Automatic organ segmentation based on novel imaging techniques and 3D deep convolutional neural networks

Reduce organ segmentation time and improve accuracy of organ at risk segmentation in a radiotherapy treatment workflow.

Executive summary
In 2016 around 108.400 patients were diagnosed with cancer in The Netherlands, of which 50% is treated with radiotherapy. Before any radiotherapy treatment, a complex treatment plan is made to calculate radiation dose to be delivered to the tumor and to spare healthy surrounding organs at risk. The dose in the organs at risk is currently evaluated by manually segmenting every organ at risk on every CT (computed tomography) slice, which the most time consuming task in radiotherapy, and is also very prone to user-variation. The tumors can have very complex shapes and are usually delineated manually by experienced oncologists. This is for now a task that cannot easily be automated.

In our research proposal we would like to investigate the combinational use of our in-house developed multi-atlas based contouring algorithm and a 3D-Unet automatic contouring algorithm to perform automatic segmentations of organs at risk on different CT datasets. Also the use of novel imaging techniques such as dual-energy CT will be investigated. This imaging technique provides a level of anatomical detail surpassing the current imaging standard: single-energy CT. The resulting contours originating from the automatic contouring algorithm will be compared with the clinical standard by an experienced radiation oncologist.

Composition research team

Brent van der Heyden (MSC - Applied medical-nuclear engineering)
Frank Verhaegen (PhD - Medical Physics engineering)
Daniele Eekers (MD, Radiation Oncology)
Cecile Wolfs (MSc, Knowledge engineering, Operations research)

Download abstract

GIANT (health management Guidance and advIce through a virtuAl patieNT avatar)

Improving health care and well-being

Executive summary
In western society 90% of cancer incidence is due to environmental causes (e.g. lifestyle, diet, pollution) and only 10% to causes beyond our control (i.e. genetics)1–3. Nonetheless, medical practice is predominantly reactive to cancer and many healthcare efforts are allocated to cancer management but lack the integration of prevention into overall health management.

The lack of systems that can integrate data to facilitate disease prevention is a major barrier in health management. This problem is rapidly growing because the amount of data, models, and biomarkers that support them is exponentially increasing.

The GIANT project provides a proof-of-concept for a virtual coach; the virtual patient avatar (VPA) (watch the animation: ). The VPA has two different interfaces, a smartphone version for the patient and a dashboard version for the doctors. At the core of this VPA is the ability to generate actionable guidance using a model catalogue, constructed from machine learning and biometric data.
Key innovative concepts featured in this project: the individualized patient decision aids merging the concept of predictive, personalized AND participative medicine, findable, accessible, interoperable and reusable (FAIR) data handling and privacy preserving data mining by using distributed learning.

Composition research team
Philippe Lambin (Prof. Dr, Radiation oncology and clinincal decision support systems)
Arthur Jochems(Dr.,Machine learning and distributed learning)
Ben Vanneste (MD, Radation oncology)
Joep van Roermund (MD, Urology)
Tatiana Deneve (Student, Health sciences)

Download abstract

Creating a Machine Readable Collection of Disease Biomarkers


Improving health care and well-being

Executive summary

Currently, a vast amount of researchers uses data from genomics, proteomics and metabolomics, in order to describe the observable characteristics (=phenotype) of an organism. Finding metabolites related to the health status of a patient (=biomarkers) is imperative to locate the underlying cause of a disease or the effect of a toxic compound. However, there is no open access, machine readable collection of metabolite biomarkers. This project’s focus lies with identifying sources where biomarker information can be found, how this information should be digitised and what tools are missing, in order to facilitate researchers to make use of the digitised information.

We want to gather information on biomarkers associated with a selection of diseases, how these biomarkers are measured in practise, and combine this information with biological pathway(s). This information will be digitised within the pathway database called WikiPathways ( - ), which is available to other researchers in various formats, e.g. GPML, OWL and RDF under a CC-Zero licence. In addition, WikiPathways relies on the user community for its input. Therefore, we want to offer a structured and user friendly method for this community to add additional biomarker information for more diseases.

Composition research team
Denise Slenter (Bioinformatics)
Richard Delava (DKE, student)
Anne Friesacher (FPN, student)
Roel Hacking (DKE, student)
Lisa Held (FPN, student)
Mzolisi Mtshaulana (LAW, student)
Hermann Ritter (DKE, student)
Egon Willighagen (Bioinformatics)

Download abstract

Delving deep into the brain – The application of machine learning in brain MRI

Accelerating scientific discovery, improving health care and well-being, and empowering communities

Executive summary
Psychiatry nowadays has to work with a classification system called Diagnostic and Statistical Manual of Mental Disorders (DSM). This system assigns a patient to a certain diagnosis based on symptoms subjectively identified by a trained psychiatrist. The major disadvantage of DSM is mis-representation of the causes underlying the mental disorder.

Evidence-based psychiatry can benefit from big data approaches applied to brain MRI and could allow for individualized prediction, treatment selection and medication adjustment. The application of deep learning techniques to large anatomical MRI datasets allows the detection of complex patterns in the brain’s structure. These patterns could then in turn be linked to distinct subgroups of patients. This classification system would be purely data-driven and will provide novel insights in changes related to a given mental disorder.

The next step would be to train and validate a machine learning (ML) algorithm based on large anatomical MRI databases with patients across the spectrum. The ML algorithm should be able to classify a newly introduced anatomical MRI dataset and give prediction for disease progression. Without assigning a diagnosis on the dataset, the algorithm should instead provide a risk profile to the psychiatrist. This profile helps to understand symptoms an individual patient exhibits.

Composition research team
Stijn Michielse (PhD student MHeNs)
Benedikt Poser (PSYCHOLOGY, Assistant Professor)
Dimo Ivanov (PSYCHOLOGY Assistant Professor)
Giancarlo Valente (PSYCHOLOGY, Assistant Professor)
Ron Megelers (MHeNs, ICT developer)
Jo Reep (MHeNs, ICT developer)
Ahmad Mohammad (DKE, student)

Download abstract

#MetooMaastricht: Mapping Sexual Harassment in Maastricht

Empowering communities to improve well-being

Executive summary
How safe are women in Maastricht? Many of those who arrive to Maastricht from other countries believe that Maastricht, being a small city, is a safe place. What if this isn’t as true as we thought? How many women in Maastricht are victims of gender violence and where are they more likely to be attacked? This project will not only answer this question, but also provide resources for the women of Maastricht.

This project will build a tool for women in Maastricht to anonymously tell their story with the location about the incidence via existing social media systems (twitter), as well as provide resources for women to report the incident or get additional help. This tool will allow local law enforcement to effectively target vulnerable areas of Maastricht and provide women with a safe space to engage and empower themselves after being victims of gender violence.

Composition research team
Mary Kaltenberg(UNU-MERIT, Statistics and Economics)
Nadia Feldkircher (GOVERNANCE, student)
Sanjay Guruprasad (MIT Media Lab, Sofware Engineer)
Julieta Marotta (GOVERNANCE, Law and Public Policy)
Gerasimos Spanakis (DKE, Data Scientist)

Download abstract

Can radiomics based on DCE-MRI predict and assess the response of neoadjuvant chemotherapy in breast cancer patients?

Accelerating scientific discovery

Executive summary
Breast cancer, the most common cancer among women in the Netherlands with more than 14.000 cases diagnosed in 2016. The percentage of patients treated with neoadjuvant chemotherapy (NAC) has risen from 7.2% to 18.8% between 2010 and 2015. Treatment response to NAC ranges from pathologic complete response (pCR) to non-response, with progression of disease at the end of the spectrum. Magnetic resonance imaging (MRI) is the preferred and best modality for tumor response assessment before and during NAC treatment. The accuracy of DCE-MRI to predict and assess NAC response in breast cancer patients is inadequate for adjusting NAC treatment, meaning; switch in case of no response or stop in case of early complete response. To date there is no adequate medical imaging technique that makes a quantitative and accurate pretreatment prediction or early (i.e. half-way therapy) assessment of NAC response. With the introduction of radiomics, a high-throughput quantitative image analysis method, standard-of-care medical images are converted into mineable textural information encompassing the entire breast tumor.

The purpose of this study; to evaluate the accuracy of radiomics imaging biomarkers extracted from dynamic contrast-enhanced MRI in the prediction and assessment of the response of NAC in breast cancer patients.

Composition research team
Renée Granzier (MD, GROW)
Sanaz Samiei (MD, AH)
Marjolein Smidt (MD, PhD oncological surgeon)
Henty C.Woodruff (PhD, MAASTRO)
Marc Lobbes (PhD Radiologist)
Thiemo Nijnatten (MD, PhD Radiologist in training)

Download abstract

Intelligent games for assessing cognitive, social, and physical capabilities of elderly and children

Improving health care and well-being

Executive summary
In the project we study the additive value of combining algorithms from data science and AI with robotics to assess the well-being in humans, and more specifically elderly and children. At the moment the number of instruments available for especially these target groups is limited and less ecologically valid. Assessments for elderly and children are difficult to apply, time consuming, and often lead to subjective results. In contrast, we evaluate the application of intelligent game devices like robotic dice, mazes, and board games for assessing cognitive, social, and physical capabilities. Monitoring these capabilities in elderly is important to detect possible deficits and to make informed decisions about required treatments. Monitoring these capabilities in children for instance in schools is important to ensure their healthy development. With the game devices we continuously obtain large amounts of data about the players that we process to infer their capabilities.

We are a team of Honours students from the Department of Data Science and Knowledge Engineering and multidisciplinary researchers of four faculties of Maastricht University. Together we develop novel assessment tools and data science algorithms that are fun for elderly and children to use, and that free caregivers and teachers from tedious work.

Composition research team
Bujnarowski, Kamil (Student DKE)
Calsius, Frederik (Student DKE)
Dàvila Mateu, Marta (Student DKE)
Gaßner, Martin (Student DKE)
Meyers, Marion (Student DKE)
Negura, Albert (Student DKE)
Ritter, Hermann (Student DKE)
Yared, Ryan (Student DKE)
Christopher, Seethu (DKE)
Dahl, Lucas (DKE)
Möckel, Rico (DKE)
Tumanov, Kirill (DKE)
Weiss, Gerhard (DKE)
Urlings, Corrie (SBE)
Borghans, Lex (SBE)
Coppens, Karien (SBE)
Hamers, Huub (Engineer at PSYCHOLOGY)
Hurks, Petra (PSYCHOLOGY)
Goanta, Catalina (LAW)

Download abstract

Generating new insights from open biomedical data using advanced multiple omics approaches

Accelerating scientific discovery

Executive summary
Multi-omics integration is one of the major challenges in current day biomedical research. An additional challenge is the reuse of experimental data, which is generally not straightforward even when data is publicly shared. This project will investigate both challenges.

The diXa project involved collecting, annotating and storing all toxicology related data generated by earlier EU projects. This included the Predtox project, which studied toxicity of 16 compounds using transcriptomics, metabolomics and proteomics. Multiple doses and time points were used and effects were measured at several molecular levels in liver, blood/serum and kidney/urine. We are working on advanced methodological workflows for the multi-omics and multi-tissue integration of this data and will further investigate the experimental results on a biological level. Currently, we only have capacity to focus on a single compound. The main objective of this proposal is to broaden our team and evaluate our developed methods on additional compounds, which will allow better identification of the extra information that is retrieved and would have been missed when only using a single omics or tissue. Also, the systematic reuse of publicly available data will provide further insights in the quality of the data and metadata in an existing resource.

Composition research team
Laura van Rooijen(Master Systems Biology, student)
Rachel Cavill (DKE, Assistant Professor)
Lars Wijssen (MHeNs, Assistant Professor)
Dennie Hebels (MERLN, Project Manager)
Andrea Nardi (DKE, student)
Jelle Bonthuis (FHML, student)

Download abstract

Improving healthcare at its source: a data-driven digital education tool to track and treat knowledge gaps in medical students

Improving health care and well – being (and empowering communities)

Executive summary
Personalized and tailor-made education based on the input and learning styles of the students have the potential to educate and train them to become high-quality health care professionals. Designing a personal education and training program as well as self-directed learning is one of the practical challenges for this era of medical students. Universities promote self-directed learning and train students to educate themselves, however the best available tools to facilitate these learning styles are either not (yet) available or not used. Knowledge gaps in specific specialties are not yet identified or cannot be addressed with the current e-learning tools.

We have developed an easy to use case-based digital tool to improve the student’s personal learning curve within the field of dermatology by passive case-based learning. The next step is the development and evaluation of the impact of active case-based learning in co-participation with experts/ teachers and peers, and also test the tool for other disciplines.

Getting personal tracking of your knowledge and gaps, being helped by professionals and peers, and having control over your personal learning strategy is a vital teaching and learning tool for future doctors, and can be of use in any professional education at any faculty within our university.

Composition research team
Lucie Wilbers (BSc, medical student)
Sylvia Heeneman (Prof. Dr, MUMC+, Educationalist)
Casper Mihl (Dr., MUMC+, Abdominal and Cardiovascular Radiologist)
Herm Martens (Drs. Dermatologist MUMC+)

Download abstract

Demystifying automated decision-making: an automated approach for evaluating the explainability of machine learning algorithms in the light of GDPR

Accelerating scientific discovery

Executive summary
Machine Learning (ML) algorithms used for prediction, profiling, and decision-making operate on big data and hyper-complex statistical methods. The rationale behind the generated outcome of these algorithms can be difficult and even impossible to explain. However, the General Data Protection Regulation (GDPR), entering into force on 25 May 2018, requires businesses to provide such explanations and, thus, force them to re-evaluate their existing algorithmically supported decision-making procedures.

This project aims to offer actionable insights on how existing decision-making procedures can be adjusted towards better compliance with the right-to-explanation. We plan to accomplish this goal by developing a standardized, automated framework for assessing the explainability of different ML techniques and, in this way, empowering practitioners to identify the right balance between transparency and decision-making power for their specific use. The deliverables will include a comprehensive report on GDPR and the explainability of common decision-support algorithms and a publicly accessible software package on the Comprehensive R Archive Network (CRAN) implementing our Explainability Assessment Framework to assist businesses in their algorithmic considerations. The project will also investigate whether a layer of automated self-explanation can be added to a subset of algorithms to improve the required transparency of the decision-making process.

Composition research team
Deniz Iren (BISS and SBE, Researcher)
Maja Brkan (LAW, Assistant Professor)
Banu Aysolmaz (SBE, Assistant Professor)
Kurt Driessens (DKE, Assistant Professor)
Arno Angerer (SBE, MSc Student)

Download abstract

Let data speak: Evidence mining tool in Health Sciences

One of accelerating scientific discovery, improving health care and well-being, and empowering communities.

Executive summary
Scientific evidence on health outcomes is growing. Meta-analyses and literature reviews are lagging behind in summarizing up-to-date evidence. Furthermore, search criteria in these studies are often narrowly defined while generalizability of the findings is limited to a specific population/country. We propose to create an evidence mining tool that would enable us to efficiently and effectively compile existing evidence from different empirical studies in Health Sciences. This tool will search abstracts upon a specified query in different search engines, such as Medline and PubMed; identify patterns in the text of the selected abstracts; organize evidence in a structured, machine-readable and accessible way, as required and specified by the user.

The tool will be open and freely available as a web-application. Users of this tool, i.e. scientists, health professionals, patients and policy makers, can quickly find what is the effect of one factor on the probability of another factor, also by adjusting their search to different criteria, e.g. certain characteristics of the population, treatment and strength of the evidence. This tool can be further developed to mine evidence from the entire text of the particular empirical paper, and not only from the abstract/summary.

Composition research team
Iryna Rud (TIER TA, Assistant Professor)
Amrapali Zaveri (IDS, PostDoc)
Olga Zvonareva (Metamedica, Assistant Professor)
Niels Hameleers (FHML, Data Scientist/Researcher)
Stijn Michielse (PhD student MHeNs)
Narek Harutyunyan (FPN, student)
Minhaz Arefin (SBE, student)

Download abstract


Improving health-care and well-being

Executive summary
Every day hundreds of UM employees and students flock into Maastricht with their cars, taking people to work and their courses—but also causing congestion, noise, pollution, and time wastage. An estimated 30% of all car traffic is only out to search for parking spaces.

Wouldn’t it be fantastic to have a parking space forecast on your smartphone, similar to a weather forecast, to predict parking space availability and current prices near their destination? Then drivers could save plenty of time and make informed decisions beforehand. Also, wouldn’t it be fantastic planners could steer demand are such that there is always a spot left for people who really need it?

We will design a system of data collection and data analytics which allows us to collect data about parking spot occupancy in the city of Maastricht. Then we will process the data to design dynamic pricing schemes for optimal occupancy rates, and efficient routing schemes which minimize travel times for drivers to reach a free parking spot in the vicinity of their desired destination. The system will be materialized by a free-to-use mobile phone app and advanced algorithmic techniques.

Composition research team
Matus Mihalak (DKE, Assistant Professor)
Matthias Mnich (SBE, Assistant Professor)
Stuti Nayak (DKE, student)
Amrapali Zaveri (IDS, PostDoc)

Download abstract

Dynamic Policy Evaluation: Measuring Public Expectations through Text Mining

Empowering communities

Executive summary
The analysis of government policy changes over time, is a challenging task, since agents anticipate its effects at the time of announcement, which precedes the policy implementation. While it is widely agreed among economists that expectation plays a fundamental role in understanding dynamic causal economic and political relationships, measuring them is still an open question. Typically, aggregated macroeconomic data may often not contain sufficient information to properly identify the role of expectations of various economic agents, on the effect of policy changes. Researchers have manually augmented macroeconomic data with narratives from periodical newspapers or congressional documents to gauge the public's expectation about future policy directions.

In this project we will use recent text mining techniques to develop an automatic, consistent and systematic way to process and analyze a large amount of unstructured data concerning public expectation. The relations and similarities between relevant terms as well as concepts of importance for public expectations will be investigated, as well as their effects across countries over time. We will develop new macroeconomic models that will explicitly incorporate public expectations. The data and models obtained will potentially open new venues and possibilities to study economic policy as well as political economy applications to areas such as country risk analysis.

Composition research team
Rui Jorge Almeida(SBE, BISS, Assistant Professor)
Lenard Lieb (SBE, PostDoc)
Kaj Thomsson (SBE)
Adam Jassem (SBE, Student)

Download abstract
  • Automatic organ segmentation based on novel imaging techniques and 3D deep convolutional neural networks

  • GIANT (health management Guidance and advIce through a virtuAl patieNT avatar)

  • Creating a Machine Readable Collection of Disease Biomarkers

  • Delving deep into the brain – The application of machine learning in brain MRI

  • #MetooMaastricht: Mapping Sexual Harassment in Maastricht

  • Can radiomics based on DCE-MRI predict and assess the response of neoadjuvant chemotherapy in breast cancer patients?

  • Intelligent games for assessing cognitive, social, and physical capabilities of elderly and children

  • Generating new insights from open biomedical data using advanced multiple omics approaches

  • Improving healthcare at its source: a data-driven digital education tool to track and treat knowledge gaps in medical students

  • Demystifying automated decision-making: an automated approach for evaluating the explainability of machine learning algorithms in the light of GDPR

  • Let data speak: Evidence mining tool in Health Sciences

  • MaastrichtPark

  • Dynamic Policy Evaluation: Measuring Public Expectations through Text Mining