Index

Thesis Description

Acute ischemic stroke (AIS) is among the leading causes of death (9%) and disability worldwide and will increase over the next 20 years [1, 2]. One of the main causes for an AIS is large vessel occlusion (LVO),accounting for more than one-third of all AIS cases [3]. Thus, making it a key component in the diagnostic process and treatment [4]. These occlusions usually occurs in the internal carotid artery, proximal middle cerebral artery or basilar artery. As a treatment for AIS thrombolysis has been used over the years in a cautious manner, due to the limited application time window and the need to asses the risks to benefits ratio individually. An alternative preferred treatment method is the mechanical thrombectomy, where patients are treated surgically [5, 6]. Thereby, doing a fast classification and possibly a detection of LVOs is crucial for a successful diagnosis and thus for the treatment of AIS [4, 6].

Computed Tomography Angiography (CTA) represents one of the most important techniques in order to detect LVOs leading to the diagnosis and treatment management of AIS, by visualizing the cerebrovascular vessel tree [7]. Detecting such LVOs manually is very time consuming, especially when handling big datasets. Hence an automated classification of patients suffering from LVOs is beneficial for the clinical workflow, by shortening the diagnostic time and improving on accurate classification. For this purpose multiple algorithms were implemented and tested for such application using simple machine learning methods, such as regularized logistic regression, linear support vector machines, and random forest, providing good results as presented by Nishi et al. in [8]. In the last couple of years more sophisticated methods like deep learning, in particular convolutional neural networks (CNNs) as seen in [9, 10] are applied and improve previous algorithms. Still traditional deep learning tends to fail to modulate irregular complex structures, such as cerebrovascular vessel trees. This is mainly due to the fact, that the topology differs between patients [11]. Thus, a plausible alternative should be tested.

Graph deep learning has emerged in recent years as a powerful tool in machine learning, using graphs to model geometrical and topological relationships. It provides an advantage over traditional deep learning regarding irregular structures, such as cerebrovascular vessel trees [11]. Another reason for cosidering GDL for detecting LVOs besides the deformations invariance is the dimension reduction of the inputdata and the ability not to use the Volume as a whole. GDL also identifies and explores a wide range of parallelization strategies using simpler neural networks compared to traditional deep learning, leading to speed and memory efficiency [12, 13].

This thesis aims to create a GDL model to automatically detect LVOs, by analyzing the topological structure using distance features to classify them into LVO left, right or no LVO. This approach will be directly
compared to a state-of-the-art LVO CNN model, like in Thamm et al [9]. In summary, the thesis deals with the following points:

Preparation of distance features for vessel graph
Training of GDL models to analyze the graph topology
Implementation of vessel-tree specific augmentation techniques
Analysis and evaluation of the GDL models

References

[1] Walter Johnson, Oyere Onuma, Mayowa Owolabi, and Sonal Sachdev. Stroke: a global response is needed.
Bulletin of the World Health Organization, 94(9):634–634A, 2016.
[2] Geoffrey A. Donnan, Marc Fisher, Malcolm Macleod, and Stephen M. Davis. Stroke. The Lancet,
371(9624):1612–1623, 2008.
[3] Konark Malhotra, Jeffrey Gornbein, and Jeffrey L. Saver. Ischemic Strokes Due to Large-Vessel Occlusions
Contribute Disproportionately to Stroke-Related Dependence and Death: A Review. Frontiers in neurology,
8:1–5, 2017.
[4] Nikita Lakomkin, Mandip Dhamoon, Kirsten Carroll, Inder Paul Singh, Stanley Tuhrim, Joyce Lee, Johanna T. Fifi, and J. Mocco. Prevalence of large vessel occlusion in patients presenting with acute ischemic
stroke: a 10-year systematic review of the literature. Journal of neurointerventional surgery, 11(3):241–245,
2019.
[5] Murugan Palaniswami and Bernard Yan. Mechanical Thrombectomy Is Now the Gold Standard for Acute
Ischemic Stroke: Implications for Routine Clinical Practice. Interventional neurology, 4(1-2):18–29, 2015.
[6] Salwa El Tawil and Ketih W Muir. Thrombolysis and thrombectomy for acute ischaemic stroke. Royal
College of Physicians, pages 161–165, 2017.
[7] M. D. Ralf W. Baumgartner and et al. Transcranial color–coded duplex sonography, magnetic resonance
angiography, and computed tomography angiography: Methods, applications, advantages, and limitations.
Journal of Clinical Ultrasound, pages 89–111, 2005.
[8] M. D. Hidehisa Nishi, Naoya Oishi, MD, PhD, Akira Ishii, MD, PhD, and et al. Predicting Clinical
Outcomes of Large Vessel Occlusion Before Mechanical Thrombectomy Using Machine Learning. American
Heart Association Journals, pages 2379–2388, 2019.
[9] Florian Thamm, Oliver Taubmann, Markus J¨urgens, Hendrik Ditt, and Andreas Maier. Detection of Large
Vessel Occlusions using Deep Learning by Deforming Vessel Tree Segmentations. EESS.IV, pages 1–7,
2021.
[10] Lucas W. Remedios, Sneha Lingam, Samuel W. Remedios, Riqiang Gao, Stephen W. Clark, Larry T.
Davis, and Bennett A. Landman. Comparison of convolutional neural networks for detecting large vessel
occlusion on computed tomography angiography. Medical physics, 48(10):6060–6068, 2021.
[11] Stavros Georgousis, Michael P. Kenning, and Xianghua Xie. Graph Deep Learning: State of the Art and
Challenges. IEEE Access, 9:22106–22140, 2021.
[12] Si Zhang, Hanghang Tong, Jiejun Xu, and Ross Maciejewski. Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):626, 2019.
[13] Federico Monti, Davide Boscaini, and et al. Geometric Deep Learning on Graphs and Manifolds Using
Mixture Model CNNs. IEEE Xplore, (2017):5115–5124.

Erstellung und Evaluierung eines Messprotokolls für die diffusionsgewichtete MRT im Kontext der Screening-basierten Risikostratifikation

1
In Deutschland erkranken rund 10 % aller Frauen im Laufe ihres Lebens an Brustkrebs
[1]. Verschiedene Risikofaktoren tragen zur Entstehung von Brustkrebs bei, wobei 20-
25 % der Frauen durch ihren familiären Hintergrund betroffen sind. Weitere Faktoren,
die zu einem erhöhten Brustkrebsrisiko führen, sind unter anderem ein dichtes Brustdrüsengewebe, ein höheres Alter bei der Geburt des ersten Kindes sowie ein höheres
Lebensalter, circa 85 % der betroffenen Frauen sind über 50 Jahre alt [2, 3].
Zur Brustkrebsfrüherkennung gibt es verschiedenen Verfahren, unter anderem die
Mammographie, die Sonographie, das Abtasten der Brust sowie die Magnetresonanztomographie (MRT).
Gegenüber der Mammographie hat die MRT den großen Vorteil, dass keine ionisierende Strahlung verwendet wird und im Vergleich zur Mammographie vier bis fünf
Krebstodesfälle mehr pro 100 Frauen entdeckt werden [4, 5].
Beim standardisierten MRT-Verfahren wird ein Kontrastmittel verabreicht, welches die
Darstellung sowohl von normalem Brustgewebe als auch von Brustläsionen verbessert. Die Anreicherung des Kontrastmittels im fibroglandulären Gewebes wird als
Background Parenchymal Enhancement (BPE) bezeichnet und ein erhöhter BPE kann
ein starker Indikator für die Wahrscheinlichkeit von Brustkrebs sein [6, 7].
Ein nicht-invasives Verfahren in der MRT, bei der kein Kontrastmittel verabreicht wird
und gleichzeitig die Auswertung der Brustdichte möglich ist, ist die diffusionsgewichtete Magnetresonanztomographie (engl. diffusion weighted imaging, DWI).
Diffusion beschreibt hierbei die zufällige Bewegung von Wassermolekülen aufgrund
von Wärmeenergie [8]. Die DWI misst diese Diffusionsbewegung im Gewebe [9].
Ein wesentlicher Bestandteil der DWI ist der Diffusionskoeffizient. In der freien Diffusion (ohne Einschränkungen der Bewegung) ist es umso wahrscheinlicher, dass die
Wassermoleküle eine längere Distanz zurücklegen, je länger die Diffusionszeit ist. Die
Diffusionsstrecke ist umso größer, je größer der Diffusionskoeffizient ist [8].
Im menschlichen Gewebe ist der Diffusionsprozess jedoch beispielsweise durch Zellwände eingeschränkt, die die Bewegung der Wassermoleküle begrenzen.
Durch diese Beschränkung ist der Diffusionskoeffizient zeitabhängig und wird ADC
(scheinbarer Diffusionskoeffizient, engl. apparent diffusion coefficient) genannt [8]. Bei
einer Brustläsion können die ADC-Werte helfen, bösartige von gutartigen Läsionen
weiter zu unterscheiden [10].
In dieser Masterarbeit geht es um die Optimierung eines MRT-Bildgebungsstandardprotokolls für die weibliche Brust hinsichtlich der Bildqualität und der Stabilität von quantitativen Parametern. Das optimierte Protokoll soll eine verbesserte multidimensionale Analyse des gesunden fibroglandulären Gewebes der Frau in der Makromorphologie ermöglichen sowie einen Einblick in die Gewebemikrostruktur geben.
Des Weiteren soll das optimierte Protokoll mit dem derzeitigen Protokollstandard verglichen werden. Hierzu werden 10 Probandinnen rekrutiert und mit beiden Protokollen
untersucht. Im Anschluss sollen die Daten miteinander verglichen werden.

2
[1] W. Hiddemann und C. R. Bartram. Die Onkologie: Teil 1: Epidemiologie-Pathogenese-Grundprinzipien der Therapie; Teil 2: Solide Tumoren-Lymphome-Leukämien.
Heidelberg: Springer-Verlag, 2. akt. Auflage, S.134, 2009.
[2] K. Rhiem und R. Schmutzler. Risikofaktoren und Prävention des Mammakarzinoms. In Der Onkologe, Vol. 21, No. 3, S. 202 ff, 2015.
[3] M. Kaufmann et al. Die Gynäkologie. Berlin Heidelberg: Springer-Verlag, 3. Auflage,
S.379, 2012.
[4] N. Hylton. Magnetic Resonance Imaging of The Breast: Opportunities to Improve
Breast Cancer Management. In Journal of Clinical Oncology, Vol. 23, No. 8, S.1678,
2005.
[5] M. Jochelson. Advanced Imaging Techniques for the Detection of Breast Cancer.
In American Society of Clinical Oncology Educational Book, Vol. 32, No. 1, S. 65, 2012.
[6] C. S. Giess et al. Background Parenchymal Eenhancement at Breast MR Imaging:
Normal Patterns, Diagnostic Challenges, and Potential for False-Positive and FalseNegative Interpretation. In RadioGraphics, Vol. 34, No. 1, S. 234, 2014.
[7] V. King et al. Background Parenchymal Enhancement at Breast MR Imaging and
Breast Cancer Risk. In Radiology, Vol. 260, No. 1, S.50, 2011.
[8] F. Laun et al. Einführung in die Grundlagen und Techniken der Diffusionsbildgebung. In Der Radiologe, Vol. 51, No. 3, S. 170 ff., 2011.
[9] M. Freitag et al. Ausgewählte klinisch etablierte und wissenschaftliche Techniken
der diffusionsgewichteten MRT. Im Kontext der onkologischen Bildgebung. In Der Radiologe, Vol. 56, No. 2, S. 137, 2016.
[10] G. Jin et al. The Role of Parallel Diffusion-Weighted Imaging and Apparent Diffusion Coeffcient (ADC) Map Values for Evaluating Breast Lesions: Preliminary results.
In Academic Radiology, Vol. 17, No. 4, S. 457, 2010.

Towards integration of prior knowledge using spatial transformers for segmentation

Description

The task of segmentation of organs or lesions in medical images such as magnetic resonance or computed tomography plays an important role for clinicians. Based on the annotations, volumes and measures can be made that characterize the patient’s disease state and track it over time. However, manual, or semi-automatic segmentation takes a substantial amount of time, requires experienced annotators, and in some cases has large inter-reader variance. Instead of manually annotating large amounts of data by hand, machine learning-based methods allow to propose segmentations without further attention required. Deep learning-based neural networks have achieved outstanding performance for cell segmentation, organ segmentation, and cardiac segmentation. However, prior information about the organ of interest is typically harder to incorporate into deep learning models. Previous works utilized active shape models (ASM) or relied on the principal component analysis to enforce constraints extracted from prior information [4,5].

In this thesis, spatial transformer networks (STN) [1] will be utilized in combination with reference shapes to obtain predictions heavily utilizing prior information. The dataset proposed for this task consists of 100 three-dimensional volumes of the heart [2] (https://acdc.creatis.insa-lyon.fr/).

[2]

Steps

This thesis can be divided into the following steps:

Training of a segmentation network (e.g., nn U-Net [3]).
Application of a rigid 2D STN to the prediction of the segmentation network + reference shape.
Extension to deformable 2D STN.
Extension of STN to 3D input and application to the segmentation prediction.

Figure 1: Example case of the ACDC dataset with overlaying segmentation masks (left) and 3D rendering of the segmentation mask (right) [2].

Requirements

Good understanding of machine learning/deep learning concepts
Strong knowledge of Python and ideally Pytorch is required

References

[1] Jaderberg, Max, et al. “Spatial transformer networks.” arXiv preprint arXiv:1506.02025 (2015).

[2] O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, et al. “Deep Learning Techniques for Automatic MRI Cardiac Multi-structures Segmentation and Diagnosis: Is the Problem Solved ?” in IEEE Transactions on Medical Imaging, vol. 37, no. 11, pp. 2514-2525, Nov. 2018

[3] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2020). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 1-9.

[4] Milletari F., Rothberg A., Jia J., Sofka M. (2017) Integrating Statistical Prior Knowledge into Convolutional Neural Networks. In: Descoteaux M., Maier-Hein L., Franz A., Jannin P., Collins D., Duchesne S. (eds) Medical Image Computing and Computer Assisted Intervention − MICCAI 2017. MICCAI 2017. Lecture Notes in Computer Science, vol 10433. Springer, Cham.

[5] Ahmadi, S.-A., Baust, M., Karamalis, A., Plate, A., Boetzel, K., Klein, T., Navab, N.: Midbrain segmentation in transcranial 3D ultrasound for parkinson diagnosis. In: Fichtinger, G., Martel, A., Peters, T. (eds.) MICCAI 2011. LNCS, vol. 6893, pp. 362–369. Springer, Heidelberg (2011)

Deriving procedure definitions from examination data by means of machine-learning

Project_Description_David (1)

Killer Whale Sound Source Localization Using Deep Learning

1 Introduction
Sound source localization (SSL) is not necessarily a new eld and much has been done in the analytical
domain using multiple microphones and utilizing the distance between the microphones as well as the Time
Delay of Arrival (TDOA) to extract position information [1]. The advances in recent years of machine
learning and deep learning techniques as well as the increasing availability of powerful hardware have opened
up new pathways to solving SSL and Sound Event Detection (SED) tasks. These methods are of particular
interest due to their reported robustness when it comes to dealing with noise as well as their performance
in comparison to conventional methods [2]. Most uses of SSL seem to be involved in human tracking and
relatively little has been done with focus on other animals and even less has been done in nature as opposed
to a closed room. This project aims to utilize deep learning SSL methods to locate orcas by using calls
received by an 8-microphone array being pulled by a boat as presented in [3].
2 Problem Description
The localization of orcas based on their emitted calls presents several problems, not the least of which is
determining the actual distance of the orcas from a certain position. While several methods such as [4] can
accurately locate the position of a sound source in relation to a unit sphere around the microphones, depth
perception presents another problem. This problem is compounded by the fact that the assumption simply
cannot be made that two dierent orcas produce calls at the same amplitude. Therefore an accurate depth
estimation can only be made with a varying degree of certainty.
This audio depth perception problem is further compounded by the fact that sound waves travel at
dierent speeds in water depending on certain circumstances like temperature, salinity, and other objects
present in the water, which, when one is in nature, is a variable which is often completely out of the control of
the observer [5]. In addition to the environmental problems lending to the diculty of audio depth perception
a further issue is encountered when dealing with orcas in that the animals can be several hundred meters
away and the produced calls will still reach the microphones with enough intensity to be registered as a call
of interest which increases the viable range and adds more uncertainty.
An additional problem ecountered in the eld is while it can be easy to see an orca if they happen to
be near the surface, it is another issue entirely to associate the produced calls of a particular animal to the
individual which created these calls.
Finally what is necessary in an environment of observation is a tool which can quickly and accurately
associate the produced calls of an animal with the location of production. If this assignment is not quick or
accurate enough it can easily be the case that the orca has since moved on and the location information is
no longer necessary nor accurate.
Master’s Thesis Proposal
A Deep Learning Toolkit for Killer Whale Localization Based On Emitted Calls Alexander Barnhill
3 Goals
The goal of this project is then to develop a toolkit which functions in conjunction with the ORCA-SPOT
toolkit; that is, when ORCA-SPOT has determined that a call has been produced and this call is of interest,
the location of the call producer will be determined quickly enough to be used during active observation.
This means that the location information should be produced quickly enough to tell researchers with enough
accuracy both in time and location to be able to say with relative certainty win which direction the orcas
being observed are currently located.
In addition to this the distance of the animal should be determined as accurately as possible by gathering
enough information to say with relative certainty, depending on the intensity of the received signal, how far
removed the producer of the signal is from the point of reception.
In addition to this, time permitting, the tool should also allow for a ner analysis of received calls, possibly
including the possibility to associate the received calls with particular animals.
4 Project Plan
In order to accomplish these goals a multi-step approach will be undertaken. First simulated data from
PAMGuard will be taken and processed in order to simulate orcas at random positions. The benet here is
that real orca samples can be used but the dataset can be expanded as much as desired in order to give not
only a large but also a representative dataset containing a wide variety of signals, positions, and amplitudes.
This dataset will be further processed by adding varying amounts of noise to the samples in order to increase
the robustness of the toolkit.
After the dataset is produced various network architectures will be tried in order to:
a) Produce a network which reliably and accurately determines the source of the sample and
b) Is small enough to quickly provide information at inference time in order to enable the localization of
orcas in real time.
During experimentation with architectures methods will also be tested to estimate distance of the orcas
including attempting to segment the distance measurements into discrete areas and then applying some
Gaussian model to these areas in the hopes of achieving a reasonable estimate of the distance of the orca
based on training data as well as amplitude of the received signal.
This toolkit will then be integrated with ORCA-SPOT to continuously accept samples from the ORCASPOT
toolkit with the goal of then localizing samples which have been deemed interesting by ORCA-SPOT.
The toolkit will have to process samples containing varying amounts of signals and samples in which the
interesting signal occurs at varying points within the sample.
4.1 Proposed Schedule
1. Preparation and generation of data for training: 2 weeks
2. Investigation of architectuers with respect to performance including experimentation and testing: 2
months
3. Analysis and implementation of depth estimation: 1 Month
4. Integration with ORCA-SPOT: 1 Month
5. Summary of results: 1 month
Master’s Thesis Proposal
A Deep Learning Toolkit for Killer Whale Localization Based On Emitted Calls Alexander Barnhill
References
[1] X. Bian, Gregory D. Abowd, and James M. Rehg. Using sound source localization to monitor and infer
activities in the home. 2004.
[2] Nelson Yalta, Kazuhiro Nakadai, and Tetsuya Ogata. Sound source localization using deep learning
models. Journal of Robotics and Mechatronics, 29:37{48, 02 2017.
[3] Christian Bergler, Hendrik Schroter, Rachael Xi Cheng, Volker Barth, Michael Weber, Elmar Noth,
Heribert Hofer, and Andreas Maier. Orca-spot: An automatic killer whale sound detection toolkit using
deep learning. Scientic Reports, 9(1):10997, 2019.
[4] Sharath Adavanne, Archontis Politis, Joonas Nikunen, and Tuomas Virtanen. Sound event localization
and detection of overlapping sources using convolutional recurrent neural networks. CoRR,
abs/1807.00129, 2018.
[5] Jens Blauert. Spatial hearing: the psychophysics of human sound localization. 01 2001.