconstruction, it is possible to recover these structure orientations. The size of the imaged structures can be considerably smaller than the resolution of the used X-ray detector. Hence, it is possible to investigate structural properties of - for example - bones or soft tissue at an unprecedented level of detail.Existing methods for 3-D dark-field reconstruction require sampling in all three spatial dimensions. For practical use, this procedure is infeasible.The goal of the proposed project is to develop a system and a method for 3-D reconstruction of structure orientations based on measurements from a practical imaging trajectory. To this end, we propose to use a helical trajectory, a concept that has been applied in conventional CT imaging with tremendous success. As a result, it will be possible for the first time to compute dark-field volumes from a practically feasible, continuous imaging trajectory. The trajectory does not require multiple rotations of the object or the patient and avoids unnecessarily long path lengths.The project will be conducted in cooperation between the experimental physics and the computer science department. The project is composed of six parts:- A: Development of a 3-D cone-beam scattering projection model- B: Development of reconstruction algorithms for a helical dark-field imaging system- C: Evaluation and optimization of the reconstruction methods towards clinical applications- D: Design of an experimental helical imaging system- E: Setup of the helical imaging system- F: Evaluation and optimization of the system performanceParts A to C will be performed by the computer science department. Parts D to E will be conducted by the experimental physics department.

The project examines the use and further development of machine learning methods for MR image reconstruction and for the classification of liver lesions. Based on a comparison model and data-driven image reconstruction methods, these are to be systematically linked in order to enable high acceleration without sacrificing diagnostic value. In addition to the design of suitable networks, research should also be carried out to determine whether metadata (e.g. age of the patient) can be incorporated into the reconstruction. Furthermore, suitable classification algorithms on an image basis are to be developed and the potential of direct classification on the raw data is to be explored. In the long term, intelligent MR diagnostics can significantly increase the efficiency of use of MR hardware, guarantee better patient care and set new impulses in medical technology.

Over the last decade, a large number of solar power plants have been installed in Germany. To ensure a high performance, it is necessary to detect defects early. Therefore, it is required to control the quality of the solar cells during the production process, as well as to monitor the installed modules. Since manual inspections are expensive, a large degree of automation is required.
This project aims to develop a new approach to automatically detect and classify defects on solar power cells and to estimate their impact on the performance. Further, big data methods will be applied to identify circumstances that increase the probability of a cell to become defect. As a result, it will be possible to reject cells in the production that have a high likelihood to become defect.

Research project in cooperation with Siemens Healthineers, Erlangen

Magnetic Resonance Imaging (MRI) is an important but complex imaging modality in current radiology. Artificial intelligence (AI) can play an important role for acclerating MR sequence acquisition as well as supporting image interpretation and diagnosis. However, there are also opportunities besides image acquisition and interpretation for which AI can play a vital role to optimze the clinical workflow and decrease costs. 

Automated Protocoling

One critical workflow step for an MRI exam is protocoling, i.e., selecting an adequate imaging protocol under consideration of the ordered procedure, clinical indication, and medical history. Due to the complexity of MRI exams and the heterogeneity of MR protocols, this is a nontrivial task. The aim of this project is to analyze and quantify challenges complicating a robust approach for automated protocoling, and propose solutions to these challenges.

Automated Billing

Moreover, reporting and documentation is a crucial step in the radiology workflow. We have therefore automated the selection of billing codes from modality log data for an MRI exam. Integrated into the clinical environment, this work has the potential to free the technologist from a non-value adding administrative task, enhance the MRI workflow, and prevent coding errors.

Research project in cooperation with Siemens Healthineers, Erlangen

A Magnetic Resonance Imaging (MRI) exam typically consists of several MR pulse sequences that yield different image contrasts. Each pulse sequence is parameterized through multiple acquisition parameters that influence MR image contrast, signal-to-noise ratio, acquisition time, and/or resolution.

Depending on the clinical indication, different contrasts are required by the radiologist to make a reliable diagnosis. This complexity leads to high variations of sequence parameterizations across different sites and scanners, impacting MR protocoling, AI training, and image acquisition.

MR Image Synthesis

The aim of this project is to develop a deep learning-based approach to generate synthetic MR images conditioned on various acquisition parameters (repetition time, echo time, image orientation). This work can support radiologists and technologists during the parameterization of MR sequences by previewing the yielded MR contrast, can serve as a valuable tool for radiology training, and can be used for customized data generation to support AI training.

MR Image-to-Image Translations

As MR acquisition time is expensive, and re-scans due to motion corruption or a premature scan end for claustrophobic patients may be necessary, a method to synthesize missing or corrupted MR image contrasts from existing MR images is required. Thus, this project aims to develop an MR contrast-aware image-to-image translation method, enabling us to synthesize missing or corrupted MR images with adjustable image contrast. Additionally, it can be used as an advanced data augmentation technique to synthesize different contrasts for the training of AI applications in MRI.

MASCARA zielt auf eine detaillierte, molekulare Charakterisierung der Remission bei Arthritis ab. Das Projekt basiert auf der kombinierten klinischen und technischen Erfahrung von Rheumatologen, Radiologen, Medizinphysikern, Nuklearmedizinern, Gastroenterologen, grundlagenwissenschaftlichen Biologen und Informatikern und verbindet fünf akademische Fachzentren in Deutschland. Das Projekt adressiert 1) den Umstand der zunehmenden Zahl von Arthritis Patienten in Remission, 2) die Herausforderungen, eine effektive Unterdrückung der Entzündung von einer Heilung zu unterscheiden und 3) das begrenzte Wissen über die Gewebeveränderungen in den Gelenken von Patienten mit Arthritis. MASCARA wird auf der Grundlage vorläufiger Daten vier wichtige mechanistische Bereiche (immunstoffwechselbedingte Veränderungen, mesenchymale Gewebereaktionen, residente Immunzellen und Schutzfunktion des Darms) untersuchen, die gemeinsam den molekularen Zustand der Remission bestimmen. Das Projekt zielt auf die Sammlung von Synovialbiopsien und die anschließende Gewebeanalyse bei Patienten mit aktiver Arthritis und Patienten in Remission ab. Die Gewebeanalysen umfassen (Einzelzell)-mRNA-Sequenzierung, Massenzytometrie sowie die Messung von Immunmetaboliten und werden durch molekulare Bildgebungsverfahren wie CEST-MRT und FAPI-PET ergänzt. Sämtliche Daten, die in dem Vorhaben generiert werden, werden in einem bereits bestehenden Datenbanksystem mit den Daten der anderen Partner zusammengeführt und gespeichert. Das Zusammenführen der Daten soll – mit Hilfe von maschinellem Lernen – krankheitsspezifische und mit der Krankheitsaktivität verbundene Mustermatrizen identifizieren.

Our senses are gateways to the past. Although museums are slowly discovering the power of multi-sensory presentations, we lack the scientific standards, tools and data to identify, consolidate, and promote the wide-ranging role of scents and smelling in our cultural heritage. In recent years, European cultural heritage institutions have invested heavily in large-scale digitization. A wealth of object, text and image data that can be analysed using computer science techniques now exists. However, the potential olfactory descriptions, experiences, and memories that they contain remain unexplored. We recognize this as both a challenge and an opportunity. Odeuropa will apply state-of-the-art AI techniques to text and image datasets that span four centuries of European history. It will identify the vocabularies, spaces, events, practices, and emotions associated with smells and smelling. The project will curate this multi-modal information, following semantic web standards, and store the enriched data in a ‘European Olfactory Knowledge Graph’ (EOKG). We will use this data to identify ‘storylines’, informed by cultural history and heritage research, and share these with different audiences in different formats: through demonstrators, an online catalogue, toolkits and training documentation describing best-practices in olfactory museology. New, evidence-based methodologies will quantify the impact of multisensory visitor engagement. This data will support the implementation of policy recommendations for recognising, promoting, presenting and digitally preserving olfactory heritage. These activities will realize Odeuropa’s main goal: to show that smells and smelling are important and viable means for consolidating and promoting Europe’s tangible and intangible cultural heritage.

                                  During the last decade, dual energy CT (DECT) became widely available in clinical routine. Offered by all major CT manufacturers, with differing hard- and software concepts, the DECT applications are alike: the acquired DECT data are used to conduct two- or multi material decompositions, e.g. to separate iodine and bone from soft tissue or to quantify contrast agent and fat, to classify or characterize tissue, or to increase contrasts (CNR maximization), or to suppress contrasts (artifact reduction). The applications are designed to work with certain organs and the user needs to take care to invoke the correct application and to interpret its output only in the appropriate organ or anatomical region. E.g. interpreting the output of a kidney stone applications in organs other than the kidney will yield a wrong classification. To obtain quantitative results the applications require to set patient-specific parameters. In order to calibrate those the user is asked to place ROIs in predefined anatomical regions. Since this is time-consuming users are often tempted to use the default settings instead of optimizing them. Here, we want to develop a DECT atlas to utilize its anatomical (and functional) information for con-text-sensitive DECT imaging and material decomposition, and to be able to automatically calibrate the open parameters without the need of user interaction. To improve quantification the initial images shall not be reconstructed separately but rather undergo a rawdata-based decomposition before being converted into image domain. A dedicated user interface shall be developed to provide the new context-sensitive DECT information - such as automatically decomposing each organ into different but reasonable basis materials, for example - and to display it to the reader in a convenient way. Similarly, user-placed ROIs shall trigger a context-sensitive statistical evaluation of the ROI's contents and provide it to the user. This will help to quantify the iodine uptake in a tumor or a lesion, to separate it from fat or calcium components, to estimate its blood supply etc. Since the DECT data display the contrast uptake just for a given instance in time and since this contrast depends on patient-specific factors such as the cardiac output, we are planning to normalize the contrast uptake with the help of the dual energy information contained in the atlas. This will minimize inter and intra patient effects and increase the reproducibility. In addition, organ-specific material scores shall be developed that quantify a patient's material composition on an organ by organ basis. The new methods (DECT atlas, material decomposition, ...) shall be tested and evaluated using phantom and patient studies, and shall be optimized accordingly.                             

X02 nutzt die in EBM erzeugten Bilddaten und mechanischen Messungen, um Deep Learning-Methoden zu entwickeln, die Wissen über Spezies hinweg transferieren. In silico und in vitro Analysen werden deutlich spezifischere Daten liefern als in vivo Experimente, insbesondere für menschliches Gewebe. Um hier Erkenntnisse aus datenreichen Experimenten zu nutzen, werden wir Transfer Learning-Algorithmen für heterogene Daten entwickeln. So kann maschinelles Lernen auch unter stark datenlimitierten Bedingungen nutzbar gemacht werden. Ziel ist es, ein holistisches Verständnis von Bilddaten und mechanischen Messungen über Artgrenzen hinweg zu ermöglichen.

The project aims to develop novel and innovative methods to improve visualisation and use of dual energy CT (DECT) images. Compared to conventional single energy CT (SECT) scans, DECT contains a significant amount of additional quantitative information that enables tissue characterization far beyond what is possible with SECT, including material decomposition for quantification and labelling of specific materials within tissues, creation of reconstructions at different predicted energy levels, and quantitative spectral tissue characterization for tissue analysis. However, despite the many potential advantages of DECT, applications remain limited and in specizlized clinical settings. Some reasons are that many applications are specific for the organ under investigation, require additional, manual processing or calibration, and not easily manipulated using standard interactive contrast visualisation windows available in clinical viewing stations. This is a significant disadvantage compared to conventional SECT.
In this project, we propose to develop new strategies to fuse and display the additional DECT information on a single contrast scale such that it can be visualised with the same interactive tools that radiologists are used to in their clinical routine. We will investigate non-linear manifold learning techniques like Laplacian Eigenmaps and the Sammon Mapping. Both allow extension using AI-based techniques like the newly developed user loss that allows to integrate user's opinions using forced choice experiments. This will allow a novel image contrast that will be compatible with interactive window and level functions that are rourintely used by radiologists. Furthermore, we aim at additional developments that will use deep neural networks to approximate the non-linear mapping function and to generate reconstructions that capture and display tissue specific spectral characteristics in a readily and universally useable manner for enhancing diagnostic value.