Index

AI based Localization of Ischemic Heart diseases using Magnetocardiography signals

Description – Magnetocardiography (MCG) is a functional imaging system that is used for medical heart diagnosis. The so-called heart current induces a magnetic field that can be measured by an MCG system.This current results from the excitation of the heart muscle cells. Thus, the signal of the magnetic field exactly maps the electrophysiological activity of the heart. Due to this passive measuring technique, an MCG does not need any electromagnetic radiation. In comparison to an electrocardiography system (ECG), the heart signal is not evaluated by different ECG lead measurements. The MCG signal is acquired in the proximity of the heart. The signal can be detected almost without inhomogeneous influences or other restrictions. In particular, the MCG is more sensitive to tangential current as an ECG. Furthermore, an MCG is additionally sensitive to the vortex current (e.g. perpendicular to the tangent space) that cannot be detected by an ECG system. The MCG reconstruction task is based on the Biot-Savart law and leads to an inverse problem. Hence, its solution is also known as pseudo current. Biomagnetik Park GmbH (bmp) provides a high data quality with its MCG technology.The Superconducting Quantum Interference Devices (SQUIDs) can detect the MCG signal in the femto Tesla range. The medical sensitivity can be achieved currently up to approximately 97%  (Park et al.1 Dobutamine stress magnetocardiography for the detection of significant coronary artery stenoses – A prospective study in comparison with simultaneous 12-lead electrocardiography).The article (Tao et al. Magnetocardiography based Ischemic Heart Disease Detection and Localization using Maschine Learning Methods) demonstrates that the potential of magnetocardiography is not exhausted analytically. Since the early detection of ischemic heart diseases based on MCG imaging is already ensured, e.g. refer to (Park and Jung Qualitative and quantitative description of myocardial ischemia by means of magnetocardiography), the functionality shall be now extended to a localization feature. i.e. it should be focused, where the stenosis is placed. The lack of morphological information in the signal results in an inverse problem that complexity increases proportionally to the accuracy of the model

Thesis objectives – To analyze and evaluate state-of-the-art deep learning methods, e.g. shown in (Maier et al. A gentle introduction to deep learning in medical image processing), in order to enable the MCG based localization of ischemic heart disease. Therefore, the publications Tao et al. should be used as a conceptual guideline. The objective of the master thesis consists in reconstructing at least the results of this article. In particular, the results shall be enhanced by the application of deep learning optimization methods and the sensitive bmp MCG data.

Rigid Registration of Bones for Freely Deforming Follow-Up CT Scans

Estimation and evaluation of a CT image based on electromagnetic tracking data for adaptive interstitial multi-catheter breast brachytherapy

The second leading cause of death worldwide is cancer. In 2018 9.6 million people died of cancer [1]. Additionally, in 2018 24.2% of the cancer incidences in women were breast cancer [2]. Common therapies for breast cancer are chemotherapy, surgery, and radiation therapy [3]. There are two options for radiation therapy: whole breast irradiation and the accelerated partial breast irradiation [4]. One well-established possibility to apply accelerated partial breast irradiation is the interstitial multi-catheter brachytherapy (iBT), a treatment technique that uses γ-radiation from enclosed radioactive sources positioned very close to the tissue to be irradiated [5]. In order to guide the radiation source there, multiple plastic catheters are inserted surgically and remain within the body throughout the course of the treatment (typically five days). A CT scan, the so-called planning CT (PCT), serves as a basis for treatment planning.
At the University Hospital Erlangen, the standard protocol for iBT consists of nine treatment fractions of 3.8 Gy per fraction, administered within five consecutive days. To date, the treatment plan remains unchanged throughout this period and hence does not account for interfractional changes. However, in a study assessing the need for treatment adaptation in the course of the treatment it was found that 4% of the patients would have benefitted from replanning [6]. In our workflow, the only measure to ensure the correct position of the catheter implant is a follow-up CT (FCT) after the fourth fraction. However, this exposes the patient to additional dose and therefore cannot be performed prior to every fraction.
The aim of this Master’s thesis is to develop a CT estimation (estCT) based on the data from the PCT and a dose-free electromagnetic tracking system. The first step is calculating the deformation vector field from the EMT data acquired immediately after the PCT (without moving the patient in between) and the EMT data acquired immediately after the FCT. In a second step, this deformation field is used to warp the PCT to yield the estCT. The estCT can then be compared to the FCT as a ground truth.
If the estCT could accurately “reproduce” the FCT, this method would allow describing interfractional changes of the catheter geometry without additional dose exposure. Since EMT measurements are easy and fast to conduct at any time point in the course of the treatment, estCTs could be acquired prior each treatment fraction. Finally, these estCTs can be evaluated dosimetrically to trigger patient-specific changes to the treatment plan, which would mean an essential step towards adaptive brachytherapy.
Master’s Thesis Description
References
[1] World Health Organization: “Cancer”, 2018, https://www.who.int/news-room/fact-sheets/detail/cancer, (Date last accessed 2018-06-19)
[2] BRAY, Freddie, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians, 2018, 68. Jg., Nr. 6, S. 394-424.
[3] WAKS, Adrienne G.; WINER, Eric P. Breast cancer treatment: a review. Jama, 2019, 321. Jg., Nr. 3, S. 288-300.
[4] STRNAD, Vratislav, et al. Practical handbook of brachytherapy. UNI-MED Verlag, Bremen-London-Boston, 2014, S. 166-183.
[5] NJEH, Christopher F.; SAUNDERS, Mark W.; LANGTON, Christian M. Accelerated partial breast irradiation (APBI): a review of available techniques. Radiation Oncology, 2010, 5. Jg., Nr. 1, S. 90.
[6] KALLIS, Karoline, et al. Is adaptive treatment planning in multi-catheter interstitial breast brachytherapy necessary?. Radiotherapy and Oncology, 2019, 141. Jg., S. 304-311

Deep Learning based Beamforming for Hearing Aids

Motivation
The human brain’s ability to focus its auditory attention on a stimulus while filtering out a range of other stimuli is a known neurological phenomenon (cocktail party problem). Reduction of unwanted environmental noises is, therefore, an important feature of today’s hearing aid devices. Most traditional noise-reduction and source-segregation schemes use limited signal properties of noisy environments, effectively only exploiting first and second order statistics. Since deep networks are able to learn a complex, nonlinear mapping function, this makes them ideal candidates for noise reduction tasks, where complex priors (speech and distortion signal properties) must be modeled.
Single-channel noise reduction aims to solve this problem making use of a single microphone only. It is however known that the signal-to-noise ratio is typically improved by making use of directional microphones, to exploit multi-channel signals.
Deep learning-based noise reduction has already been explored yielding good results on single-channel signals. We aim to support the hearing impaired in a noisy environment by improving an already existing deep-learning based noise reduction framework using multi-channel signals, which enables to exploit directional information.

Topic
To incorporate multi-channel signals in a deep-learning framework for noise reduction, we aim to use beamforming. Beamforming is a signal processing technique for directional signal transmission or reception and works by eliminating undesirable interference sources and focusing transmitted signals on a specific location.
We propose to use as data multi-channel noise signals from hearing aids. These speech signals are cleaned using some signal processing and transformed with HRTFs (head-related transfer function). Multiple channels and positional information of the microphones can be used to estimate the beamforming coefficients.

Modelling context transitions in picture descriptions of Alzheimer patients

Deep keyword recognition in speech exercises for aphasia patients

Solar Cell Aging Prediction using Deep Learning Image2Image Translation

Cracks in solar cells, caused in production or assembly, can considerably affect the degradation of a cell in the field [1]. Predicting the impact of these cracks improves quality control and helps to cope with degradation throughout the lifetime of a solar cell.

Although information about photovoltaic module degradation has been available since the early 1970s, predictions for different types of degradation are still poorly studied [2].

This thesis aims at developing a new approach to predict the aging of solar cells using Deep Learning, given an initial electroluminescense (EL) measurement of the latter. The data used in this thesis consists of 2 measurements of 94 modules at different points in time. We will

1. use this dataset to train an unpaired Image2Image approach (e.g. CycleGAN [3]) to assess, if the network is capable of learning the relationship between initial and aged measurements from data, using unpaired datasets only
2. extend the approach in I. to incorporate the additional information available from using pairs of initial and aged measurements.

Since the initial and aged measurements are not registered exactly, we aim to design a custom loss function in 2. that is invariant to small registration mismatches, but enforces consistency between cracks in generated and real aged measurements. We want to assess, if the weakly supervised crack segmentation by Mayr et al. [5] can be used for that purpose. To this end, we plan to enforce consistency between the coarse segmentation maps of real and aged measurements. This can be seen as an extension to the common combination of adversarial loss with L1/L2 distance between fake and real target image [3].

The purpose of the L1/L2 distance in CycleGAN can be seen as enforcing consistency between input and output of the generative network. Since our custom loss compares the generated fake cell to the real aged cell, the input/output consistency for the generative network can possibly be ensured without taking into account the L1/L2 distance between generator input and output. Apart from combining the two common losses with our custom loss, we therefore want to evaluate whether we can get better results by replacing the L1/L2 distance altogether.

A prototype of this network will be realized in Pytorch, based on implementations of [3] and [4].

References:

[1] Quintana, M.A., King, D.L., McMahon, T.J., Osterwald, C.R., 2002. Commonly observed degradation in field-aged photovoltaic modules. In: Proc. 29th IEEE Photovoltaic Specialists Conference, pp. 1436–1439.
[2] Ndiaye, A., Charki, A., Kobi, A., Kébé, C.M., Ndiaye, P.A. and Sambou, V., 2013. Degradations of silicon photovoltaic modules: A literature review. Solar Energy, 96, pp.140-151.
[3] Zhu, Jun-Yan, Taesung Park, Phillip Isola, and Alexei A. Efros. “Unpaired image-to-image translation using cycle-consistent adversarial networks.” In Proceedings of the IEEE international conference on computer vision, pp. 2223-2232. 2017.
[4] Mayr, Martin, Mathis Hoffmann, Andreas Maier, and Vincent Christlein. “Weakly Supervised Segmentation of Cracks on Solar Cells Using Normalized L p Norm.” In 2019 IEEE International Conference on Image Processing (ICIP), pp. 1885-1889. IEEE, 2019.

Predicting Hearing Aid Fittings Based on Audiometric and Subject-Related Data: A Machine Learning Approach

Hearing aids (HA) are configured to the wearer’s individual needs, which might vary greatly from user to user. Currently, it is common practice, that the initial HA gain settings are based on generic fitting formulas that link a user’s pure-tone hearing threshold to amplification characteristics. Subsequently, a time-consuming fine-tuning process follows, in which a hearing care professional (HCP) adjusts the HA settings to the user’s individual demands. An advanced, more personalized gain prescription procedure could support HCPs by reducing fine-tuning effort and facilitate over-the-counter HAs. We propose a machine learning based prediction for HA gain to minimize subsequent fine-tuning effort. The data-driven approach takes audiometic and personal variables into account, such as age, gender, and the user’s acoustical environment.

A random forest regression model was trained on real-world HA fittings from the Connexx database (fitting software provided by Sivantos GmbH). Three months of data from Connexx version 9.1.0.364 were used. A data cleaning framework was implemented to extract a representative data set based on a list of machine learning and audiological criteria. These criteria include, for instance, using only ‘informative’ HCPs who perform fine-tuning for at least some patients. Furthermore, ‘informative’ HCPs are those who perform diagnostics beyond air conduction audiograms, use new technologies and special features. The resulting training data comprised 20,000 HA fittings and a 10-fold cross validation was used to train the random forest.

Deep Learning-based Spectral Noise Reduction for Hearing Aids

The great success of deep learning-based noise reduction algorithms makes it desirable to also use them for hearing aid applications. However, neural networks are both computationally intensive and memory intensive, making it challenging to deploy on an embedded system with limited hardware resources. Thus, in this work, we propose an efficient deep learning-based noise reduction method for hearing aid applications. Compared to a previous study, where a fully-connected neural network was used to estimate Wiener filter gains, we focus on using Recurrent Neural Networks (RNNs). Additionally, convolutional layers were integrated. The neural networks were trained to predict real-valued Wiener filter gains to denoise the noisy speech spectrum. Normalizing the input of the neural network is essential. Therefore, various normalization methods were analyzed, allowing low-cost real-time processing. The presented methods were tested and evaluated on unseen noise and speech data. In comparison to the previous study, the computational complexity and the memory requirements of the neural network were reduced by a factor of more than 400, the complexity of the normalization method by a factor of over 200, while even reaching a higher denoising quality.

Multi-Task Learning for Speech Enhancement and Phoneme Recognition

For speech intelligibility, consonants have a fundamental importance. Unfortunately, when reducing noise in speech, consonants are often also degraded while vocals are easier to preserve/enhance. To improve the detection and enhancement of consonants, we want to use multi-task learning to reduce the noise in the signal and furthermore detect phonemes (smallest acoustic unit in speech).