MSc Defences
See the list of MSc defences at DIKU. The list will continuously be updated.
If the defences are announced as ‘online defence’, the student has to be alone in the room during the examination and assessment. Guests can participate online but the links for the defences are not public. If you want to be present during the defence, please contact uddannelse@diku.dk or the supervisor for a link.
Computer Science
June
Title
Plant health and food quality from RGB images using deep learning
Abstract
There is little work done on plant growth prediction or health through cheap RGB cameras, while such images are often used for segmentation and classifications of plants. For this project data has continuously been collected from 2 different basil plants grown in vertical farms to predict the age of these plants. The best results show that an absolute prediction error of 2.3 days with greater than 68% certainty and an absolute prediction error of less than 3.34 days with 95% certainty was achieved on a small dataset.
We find that more data is required to achieve better results and for the prediction error to be useful real or close to real time estimation of plant health and age. To achieve these results a transfer learning technique was used to implement and convert a staple in classification models called ResNet50 to a regression model.
Additionally, an assortment of augmentations, such as warping, translating and down sampling, were used for preprocessing the images, which are relevant for the vertical farm setup that these tests were performed on.
Time and place
3 June 2021 at 09:00
Online
Supervisor(s)
Kim Steenstrup Pedersen
External examiner(s)
Rasmus Reinhold Paulsen
Title
Semi-supervised classification of grain kernel images
Abstract
This thesis presents an analysis of the viability of semi-supervised deep learning models for the classification of grain kernel images. Through the experiments performed, it is shown that the semi-supervised model is capable of not only matching but even outperforming a baseline supervised model while using, at most, 10% of the amount of labeled data that the baseline uses. An increase of as much as 3.17% in weighted F1 score was observed, which, together with the reduction in the amount of labeled data required, presents a strong argument in the favour of semi-supervised models.
Throughout the thesis, experiments are performed to determine the effects that various aspects, such as the size and composition of the labeled data-set, have on the capabilities of the models. Through these experiments, areas that pose greater challenges to the models, such as classes with low separability between them, are identified and explored to determine their causes and potential solutions.
Keywords: Semi-supervised, Noisy Student, Image Classification, Grain Classification, Deep learning, Convolutional Neural Networks
Time and place
3 June 2021 at 11:20
Online
Supervisor(s)
Erik Schou Dreier and Kim Steenstrup Pedersen
External examiner(s)
Rasmus Reinhold Paulsen
Title
Deep-learning methods on taxonomic beetle data
Abstract
This thesis investigates U-Net (Ronneberger et al. 2015) segmentation performance and data requirements for satisfactory segmentation of beetles from the natural history museum of Denmark. We investigate further the multi task learning capabilities of U-Net by introducing several classi cation heads to the model for taxonomic classi cation. U-Net is compared to HR-net (Wang et al. 2019) in both segmentation and multi task learning and parameter sharing capabilities. U-Net has proven
superior to HR-net during our investigation in both pure segmentation tasks but also multi task learning and feature transfer. Training the multi task U-Net on specialized data sets has increased classi cation performance signi cantly, but it still is inferior to specialized models such as from our bachelors thesis (Gutschenreiter & Bech 2019).
Time and place
3 June 2021 at 12:45
Supervisor(s)
Kim Steenstrup Pedersen
External examiner(s)
Rasmus Reinhold Paulsen
Title
Exploring Semantic Parsing for Process Models
Abstract
Business process models are an important part of managing a business. The field of meaning representation parsing is centered around automating parsing from natural language sentences. In this thesis we explore how to combine these two fields in order to automate the construction of business process models given natural language process descriptions.
We present applications for translating data between the two fields. Furthermore we explore an available meaning representation parser as a possible tool for automating the process description to process model parsing.
Time and place
3 June 2021 at 11:00
Online
Supervisor(s)
Daniel Hershcovich
External examiner(s)
Søren Debois
Title
ECS Using Rust
Abstract
Entity Component System(ECS) is a architectural pattern which has gain increasing traction in the Game industry over the recent years. ECS is used to better utilize the hardware running in parallel.
Rust is a programming language which goals is fast performance like C while restricting memory to avoid common programmer mistakes. To implement a fast parallel ECS using Rust I will investigate what techniques are used in the industry and compare them to know which method is preferable in different scenarios. ECS in Rust is still being developed across many libraries in the Rust community where they still finds new ways to improve upon the ECS pattern. There are many ways currently being used to implement the ECS pattern all which has benefits and penalties given different scenarios.
Time and place
8 June 2021 at 13:30
Online
Supervisor(s)
Ken Friis Larsen
External examiner(s)
Peter Sestoft
Title
Approximate Triangle Counting via Sampling and Fast Matrix Multiplication
Abstract
There is a trivial O(n^3/T) time algorithm for approximate triangle counting where T is the number of triangles in the graph and n the number of vertices. At the same time, one may count triangles exactly using fast matrix multiplication in time (Õ(n^w). Is it possible to get a negative dependency on the number of triangles T while retaining the n^w dependency on n? We answer this question positively by providing an algorithm which runs in time O(n^w/T^w-2)*poly(n^o(1)/€). This is optimal in the sense that as long as the exponent of T is independent of n,T, it cannot be improved while retaining the dependency on n; this as follows from the lower bound of Eden and Rosenbaum [APPROX/RANDOM 20218]. Our algorithm improves upon the state of the art when T = w(1) and T = o(n).
We also consider the problem of approximate triangle counting in sparse graphs, parameterizing by the number of edges m. The best known algorithm runs in time Õ(m^3/2 /T) [Eden et al., SIAM Journal on Computing, 2017]. There is also a well-known algorithm for exact triangle counting that runs in time Õ(m^2w/(w+1). We again get an algorithm that retains the exponent of m while running faster on graphs with larger number of triangles. Specifically, our algorithm runs in time O(m^2w/(w+1) / T^2(w-1)/(w+1)*poly(n^o(1) / €). This is again optimal in the sense that if the exponent of T is to be constant, it cannot be improved without worsening the dependency on m. This algorithm improves upon the state of the art when T = w(1) and T = o(v~m).
Time and place
8 June at 11:00
Online
Supervisor(s)
Mikkel Thorup
External examiner(s)
Inge Li Gørtz
Title
Security Issues in eBPF verifier
Abstract
Extended Berkeley Packet Filter (eBPF) is widely used by companies, such as Facebook and Netflix to implement performance analysis and security measures. The technology has gained popularity as it allows execution of dynamically loaded user-defined programs inside the kernel, and as a result of this unsafe programs can have significant consequences.
To avoid these consequences, the programs have to pass various safety checks, which is handled by the in-kernel verifier. The checks are aimed at keeping the programs safe, but some enforce security constraints as well, for instance by preventing kernel pointer leaks. Currently, the verifier allows programs to leak almost anything to user space as it has little to none information security measures. To guarantee information security of a program, information flow can be utilised to prevent the program from leaking confidential information either directly or indirectly.
In this thesis, we propose a method that systematically uses information flow analysis to prevent eBPF programs from leaking confidential information. Using our proposed helper function approach, security policies and annotation techniques, we created a sound type system and implemented a type checker based on seven insecure eBPF programs found in the wild. The type system was evaluated using the seven insecure programs, two secure programs and a collection of unit tests.
We found that using information flow to prevent confidential information from being leaked looks promising, however, further work is needed to make a definite conclusion. The results are discussed in terms of validity and we make suggestions for further work.
Time and place
9 June 2021 at 14:00
Online
Supervisor(s)
Ken Friis Larsen
External examiner(s)
Philippe Bonnet
Title
Measuring Orderliness using video games
Abstract
Personality tests are used in psychology research for quantifying human differences. Usually, they rely on people describing themselves, by presenting a standardized set of questions that people can fill in answers to. These have a number of problems. We instead investigate the possibility of measuring personality by modifying a video game to permit expression of personality and record personality data. We choose to investigate the trait of Orderliness, and modify the game of shapez.io to yield several variables which we take to be indicators of Orderliness. Comparing the results of this test to a self-report questionnaire, we find no clear connection between the game behavior and personality, though our low sample size prevents definite interpretations of this result.
Time and place
9 June 2021 at 14:00
Supervisor(s)
Henning Pohl
External examiner(s)
Timothy Merritt
Title
Proof-Carrying Code for WebAssembly
Abstract
WebAssembly is a popular compilation target that has seen wide adoption by all major browser vendors as well as other platforms. It is a fast and secure low-level target, which achieves safety via both static validation checks as well as runtime checks. Proof-carrying code (PCC) is a scheme that enables safe execution of untrusted code. In the scheme, a given program is annotated such that the code consumer can statically verify the safety of the program. A set of safety properties called the
safety policy defines what a program must satisfy in order to be considered safe. In this report, I present the design of my adaptation of the PCC scheme to WebAssembly, which allows for modular verification. I explore the different possibilities of expressiveness and extensibility gained from separating the core of the PCC algorithm from both the modelling of
the language’s semantics and the safety policy. I also explore the design space for encoding high-level types for this scheme. I present my prototype implementation of this scheme implemented as an extension of the WebAssembly reference interpreter, using the Why3 platform for verification. Finally, I show the experimental results of sucessfully verifying non-trivial programs using the prototype.
Time and place
10 June 2021 at 14:00
Online
Supervisor(s)
Ken Friis Larsen
External examiner(s)
Carsten Elmar Schürmann
Title
Implementation of a blockchain with native DCRgraphs smart-contracts
Abstract
Today, data integrity is more important than ever. By having a distributed system, data across many peers are challenging to keep safe because of the many end-points. Modern technologies, such as blockchain, make this problem obsolete by having data integrity protection built in. This work demonstrates a blockchain built with the Substrate-framework with smart contracts capability encoded in heavily inspired DCR graph notation. Furthermore, a comparison between an existing implementation of a DCR Graph engine in a blockchain is presented. The implementation and upkeep cost is discovered for a novel blockchain. Finally, this work concludes by showing the real-world cost and results of using a private permissionless blockchain.
Time and place
10 June 2021 at 10:00 - 11:00
Supervisor(s)
Tijs Slaats
External examiner(s)
Søren Debois
Title
Regular Segmented Single-pass Scan in Futhark
Abstract
This thesis presents an implementation of a single-pass scan algorithm described by researchers Merrill & Garland, as an extension to the functional language Futhark. The work consists of a generalization of a prior implementation made by Persson & Nicolaisen, modified to work on regular segments.
In addition to generalization of the implementation, two major contributions in the form of an analytical model to estimate optimal workload per thread based on type analysis and safe rewriting of index arithmetic to computationally cheaper calculations are made.
The implementation and contributions are tested, and their respective benefits are documented using Futharks built-in benchmarking system.
Time and place
10 June 2021 at 15:00 - 16:00
Online
Supervisor(s)
Cosmin Eugen Oancea
External examiner(s)
Patrick Bahr
Title
Demographic bias in public neuroimaging databases, and its effect on AI systems for computer-aided diagnosis
Abstract
Deep learning is thriving, and neural networks are currently being developed for all kinds of computational problems. Recently it has been shown that a convolutional neural network performs very well on diagnosing Alzheimer’s Disease (AD) patients based on their MRI scan [1]. However, the focus is often placed on computational time and performance when developing such a network. Little attention is often paid to the subjects’ demographics and whether the training data is representative and balanced according to the demographics
[2]. This project investigates whether a bias in ethnicity in the publicly available dataset ADNI will propagate as a bias in a model for a convolutional network.
A state-of-the-art convolutional neural network used to diagnose AD [1] is explored. The network is described and implemented in Python. Furthermore, a simpler network with comparable performance and a similar structure is presented. They were trained on different splits of preprocessed magnetic resonance (MRI) images from the American
Alzheimer’s Disease Neuroimaging Initiative (ADNI) [3] dataset. They yielded around 80-85 % accuracy and sensitivities above 65 % depending on the split and model. Using the models to predict AD on a test set from the Australian Imaging, Biomarker Lifestyle Flagship Study of Ageing (AIBL) [4] yielded a performance comparable to the AD prediction on the ADNI images. However, when classifying images from the Italian Alzheimer’s Disease Neuroimaging Initiative dataset (I-ADNI) [5], all the models showed a worrying drop in performance. The best model could only achieve 78.0 % accuracy and sensitivity (almost no CN patients were present in the Italian dataset). In the worst case, the model only had the ability to classify the Italian patients randomly. Subsequently, it was found that the models in almost all cases have higher performance on MRI images
taken at field strength 1.5T images varying up to over 10 % and that there were no significant differences in performance for the genders.
This project was challenged by the low amount of data from AIBL and I-ADNI and that the sites were very imbalanced concerning healthy controls (CN) and AD patients. Furthermore, it can be hard to ensure that no hidden bias in the different sites unexpectedly affects the results.
Conclusively the results suggest that the ethnic bias in the training dataset did propagate as a bias in the convolutional neural networks leading to a concerning drop in performance. More data should be included to investigate this further; however, this can have crucial consequences for how the ADNI data should be applied to deep learning in the future.
Time and place
11 June 2021 at 10:30
Online
Supervisor(s)
Aasa Feragen
External examiner(s)
Morten Pol Engell-Nørregård
Title
Political astroturfing on Twitter: Identifying spam-campaign collaborators on social media
Abstract
Astroturfing in the context of political disinformation/spam campaigns on social media is a rising issue. Paid groups of workers are being deployed on social media by malicious actors in order to spread disinformation to destabilize countries, elections, companies, and products. In general, social media platforms and other tech companies need the required knowledge and tools to implement countermeasures and detection techniques to tackle such novel threats. In this project, we study the application of machine learning techniques to the domain of spam-campaigns in order to automatically detect the malicious users. We show that a typical inauthentic account reveals several significant indicators, such as hashtag similarities, account lifespan and the sentiment of the tweet. We show that it is possible to achieve high-precision results when using these indicators in a ML detection model. Furthermore we discuss some possible techniques that the spammers can use to avoid detection, using the capabilities that the Twitter platform provide.
Time and place
11 June 2021 at 10:00 - 11:00
Online
Supervisor(s)
Boris Düdder
External examiner(s)
Mads Rosendahl
Title
Code Generation for Stencils in Futhark
Abstract
In this thesis, I investigate two different optimizations – tiling and partitioning – for stencil computations on parallel architectures through prototyping in CUDA and C, and implement code
generation for stencils in the Futhark compiler’s multicore backend. The prototyping reveals that the two optimizations in combination yield significant performance increase for multithreaded C programs, compared to an unoptimized version. When benchmarking the code generated by the Futhark compiler implementation, the fully optimized version again far outperformed the non-optimized version, but did not outperform implementations of the same stencils in terms of non-stencil-specific Futhark constructs.
Time and place
14 June
Online
Supervisor(s)
Cosmin Eugen Oancea
External examiner(s
Mads Rosendahl
Title
Segmenting the Left Ventricle of Hearts using U-Nets
Abstract
This project provides the groundwork for segmenting out the entire heart in the future and automating the segmentation task with a sub-goal of lowering the cost of medical diagnosis.
We also consider this project a further approval of the MPUnet [1] article’s claims of having a well-generalized model that will perform well on any medical segmentation task.
As a result of segmenting the heart, we will determine how well the MPUnet model generalizes towards another problem which is the left ventricle on the heart in PET images. Once confirmed, we will compare the MPUnet to other relevant models. To maximize performance, it is standard practice to evaluate numerous pipelines with varying model topologies, optimization
parameters, pre-processing and post-processing steps, and even model cascades.
The MPUnet paper [1] proposes a model that generalizes well over different scan modalities and body parts to handle the medical segmentation task in 3D. This project aims to test MPUnet’s ability to generalize over our PET scan dataset and to see how much further performance can achieve with other relevant models. We will be comparing with the MPUnet[1] as a baseline throughout the process.
The problem we aim to solve with various models is segmenting the heart’s left ventricle on PET scans of real patients. We have looked at the output segmentation from the model, and
we can clearly see that the model also recognizes the heart ventricle shape, size, and location.
We found that the best model to use is the MPUnet. It provides significantly better results on the final evaluation dataset than all the other models we used, which did not always generalize as well as the MPUnet.
Time and place
14 June at 13:50
Online
Supervisor(s)
Erik Bjørnager Dam
External examiner(s)
Dan Witzner Hansen
Title
Data-Parallel Coherency Sensitive Hashing for Approximate Nearest Neighbour Fields
Abstract
Finding similar patches in two images is an often used technique in computer vision. Methods that achieve this purpose, known as Nearest Neighbour Fields, are slow and expensive and
modern demands have only served to increase the resolution of the images these techniques are used on. To alleviate this issue approximate solutions are computed, which trade exactness
for speed. Recent research has shown that highly parallel implementations that use GPUs for computing can be used to achieve large speedups with no impact on accuracy. In this thesis
I propose a data-parallel implementation of Coherency Sensitive Hashing(CSH). CSH, when using CPUs, has been observed to achieve good performance by exploiting the coherency of
images and similarity between patches. The data-parallel implementation shows a significant increase in speed at no cost to accuracy and is quick enough to be run in real time on high
quality images. The performance of data-parallel CSH has been verified using an image set of 133 image pairs.
Time and place
14 June
Online
Supervisor(s)
Cosmin Eugen Oancea
External examiner(s
Mads Rosendahl
Title
Optimisation and GPU code generation of Stencils for Futhark
Abstract
Stencils are a common problem in the area of scientific computing.
Exploitation of parallel computing is a central part when optimising for faster execution times of stencils running on large amounts of data. For this reason stencils are well suited to be run in a GPGPU setting. However, programming stencils to run on massively-parallel hardware is a time-consuming and error-prone exercise. For this reason it is useful to be able to express these stencils in a more abstract form, in a high-level programming language. Then a compiler will translate the stencil into
more efficient and parallel computations in a GPGPU setting. Futhark is a high-level programming language, with the purpose of producing efficient multi-threaded CPU, CUDA and OpenCL programs. However, it has no native support for stencils. This thesis concerns the implementation of code generation for a stencil construct for the Futhark OpenCL and CUDA back-ends of the compiler. We investigate many designs for running stencils in a GPGPU setting, and analyse these different designs. We then choose the most efficient and robust prototype, to guide
our implementation of code generation of the stencil construct in the Futhark compiler. The implemented stencil construct provides significant speedups compared to what could already be done with a nested map implementation in Futhark. For some hardware and stencils we achieve up to three times speedup.
Time and place
14 June at 16:00
Online
Supervisor(s)
Troels Henriksen
External examiner(s)
Mads Rosendahl
Title
Registering Mouse Brain Slices to a Reference Atlas with Convolutional Neural Networks
Abstract
Registering mouse brain slices to a reference atlas can help to determine the location of anatomical structures which is an essential step for understanding the function of brain
circuits. However, tools that are based on conventional registration methods, such as Elastix, can be slow, limiting their application in real life. This thesis explored Convolutional Neural
Networks (CNNs) as an alternative to conventional registration methods. The problem of registering partial 2D mouse brain slices to a reference atlas is investigated by taking a twostep approach. In the first step, 2D brain slices are identified by finding the corresponding 2D atlas plates with Siamese Networks. In the second step, brain slices are affinely registered
to the corresponding atlas plates using a CNN-based network. In the identification task, Siamese CNN achieved TOP-1 and TOP-5 accuracy of 25% and 100%, respectively, taking only 7.2 seconds to identify 29 images. In the affine registration task, CNNs outperformed SimpleElastix in both accuracy and speed. This study shows that CNNs have a big potential in partial image registration to a reference atlas.
Keywords: affine image registration, mouse brain, reference atlas, deep learning.
Time and place
14 June 2021 at 14:00
Online
Supervisor(s)
Raghavendra Selvan
External examiner(s)
Rasmus Reinhold Paulsen
Title
Enforcing data consistency in event-driven microservices through event-based constraints
Abstract
Microservices are an emerging architectural style for designing software applications as a set of self-contained and loosely coupled units with private data stores. Although the individual services are developed and maintained independently, in order to provide certain system functionalities, such as a correlation of distinct events, they often depend on each other. One of the inter-service communication models that are increasingly being employed are asynchronous events. Event-driven microservices
generate and publish events that reflect changes in their internal state to other parties that necessitate operating over such data or react to such external changes.
Decentralizing data ownership across microservices has its implications for responding to changes.
In order to preserve the flexibility and availability of microservice systems, strong consistency guarantees are usually replaced by the eventual consistency model. This introduces new challenges
as application developers need to implement a significant amount of data management logic and validations at the application layer.
This leads to difficulties in reasoning about application invariants and preserving them under diversified event schedules and arbitrary interleavings. To address these challenges, we start by analyzing popular open-source microservice repositories to identify common patterns and shortcomings in
application data management that lead to inconsistencies. Based on detected patterns in application encodings, we propose three categories of event-based constraints that serve as an easy-to-use guideline to developers for reasoning about data integrity.
We introduce Stream Constraints, a system-level stream processing solution created on top of Kafka Streams that allows developers to specify event-based constraints and enforce them on a stream of events. The evaluation of Stream Constraints conducted on a popular open-source eventdriven microservice application shows that explicit definition of application invariants eliminates or significantly decreases the number of data integrity anomalies and introduces lower overhead compared to baseline solutions.
Time and place
14 June at 14:00
Online
Supervisor(s)
Yongluan Zhou
External examiner(s)
Philippe Bonnet
Title
Identifying and Quantifying Feral Anomalies in Event-Driven Applications
Abstract
The advent of cloud computing and large-scale data-driven applications has seen developers shifting towards microservices for higher availability, scalability, and faster deployment cycles. Microservices are deployed as lightweight independent components to perform a particular task, unaware
of each other’s underlying databases. Practitioners rely on persistent message queues for asynchronous event-based communication between the decoupled components, thereby adopting event-driven architecture. The cross
dependency between multiple microservices forces developers to implement application-level (feral) constraints to safeguard the system’s data integrity spanning different database models. These feral mechanisms employed in event-driven applications may lead to data inconsistencies if they fail to ensure data integrity.
To support this claim, we first catalog the feral invariants susceptible to failure by investigating two popular open-source event-driven reference applications. Then devised a novel tool, HawkEDA, to quantify and monitor data integrity violations caused by the classified feral invariants in realtime using complex event processing. We examine the gravity of the data
anomalies using the supplied experiments and prove that the feral invariants lacking the necessary coordination to gain higher performance exhibited data inconsistencies with concurrent workloads.
Unlike monolithic applications, it is far more complicated to rationalize the convoluted data flow encompassing multiple microservices, making it difficult to choose the right balance between application safety and performance trade-offs. Therefore, HawkEDA provides easy integration and great
extensibility for enabling practitioners to proactively review their feral invariants’ design on a case-to-case basis.
Keywords: microservice, event-driven architecture, data integrity
Time and place
14 June at 15:00
Online
Supervisor(s)
Yongluan Zhou
External examiner(s)
Philippe Bonnet
Title
Real-time crowdsourcing controller: Utilizing reliable agents in continuous crowdsourcing tasks
Abstract
In recent years crowdsourcing has proven useful in a variety of tasks, from labeling images and captioning videos to solving scientific problems and writing short novels. A relatively new branch of crowdsourcing concerns real-time and continuous systems, which engage workers in longer sessions,
enabling them to experience an instant feedback loop and keep adapting their strategy for the given task.
This thesis presents a new use case for continuous crowdsourcing, which is a process of collectively
controlling a single character in a video game. Similar gameplay mechanics have already been proposed, but they suffer from certain limitations. They usually incorporate a type of network voting process that does not work well in real-time. Furthermore, they do not utilize approaches commonly implemented in crowdsourcing systems, such as estimating player reliabilities.
We introduce an implementation of a multiplayer game with a crowdsourcing system that addresses the above-mentioned limitations. Multiple network players interact with the system simultaneously by issuing standard game commands. Their input, however, is processed using two novel approaches that
follow the wisdom of the crowd principle. Dynamic Input Frame aggregates incoming player input in real-time, such that each of the resulting groups corresponds to one action performed by the character at a specific time. Continuous Reliability System estimates and utilizes differences in players’ skills
by assigning them a reliability score which translates to the weight of their input.
Proposed systems are evaluated on a series of experiments, which use both human and simulated AI-controlled players. Their performance is compared to well-known approaches commonly used in crowdsourcing systems. Dynamic Input Frame proves to be a more effective method in discrete command analysis, compared to a state-of-the-art static approach. Continuous Reliability System also performs better when compared to the widely used majority voting, especially on sets of players with
a significant variety of skills.
Time and place
15 June at 08:00
Online
Supervisor(s)
Maria Maistro
External examiner(s)
Timothy Merritt
Title
Mechanized formalization of a propositional calculus for contract specification
Abstract
The successful management of commercial contracts is vital for businesses. Improper management is a costly affair, at worst leading to unintended contract breaches with hefty legal fees. Tools that support proper management of contracts are therefore highly desirable. One such tool is the contract specification language CSL,
developed by Andersen et al. [1], supporting compositional specification of contracts. In this thesis, we formalize and mechanize a calculus for a restricted variant of CSL, with the mechanization carried out in the proof-assistant Coq. The calculus presented here will be used in the later formalization and mechanization of a calculus for CSL2, the successor of the CSL language.
Time and place
16 June at 14:00
Online
Supervisor(s)
Fritz Henglein and Agata Anna Murawska
External examiner(s)
Patrick Bahr
Title
Prototype Implementation and Application of Contract Specification Language CSL2
Abstract
With the increasing popularity of distributed ledger technology, financial contracts may now be specified using virtual contracts maintained over such a distributed ledger; these virtual
contracts are also known as smart contracts. However, a communication gap presents itself between the domain experts that specify the equivalent ”paper” contracts, and the programmers who must implement them. Domain specific languages for writing smart contracts may help bridge this gap. Based on CSL (Contract Specification Language), a domain specific language developed by Deon Digital, we develop CSL2, which extends CSL with output states, a useful property for contract specification. We first define the operational semantics and
type system for the language, and then implement the language as a deeply-embedded DSL in Haskell. We also introduce a novel algebraic data type for expressing dates. Finally, we
present a variety of financial contracts specified using CSL2, and argue for their readability.
Time and place
16 June at 15:05
Online
Supervisor(s)
Fritz Henglein
External examiner(s)
Patrick Bahr
Title
Estimating Web Credibility: A Reproducibility Study
Abstract
News that gives incorrect information to readers, better known as fake news or misinformation, is a rapidly growing problem due to the nature of the internet and social media websites. This brings
about a need for systems that predict the credibility of different types of web content (e.g., Webpages, News Articles, etc.). In this thesis, we explore the existing research done in the domain of credibility estimation. First, we generate a comprehensive list of features based on previous literature. Next, we use these features as input for supervised learning models to estimate the credibility. Specifically, we reproduce the models proposed by Shrestha and Spezzano 2021 and Esteves et al. 2018. We validate
their findings across multiple datasets of varying sizes. Finally, we analyse the experimental results to find what group of features are the most important when it comes to credibility estimation. We find that the feature group varies based on the type of news content (e.g., political, entertainment, sports, etc.) being classified.
Time and place
17 June at 14:00
Online
Supervisor(s)
Maria Maistro
External examiner(s)
Jacob Nordfalk
Title
Banking the unbanked: Future-proofing the least developed countries as they go from cash to online-payment
Abstract
Banking is a necessity for everyone, it is a key factor to reduce poverty and is a focal point for many organizations around the world. Unfortunately, 1.7 billion people remain unbanked. We take an example-driven approach to explore the reasons for why this is the case, where this is the case, and how we
can bank the people of these countries. We introduce a banking model based on M-Pesa that circumvents some of the complications of the M-Pesa model.
In these regions, cash is king. As the digital divide lessens we implement two systems based around this model. One for the current generation, based on the technology already available, and one for future generations, based on technology that will become available. We find that converting from a static
agent model to a dynamic one, multiple benefits can appear: The distance to banks is reduced, fees might be reduced, new job opportunities are made, and lack of identification might no longer be a limiting factor.
Time and place
17 June at 14:00
Online
Supervisor(s)
Fritz Henglein, Søren Terp Hørlück Jessen
External examiner(s)
Mads Rosendahl
Title
Using Graph Neural Networks To Learn Node Embeddings For Spatial Transcriptomics Neighborhood Graphs
Abstract
Recently, spatial transcriptomics methods have emerged and become more accessible. However, the number of computational methods that make use of the spatial information is limited. Existing
machine learning methods either do not incorporate spatial aspects or work on regular structures. My aim with this thesis is to present a machine learning approach that makes use of the true strength of
the spatial transcriptomics technology: spatiality. By turning spatial data into neighborhood graphs, we abstract the spatial information and make it possible to work with Graph Neural Networks. With these, we learn how to aggregate spot information with neighboring spot information and use these aggregations for machine learning predictions. To facilitate this process, I provide a user-friendly pipeline that assists with the graph construction, model creation and -tuning, and the extraction of the node embeddings, the aggregated spot information. I compare the results with a benchmark model that does not factor spatial information to compare the method to neighborhood-agnostic approaches.
I found that our approach outperforms other machine learning methods that don’t factor spatial information by 7% in prediction accuracy in a supervised machine learning task classifying multiple
annotated brain regions within a mouse brain atlas with an overall score of 79.01%. Furthermore, I present how the node embeddings serve downstream data analysis tasks like clustering and anomaly
detection. Applying my method to another use case, detecting Alzheimer’s diseased brain tissue spots shows that our approach works across different datasets and use-cases.
Time and place
21 June at 15:00
Online
Supervisor(s)
Anders Krogh, Tune Pers, Petar Todorov
External examiner(s)
Jes Frellsen
Title
"Sacrifice to the Network God": A Critical Design Approach to Rethink Privacy
Abstract
The main goal of the thesis was to challenge conceptions about data privacy through a critical design approach that explores how people conceptualize the value of personal data when using privacyinvasive technologies. An application is designed to imitate a religious sacrifice of personal data to promote awareness of the mundane data disclosure process. This privacy-invasive application blocks all internet access and will only return it if they sacrifice their personal data. The application aims to start
a conversion about our data value and what situations drive us to give up our personal data. Qualitative
data was gathered from user tests and interviews and thematically analyzed. The results showed that the
critical design approach successfully engaged the participants in critical reflection of data disclosure.
The findings also indicated a discrepancy in their privacy attitude and behavior, raising awareness about the context they disclose their information and general expectations for privacy-invasive technologies.
Keywords - Critical Design, Privacy, Rethinking Privacy, Privacy Paradox, Data value, Data Disclosure.
Time and place
22 June at 09:00
Online
Supervisor(s)
Irina Shklovski
External examiner(s)
Eve Elizabeth Hoggan
Title
Inversion of Tail-recursive Term Rewriting Systems
Abstract
Program inversion is a useful program transformation technique and can be used to automatically derive new programs from manually written programs. Many well-known problems are each others inverses, e.g. zip and unzip, and encoding and decoding.
A particular challenge in program inversion is the inversion of tail-recursive programs. It is difficult to automatically generate a functional inverse of a tail-recursive program, even if such an inverse is known to exist. The problem has been studied previously and approaches to solving it have been proposed.
In this thesis we present and evaluate a new method for inverting tail-recursive programs and compare it to approaches in the literature.
We represent programs as term rewriting systems, which are simple but effective systems that are used in many areas of computer science. Since program inversion does not respect programming language paradigms, using term rewriting systems is especially useful because they can model different
programming paradigms, like reversible, functional, and logic programming.
Our method includes a pre-processing phase with a transformation to a symmetric representation of a program. The method then uses a straightforward approach for inversion of the symmetric program. Experiments show that the presented method performs well in many cases, often equally well
or better than existing methods, though the different methods have their advantages and disadvantages.
Time and place
22 June at 13:15
Online
Supervisor(s)
Robert Glück
External examiner(s)
Jan Peter Schneider-Kamp
Title
The need of automation in stock trading, a framework proposal
Abstract
There is a clear distinction of manual stock trading, performed by
professional analysts, and automated trading performed by
computer software. As it would be expected, achieving absolute
efficiency in trading requires a great deal of effort for a person as
the time and attention span is often limited. The various trading
bots were designed to alleviate this issue and provide an automated response to the changes in the stock market. The idea presented in this paper comes on one hand as an analysis of this discrepancy between how humans react to market fluctuation compared to bot, and on the other hand to provide a solution that could merge the needs of a user with the efficiency and the awareness of a bot in an interactive way and personalized behaviour.
Artificial intelligence can be used to enhance the performance of
such bots but it has been observed that such attempts lead to poor results as often the value of such assets is influenced by external factors which can be estimated by machine learning. The user still should have a clear and impactful role in a potential solution. Thus, data should be presented in a clear and meaningful way. On top of that, there should be the option to choose or combine several trading strategies provided by the platform such that one archives the goal either on the short or long term. The user should be allowed to either
trust the bot completely or to get just trading insights and follow up with his own decisions.
A lot of bots nowadays are centered around specific brokers
thus offering a limited customer involvement. This study will be
centered more towards what a user needs rather than what a broker can offer. Through the analyser module, our solution will provide a uniform interface, independent of the broker choice.
Furthermore this study provides a production-ready prototype
that can serve as backbone for future studies in this area.
Time and place
22 June at 15:00
Online
Supervisor(s)
Tariq Andersen
External examiner(s)
Maksim Sorokin
Title
Investigating the Reversibilization of Irreversible Algorithms
Abstract
The area of designing reversible algorithms is largely uncharted as the reversible paradigm only recently gained momentum. With the introduction of reversible programming languages, such as Joule and ROOPL, it became possible to design and validate more sophisticated reversible algorithms [26]. We start from a
set of injective and non-injective irreversible algorithms and investigate their reversibilization into reversible algorithms. For the injective area, we selected compression algorithms for delta and Huffman, and for the non-injective area,
we selected algoritms for breadth-first search and maximum flow. That goals was that the reversible algorithms should perform within the same complexity class as the irreversible ones. We aim to provide findings that can improve the design and development of reversible algorithms. The presented reversible
algorithms will are written in ROOPL and provide the basis for a discussion of the usability and practicality of reversible programming.
Time and place
23 June at 11:00
Online
Supervisor(s)
Robert Glück
External examiner(s)
Ulrik Pagh Schultz
Title
Super-Resolution of Sentinel-2 satellite imagery
Abstract
Image super-resolution is an important class of computer vision techniques which refers to the process of upscaling and enhancing a low resolution (LR) image to a Higher resolution(HR). Traditionally super resolution algorithms are divided into two types: Reconstruction based and Learning based algorithms. However, with recent advances in GPU computation, the last couple of decades have seen a significant amount of progress in the development of deep learning-based super resolution algorithms.
In this thesis, we explore Super-Resolution techniques for enhancing the spatial resolution of Sentinel-2 satellite images. This problem presents its own unique challenges, since the current State-of-Art Super-Resolution algorithms were developed for RGB images, while Sentinel-2 captures multi-spectral images, available at different Spatial Resolutions.
Time and place
23 June at 13:00
Online
Supervisor(s)
Søren Ingvor Olsen
External examiner(s)
Morten Pol Engell-Nørregård
Title
Game Physics Engine for Interactive Fluid Effects
Abstract
There exists a wide variety of different fluid simulation techniques which falls into categories such as Lagrangian, Eulerian, Hybrid or even machine learning methods. The challenge is that a lot of these method were not developed with real time simulations in mind. This thesis investigates current state of the art
methods for fluid simulations. Out of the investigated methods the material point method was selected to be implemented as a prototype in the Unity game engine.
The implemented prototype was tested in terms of its real time performance and compared to an existing fluid simulation library developed by NVIDIA called Flex.
Experiments were performed that showcased that both NVIDIA’s Flex and the prototype were able to run in real time in Unity. Based on these experiments concerns were raised when it comes to application of these methods in actual video games. The primary concern was whether there would be enough computational resources available in actual video games to allow for these sorts of simulations.
This likely means that the only use case for these sorts of methods would be in video games which are very focused on during these types of simulations such that more computational resource could be allocated towards the simulations.
Time and place
25 June at 10:00
Online
Supervisor(s)
Kenny Erleben
External examiner(s)
Morten Pol Engell-Nørregård
Title
Articulated 3D Human Pose Estimation in Climbing
Abstract
The rising popularity of sports climbing has resulted in increased interest in the professional analysis of climbing technique. A 3D model of human pose can be used as a foundation for such in-depth analysis, as well as for comparison between climbers.
There are many different computer vision models for creating 3D human models, and many have been created in just the last few years, largely due to the publication of a large dataset mimicking real-life scenarios, called 3D Poses in the Wild. One such
pose estimation model is called MEVA [1], and it uses a learned latent space of human motion to process predictions from a per-frame model, SPIN [2]. In this work, I train and evaluate versions of SPIN and MEVA on a brand new 2D labeled climbing dataset.
The results are promising and show how MEVA is able to learn complex 3D information, even when only 2D annotations are available. The predictions made by the best model,
MEVA90-Climb, are both accurate and smooth, giving natural-looking motions that can be used for climbing analysis.
Time and place
28 June at 09:00
Online
Supervisor(s)
Kim Steenstrup Pedersen
External examiner(s)
Morten Pol Engell-Nørregård
Title
How to prototype? A study on characteristics and the fidelity of shape-changing interface prototypes
Abstract
Constructing and utilizing prototypes plays an essential role in the advancement of the research field of shape-changing interfaces (SCIs). Currently, not enough work has been done to comprehensively outline the design and focal points of the prototyping procedure. The thesis provides useful insights and
recommendations on how to develop better SCI prototypes. This is accomplished by firstly creating six prototypes that revolve around the concept of a height-changing smartwatch and then conducting a study. Sixteen participants volunteered, while quantitative and qualitative data was collected
on the fidelity and overall performance of the prototypes, along with how they were perceived by the participants. Afterward, the data from the study were analyzed using different techniques and the results unveiled a number of characteristics that affect the fidelity of the prototypes. The most important of these characteristics are the device’s shape-change, the perceived
aesthetic value, the movement resolution, the rigidness and the sounds of the prototypes. Additionally, the findings suggest that users cannot accurately estimate the exact extent of a shape-change and supplementary help is required to assist with the estimations, an observation that also arose from
the participants feedback which advocated for a height-scale. A significant majority of the participants also claimed that future smartwatches should incorporate shape-change functionalities into their designs despite the fact that most of them were unaware of the existence of SCIs beforehand.
Keywords: shape-changing interfaces, prototype, fidelity, study
Time and place
28 June at 10:00
Online
Supervisor(s)
Daniel Lee Ashbrook
External examiner(s)
Timothy Merritt
July
Title
WebAssembly Backends for Futhark
Abstract
Futhark is a high-performance purely functional data-parallel array programming language targeting parallel compute hardware. Futhark has backends for several compute architectures and this thesis adds browsers by targeting WebAssembly and threaded WebAssembly. These are browser technologies which map better to the underlying hardware of devices, including multicore CPUs.
A JavaScript API is developed for easily calling compiled Futhark WebAssembly libraries in the browser. The implementation and generated WebAssembly code is benchmarked for both browsers and Node.js, against the Futhark sequential C and multicore C backends. The sequential WebAssembly performs close to sequential C speeds. The parallel execution of threaded WebAssembly speeds up some example programs by a factor equal to the number of physical CPU cores.
Time and place
09-07-2021 @ 09:00
Online
Supervisor(s)
Troels Henriksen
External examiner(s)
Maja Hanne Kirkeby
August
Title
3D Reconstruction of Transparent Objects
Abstract
In this paper we are going through the theory of Qian et al.[5] for 3D reconstruction of transparent objects, and what we need to implement their setup. We will show how to implement most of Qian et al.’s method, though with some key differences in our setup, which will also
be explained. We will show that the method shows promise with simple objects, like a glass filled with water. In the case of objects with view obstructions, such as a blue dolphin inside a transparent object, we find that we are able to reproduce a lot of the required steps needed for
the reconstruction, but that determining the exact position falters.
Time and place
03-08-2021 @ 9:00 - 10:00
Online
Supervisor(s)
Kim Steenstrup Pedersen
External examiner(s)
Morten Pol Engell-Nørregård
Title
Diagnosis and Prognosis Prediction for Ebola Virus Disease Using Machine Learning Methods
Abstract
Background. Ebola Virus Disease (EVD) is a neglected and deadly, emerging hemorrhagic viral infection with epidemic potential. Due to its incidence in resource-limited settings, diagnosis and management often rely on probabilistic decision-making. However, currently available clinical decision support tools are trained on small datasets fragmented across heterogeneous populations and, as a result, have limited statistical
predictive performance and generalizability.
Aim. In this work, we produce the largest standardized and centralized clinical EVD dataset on which we build and compare diagnostic and prognostic models using a range of machine learning methods, with the aim of creating adaptable models that are better able to adapt to incoming data.
Methods/Findings. The data is derived from the Ebola Data Platform (EDP) of the Infectious Disease Data Observatory which comprises 13558 patients triaged and/or treated for suspected EVD at one of 13 Ebola treatment centers established during the 2014 – 2016 West African Ebola epidemic. These tabular clinical datasets include demographics, clinical signs, symptoms and laboratory values as well as the diagnostic
label of RT-PCR (EVD+/EVD-) and the prognostic label (survival/death).
We 1) construct a standardized data cleaning and alignment pipeline to aggregate all the EDP datasets and 2) perform detailed data understanding on the resulting centralized dataset and its constituent local subsets. We then 3) develop a series of ML models (logistic regression (LogReg), k-nearest neighbors (kNN), support vector machine (SVM)
and random forest (RF)) for the tasks of diagnosis and prognosis. We evaluate 4) the local and central model performances for both predictive tasks. Finally, we 5) determine the most important clinical characteristics for each task.
The diagnostic central model has an average AUC evaluated across local datasets of 0.66 (LogReg), 0.74 (kNN), 0.76 (SVM) and 0.77 (RF). The diagnostic local models has an average AUC of 0.72 (LogReg), 0.76 (kNN), 0.74 (SVM) and 0.79 (RF). The most important diagnostic predictors are determined to be EVD contact history and diarrhea.
Similarly, the prognostic central model has an average AUC: 0.72 (LogReg), 0.71 (kNN), 0.75 (SVM) and 0.72 (RF). The prognostic local models has an average AUC of 0.75 (LogReg), 0.74 (kNN), 0.75 (SVM) and 0.72 (RF). The most important prognosis predictors are determined to be RT-PCR value and patient age.
Conclusion. This work is the first to produce diagnostic and prognostic models for Ebola on a dataset of this size. We also provide an analysis-ready dataset to facilitate further research.
Keywords: Ebola virus disease, epidemology, data pre-processing, machine learning
Time and place
13-08-2021 @ 10:00
Online
Supervisor(s)
Christina Lioma
External examiner(s)
Troels Andreasen
Title
Cascading Abort of Pre-Scheduled Actor Transaction
Abstract
Nowadays, the actor model is widely adopted in building stateful middle-tiers for large-scale interactive applications, where ACID transactions are essential to ensure application correctness. Snapper, an on-going research project conducted in DMS Lab of DIKU, has employed deterministic transaction execution where concurrent cross-actor transactions are pre-scheduled. Snapper improves transaction throughput significantly comparing to conventional dynamic concurrency control method, especially under high contention.
However, Snapper applies speculative execution where a transaction can be executed without waiting for transactions it depends on to commit. Thus, the abort of a transaction will cause cascading abort. Even though Snapper tightly pre-schedules all concurrent transactions, it is inevitable to have transactions abort due to user-defined transaction logic, erroneous transaction input and different types of failures. The aborts of transactions could cause the system result in an inconsistent state and block Snapper to proceed the pre-determined schedules. This thesis analyzed different scenarios of the abort of pre-scheduled transactions in Snapper, provided a formalized definition of the scope of cascading abort, proposed a cascading abort protocol for Snapper which can correctly handle aborts both in single-server and multi-server deployments, and implemented the protocol on Snapper.
An evaluation has been conducted to reveal the characteristics of the proposed cascading abort protocol. The experimental results show that in the singleserver deployment, submitting 10% transactions with user aborts will cause 60% transactions to abort, and get 30% throughput degradation comparing to the one with no user aborts. And Global transactions suffer much more from cascading abort than local transactions.
Supervisor
Yongluan Zhou
External examiner
Philippe Bonnet
Time and place
25 August 2021 at 10:00
Online
Title
Benefits of auxiliary information in automatic teeth segmentation
Abstract
This paper evaluates deep learning methods on segmentation of dental arches in panoramic radiographs. Our main aim is to test whether introducing auxiliary learning goals can improve image segmentation. We implemented three multi-output networks that detect (1) patient characteristics (e.g missing teeth, no dental artifacts), (2) buccal area, (3) individual teeth, alongside the dental arches. These design choices may restrict the region of interest and improve the internal representation of teeth shapes.
The models are based on the modified U-net (Ronneberger et al., 2015b) architecture and optimized with Dice loss. Two data sets, of 1500 and 116 samples, collected at different institutions (Silva et al., 2018; Abdi and Kasaei, 2020), were used for training and testing the methods. Additionally, we evaluated the networks against various patient conditions, namely: 32 teeth, 6= 32 teeth, dental artifacts, no dental artifacts.
The standard U-net architecture reached the highest Dice scores of 0.932 on the larger data set (Silva et al., 2018) and 0.946 on the group of patients with no missing teeth.
The model that outputs probability masks for individual teeth reached the best Dice score of 0.903 on the smaller data set (Abdi and Kasaei, 2020). We observe certain benefits in augmenting teeth segmentation with other information sources, which indicate the potential of this research direction and justifies further investigations.
Keywords: Computer vision, Deep learning, Segmentation methodologies
Supervisor
Bulat Ibragimov
External examiner
Rasmus Reinhold Paulsen
Time and place
30 August 2021
Online
September
Title
Tree count and canopy cover estimation using deep learning with remote sensing imagery
Abstract
Trees are an essential natural resource with ecological and economic importance. They serve as habitat for other animal and plant species, prevent against soil erosion, protect water bodies and serve as
important crops. The availability of very high-resolution satellite images has recently sparked a growing interest in mapping out individual trees at a large scale, using images to understand better their distribution, size and count in regions of interest. While most of the previous approaches have focused on segmenting individual trees, in this work, we focus on two derived quantities, namely, the canopy cover and the count of trees.
Canopy cover estimation is the problem of quantifying the presence of trees in a given area, and it is important for evaluating the effectiveness of forest conservation efforts. At the same time, tree counting
is relevant because it allows collecting information on individual trees, which provides additional insights beyond the canopy cover. For example, the tree count can be used to investigate the role of forests as
carbon sinks, because most of the carbon is stored in the tree trunks.
This work investigates end-to-end trainable deep learning models for segmentation and density estimation to predict the canopy cover and tree count. We consider the case of individual trees in the Sahara
and Sahel-Sudan regions and more dense scenes from Rwanda, using sub-meter high-resolution satellite and aerial imagery. In particular, we analyze ground truth generation using per-pixel and point supervision, various loss functions, and different blocks in a U-Net architecture. Finally, we integrate
the findings for each task in isolation to create a multi-task model that simultaneously attempts to learn canopy cover and count estimation to see the impact on performance in both of the tasks.
We train models that can predict canopy cover and tree count simultaneously with comparable accuracy to models explicitly trained for each task for the Sahara and Sahel-Sudan datasets. Furthermore,
experiments with density models trained on the Rwanda dataset suggest that using per-pixel supervision achieves better count performance for datasets that mostly present very dense scenes.
We perform an experiment to observe the effect of using dilated convolutions in the decoder of an U-Net model that targets canopy cover estimation. The results suggest that this block may allow possible
improvements, but further experimentation on tuning the dilation factors is necessary.
Supervisors
Christian Igel and Ankit Kariryaa
External examiner
Morten Pol Engell-Nørregård
Time and place
3 September 2021
Place TBD, contact uddannelse@di.ku.dk for further information.
Title
Unsupervised Clustering of Sparse Data in Futhark
Abstract
K-means is a basic building block of modern machine learning. As
such, its performance has a critical impact on workflows and explorations
it is involved in. In this thesis, we focus on application domains that involve large sparse datasets and investigate the feasibility of the Futhark
programming language to map k-means and its generalization, mixture
models, to efficient GPU code. We propose a framework that abstracts
from the (possibly sparse) representation of the data while maintaining
the efficiency of sparse representations, where they are used. We demonstrate, that the implementation of k-means, spherical k-means, Gaussian mixture models and von Mises-Fisher mixture models through our framework is possible without the need to explicitly address the underlying data representation. Our k-means implementation yields performance speedups of at least factor 10 over the multicore CPU implementation of scikit-learn and our implementation of Gaussian mixture models with diagonal covariance matrices achieves a speedup of factor 1893 over a single-core CPU implementation that does not support sparse data representations.
Supervisor
Cosmin Eugen Oancea
External examiner
Patrick Bahr
Time and place
3 September 2021 at 09:00
Online
Title
Generative Neural Networks for Ecosystem Simulation
Abstract
Climate change is one of the greatest threats humankind has ever faced. Remote sensing data can be used to monitor changes in the climate and the effect it has on ecosystems.
The aim of this project is to develop a system that can simulate changes in ecosystems using remote sensing data. Generative adversarial networks (GAN) have seen a rapid development in image generation and image translation over the past few years. This type of model has been used before to simulate ecosystem changes, for post-flood scenarios and by using environmental variables to generate realistic images, but these models are either only able to simulate one ecosystem change (flooding) or they do not use contextual information in satellite images.
We have created a deep learning model, building on state of the art loss functions and network architectures. By training the model on remote sensing data from Sentinel-2 we demonstrate that it learns to generate multiple types of realistic changes in a satellite image. Furthermore, it uses contextual information and is able to convincingly preserve objects and realistically simulate changes in the ecosystem.
Supervisors
Stefan Oehmcke and Christian Igel
External examiner
Morten Pol Engell-Nørregård
Time and place
3 September 2021
Place TBD, contact uddannelse@di.ku.dk for more information.
Title
Continuous Collision Detection Using Discrete-oriented Polytope Bounding Volume Hierarchies and Conservative Advancement
Abstract
Collision detection is the concern of this thesis and treated as a three-staged problem consisting of a broad-phase, a mid-phase and a narrow-phase. Alongside the introduction of continuous collision detection, background is given to existing research on methods applicable in each phase but with emphasis on the last two with surrounding discussion of the strengths and weaknesses of known approaches.
Based on the conducted research, theory is turned into practice and a prototype implementation of continuous collision detection is devised for an early staged interactive simulation library and its requirements are identified and taken into consideration. Guided by a testing strategy, the functionality of the implementation is first verified in isolation and then integrated into the library as a seemless, optional feature.
A bounding volume hierarchy of discreteoriented polytopes is used for mid-phase and conservative advancement is performed in the narrow-phase using local optimisation over signed distance fields for an accurate distance tracking routine. Experiments reveal that the prototype implementation works as expected with the assumptions made but lacks in performance for an interactive experience.
Through modular design, this was to be expected as inefficiencies are introduced along the way but opportunity for great acceleration in performance by means of parallel execution becomes possible which is left for future work.
Supervisor
Kenny Erleben
External examiner
Morten Pol Engell-Nørregård
Time and place
3 September 2021 at 10:00
Online
Title
MMLA-Investigating conversational characteristics with waveforms and spectrograms
Abstract
Conversation, defined as the communication between two or more people, is an indispensable part of interpersonal communication and teamwork in our daily lives. During the conversations, people can exchange their thoughts and ideas by hearing each other while observing body languages. As human activities are usually accompanied by conversations, therefore, it is plausible and meaningful to evaluate the activity’s effectiveness by analysing the conversations.
We designed a IoT system data pipeline integrating data collection, data analysis and instructive feedback to evaluate participants’ engagement levels of learning activities (app.9.1). This project concentrates on the feature extractor of the data pipeline which accepts audio data collected from sensors and outputs several conversational characteristics in segments.
We used THCHS-30 and MULTISIMO dataset for training. The THCHS-30 dataset involves 11043 utterances from 25 speakers. Based on the that, we developed a speaker identification model that can predict the speaker of a given utterance from a list registered speakers. The speaker identification
model achieves an accuracy of 94.8% on test set, referring to the probability of correctly identifying speaker. The MULTISIMO dataset contains 23 sessions of group conversations with an average
duration of 10 minutes. With this dataset, we manually labeled the slicing segments with emotional level and overlapped degree according to a pre-defined coding scheme. After that, we created the emotional predictor with the use of random forest, and an overlap detector with the structure of convolution neural network and bidirectional long short-term memory. The emotional predictor has an AUC of 0.88 under the micro-average ROC curve with classifier, and has a MAPE of 17.89% with regressor. The overlap detector (Unaugmented, Weighted, ZCR-Enhanced) achieves an overall accuracy of 0.7290 and an overlapped class f1 score of 0.6824.
In this project, we created three feature extractors, which are speaker identification model, emotional predictor and overlap detector. By giving more time, we believe that those models’ performance can be improved by refining the model structure and augmenting the training dataset in a proper way. Furthermore, we may use the tensorflow lite converter to deploy those models onto microcontrollers and complete the entire data pipeline in the future.
Supervisor(s)
Daniel Spikol
External examiner(s)
Andrea Corradini
Time and place
6 September 2021 at 14:00 - 15:30
Universitetsparken 1, 2-0-04
Title
Fuzziing as a Means of Bug Detection in eBPF
Abstract
In this thesis I examine fuzzing as a means of testing the In-Kernel eBPF Verier. I move the eBPF Verier to User-Space in order to obtain better fuzzing performance, and examine the limits on what fuzzing can ectively do, in the context of ensuring safe behavior of generated inputs. Dierent variations of fuzzers and strategies were applied to the Verier and used to discuss the positive and negative aspects of utilizing fuzzing to test the Verier in isolation and what would be required to utilize fuzzing to test the Verier and BPF as a whole.
Time and place
28-09-2021 at 14:00
01-0-S29 at the PLTC section, HCØ
Supervisor(s)
Ken Friis Larsen
External examiner(s
Philippe Bonnet
Title
Embedding Threads Into 3D-Printed Models
Abstract
3D printing is a fairly new technology that has been revolutionizing the manufacturing industry. It offers the creation of complex objects with various shapes, mostly using only one type of material in the form of filament. However, combining different materials (e.g. thread) during a 3D print can extend the range of capabilities of the printed objects.
This thesis introduces a new type of 3D printing system for effectively embedding threads into 3D objects. The threads can be easily manipulated within the print layers and can be fixed or loose inside a small pipe. This offers additional functionalities to 3D printed objects by taking advantage of thread properties such as elasticity, flexibility and thin shape. At the core of this system is a modified off-the-shelf fused deposition modeling (FDM) 3D printer that has an attached gear ring on the x-axis. There is a thread spool at the exterior of the ring, whose position can be controlled by the system. A python script parses the g-code file of the sliced model and prepares it for the printing process.
An essential contribution of the system is the support of various types of thread (macrame, elastic). The design space of the system is demonstrated by following applications: self-assembling boxes, actuated puppets, an abacus, and a hook. In addition, it contributes open-source firmware, hardware specifications, and 3D models for replication.
Keywords: 3D printing, rapid prototyping, textile
Supervisor(s)
Daniel Lee Ashbrook & Hyunyoung Kim
External examiner(s)
Mikael Brasholt Skov
Time and place
30-09-2021 09:00 - 10:30
Sigurdsgade 41, room 0-11
Title
ClipWidgets: 3D-printed Modular Tangible UI Extensions for Smartphones
Abstract
Touchscreens provide a platform for adaptable and versatile user interfaces making them a popular choice for modern smart devices. However, touchscreens lack physicality. Existing solutions to add tangible user interfaces to smart devices often require complicated assembly or occlude part of the touchscreen. To address the problem, I propose ClipWidgets: 3D-printed modular tangible UI extensions for smartphones. ClipWidgets uses a conical mirror and a custom phone case to redirect the field of view of the rear camera of a smartphone to the phone’s periphery. This allows the phone to optically sense input from modular passive 3D-printed widgets that are attached to the phone case. I developed three different widget types (button, dial and slider) that require no calibration and minimal assembly. To demonstrate the functionality of ClipWidgets I used it to prototype three different applications: a game controller, a music interface and an interactive graph tool.
Keywords: 3D printing, user interfaces
Supervisor(s)
Daniel Lee Ashbrook
External examiner(s)
Mikael Brasholt Skov
Time and place
30-09-2021 11:00 - 12:30
Sigurdsgade 41, room 0-11
Bioinformatics
Title
Image Segmentation Analysis of Small Intestinal Epithelial 3D Confocal Images with Deep Learning
Abstract
The quantitative analysis of cells at different time points of turnover in small intestinal epithelium is of great significance for understanding its population homeostasis and functions. The cell population dynamics are largely regulated by the migrations, proliferation, and apoptosis of cells. Tracing the number of nuclei is a feasible way to investigate the dynamics of thesis cells. By a lineage tracing method and confocal imaging, we capture three-dimensional confocal images of fluorescently marked
stem cells and their progeny. We construct seven U-Net-based deep learning models including two generative models to segment the cell membrane, cell nucleus and fluorescently regions to track the population sizes of cell nuclei located at different three-dimensional targeted regions. Given unified
computation resources and labelled segmentation objects, we train and test these models among which the pixel-to-pixel model generate the best predictions, 95.8 per cent of prediction accuracy, and 88.4 per cent of Dice score on the test set. With pre-processing and post-processing, the model segment
predict that there are 1788.5 24.6 nuclei in the volume of interest.
Time and place
18 June at 10:00
Online
Supervisor(s)
Jon Sporring
External examiner(s)
Morten Pol Engell-Nørregård
Title
Mining the literature to detect connections between lifestyle and diseases
Abstract
Background and Methodology: Text mining is a flexible technique that can be applied
to various tasks in the biomedical field. The association between diseases and genes is
well established in the literature and as such it has been extensively mined and stored
in dedicated databases. However, another factor related to the onset and development
of diseases – lifestyle – is still hidden in the vast sea of texts, and there is no dedicated
database with this information integrated.
In this thesis, I fine-tuned the BioBERT model of natural language processing to
identify lifestyle factors, thereby extending a prototype lifestyle factors ontology. After
completing the expansion, I used the JensenLab dictionary-based tagger to extract
Disease-Lifestyle associations from PubMed. Tagger, an efficient dictionary-based text
mining software, is used both to identify lifestyle factors and diseases in text, and to
find the association between them by considering their co-occurrences within and
between sentences.
Results: After fine-tuning the pre-trained BioBERT model, the model’s prediction
accuracy for the named entity recognition task was 94.61%. This model was used to
predict whether Wikipedia titles with over 1000 matches in PubMed are also lifestyle
factors. After assigning proper thresholds for inclusion and extensive manual
annotation, 447 new terms from Wikipedia titles were added to the prototype ontology
of lifestyle factors. Finally, 501,952 pairs of Disease-Lifestyle associations were
obtained, by running tagger, out of which 50,997 were of high or very high confidence.
Conclusion: This project enriched the lifestyle factors ontology and detected
associations between diseases and lifestyle factors. The manual inspection of results
suggests to a certain extent that when the confidence level is high, the Disease-Lifestyle
associations found through text mining are credible, but further testing is needed to
avoid false positives.
Time and place
22 June at 09:00
Panum, Room 6.2.09
Supervisor(s)
Lars Juhl Jensen, Aikaterini Despoina Nastou, Anders Krogh
External examiner(s)
Jes Frellsen
Title
Using machine learning as a weapon to fight scientific fraud by detecting paper-mill publications
Abstract
With the rapid development of society and economy, the increasingly serious problem
of scientific fraud has attracted public attention. The shadowy companies that fabricate
papers in bulk, the so-called paper mills, are gradually being noticed. In this thesis,
different machine learning-based methods were implemented to detect paper mill
publications. Some known paper mills were collected, and the biggest one called the
Tadpole paper mill is the one mainly used. Through the application of named entity
recognition from text-mining, all papers mentioning non-coding RNA in the Tadpole
paper mill were used as the input data to train supervised machine learning methods,
namely support vector machine, logistic regression, multinomial naive bayes, stochastic
gradient descent, passive aggressive classifier, random forest and XGBoost. Text was
vectorized using the TF-IDF approach and after hyperparameter optimization, the
trained classifiers were applied to other paper mills and papers from 2021 for prediction.
Almost all classifiers achieved good performance with approximately F1-scores of 90%,
proving that they can learn from the specific fraud style rather than the theme. From
prediction results, the classifier shows the ability to only identify fake papers belonging
to the paper mill it was trained on, and also does not have journal bias even if the paper
mill publications concentrate on some specific journals. In addition, the paper mills
seem to have fraud templates or patterns. According to their preference for combining
non-coding RNA and disease as main contents, the function of relationship extraction
was used to obtain papers mentioning such pairs for association analysis. After scoring
for confidence, the results show fake papers mainly focus on under-studied pairs. Such
fake studies linking ncRNA to disease represent a significant threat to science, because
it will pollute under-investigated fields and thereby mislead further research. In
conclusion, the paper mills may have and will definitely continue to seriously damage
the research ecosystem, while it is probable that the machine learning classifiers
working with detecting image duplication could better detect fraud and protect
scientific integrity.
Time and place
22 June at 10:30
Panum, Room 6.2.09
Supervisor(s)
Anders Krogh, Lars Juhl Jensen, Aikaterini Despoina Nastou
External examiner(s)
Jes Frellsen
Title
Rationalizing Clinical Text Makes ICD Codes Assigning More Interpretable
Abstract
The International Classification of Disease (ICD) is a standard disease classification system. Given a clinical text, it is laboring,
time-consuming, and prone to errors to assign the correct ICD codes even for an experienced physician. Many models have been proposed to address the requirement of automatically ICD codes assigning. However, medical texts require explainability for predictions, and there is still debate about whether attention is the explanation. We proposed a Binary Soft Mask (BSM) model, which rationalizes the input text and uses the rationalized text
(rationale) to make predictions. After experiments on MIMIC-III top 50 labels test set, the BSM model performs very closely to some attention models. On average, the BSM model
only needs 38.8% of words per document for
predictions.
Time and place
23 June
Online
Supervisor(s)
Anders Søgaard and Mareike Hartmann
External examiner(s)
Zeljko Agic
Title
Evaluating task- and resting-state functional connectivity differences in major depressive disorder and prediction of antidepressant treatment response
Abstract
Major Depressive Disorder (MDD) is one of the leading mental disorders often characterized by low mood, lack of interest and low self-esteem. It is commonly treated with antidepressant medications such as selective serotonin reuptake inhibitors (SSRIs). However, 30-50% of patients do not respond to the treatment, leading to a prolonged path to recovery and an inefficient allocation of resources. Therefore, discovering the
neurobiological causes underlying the condition and advancing its treatment have been topics of high interest among scientists for years. Hereby we have analysed whole-brain connectivity of healthy controls and depressed individuals, acquired via
functional Magnetic Resonance Imaging (fMRI), in search for biomarkers characterizing MDD or predicting the treatment response. Treatment response was analysed
separately for (non-)remitters (patients who have reached full recovery by week 8) and (non-)responders (patients who are responding to the treatment but may or may
not have reached full recovery by week 8). We ran experiments with Linear Support Vector Classifier, Random Forest Classifier, and Radial Basis Function Support Vector Machine on data combined from resting-state and two task-based fMRI scans (faces
paradigm representing visual and emotional stimuli, and reward paradigm targeting regions involved in reward processing). Our analysis yielded no results, with ROC AUC
scores staying around 50% (i.e. by random) for all prediction goals both in the original feature space, as well as after dimensionality reductions via Principal Component Analysis or feature selections. Visualizing the data revealed no clear group differences
between the healthy and the depressed group, or for either of the treatment response groups. Furthermore, the analysis failed to reproduce findings reported by existing studies, which claim alternated connectivity levels for Amygdala and Hippocampus and within the Default Mode Network.
Time and place
24 June at 13:30
Online
Supervisor(s)
Melanie Ganz-Benjaminsen
External examiner(s)
Rasmus Reinhold Paulsen
Title
Predictive Protein Property Modeling through Gaussian Processes and Encoding Methods
Abstract
Efficient protein design relies on properties of protein variants and plays a major role in the development of bio-technical solutions or pharmacological research. The protein-sequence space is vast, and experimental assessments can be expensive and timely. Therefore, efficient computational methods are required to guide experimental research and derive additional insights in protein design. In this work we integrate two state of the art approaches: the Gaussian Process (GP) based mGPfusion method and the sequence-alignment (VAE) based Deep Sequence method [38, 68]. We propose two novel approaches to incorporate the signal of a protein family latent representation into
a GP regression: (1) by adding the log-ratio of the evidence lower bound as in-silico data and (2) by computing a covariance function from the VAE’s latent representation. We show that mGPfusion is applicable to a range of protein property predictions, apart
from thermodynamic stability, while accounting for noise in the data and epistemic uncertainties. We propose a covariance function that uses the information from
learned representations, which lets us derive a VAE-protein-specific substitution matrix.
The results of this work suggest: that (1) adding a transformed log-likelihood ratio, can improve property predictions, (2) a numerically stable covariance function, can compute log-likelihoods for protein sequences with respect to individual residues from the underlying data and representation. Finally, we evaluate experimentally, how the signal from a latent representations affects GP model predictions.
Time and place
24 June at 10:00
Online
Supervisor(s)
Wouter Boomsma
External examiner(s)
Jes Frellsen
Physics
Title
Cropping MRI scans in cartilage segmentation using Multi-Planar U-Net on knee scans
Abstract
Osteoarthritis is a disease with no known cure that decomposes the cartilage in the joints over time. Automatic segmentations of knee MRI scans are possible to make through the use of Convolutional Neural Networks, but can require a great amount of computer resources and time. Moreover, the cartilage is thin and thus constitutes a very small percentage of a scan. In this thesis I have investigated whether cropping the scans around the cartilage can improve the segmentations as well as saving computational time and resources. The results pointed towards that cropping made no significant difference in segmentation performance though more data and better balanced
validation and test sets are needed to strengthen or weaken this conclusion.
Time and place
14 June at 13:00
Online
Supervisor(s)
Erik Bjørnager Dam og Troels Christian Petersen
External examiner(s)
Dan Witzner Hansen
Statistics
Title
Stochastic Image Deformation
Abstract
TBA
Time and place
9 June at 13:00
Online
Supervisor(s)
Stefan Horst Sommer
External examiner(s)
Steen Markvorsen
Sundhed og informatik
Title
En kvalitativ undersøgelse af mulighederne med NFC-tags på MR-afdelinger i forbindelse med identifikation af indopererede implantater
Abstract
Background: Radiographers have a difficult time identifying implants in patients that need an MRI. This, and the emergence of diagnostic imaging, could cause a bottleneck in patient flow and result in longer diagnostic times. RFID technology has been studied in the identification of implants but not in a real-world setting. Therefore, there is a need to study how this technology could fit in a healthcare setting.
Aim: This study aims to learn and find patterns on how RFID can help in the process of identifying implants and procedures in patients that need an MRI in a clinical setting.
Methods: This study is using Lean Startup as an overall agile method to develop prototypes and tests. A group of 13 radiographers from 3 different hospitals was invited to a participatory design workshop using a scenario-based approach to engage the users and learn about their needs. After the workshop a list of insights was analyzed using a Diagnostic Map and we created an initial User Experience map to visualize touchpoints, thoughts & feelings, emotions, and implications. After
the introduction of our solution, we created a second version of a User Experience map to illustrate and visualize how the change impacted the radiographers. To validate our proposed solution, we conducted two solution interviews with one hospital and a Medical Device Regulative expert.
Results: 11 insights were found during the workshop and the top 3 insights were: Great communication barriers across departments, radiographers need information before booking the patient and radiographers cannot depend on patients providing information. Steps in the User Experience map that generated friction were: Step one: Booking patients and step three: Finding information. Using our insights from the workshop and User Experience mapping, we created a workflow concept called Deep Impact. At the solution interview the radiographers were overall positive and had little to no negative feedback. The Medical Device Regulative expert did express concerns regarding the Medical Device Regulative territory and getting the implant producers to implement the idea.
Conclusion: In conclusion, it seems that NFC technology could be a potential solution in identifying implants in patients needing an MRI. The current workflow is limited by clinicians that do not provide the necessary information that radiographers need to book and safely scan patients.
This, and the fact that patients cannot be expected to provide the information necessesary, results in a cumbersome workflow. Deep Impact seems to solve the issues that radiographers are facing by providing the radiographers with the information prior to booking. However, this has not been validated by key stakeholders as implant manufactures. Deep Impact is therefore still in it’s infancy and requires validation with key stakeholders and more in-depth tests with the end-users.
Time and place
14 June at 10:00
Sigurdsgade, 1. sal
Supervisor(s)
Ditte Ibfelt Andersen
External examiner(s)
Mikael Brasholt Skov
Title
Agil IT-Projektstyring i Region Hovedstaden - En empirisk undersøgelse af Region Hovedstadens Fokusområde Medicins agile projektstyring i forbindelse med udvikling og drift af Sundhedsplatformen
Abstract
With a qualitative method approach this master’s dissertation seeks to explore why the Capitol Region of Denmark chose to implement agile methodology in software development and
maintenance of the electronic health record Sundhedsplatformen. This dissertation furthermore seeks to identify driving as well as restraining forces in regards to keeping (freezing) the agile
methodology as a method in the governance of the electronic health record Sundhedsplatformen.
The aim of this dissertation was to get an overall view and understanding of agile methodology in general as well as in the specific context of Fokusområde Medicin’s use of agile methodology in developing and maintaining Sundhedsplatformen.
This dissertation performed 5 interviews with people from Fokusområde Medicin, with the aim of identifying driving and restraining forces that contribute to the freeze of change, the
implementation of agile methodology introduced.
The findings indicate that the main reasons for the Capitol Region of Denmark to implement agile methodology in software development and maintenance of the electronic health record
Sundhedsplatformen, is because of the large amount of criticism towards the electronic health records system as well as the poor and stagnating user satisfaction of it.
A taskforce of experts were called to propose solutions to the challenges that arose during and after the implementation of Sundhedsplatformen back in 2016. This dissertation finds that it is on the basis of the abovementioned task force’s report on proposed solution, that the Capitol Region of Denmark chose to reorganize the governance of Sundhedsplatformen as well as implement agile methodology to do so.
Furthermore the findings of this dissertation indicates that some of the driving forces for keeping and freezing of the use agile methodology in Fokusområde Medicin, is the motivation to change
caused by the large amount of criticism and poor user satisfaction of Sundhedsplatformen and increased communication between developers and end users.
Some of the restraining forces against keeping and freezing the use of agile methodology are the organizational changes, the changes of workflows and working in teams that for some of the
developers poses difficulties in regards collegial codependency and not being able to independently decide what tasks to prioritize when it comes to developing and maintaining
Sundhedsplatformen.
Time and place
16 June at 13:00
Online
Supervisor(s)
Erling Carl Havn
External examiner(s)
Jens Pedersen
Title
App for patients with anxiety disorder
Abstract
This project examined how patients affected by anxiety disorder deal with their anxiety and how they perceive their treatments and the results thereof. It is well known that the treatments and
therapies offered in the public health care systems are generally inadequate, so the purpose of the project was to use the findings to develop an app addressing the shortcomings to act as a
supplement to conventional treatments.
Anxiety disorder is a complex psychiatric disease with many subdivisions and degrees of severity that can strongly debilitate and incapacitate the patients affected. Thus, all the efforts and
resources invested in search for better treatments are well justified. No two anxiety patients are alike, and the disorder can both build up over a period of time and kick in acutely, and relapses are common after dormant periods. Hence close user involvement and co- creation with a number of
patients was employed in the current project in all stages from initial interviews over defining key daily-life challenges to prototype testing.
The project started with a systematic literature review covering both the disorder and possible apps developed elsewhere. It became clear that development of information-technology supported tools like apps in treatment of anxiety disorder is still in its infancy, and more research is needed to optimize the specific tools in an app to meet the exact needs of anxiety patients and the diversity of the disorder.
The findings from qualitative research, systematic review, and analysis of the most popular apps in Playstore and the data from patient interviews lead up to the prototype app that still within the
project was tested on a small group of patients. Their feedback from use of the app is valuable data for the next iteration of the app in a follow-up project. It is certainly concluded that the app was well received in the patient group, and it holds potential to become a real valuable addition to current treatments for anxiety disorder.
Finally, as a curiosity not to underestimate, it was observed that the mere participation in the project had a therapeutic effect on the patients involved. That observation is in line with the
hypothesis that any activation of anxiety patients out of their passive and inactive state toward productive contributions is highly beneficial. This is a strong motivation to involve patients even more intensely in a possible continuation of the project.
Time and place
22 June at 15:30
lokale 4-0-24 på Biocenteret
Supervisor(s)
Henriette Mabeck
External examiner(s)
Yutaka Yoshinaka
Title
Implementation of welfare technologies in municipalities - an analysis of experiences and implications
Abstract
Problem area: A large number of welfare technologies are implemented every year around Denmark's 98 municipalities and 5 regions, but not all succeed. The implementation depends on several factors and points of attention in the pursuit of a successful implementation. E.g., where and who the technologies are to be implemented for, how many stakeholders are involved and the individual municipalities' amount of resources. In
addition, the complexity and context of the technology can also influence the implementation.
Purpose: This study seeks to identify relevant success and criticism points by implementation of welfare technologies in Denmark's municipalities. The thesis wants to shed light
on positive and negative experiences from two municipalities' welfare technologies implementation projects. This should be used as a recommendation for future implementation projects.
Methods: Qualitative interviews were conducted to examine what experiences the selected projects have dealt with in connection with the implementation. A systematic literature research was performed to uncover the existing knowledge regarding implementation processes in municipalities. Finally, document analysis of written documents was used to supplement the identified patterns and themes.
Results: Various essential insights were associated with the implementation process. It is presented how challenges such as legal barriers, the functionality of the technology in
practice, lack of agility, large vs. small municipalities, varying agendas and diversity among actors can complicate the process. Standardization is hereby mentioned as a possible solution if used with precaution. In addition, it is presented that users should be involved early and preferably throughout the development and implementation process, as implementation of new technology can cause severe impact on existing workflows and
make the user experience for involved actors such as employees and citizens poor.
Conclusion: In order to implement welfare technologies in Denmark's municipalities as best as possible, it is recommended to: 1) Use qualitative methods to uncover users' needs
and identify hidden workflows 2) Work agile and follow a systematic development method 3) Users must be involved throughout development and the implementation
process 4) A socio-technical and information ecology understanding can contribute to a successful implementation. By following these recommendations, it is expected that the
danish municipalities can achieve successful implementation of welfare technologies.
Time and place
22 June at 13:00
lokale 4-0-24 på Biocenteret
Supervisor(s)
Henriette Mabeck
External examiner(s)
Yutaka Yoshinaka
Title
Optimization of the Health Platform through local change - a study of the physician builder program
Abstract
Background:
This master thesis in Health Informatics examines the physician builder program after it was implemented in 2016 as a strategy to improve the condition of the Health Platform. A study conducted in 2019 showed 69% of physicians find the Health Platform does not facilitate their
work (Bansler, J. P, 2021 s. 12). A new study conducted in 2021 show that more than half of the physicians remain unsatisfied with the Health Platform (T. Jensen, F, 2021). This highlights that the physician builder program has not yet had the desired effect, however the reason is unclear.
Aims:
Based upon the experience of the physician builders this study aims to identify problem areas within the program to optimize workflow at the hospitals.
Methods:
This project examines the aims raised using qualitative methods. We conducted 10 interviews with application coordinators and physician builders which make up our entire data collection.
Results:
The results indicate that there are various problem areas throughout the physician builder program. The physician builders have been experiencing issues in terms of working conditions. Furthermore, cooperation between multiple actors in the program has been ineffective, resulting
in delays for the physician builders. Approval procedures has been slow, tedious, and frustrating for the physician builders and organizational changes implemented to the physician builder program has shown unintended effects.
Conclusion:
The findings indicate that the physician builder program is currently running suboptimal and that designing a successful physician builder program is challenging. Organizational changes implemented to the physician builder program has not proven as effective as expected. The
physician builders find it especially frustrating to deal with external factors, approval procedures and collaboration between actors all of which influence the outcome of the builds.
Supervisor(s)
Jørgen P. Bansler
External examiner(s)
Troels Mønsted
Time and place
8 September 2021 at 13:00
Room 2.03 in Sigurdsgade
Title
Agile Project Management in the Capitol Region of Denmark: an empirical study of the Capitol Regions' agile project management in regards to software development and maintenance of Sundhedsplatformen.
Abstract
With a qualitative and phenomenological method approach this master’s dissertation seeks to explore why the Capitol Region of Denmark chose to implement agile methodology in software development and maintenance of the electronic health record Sundhedsplatformen. This dissertation furthermore seeks to identify driving as well as restraining forces in regards to keeping (freezing) the agile methodology as a method in the governance of the electronic health record Sundhedsplatformen.
The aim of this dissertation was to get an overall view and understanding of agile methodology in general as well as in the specific context of Fokusområde Medicin’s use of agile methodology in developing and maintaining Sundhedsplatformen.
This dissertation performed 5 interviews with people from Fokusområde Medicin, with the aim of identifying driving and restraining forces that contribute to the freeze of change, the implementation of agile methodology introduced.
The findings indicate that the main reasons for the Capitol Region of Denmark to implement agile methodology, specifically Scaled Agile Framework for Lean Enterprises (SAFe), in software development and maintenance of the electronic health record Sundhedsplatformen (EPIC), are the incredibly large amount of criticism towards the electronic health records system as well as the poor and stagnating user satisfaction of it.
A taskforce of experts were called to propose solutions to the challenges that arose during and after the implementation of Sundhedsplatformen back in 2016. This dissertation finds that it is on the basis of the abovementioned task force’s report on proposed solution, that the Capitol Region of Denmark chose to reorganize the governance of Sundhedsplatformen as well as implement agile methodology to do so.
Furthermore the findings of this dissertation indicates that some of the driving forces for keeping and freezing of the use agile methodology in Fokusområde Medicin, is the motivation to change caused by the large amount of criticism and poor user satisfaction of Sundhedsplatformen and increased communication between developers and end users.
Some of the restraining forces against keeping and freezing the use of agile methodology are the organizational changes, the changes of workflows and working in teams that for some of the developers poses difficulties in regards collegial codependency and not being able to independently decide what tasks to prioritize when it comes to developing and maintaining Sundhedsplatformen.
Supervisor(s)
Erling Carl Havn
External examiner(s)
Jens Pedersen
Time and place
28-09-2021 at 13:00
UP1 2-0-04