Jump to contentJump to search

Abschlussarbeiten und Lab Rotations

Aufgrund des großen Interesses an Abschlussarbeiten können wir für das Sommersemester 2023 leider keine weiteren Projekte anbieten.

Wir bieten Bachelor- und Masterarbeiten sowie lab rotations in unserer Forschungsgruppe am Forschungszentrum Jülich an. Im Folgenden finden Sie eine Liste aktuell offener Projekte. Sprechen Sie uns gern an.

Factorization of learned cytoarchitectonic features into disentangled features

We have deep neural networks to learn features from high-resolution microscopic images. These features encode the cytoarchitecture of the tissue, but they are inherently entangled. This makes them difficult to interpret. The goal of this project is to evaluate factorization/disentanglement methods to learn disentangled deep learning features.

Self-supervised Segmentation of Blood Vessels in 3D-PLI Images

3D-PLI images provide a comprehensive view of brain tissue with high resolution. In addition to nerve fibers, blood vessels and even the smallest capillaries that interfere with the signal of the nerve fibers can be observed in the images. For an automatic analysis of 3D PLI images, the identification of such vascular structures is a crucial step. Initial results show that the structure of larger blood vessels can be consistently detected using deep learning segmentation models. However, due to the high annotation overhead, the recognition of the small capillaries remains an open challenge. This thesis will investigate recent advances in self-supervised pre-training of deep neural networks for image segmentation in order to obtain a segmentation of the vascular structures in the brain down to the smallest capillaries with minimal human effort.

Tackling the open set problem in cytoarchitectonic mapping with deep learning

We have deep neural networks to predict cytoarchitectonic brain areas from microscopic images. However, the set of classes is incomplete: There are brain areas which are not yet named or studied. Models need to account for this, e.g., by being able to let users know when they "don't know" an observed area. The goal of this project is to evaluate methods from the field of open-set classification for the task.

Unsupervised deep learning for artifact detection in histological brain images

Preparing and scanning histological very thin brain sections is a sensitive task. Many, but not all, artifacts (for examples tears in the tissue or crystallization of the staining) can be avoided. In order to get the best results these artifacts need to be detected (and later on removed). The goal of this project is to evaluate unsupervised deep learning methods for the detection of artifacts.

Responsible for the content: