Frei zugänglich
Refine
Year of publication
Document Type
- Master's Thesis (19)
- Bachelor Thesis (4)
- Preprint (2)
- Diploma Thesis (1)
Language
- English (26) (remove)
Has Fulltext
- yes (26)
Is part of the Bibliography
- no (26)
Keywords
- Neuronales Netz (3)
- Application (2)
- Bildgebung (2)
- MRI (2)
- Maschinelles Lernen (2)
- Simulation (2)
- Software (2)
- stroke (2)
- web application (2)
- 3D Slicer (1)
Institute
- Informatik (13)
- Medizinische Informatik (12)
The e-commerce turnover has a constant growth rate of about 10%. An additional increase
in complexity and traffic spikes clarify the need for a scalable software architecture to prevent
a potential technical debt, higher financial cost, longer maintenance, or a reduced reliability.
Due to the fact, that existing approaches like the Palladio Approach require a high modelling
overhead and the importance of dropping this overhead was identified this master thesis is
focused on the modelling and simulation of e-commerce web application architectures using
a high-level approach to provide a faster, but possibly more inaccurate prediction of the
scalability.
This is done by the usage of the Design Science Research Process as a frame, a scientific
literature review for use of the existing knowledge base and the Conical Methodology for the
artefact creation. The artefact is a graphical model which is evaluated using a simulation
developed with Python and its framework SimPy. For model creation and evaluation a total
of twelve papers investigating the scalability of e-commerce web application architectures is
split into a test and train group. The training group and parts of the scientific research are
used to identify the components load balancer, application server, web tier, ERP system,
legacy system and database as well as some general characteristics that need to be considered.
The components with the most modelling variables are the application server and web
tier with a total of thirteen, while the ERP and legacy system only required five.
The model is evaluated using a total of three papers from the test group, where an average
throughput error of 5.78% and a response time error of 46.55% or 26.46% was identified. An
additional evaluation based on two non-e-commerce architectures shows the usability of the
model for other types of architectures. Even though the average error gives the impression,
that the model is not providing a good estimation, the graphical results show, that the model
and its simulation can be used to provide a faster scalability prediction. The model is least
accurate for the prediction of the situation, where the response time increases exponentially,
as this is the point, where variables, only accountable for some percentage and thus ignored
for the model, have the highest influence.
Future research can be found in the extension of the model by either adding or investigating
additional components, adding features ignored within this work or applying it to other
types of web application architectures. Additionally, both the low-level and the high-level
approaches can be brought together to combine the advantages from both approaches.
Development and validation of a neural network for adaptive gait cycle detection from kinematic data
(2020)
(1) Background: Instrumented gait analysis is a tool for quantification of the different
aspects of the locomotor system. Gait analysis technology has substantially evolved over
the last decade and most modern systems provide real-time capability. The ability to
calculate joint angles with low delays paves the way for new applications such as real-time
movement feedback, like control of functional electrical stimulation in the rehabilitation
of individuals with gait disorders. For any kind of therapeutic application, the timely
determination of different gait phases such as stance or swing is crucial. Gait phases are
usually estimated based on heuristics of joint angles or time points of certain gait events.
Such heuristic approaches often do not work properly in people with gait disorders due to
the greater variability of their pathological gait pattern. To improve the current state-ofthe-
art, this thesis aims to introduce a data-driven approach for real-time determination
of gait phases from kinematic variables based on long short-term memory recurrent neural
networks (LSTM RNNs).
(2) Methods: In this thesis, 56 measurements with gait data of 11 healthy subjects,
13 individuals with incomplete spinal cord injury and 10 stroke survivors with walking
speeds ranging from 0.2 m
s up to 1 m
s were used to train the networks. Each measurement
contained kinematic data from the corresponding subject walking on a treadmill for 90
seconds. Kinematic data was obtained by measuring the positions of reflective markers on
body landmarks (Helen Hayes marker set) with a sample rate of 60Hz. For constructing a
ground truth, gait data was annotated manually by three raters. Two approaches, direct
regression of gait phases and estimation via detection of the gait events Initial Contact
and Final Contact were implemented for evaluation of the performance of LSTM RNNs.
For comparison of performance, the frequently cited coordinate- and velocity-based event
detection approaches of Zeni et al. were used. All aspects of this thesis have been
implemented within MATLAB Version 9.6 using the Deep Learning Toolbox.
(3) Results: The mean time difference between events annotated by the three raters
was −0.07 ± 20.17ms. Correlation coefficients of inter-rater and intra-rater reliability
yielded mainly excellent or perfect results. For detection of gait events, the LSTM RNN
algorithm covered 97.05% of all events within a scope of 50ms. The overall mean time
difference between detected events and ground truth was −11.62 ± 7.01ms. Temporal
differences and deviations were consistently small over different walking speeds and gait
pathologies. Mean time difference to the ground truth was 13.61 ± 17.88ms for the
coordinate-based approach of Zeni et al. and 17.18 ± 15.67ms for the velocity-based
approach. For estimation of gait phases, the gait phase was determined as a percentage.
Mean squared error to the ground truth was 0.95 ± 0.55% for the proposed algorithm
using event detection and 1.50 ± 0.55% for regression. For the approaches of Zeni et al.,
mean squared error was 2.04±1.23% for the coordinate-based approach and 2.24±1.34%
for the velocity-based approach. Regarding mean absolute error to the ground truth, the
proposed algorithm achieved a mean absolute error of 1.95±1.10% using event detection
and one of 7.25 ± 1.45% using regression. Mean absolute error for the coordinate-based
approach of Zeni et al. was 4.08±2.51% and 4.50±2.73% for the velocity-based approach.
(4) Conclusion: The newly introduced LSTM RNN algorithm offers a high recognition
rate of gait events with a small delay. Its performance outperforms several state-of-theart
gait event detection methods while offering the possibility for real-time processing
and high generalization of trained gait patterns. Additionally, the proposed algorithm
is easy to integrate into existing applications and contains parameters that self-adapt
to individuals’ gait behavior to further improve performance. In respect to gait phase
estimation, the performance of the proposed algorithm using event detection is in line
with current wearable state-of-the-art methods. Compared with conventional methods,
performance of direct regression of gait phases is only moderate. Given the results,
LSTM RNNs demonstrate feasibility regarding event detection and are applicable for
many clinical and research applications. They may be not suitable for the estimation
of gait phases via regression. For LSTM RNNs, it can be assumed, that with a more
optimal configuration of the networks, a much higher performance is achieved.
Implementation of an interactive pattern mining framework on electronic health record datasets
(2019)
Large collections of electronic patient records contain a broad range of clinical information highly relevant for data analysis. However, they are maintained primarily for patient administration, and automated methods are required to extract valuable knowledge for predictive, preventive, personalized and participatory medicine. Sequential pattern mining is a fundamental task in data mining which can be used to find statistically relevant, non-trivial temporal dependencies of events such as disease comorbidities. This works objective is to use this mining technique to identify disease associations based on ICD-9-CM codes data of the entire Taiwanese population obtained from Taiwan’s National Health Insurance Research Database.
This thesis reports the development and implementation of the Disease Pattern Miner – a pattern mining framework in a medical domain. The framework was designed as a Web application which can be used to run several state-of-the-art sequence mining algorithms on electronic health records, collect and filter the results to reduce the number of patterns to a meaningful size, and visualize the disease associations as an interactive model in a specific population group. This may be crucial to discover new disease associations and offer novel insights to explain disease pathogenesis. A structured evaluation of the data and models are required before medical data-scientist may use this application as a tool for further research to get a better understanding of disease comorbidities.
In this bachelor thesis, different models for predicting the influenza virus are
examined in more detail.
The focus is on epidemiological compartmental models, as well as on different
Machine Learning approaches.
In particular, the basics chapter presents the SIR model and its various extensions.
Furthermore, Deep Learning and Social Network approaches are
investigated and the applied methods of a selected article are analysed in more
detail.
The practical part of this work consists in the implementation of a Multiple
Linear Regression model and an Artificial Neural Network. For the development
of both models the programming language Python was chosen using the
Deep Learning Framework Keras.
Tests were performed with real data from the Réseau Sentinelles, a French
organisation for monitoring national health.
The results of the tests show that the Neural Network is able to make better
predictions than the Multiple Linear Regression model.
The discussion shows ideas for improving influenza prediction including the
establishment of a worldwide collaboration between the surveillance centres as
well as the consolidation of historical data with real-time social media data.
Therefore, this work consists of a state-of-the art of models regarding the
spread of influenza virus, the development and comparison of several models
programmed in Python, evaluated on real data.
Alzheimer’s Disease affects millions of people worldwide, but till today, the gold standard
for definitive diagnosis of this disease is a biopsy. Nevertheless, with the progress
of the disease, a volume loss in the Hippocampus can be observed. Therefore, good
segmentation methods are crucial to facilitate quantification of this loss.
The focus of this work is on the development of a Machine Learning algorithm, more
precisely a Generative Adversarial Network, for the automated segmentation of the
human Hippocampus and its substructures in Magnetic Resonance Images. In particular,
the task is to determine if the integration of a pre-trained network that generates
segmentations into a Generative Adversarial Network scheme can improve generated
segmentations. In this context, a segmentation network in form of a U-net corresponds
to the generator. The discriminator is developed separately and merged in a second
step with the generator for combined training.
With a literature review regarding the automated segmentation of the Hippocampus,
current methods in this field and their medical and technological basics were identified.
The datasets were preprocessed to make them suitable for the use in a neural
network. In the training process, the generator was trained first until convergence.
Then, the Generative Adversarial Network including the pre-trained generator was
trained. The outcomes were evaluated via cross-validation in two different datasets
(Kulaga-Yoskovitz and Winterburn). The Generative Adversarial Network scheme
was tested regarding different architectural and training aspects, including the usage
of skip-connections and a combined loss function.
The best results were achieved in the Kulaga-Yoskovitz dataset with a Dice coefficient
of 90.84 % after the combined training of generator and discriminator with a joined
loss function. This improves the current state of the art method in the same task and
dataset with a Dice index of 88.79 % by Romero [Rom17]. Except of two cases in the
Winterburn dataset, the proposed combined method could always improve the Dice
results after the training of only the generator, even though only by a small amount.
Medical imaging produces many images every day in clinical routine. Keeping up with the
daily image analysis task and this vast amount of data is quite a challenge for radiologists.
However, these analysis tasks can be automated with well-proven automatic segmentation
methods. Segmentation reviewing of an expert is necessary because learningbased
automatic segmentation methods may not perform well on exceptional image
data. Creating valid segmentations by reviewing them also improve the learning-based
methods.
Combining established standards with modern technologies creates a flexible environment
to efficiently evaluate multiple segmentation algorithm outputs based on different metrics
and visualizations and report these analysis results back to a clinical system environment.
The presented software system can inspect such quantitative results in a fast and intuitive
way, potentially improving the daily repetitive segmentation review and rework of a
research radiologist. The presented system is designed to be integrated into a virtual
distributed computing environment with other systems and analysis methods. Critical
factors for this particular environment are the handling of many patient data and routine
automated analysis with state of the art technology.
First experiments show that the time to review automatic segmentation results can be
roughly divided in half while the confidence of the radiologist is enhanced. The system
is also able to highlight individual slices which are essential for the expert’s review
decision. For this highlighting, different metric scores are compared and evaluated.
Spinodal Zr0.4Hf0.6Ni1.15Sn half-Heusler thermoelectrics are synthesized and aged. The complex microstructure due to the spinodal decomposition is investigated by powder X-ray diffraction (XRD), scanning electron microscope (SEM) and transmission electron microscope (TEM). By Rietveld refinements, it is confirmed that excess Ni atoms in the arc-melted and spark-plasma-sintered half-Heusler matrix prefer to form nanoscale/ submicron Heusler precipitates via spinodal decomposition and growth during the aging at 1173 K. Such phase separation changes the band gap of semiconductors, reduces the thermal conductivity drastically and improves the performance of thermoelectrics. As a result, a more than 50 % improvement of the ZT value on the unaged specimens was achieved.
Initial results of an ongoing research in the field of reactive mobile autonomy are presented. The aim is to create a reactive obstacle avoidance method for mobile agent operating in dynamic, unstructured, and unpredictable environment. The method is inspired by the stimulus-response behavior of simple animals. An obstacle avoidance controller is developed that uses raw visual information of the environment. It employs reinforcement learning and is therefore capable of self-developing. This should result with obstacle avoidance behavior that is adaptable and therefore generalizes on various operational modalities. The general assumptions of the agent capabilities, the features of the environment as well as the initial result of the simulation are presented. The plans for improvement and suitable performance evaluation are suggested.
A considerable amount of research in the field of modern robotics deals with mobile agents and their autonomous operation in unstructured, dynamic, and unpredictable environments. Designing robust controllers that map sensory input to action in order to avoid obstacles remains a challenging task. Several biological concepts are amenable to autonomous navigation and reactive obstacle avoidance.
We present an overview of most noteworthy, elaborated, and interesting biologically-inspired approaches for solving the obstacle avoidance problem. We categorize these approaches into three groups: nature inspired optimization, reinforcement learning, and biorobotics. We emphasize the advantages and highlight potential drawbacks of each approach. We also identify the benefits of using biological principles in artificial intelligence in various research areas.
eHMIS is a Ugandan Hospital Information System (HIS), which targets the Sub-Saharan market. In its first version all forms were programmed statically and adaptations were done by code modifications. In 2014 the development of a second version of eHMIS based on Java started.
This work aims at introducing dynamic forms to this new version. While forms that are significantly important to the workflow of the application will remain static, others are replaced by forms that are dynamically designed by the user. By that, the application will become more flexible and local and situational tailoring will be possible without inducing extra costs.
In this thesis the design, implementation and testing of dynamic forms in eHMIS is discussed. The architecture is based on the questionnaire resource of FHIR®. The module enables the user to create questions and group them into sections and questionnaires. For each question the type of answer expected and other constraints can be defined. A user interface covering all functions was designed, so that no programming skills are required. In a first step dynamic forms were integrated in the application's workflow for recording symptoms, though other fields of application are possible. For testing, a usability experiment was conducted in Tororo Hospital in Eastern Uganda, using the thinking aloud method. Results were analysed and evaluated to detect usability problems and gain a general impression of user satisfaction.