We perform teaching and research in machine learning strategies for the pattern analysis of various kinds of data. This comprises statistical models for clustering, graphical models for network inference and algorithmic methods to efficiently find these structures in the data.

Contact Info
CAB F 61.1
Universitaetstrasse 6,
8092 Zurich
Schweiz

+41 44 632 64 96

Follow Us

Opportunities

For more information please contact the corresponding advisor(s)

Data Driven Sleep Stage Classification

Ami Beuret Prof. Dr. Joachim Buhmann

Sleep phase classification is an important step to study the biology of sleep in both humans and animals. Previously, sleep phase identification has been done by tedious visual inspection of the recorded electroencephalogram (EEG) data. In addition to be a laborious task, this visual inspection is prone to disagreements between experts. Recently, machine learning approaches have automated the classification of sleep phases from recorded EEG and electromyogram (EMG) data, accelerating the speed of analysis while being a more objective assessment of the sleep states.
Previous state-of-the-art has employed convolutional neural networks (CNNs) on spectograms of EEG and EMG data as a first step. Then in order to impose physiologically plausible state transitions, either hidden Markov models (HHMs) [Miladinovi ́c et al., 2019] or long short-term memory (LSTM) modules ([Li et al., 2022]) are employed. Unfortunately there are still shortcomings with current methods. The achieved accuracy of these methods are still not high enough for many practical research purposes [Yamabe et al., 2019]. Additionally, these methods fail quickly in presence of minor distributional shifts [Barger et al., 2019]. Furthermore, In many practical cases, only one single EEG channel might be available. A more useful algorithm should achieve the same accuracy with only one EEG channel in contrast to many existing methods.
In this proposal we would like to explore alternative architectures such as vision transformers and multi-head attention modules in order to both extract features from spectograms and learn the temporal consistency between different sleep stages [Vaswani et al., 2017, Dosovitskiy et al., 2021]. Recent advances in attention modules have demonstrated equal or superior performance to the CNNs in various vision tasks.
**This master thesis will be the continuation of our previous work https://sleeplearning.ethz.ch. No special skills are required for the thesis beside the usual ptyhon skills and specially very high motivation.**


Proposal

Deep-Learning based detection and 3D localization of endometriosis in ultrasound videos

Fabian Laumer, João Borges de Sá Carvalho Prof. Dr. Joachim Buhmann

Endometriosis is a common gynecological condition affecting an estimated 10 percent of women of childbearing age. Women with endometriosis develop tissue that looks and acts like endometrial tissue outside of the uterus, usually on other reproductive organs inside the pelvis or in the abdominal cavity. Each month, this misplaced tissue responds to the hormonal changes of the menstrual cycle by building up and breaking down just as the endometrium does, resulting in small bleeding inside of the pelvis. This leads to inflammation, swelling and and surrounding tissue can become irritated, eventually developing scar tissue and adhesions.
Laparoscopy, a minimally invasive surgical procedure, can be used to definitively diagnose and treat endometriosis. On the other hand, pelvic ultrasound is a noninvasive diagnostic exam that uses sound waves to produce images that are used to assess organs and structures within the female pelvis. A pelvic ultrasound allows quick visualization of the female pelvic organs and structures, including the uterus, cervix, vagina, fallopian tubes and ovaries. In this work, we are interested to develop an algorithm that is able to automatically and reliable detect endometriosis in pelvic ultrasound videos by analysing the sliding motion of the uterus. Additionally, we want to support medical doctors during the examination by visualizing - in a surrogate 3D model of the uterus - which parts of the uterus were already examined and which not, and where potential adhesions are located.
In order to evaluate the quality of our approach, we will use a in-house dataset consisting of over 100 patients that underwent pelvic ultrasound examination.
More details can be found in the link bellow.


Proposal