Doctoral researcher: Alexander Jaus

Institution: KIT

Data science PI: Rainer Stiefelhagen

Life science PI: Jens Kleesiek

While there has been impressive progress in the field of biomedical image segmentation, current approaches hardly incorporate anatomical knowledge or common sense. Whereas radiologist have an explicit knowledge about the human body, this knowledge is only implicitly learned by current models which mostly focus on specific regions of the human body.

To alleviate this problem, we aim to develop a dataset spanning multiple modalities for the entire human body, allowing models to learn the entire anatomy. Furthermore, we will leverage a dataset consisting of radiological reports which are weak labels on how the described anatomy matches or differs from a certain norm. Combining these two approaches, we will train a model which incorporates prior knowledge of the human anatomies and pathologies.

While the described model offers many possible downstream tasks, one focus applications within this thesis will be to develop a framework which makes anatomical databases searchable beyond standard queries. By indexing the databases with the help of the developed model, fine grained semantic searches such as “retrieve all CT scans with left lungs with a diameter of X with pulmonary embolism” should be possible.

Doctoral researcher: Luca Deininger

Institution: KIT

Data science PI: Ralf Mikut

Life science PI: Sabine Jung-Klawitter

Organoids are self-assembled three-dimensional aggregates generated from human pluripotent stem cells (hPSC) with cell types and cytoarchitectures that resemble human organs and tissues. Cerebral organoids are a growing field of interest to understand and model the development of the human brain for health and disease, in particular common and rare inherited diseases. Data of cerebral organoids can include heterogeneous data sources, such as longitudinal Magnetic Resonance Imaging (MRI), immunofluorescence analysis of organoids as well as metabolomic and transcriptomic data at bulk (whole organoids) or single cell level. A main bottleneck of organoid research is the lack of systematic analysis pipelines covering the full range of acquired data. The aim of the proposed project is to establish such a standard analysis pipeline for cerebral organoids and to show the potential of this pipeline for the characterization of human diseases, especially for inborn errors of neurotransmitter metabolism.

Doctoral researcher: Nico Disch

Institution: DKFZ

Data science PI: Klaus Maier-Hein

Life science PI: Jens Kleesiek, Rainer Stiefelhagen

Current image analysis of patient data only uses single images, with previous measurements not being incorporated into the model. Furthermore, image time series in other domains rely on uniformly sampled times, such as videos, in order to find anomalies or to predict time series progression. Other anomaly detection methods learn the distribution of images, however distribution shifts are difficult to encompass. Therefore a model has to be developed that is able to deal with sparse image time series and find anomalies in these time series. The developed models will help to distinguish regular and anomalous progression in patients’ data.

These insights might lead to a better understanding in the human process of aging and disease progression, and possibly aid the decision making of medical practitioners.

Doctoral researcher: Zdravko Marinov

Institution: KIT

Data science PI: Rainer Stiefelhagen

Life science PI: Jens Kleesiek

Annotated medical data is a prerequisite for successful and robust deep-learning models. However, the curation of labels for medical images is a difficult challenge due to the complexity of the data and the need of expert knowledge. Interactive annotation alleviates the burden of manual labeling by reducing the annotation time significantly and allowing the annotators to iteratively refine the labels until they are satisfied with their quality. This project aims to explore how to integrate interactive segmentation models for personalized diagnostics on multimodal medical data, e.g., PET-CT, using active learning and sparse labels. Medical reports will also be integrated to ease and enhance the annotation process and improve the segmentation by combining textual and visual information.

The simulation of user interactions during training and suggesting slices for annotation with active learning are open challenges for interactive segmentation. Typically, user interactions, e.g., in the form of clicks, are simulated in locations where the model has produced segmentation errors. However, in multimodal imaging, the sources of error are two-fold, and errors from one modality can be used for co-training of the other one. Thus, this project aims to explore various multimodal fusion paradigms to leverage the mutual information between the modalities and investigate ways to simulate interactions for multimodal imaging.

Doctoral researcher: Piotr Kalinowski

Institution: DKFZ

Data science PI: Lena Maier-Hein

Life science PI: Hannes Kenngott

Death within 30 days after surgery has recently been found to be the third-leading cause of death worldwide [1], with research suggesting that a large proportion of these deaths are due to surgical error. The field of Surgical Data Science [2] aims to address this issue with data-driven methods. However, a large international consortium of experts [3] revealed a lack of clinical success stories and attributed this issue to the lack of large, annotated databases. In this project, we investigate an entirely new approach to address this roadblock. Specifically, we propose the encoding of existing medical knowledge in ontologies and integrating this prior knowledge in a Graph Neural Network (GNN) based approach to surgical decision support. The method will be validated based on existing data sets from different surgical disciplines.

References
[1] D. Nepogodiev, J. Martin, B. Biccard, A. Makupe, A. Bhangu, and National Institute for Health Research Global Health Research Unit on Global Surgery, “Global burden of postoperative death,” The Lancet, vol. 393, no. 10170. p. 401, Feb. 02, 2019.
[2] L. Maier-Hein et al., “Surgical data science for next-generation interventions,” Nat Biomed Eng, vol. 1, no. 9, pp. 691–696, Sep. 2017.
[3] L. Maier-Hein et al. “Surgical data science - from concepts toward clinical translation.” Medical image analysis vol. 76 (2022).

Doctoral researcher: Sebastian Pirmann

Institution: DKFZ

Data science PI: Benedikt Brors

Life science PI: Daniel Hübschmann

Pharmacogenomics (PGx) studies how variations in the genome affect drug response in patients. There are genomic variants in so-called pharmacogenes which have an impact on a drug’s pharmacokinetics, influencing drug response and side effects. Although this link is well known, it is still not clinical practice to comprehensively determine a patient’s PGx profile prior to drug treatment. Especially in cancer therapy, PGx profiling could increase treatment success and reduce potential side effects.

In this project, PGx analysis is done retrospectively for large cancer patient cohorts by using computational methods for genotyping from next-generation-sequencing data. This will allow to create distributions of PGx variants, genotypes, and phenotypes as well as to decipher the link to the observed treatment outcomes. In addition, sequencing data of matched tumor samples will be used to capture PGx differences acquired by e.g. somatic resistance mutations associated with metabolizer phenotype changes. Other somatic effects contributing to the development of resistance, such as allele-specific expression or epigenetic changes of pharmacogenes, will also be assessed. Finally, this project aims to develop supervised models for prediction of treatment outcome and side effects based on the aforementioned PGx data.

Doctoral researcher: Sonja Adomeit

Institution: UNI HD

Data science PI: Stefan Riezler

Life science PI: Daniel Durstewitz

In order to build machine learning models for predictive healthcare, the standard approach is to learn a unified prediction model on data pooled from several patients. Instead, we propose to learn predictive models on data labeled by experts for individual patients and combine these patient-specific models by ensembling techniques. However, working with individual patient information, poses the need for improved privacy protection of the training data. The goal of the proposed project is to find the optimal trade-off between accuracy and privacy both from a theoretical perspective, and in experimental applications for predictive models for sepsis and diagnosis tasks from psychiatry.

Doctoral researcher: Daniel Wolffram

Institution: KIT

Data science PI: Melanie Schienle

Life science PI: Johannes Bracher

During the ongoing COVID-19 pandemic, probabilistic forecasts of infectious disease spread have become a research priority, because they allow for a nuanced assessment of an uncertain future and can guide policymakers to informed decisions. Several major collaborative forecasting projects have been launched to systematically compare, evaluate and aggregate forecasts from different models. To make forecasts comparable, they are stored in a standardized quantile format. This approach raises numerous statistical challenges, which we will address in our project.

Statistically sound evaluation is a prerequisite for model improvement. However, many existing methods are not suitable for quantile forecasts. We aim to develop novel visual tools and statistical tests to assess forecast calibration and to make score-based evaluations more interpretable.

To improve forecast performance and reliability, combining forecasts into ensembles has shown promise in disease forecasting and is already common practice in fields like weather forecasting. We will study aggregation methods for quantile forecasts, which will involve state-of-the-art quantile regression methods, non-parametric approaches like isotonic distributional regression, and recent machine learning approaches. Particular focus will be given to common characteristics of disease spread like geographical heterogeneities and regime shifts, which could potentially be leveraged by meta-learning approaches.

Doctoral researcher: Julia Münch

Institution: DKFZ

Data science PI: Benedikt Brors

Life science PI: Anne-Kristin Kaster

During the last decade, single-cell omics technologies such as single-cell transcriptomics (SCT) have rapidly emerged. As a complement to cultivation-based meta-transcriptomic approaches, SCT allows direct access to information on gene expression from individual cells rather than bulk samples, and thus tremendously improves our understanding of numerous biological processes.

While single-cell technologies are mainly applied to and consequently vastly advanced for eukaryotes, adapting these methods to prokaryotic organisms has been proven challenging for several reasons. First, the stiff but yet highly diverse cell wall architecture among bacterial species hampers effective permeabilization and lysis. Furthermore, the low RNA amounts are insufficient to be directly analyzed via current sequencing technologies. Hence, most protocols comprise an essential amplification step before sequencing. However, this step might introduce a bias to the data set veiling the genuine expression profiles by incomplete or uneven amplification, chimera formation, and biases against templates with high GC content. Although various experimental techniques have been developed to deal with these issues, no systematic approach has been taken to evaluate this bias. Nonetheless, studying the microbial world at single-cell level is of fundamental importance, particularly regarding the enormous potential of (yet unknown) microbial species for biotechnological applications, but also in the context of infectious diseases and cancer progression and therapy.

Therefore, this project aims at addressing the issues of amplification bias in SCT. To do so, prokaryotic single-cell RNA-seq data will be generated and used to train suitable regression models. By using modern methods from machine learning, technical noise and biological variation will systematically be discriminated to eventually correct for the bias.

Doctoral researcher: Tim Ortkamp

Institution: KIT

Data science PI: Martin Frank

Life science PI: Oliver Jäkel

Half of all cancer patients receive radiation therapy which delivers high-energetic, ionizing radiation to target cancerous tissue while sparing healthy tissue. The planning of such a radiotherapeutic intervention is highly personalized due to the application of medical imaging techniques like computed tomography or magnetic resonance imaging followed by simulation and optimization of the radiation dosage. In the classical approach to treatment planning, the underlying trade-off between tumor control probability (TCP) and normal tissue complication probability (NTCP) is usually taken into account only in an indirect manner by using surrogate planning objectives depending on empirical dose prescriptions and tolerances, instead of performing a „biological optimization“ based on these two quantities. However, in the current literature there are technically feasible approaches to integrate TCP and NTCP directly into the plan optimization process. Unfortunately, these approaches rely on simple analytical models with low-dimensional parameter spaces, e.g. the logistic LKB model, and are therefore consistently questioned with respect to accuracy and usability. Therefore, the proposed project aims at superseding these approaches by integrating state-of-theart Machine Learning (ML) models for TCP and NTCP into the radiotherapy treatment plan optimization process, which poses multiple mathematical, computational, and clinical challenges. In summary, the following project objectives are set:

• Investigation and definition of ML-based model stability within the dose optimization environment given untypical input during optimization iterations
• Development of computationally efficient optimization objective and constraint functions based on ML models, especially with respect to differentiability and respective efficient gradient computations
• Interpretation of generated treatment plans and definition of clinically acceptable optimization strategies

Additional output is planned to be a (partly open-source) software module for further research and pre-clinical testing.

Doctoral researcher: Simon Warsinsky

Institution: KIT

Data science PI: Ali Sunyaev

Life science PI: Martin Wagner

Machine Learning (ML) models are increasingly diffusing in healthcare. ML models can, for example, come into play in cognitive surgical robots by helping the robot to understand the surgery context, comprehend the surgery procedure and eventually generate safe trajectories to assist during the surgery. To train ML models sophisticated enough to assist in surgery, large amounts of annotated image data are required. Due to expertise often being necessary to interpret medical images (e.g., surgery images, MRI images), medical image annotation is often manually conducted by healthcare professionals. Manual annotation is prone to human errors since it can be tedious, monotonous, and exhausting. Accordingly, poor label quality is a common problem. However, for surgical robots to improve surgical procedures, sufficient data quality of annotated images is a decisive factor. If ML models for surgical robots are trained based on poorly labeled image data, this may negatively influence patients’ health since the robots cannot be utilized to their full potential. In this project, we tackle the problem of poor label quality in medical image annotation through gamification (i.e., the use of game design elements in non-game contexts). In this project, we address the problem of poor label quality of surgical image data by augmenting the annotation process with gamification. The overarching objective of the project is to design, implement, and evaluate a data-driven ML-based gamification concept to augment the annotation process, foster annotators’ engagement and, thereby, ensure high quality data labeling. By drawing on ML methods, the gamification concept is able to adapt to individual user preferences and overcome the weaknesses of one-size-fits-all gamification approaches.

Doctoral researcher: Abdul Moeed

Institution: DKFZ

Data science PI: Oliver Stegle

Life science PI: Pavlo Lutsik

The goal is to develop machine learning strategies for deciphering cell-of-origin related cancer heterogeneity. Building on variational autoencoders and related dimensionality reduction methods, we will derive generative representations of population-scale single cell data from thousands of individuals. Leveraging these trained models, we will then interpret disease sample in a reference latent space, thereby enabling patient/cancer specific cell-of-origin identification at unpresented resolution. We will apply these methods to blood cancer sand specifically Acute Myeloid Leukemia (AML) and Chronic Lymphocytic Leukemia (CLL), for which we have access to pertinent data resources. The ultimate goal is the generation of model-driven insights for clinical guidance, enabling early detection, as well as patient-specific strategies for treatment and disease management.

Doctoral researcher: Marlen Neubert

Institution: KIT

Data science PI: Pascal Friederich

Life science PI: Frauke Gräter

Collagen is the most abundant protein in our body and performs a variety of functions, including strengthening and supporting skin, tendons, and bone tissue. Recent research of the Gräter research group revealed that collagen produces highly reactive radicals when subjected to mechanical stress, which can damage surrounding tissue. At the same time, it is shown that collagen can mitigate radical damage. In light of these findings, this project aims to investigate the relation between collagen-related diseases and defenses against radical damage, as there are more than 1000 known genetic collagen mutations. The influence of these mutations and their consequences on the function of collagen on a molecular level are not well understood. Since microscopes cannot be used to observe interactions on an atomic scale, computer simulations are more suited in helping to understand these dynamic molecular processes. In order to simulate changes of a molecular system over time, an accurate description of interatomic energies and forces is needed. Quantum mechanical energy calculations are complex for large systems such as collagen fibrils since their computational cost scales roughly cubic with the system size. Machine learned potentials are a data-driven energy calculation method which maps the atomic coordinates to the potential energy of the system using a machine learning model. These models are computationally less expensive than quantum mechanical calculations, but can reach similar accuracies, thus offering the ability to speed up large scale molecular dynamic simulations.

Doctoral researcher: Marcel Meyer

Institution: UNI HD

Data science PI: Ullrich Köthe

Life science PI: Rebecca Wade

Protein-biomolecule-interactions are ubiquitous in life, from DNA-replication to regulating the heartbeat. The stability of biomolecular complexes depend heavily on the involved proteins’ amino acid sequence. Changes in this sequence and their effects on complex formation are of interest in many diseases, but also in protein design. However, predicting how two molecules will interact remains a challenge due to the high degrees of freedom both within and between the interaction partners. In the recent past, deep neural networks have proved useful for structure prediction of biomolecules. We investigate the suitability of invertible neural networks (INNs) for predicting the structures of molecular complexes, as well as the effects of mutations in sequence. INNs are interesting candidates because they model the full probability distribution of molecular interactions. Also, they do not suffer from the curse of high dimensionality and can therefore be scaled to very large problems like protein-protein-interactions. Major tasks at hand are finding the right representation of the molecular structure and building suitable network architectures. Once we can reliably predict molecular interactions we will integrate sequence information to learn how sequence alterations influence the structure of biomolecular complexes.

Doctoral researcher: Stefan Haller

Institution: UNI HD

Data science PI: Carsten Rother

Life science PI: Gerd Ulrich Nienhaus

Cell segmentation and tracking is the problem of processing a time series of (3D) images showing development of an organism (e.g. drosophila) on the cellular level, that is, growth, movement, division and death of cells. Existing methods for this problem work reasonably well in simple cases with relatively few cells per time frame and relatively small temporal changes between them. The goal of the project is to develop a new cell-tracking method which is able to cope with the problem at later development stages, where existing approaches hit their limit. The new key components of the method are: (a) a scalable combinatorial solver which is able to efficiently deal with millions of variables and constraints,as cell-tracking can be seen as a large-scale combinatorial problem; and (b) a training technique to learn the parameters of the solver to fine-tune it to different types of input data.

Doctoral researcher: Verena Bitto

Institution: DKFZ

Data science PI: Benedikt Brors

Life science PI: Michael Baumann

Tumour hypoxia, a state of low oxygen levels in certain tissue regions, seems to play a prognostic role for loco-regional tumour control. Hypoxic cells are associated with resistance to radiotherapy, which allows them to survive standard treatment. Both phenomena, tumour hypoxia and radioresistance, however, are highly heterogeneous in between patients and within one tumour.Therefore, the aim of the project is to study both, tumour hypoxia and radioresistance, in more detail. We plan to integrate different omics data in order to find regulatory pathways and combine them with results from analysis of MRI scans.

Doctoral researcher: Philipp Wimmer

Institution: UNI HD

Data science PI: Filip Sadlo

Life science PI: Daniel Durstewitz

Since reliable biomarkers for psychiatric diagnosis and treatment indication are lacking, such diagnostics and prognostics is mainly based on structural interviews with the patient. Although abnormalities in non-invasive measurements, including structural and functional magnetic resonance imaging (fMRI), are associated with psychiatric conditions, the effects are often small and very heterogeneous, and thus do not directly allow for medically reliable classification of psychiatric conditions or subtypes. The group of PI Durstewitz has suggested that derailed cortical attractor dynamics, rooted in known biophysical (synaptic) alterations, could be a hallmark feature of many psychiatric illnesses. Identification of attractor dynamics in brain recordings is, however, a highly challenging topic, and only recently progress in statistical machine and deep learning has enabled to extract crucial dynamical systems features from neural, including fMRI, recordings. Major challenges for translating these methods and observations into clinical practice include their extension to additional data sources for extended robustness, effective and efficient analysis of the resulting very highdimensional state spaces, identification and extraction of those dynamical features most crucial to psychiatric diagnosis and treatment, and, last but not least, presentation of the results, together with context information, in a format that is accessible to clinicians. The goals of this project is to develop advanced visual data analysis techniques addressing these challenges. On the one hand, we develop techniques that reveal topological features, e.g., invariant manifolds in the high-dimensional phase space of the inferred dynamical systems, support navigation of these spaces, and bring them into context with additional data sources. On the other hand, we include structural magnetic resonance imaging data, develop respective local feature definitions, and investigate the integration of these spatial data with the high-dimensional phase space of the inferred dynamical systems. Last but not least, we combine the obtained techniques into an overall approach that enables effective application and interpretation by clinicians.

Doctoral researcher: Alexandra Walter

Institution: KIT / DKFZ

Data science PI: Martin Frank

Life science PI: Oliver Jäkel

The precise spatial delineation of cancerous and healthy tissue in radiation therapy is necessary to prevent side effects and the reoccurrence of the tumor. The sides, sizes and shapes of tumors vary widely, which make the identification of the clinical target volume (CTV) difficult for humans as well as for machine learning algorithms. Expert guidelines were formulated to establish a best in class delineation standard. The goal of this project is to establish a reliable, fully automated pipeline for CTV delineation by combining the achievements of convolutional neuronal networks on medical image segmentation tasks with human expert guidelines.

First, state-of-the-art ML solutions will set a baseline for the delineation accuracy on the CT scans of the investigated head and neck cancer cohort. A divide-and-conquer strategy will be implemented to realize expert guidelines as single constraints. Finally, these constraints will be built into the learning routines by either translating them into the target function or the networks architecture.

Doctoral researcher: Paul Maria Scheikl

Institution: KIT

Data science PI: Franziska Mathis-Ullrich

Life science PI: Martin Wagner

Laparoscopic surgery is a team effort. A surgeon and her assistant(s) collaborate to solve a shared task working individually and as a team for a successful surgical outcome. However, our society faces increasing shortage of skilled surgeons and assistants, especially in rural areas. This shortage may be resolved by providing cognitive surgical robots that automate certain tasks. Our project addresses the highly interdependent behavior of surgeon and assistant(s) as a multi-agent system problem of human and artificial agents using methods of cooperative multi-agent reinforcement learning (cMARL). In contrast to previous work, this project aims to train multiple, decentralized artificial agents that cooperatively solve a shared, robot-assisted laparoscopic task.

Doctoral researcher: Lucas-Raphael Müller

Institution: DKFZ

Data science PI: Lena Maier-Hein

Life science PI: Hannes Kenngott

Death within 30 days after surgery has recently been found to be the third-leading cause of death worldwide. In this context, anastomotic leakage (AL) has been recognised as one major cause for intraoperative or postoperative complication. Anastomoses in medicine refer to the surgical connection between two diverging tubular structures, such as blood vessels or intestines, while leaks may occur as late as days or even weeks postoperatively. Recent results suggest that the risk of anastomosis leakage varies crucially from hospital to hospital and can reach up to 16% (1). The long-term goal of this project is to develop the first fully-automatic video-based approach to anastomosis leakage prediction based on laparoscopic video and multi-modal sensor data recorded throughout the surgical intervention. The core methodological challenge is to develop a representation of the multi-modal data that generalises across specific acquisition hardware and settings and thus serves as a fundament for clinical translation and multi-centre application of our methodology.

Image: Edited, original by Mariana Ruiz et al., image in public domain, Wikipedia Digestive System
(1) Kingham, T. P. & Pachter, H. L. Colonic Anastomotic Leak: Risk Factors, Diagnosis, and Treatment. Journal of the American College of Surgeons 208, 269–278 (2009).

Doctoral researcher: Philipp Toussaint

Institution: KIT

Data science PI: Ali Sunyaev

Life science PI: Matthias Schlesner

As it is becoming progressively challenging to wholly analyse the ever-increasing amounts of generated biomedical data (e.g., CT scans, X-ray images, omics data) by means of conventional analysis techniques, researchers and practitioners are turning to artificial intelligence (AI) approaches (e.g., deep learning) to analyse their data. Although the application of AI to biomedical data in many cases promises to deliver improved performance and accuracy, extant AI approaches often suffer from opacity. Their sub-symbolic representation of state is often inaccessible and non-transparent to humans, thus limiting us in fully understanding and therefore trusting the produced outputs. Explainable AI (XAI) describes a recent trend in AI research with the aim of addressing the opacity issue of contemporary AI approaches by producing (more) interpretable AI models whilst maintaining high levels of performance and accuracy. The objective of the XAIOmics research project is to design, develop, and evaluation an XAI approach to biomedical (i.e., omics) data. In particular, we will identify biomedical use cases and current, viable approaches in the domain of XAI and apply and adapt them to the identified use cases. With regards to the highly interdisciplinary field, a central research hurdle will be the development of an understanding for the different kinds of biomedical data and the subsequent feature engineering in the context of the design of the AI algorithms. In doing so, this project will not only aid researchers and physicians in obtaining a better understanding of the outputs of contemporary AI approaches for biomedical data but also create more transparency, which will support the building of trust in AI-based treatment and diagnosis decisions in personalized medicine.

Doctoral researcher: Alejandra Jayme

Institution: UNI HD

Data science PI: Vincent Heuveline

Life science PI: Heinz-Peter Schlemmer

In recent years, biomedical data have been increasingly available. In particular, the costs for procuring omics data, including genome and exome sequences as well as protein and transcriptomicdata, have dramatically decreased. Together with histological images of tumors, these are important sources of high-dimensional data beneficial for cancer diagnostic procedures and treatment optimization.

We look to big data analysis to derive clinical knowledge on cancer evolution: identifying patterns and structures that signal critical events in cancer development to aid in developing adequate diagnosis, prevention and treatment for all kinds of tumors. This concept of data-driven modeling (DDM) unifies statistics, data analysis and machine learning in order to understand the actual phenomena with data. Moreover, we couple the model with uncertainty quantification (UQ) methods to address the impact of uncertainties inherent in data and mathematical models.

The aim of this project is to develop and analyze new and innovative computational approaches, combining DDM and UQ, for solving current challenges in the field of hereditary cancer research. The derived model should enhance existing diagnostics and prevention strategies and allow for better understanding of cancer evolution.

Doctoral researcher: Julian Herold

Institution: KIT

Data science PI: Achim Streit

Life science PI: Alexander Schug

A deep understanding of tissue growth as emerging behavior from single-cell events might lead to new insights for different scientific fields, ranging from fundamental biology to in-silico testing of treatment regimes for personalized medicine. We aim at simulating mm-sized virtual tissues such as embryogenetic brain tissue or cancerous tumors with more than a million µm-resolved individual cells. Such a model requires many parameters, which have to be adjusted accordingly in order reproduce experimental data. We deploy a combination of traditional optimization techniques and deep learning approaches in order to find a reliable and flexible way to adapt those parameters.

Doctoral researcher: Elaine Zaunseder

Institution: UNI HD

Data science PI: Vincent Heuveline

Life science PI: Stefan Kölker

The accurate diagnosis of patients with rare diseases is important but challenging, since the prevalence of them is very low. This is especially relevant for diseases where quick identification can lead to efficient therapies and treatments which can positively change the outcome and severity of the disease. In this respect, the analysis of patient-specific data from disease and control cohorts can support to improve the diagnosis. Thanks to advances in data mining and machine learning as well as the computing landscape in recent years, new opportunities to examine large data sets with high dimensional feature spaces have been developed. In particular, Knowledge Discovery in Databases (KDD) methods can be helpful to discover new clues for unknown causal relations as well as novel metabolic patterns. A classical approach is to consider classification schemes, which rely on mathematical models describing the underlying relationships. When developing these classification models real-world data has to be processed, so the medical input data and the mathematical models are subjected to measurement inaccuracies and uncertainties.

In our project, we want to analyze and develop innovative methods for mathematical uncertainty quantification (UQ) to describe and quantify noise in the data as well as to assess model inaccuracies, in order to obtain reliable classification results. In that context, it is important not only to ensure the reliability of the model, but also to make the model understandable and interpretable for practitioners. Specifically, we want to extend techniques of Explainable AI (XAI) in the context of disease diagnosis. The overall goal is to analyze and develop reliable and interpretable models with high diagnostic prediction to support the diagnosis of rare diseases.

Doctoral researcher: Constantin Seibold

Institution: KIT

Data science PI: Rainer Stiefelhagen

Life science PI: Heinz-Peter Schlemmer

While Deep Learning approaches, in particular Convolutional Neural Networks (CNNs), are now also being widely adopted in medical image analysis, only very few publications exist that deal with the integration of visual and textual features or the generation of explanations for findings in images. So far, none of the state-of-the-art image-captioning approaches have been applied to 3D medical images thereby utilizing the available medical reports as the sole training input. The goal of this project is investigate state-of-the-art machine learning methods in order to automatically generate clinically meaningful image descriptions (medical reports) for radiological CT images. The resulting model shall be able to describe radiological findings, eg. lesions in an image, their location and descriptive relevant attributes, such as they are usually used by doctors in medical reports.

Doctoral researcher: Jan Sellner

Institution: DKFZ

Data science PI: Lena Maier-Hein

Life science PI: Felix Nickel

Death within 30 days after surgery has recently been found to be the third-leading cause of death worldwide. One of the major challenges faced by the surgeons is the visual discrimination of tissues, for example to distinguish pathologies or critical structures from healthy remaining tissue. In this context, multispectral imaging has been proposed to overcome the arbitrary restriction of conventional camera systems (and the human eye) which capture only red, green and blue colors. Instead, the spectral dimension consists of multiple specific bands of light (e.g. 16 or 100). As different tissue types have different absorbance patterns, multispectral cameras provide additional information about the tissue and corresponding functional properties, such as oxygenation. In our research, we leverage this additional information for machine learning-based classification of tissue in real-world surgery settings. Compared to non-medical machine learning algorithms, however, the data available for training our models is comparably small (e.g. only hundreds of images instead of millions). Furthermore, due to the novelty of the technique, the available hardware undergoes rapid developments as well as application-specific adaptations (e.g. open surgery vs. minimally invasive surgery). In consequence, the available captured data sets feature large variations with respect to the number of bands, illumination conditions and other factors. The primary goal of this project is therefore to develop novel concepts for tissue classification that avoid losing valuable knowledge encoded in the data of one camera when applying a different camera. More specifically, we aim to develop a camera invariant tissue classification framework that leverages all the spectral imaging data available, irrespective of the number of bands and the specific camera configuration.

Doctoral researcher: Hamideh Hajiabadi

Institution: KIT

Data science PI: Anne Koziolek

Life science PI: Lennart Hilbert

Fluorescence microscopy has several inherent limitations, which are dictated by basic optical and physical laws as well as compromises arising from the fact that biological materials are being imaged. Recent publications demonstrated how machine learning approaches can substitute image information in situations where it is otherwise difficult or even fundamentally impossible to obtain directly (e.g., a neural network trained on separate recordings of samples prepared with different fluorescence markers that predicts “all” the markers for new recorded images). Another broader challenge that affects not just the microscopy community concerns the broad exploitation of cutting-edge algorithms. However, the understanding which structural, organizational, technical and design decisions result in effective exploitation of expert-developed microscopy software by non-expert users is only understood at an anecdotal level. Also from the software engineering perspective, situations where non-experts build new machine learning applications from existing algorithms is not yet well understood and supported. The goals of the project are to provide a machine learning-based image analysis tool that monitors more different nuclear components than possible with current fluorescence microscopes and to provide a software engineering method to support non-expert developers when transferring neural network-based solutions from one image analysis setting to another.

Doctoral researcher: Katharina Löffler

Institution: KIT

Data science PI: Ralf Mikut

Life science PI: Uwe Strähle

Many biomedical imaging devices are able to collect time series of high resolution 3D data (3D+t). A typical task in these datasets is the tracking of objects over time, e.g. cells or sub-cellular structures. An example for such imaging tasks is the analysis of light sheet microscopy data by imaging fluorescently labeled cells. Our initial focus will be on the embryonic development of zebrafish (Danio rerio) as a preferred model animal for vertebrates. Our investigations aim at quantitative descriptions of cell death, divisions, movements, and interactions. Usual methods for basic tasks like identifying cell nuclei in the images have accuracies around 95 %. This sounds good by the standards of conventional image processing. However, this implies that in every image the system will lose sight of so many nuclei that reconstructing the overall cellular lineage (nuclei movements and cell divisions) becomes an intractable jigsaw puzzle even for the simplest task of cell nucleus tracking.The aim of the project is a significant improvement of 3D+t tracking algorithms for the case of dense objects and limited imaging quality. Our initial goal in light-sheet microscopy is to follow the movement of more than 10 000 cells and then scale the solution to hundreds of experiments.

Doctoral researcher: Vojtech Kumpost

Institution: KIT

Data science PI: Ralf Mikut

Compared to man-made technical control systems, biological control systems exhibit a remarkable fault tolerance over wide ranges of environmental cues as well as in face of variability at the level of the system-encoding genes. One of the striking features of the biological control systems is a high level of regulatory redundancy, however, it is still unclear how such features might specifically contribute to the wide-ranging fault tolerance of biological control systems. This project aims to investigate whether this fault tolerance of biological control systems is rooted in design principles that differ from man-made control systems. In particular, we investigate how far and what kind of redundancy contributes to the wide range of genetic and environmental variability that biological control systems can cope with.

Doctoral researcher: Vahdaneh Kiani

Institution: UNI HD

Data science PI: Holger Fröning

Life science PI: Oliver Jäkel

The image shows the challenge of the project in my view: the first part shows a volume rendering of a lung with a tumour inside, the arrows and the red area indicate the motion of the tumour with the respiration. In RT (radiation therapy), we need to track this motion to be able to adapt the irradiation. The second image shows the three components of image registration process (a deformable transformation in the grid, the similarity measure in the histogram, and the optimizer strategy this 2D lines and the sampling of a path within this objective function). Even if everything is running on a one (blue) GPU it's not real-time. The aim of the project is to find ways to redistribute the calculation load to multiple (green) GPUs.

Doctoral researcher: David Zimmerer

Institution: DKFZ

Data science PI: Klaus Maier-Hein

Life science PI: Heinz-Peter Schlemmer

An assumption-free automatic check of medical images for potentially overseen anomalies would be a valuable assistance for a radiologist. Deep learning and especially generative models such as Variational Auto-Encoders (VAEs) have shown great potential in the unsupervised learning of data distributions. By decoupling abnormality detection from reference annotations, these approaches are completely independent of human input and can therefore be applied to any medical condition or image modality. In principle, this allows for an abnormality check and even the localization of parts in the image that are most suspicious.

Doctoral researcher: Silvia Seidlitz

Institution: DKFZ

Data science PI: Lena Maier-Hein

Life science PI: Hannes Kenngott

Machine learning-based decision support can potentially improve the quality of healthcare by providing physicians with the right information at the right time. An important prerequisite for context-awareness and autonomous robotics in surgery is the fully-automated real-time analysis of medical images. Relevant tasks include semantic segmentation for surgical scene understanding as well as the automated image-based classification of medical conditions, such as sepsis. Relying on conventional RGB imaging, however, is problematic, as important morphological and functional information is invisible on such widely acquired surgical images. Spectral imaging which unlike conventional RGB imaging and the human eye, captures multiple specific bands of light instead of being restricted to the red, green, and blue bands, has the potential to improve tissue classification. Due to the high dimensionality of spectral data, interpretation by humans is infeasible. While machine learning has the potential of addressing this bottleneck, several open research questions remain.

Spectral data representation: It has so far not been determined what constitutes an adequate representation of spectral data for machine learning-based fully automated tissue classification, especially with respect to the spatial granularity of the data (pixels vs. superpixels vs. patches vs. full images) and the processing of the spectral data. A primary goal of this thesis, is therefore to close this gap.

Confounders in spectral image classification: While machine learning-based classification has generated several clinical success stories during the past years, recent research also revealed a large number of algorithms for which performance is overestimated due to biases. To date, however, this important issue has been completely ignored in the context of medical spectral imaging. The second major goal of this thesis, is therefore, to automatically identify confounders in medical spectral data and to achieve a bias-free representation of spectral data for machine learning-based tissue classification.

Based on the methodological contributions related to the handling of spectral data in neural networks and the bias awareness, the project will determine, whether automated sepsis diagnosis is feasible with medical spectral data and whether there is a benefit of using spectral data compared to other modalities (e.g. RGB imaging) for automated organ segmentation.

Doctoral researcher: Pia Stammer

Institution: KIT

Data science PI: Martin Frank

Life science PI: Oliver Jäkel

Radiation therapy is one of the cornerstones in modern cancer treatment being applied in 50 % of all patients. It is a highly personalized intervention relying on a computational patient model (provided by patient-specific computed tomography and/or magnetic resonance imaging) and, based on this image data, a quantitative, spatial simulation of the radiation dose delivered into the patient’s body. This digitalization enables a computer-based optimization of every individual radiation treatment before the onset of therapy. In clinical practice, however, treatment optimization only considers one unique “best guess” treatment scenario, neglecting a number of sources of uncertainty along the workflow. Sources of uncertainty in radiation therapy are myriad: Physical limitations to patient immobilization cause misalignment relative to the treatment beam, changes in the patient’s anatomy both intrafractional (i.e. during the treatment session, for example due to breathing) and interfractional (in between treatment sessions, for example due to weight loss) compromise the validity of the original CT. Measurement errors during CT acquisition cause uncertainty about radiological depths. We want to build an efficient computational pipeline for uncertainty management in ion beam therapy based on Monte Carlo dose calculation algorithms. Therefore, we want to reduce the arithmetic load of the involved sampling processes by orders of magnitude leveraging state-of-the-art methodology from the mathematical uncertainty quantification community. Specifically, we pursue the following two objectives:

1. We want to enable efficient and highly accurate dosimetric uncertainty quantification considering range, setup, and motion uncertainties during forward dose calculation.
2. We want to use information about treatment uncertainties to systematically minimize their potential negative impact onto treatment outcome.

Doctoral researcher: Patrick Godau

Institution: DKFZ

Data science PI: Lena Maier-Hein

Current Machine Learning algorithms, especially Deep Neural Networks (DNNs), have shown to be successful tools in areas particularly relevant for Life Science, for example image analysis. However they feature some drawbacks preventing the transition from research to real applications. First and foremost these algorithms rely heavily on huge amounts of data to train on, which is often not given in the medical context because of data protection and privacy policy. Missing this data, the full potential of the used DNN is not accomplished and results of the same DNN architecture appear poor in contrast to non-medical databases. Second and closely connected to the lack of data is the missing ability to generalize learned concepts. Algorithms trained on outputs produced by a specific medical device in a fixed location often suffer from significant performance loss when using a device from another vendor or switching the hospital/surgeon/procedure. Well-known techniques to overcome missing training data come from Transfer Learning: One implementation of it involves a pre-trained DNN and focuses mainly on slight adjustments by removing final layers and retraining them on the specific medical data. Although this method provides noticeable improvements in some cases there is often no clear methodology other than trial and error for choosing the optimal pre-training data, the correct DNN architecture or suitable re-training hyper-parameters. To better address these shortcomings, we propose to use ideas from Meta Learning and implement a lifelong learning algorithm in an expandable framework of available data sets. The objective of the project is a meta-algorithm that for a given medical task chooses the optimal learning conditions to generate a model solving the problem (we will focus on minimally invasive surgery as primary use case). This includes the choice of relevant training data, compatible model architecture and desirable hyper-parameters. Because the meta-algorithm itself should be adaptive and learn to better supervise the model production during run time we explicitly address the problem of scalability. We aim at identifying the relevant parameters to predict such learning conditions, which may include: similarity of data, dependencies between tasks or functionality of network sub-components. Moreover the goal is to overcome the limitation of solving a single task in an isolated fashion, but to reuse existing knowledge/models to improve on new problems.

Doctoral researcher: Melanie Schellenberg

Institution: DKFZ

Data science PI: Lena Maier-Hein

Life science PI: Hannes Kenngott

Photoacoustic imaging (PAI) is an emerging modality that has the potential to provide tomographic images of blood oxygenation - an important biomarker for the diagnosis and staging of a variety of diseases. The generation of photoacoustic images follows a well-known forward process: Following the illumination of the tissue by multi-wavelengths laser pulses, photons are absorbed by various molecules according to their characteristic absorption spectra. The absorption follows a thermo-elastic expansion of the tissue which leads to a local pressure rise that generates an acoustic shock wave. The propagated acoustic wave can be measured as the PAI signal at the surface.

A core unsolved problem in PAI is the accurate quantification of the optical properties based on the acquired measurements. In this project, we tackle this ill-posed inverse problem with neural networks. Due to the lack of annotated reference data, the core methodological challenge is the generation of adequate training data. In this context, we go beyond the already established concept of learning from simulations with a novel approach to training data generation that we refer to as learning to simulate. Specifically, we investigate a machine learning-based approach to continuously improve the quality of in silico PAI data based on acquired real-world data.

Figure: The principle of photoacoustic imaging (PAI). (a) Based on the photoacoustic (PA) effect, the pressure time series can be measured as (b) raw PA signals and reconstructed into a (c) PA image or further processed, e.g. into (d) semantic PA image annotations.

Doctoral researcher: Mikael Beyene

Institution: KIT

Data science PI: Ali Sunyaev, Benedikt Brors

Modern life sciences with their highly sensitive omics data sets face several challenges with regard to data storage and sharing. On the one hand data must be protected in order to preserve the privacy of those individuals who contributed their data to research. On the other hand, the true value of omics data is only to be realized if shared with as many researchers as possible. In an ideal world, data subjects (i.e., patients) should be able to control access to their data directly. However, granting and revoking access to data is a slow and tedious process within the current life sciences research paradigm, where most data is either stored on central controlled-access data repositories or kept locally within the respective research groups. Adding to this, centralized data repositories create single points of failure in terms of data availability and integrity. Distributed ledger technology (DLT; e.g., blockchain) enables immutable transactions between untrustworthy parties, which are kept in a consistent state through automated, algorithm-based consensus building mechanisms, thus eliminating the need for third-party trust enforcement. The objective of this doctoral thesis is to develop a secure and privacy-preserving, DLT-based data sharing infrastructure for life sciences. Such a data sharing infrastructure with DLT-based access control will a) give patients flexible, and individual control over the flow of their personal information and b) ease comprehensive data sharing among researchers.

Doctoral researcher: Paula Breitling

Institution: KIT

Data science PI: Michael Beigl

Life science PI: Frank Ückert

With the possibility of whole genomic sequencing for oncologic patients, many processes in their treatment have to be adapted. Physicians in oncology have significantly more data to consider in order to tailor diagnostic or therapy approaches to the need of a specific patient. The National Center for Tumor Diseases (NCT) in Heidelberg established workflows for physicians to utilize genomic and clinical data to assess second or third line therapy options for patients. Within the workflows, a lot of external knowledge from databases is needed. Hence, for each database (or database portal), many queries have to be formulated and queried in order to obtain needed information. The steps involved to research information for a single patient are identified for the majority of cases. However, the steps can vary in order or only be necessary depending on the result of previous steps. Hence, a precise detection and discrimination is necessary to infer plausible next steps. The overall goal is to support a physician in researching options for diagnosis or therapy for patients based on genomic and clinical information by preemptively supplying information needed at the particular workflow step. To achieve that, we aim to precisely identify workflow steps based on usage signals (screen video, click events, time events, …) and possible triggers for further informational needs. A model will be trained to identify steps currently worked on and predict possible next steps (ideally multiple alternatives with likelihoods). These steps of the model will be evaluated for accuracy and (expected) utility in the process at hand.

Doctoral researcher: Rene Snajder

Institution: DKFZ

Data science PI: Oliver Stegle

Life science PI: Matthias Schlesner

Multi-omics, the generation of omics-profiles with multiple assays on the same set of biological samples, is a fundamental experimental design pattern in biomedical research and basic biology. In biomedical applications, objectives comprise biomarker discovery and disease stratification; in developmental and basic biology, understanding cell states, differentiation and cell type evolution, now increasingly at single cell resolution. However, there is a lack of computational methods for the analysis of such data. We will develop methods and software to address this need. In this project, we seek to deliver new computational methods and applications of supervised and unsupervised multi-omics analytical methods. Our approach builds on the multi omics factor analysis (MOFA), an unsupervised factor analysis method based on a sound and versatile mathematical foundation. MOFA is already a widely-used model in the field and is being applied in cancer genomics, single-cell biology and for other data integration tasks. The goal of this project is to extend and build on MOFA to deliver new functionality and to enable novel applications. A particular interest is the integration of imaging data modalities (e.g. FMRI, cellular imaging) with other omics data types, thereby exploiting large volumes of datasets available in Heidelberg.