Medical Image Processing Lab

Completed diploma projects

[Spring 2012] Optimal graph embeddings for structural connectivity decoding

Student: Ahmed Abdulkadir
Responsible: Jonas Richiardi

We have recently proposed several vector-space representations of functional connectivity grahs to enable machine learning on brain connectivity graphs. The goal of this project is to understand the extent to which structural connectivity must be modelled differently. We will explore other basic representations than the adjacency matrix, and experiment with multiple vertex and graph attributes embedding, including fused vertex attributes. We will also explore principled selection and combination of vertex attributes based on a convex optimisation formulation, as well as various sparse and non-negative matrix factorisation techniques.

[Spring 2012] Hybrid anatomical/data-driven atlasing techniques for improved connectivity decoding

Student: Romain Pirson
Responsible: Jonas Richiardi

Modelling fMRI brain connectivity as a graph involves choosing a spatial level between single voxels and the whole brain to represent abstract vertices in the graph. While atlas-based techniques have been very successful, a recent trend is to use a data-driven technique to derive more functionally coherent graph nodes. However, this often comes at the price of interpretability. The goal of this project is to develop another option in which an anatomical atlas is used as prior information, from which functional subdivisions are then obtained. The problem is formulated as an optimisation procedure, either to obtain partitions maximising the percentage of explained variance ascribed to each regional representative time course, or to obtain partitions which are maximally discriminative between two classes of choice while retaining strict inter-subject alignment.

[Spring 2011] Wave-based analytic sensing and its applications

Student: Vagia Tsiminaki
Responsible: Ivana Jovanovic and Zafer Dogan

We want to advance our insight on analytic sensing for systems governed by the wave equation for localizing and reconstructing acoustic sources inside a bounded acoustic systems from the boundary measurements. Before, we showed that when some special "sensing operators" are applied on the boundary measurements the problem of localizing and reconstructing the acoustic sources can be simplified to an equivalent root-finding problem.
Here, we like to explore different applications of this result (and provide necessary extensions of the theory). We envision two potential applications. First, in optoacoustic tomography lasers are used to heat certain points inside a body; these focal points start to expand and to oscillate. Equivalently, they act as Dirac sound sources and we can sense the sound that they produce on the boundary of the system (on the surface of the body). The goal is to localize these points and to estimate the amount of heat that is absorbed, which is proportional to the absorption coefficient and different tissue can be characterized by this coefficient. Second, in reflection tomography, inside the insonified body, strong reflectors can be directly localized using our "sensing operators". We want to explore the focusing characteristic of our method; i.e., how accurate we can determined small reflectors.

[Spring 2011] Normalization of connectivity measures for the analysis of resting state fMRI

Student: Jose Antonio Lopez Moreno
Responsible: Jonas Richiardi

Functional connectivity measures give access to important aspects of integration between brain regions; i.e., how functional networks are organized and interact. We have developed a data processing pipeline for classification of cognitive states and patient status based on the connectivity patterns as measured by fMRI. One important aspect is proper normalization of the connectivity measures. The project investigates the use of subspace projection methods to remove confounds of the data that are not related to the task at hand. We will also look into alternatives to temporal correlation as functional connectivity measures and evaluate their robustness.

[Spring 2011] Compressed sensing for accelerated diffusion magnetic resonance imaging

Student: Maryna Babayeva
Responsible: Yves Wiaux

---

Completed semester projects

[Fall 2011] Data adaptation for transfer learning in multi-centric functional neuroimaging studies

Student: Ahmed Abdulkadir
Responsible: Jonas Richiardi

Neuroimaging data collection is very costly. There is now an increased tendency to share data and distribute dataset collection across multiple sites. However, MR physics is such that even two scanners from the same brand, model, and OS will not yield identical images. The goal of this project is to develop machine learning techniques to adapt data between sites so that inter-site difference is minimised, and dataset size can be increased to yield more meaningful results. We will acquire an original phantom functional dataset on twin scanners, and implement extractors for several state-of-the-art quality measures for fMRI signals. Using this data, we will derive a convex optimisation problem that allows us to learn a data mapping minimising the differences between scanners, either at the voxel level if perfect alignment can be attained, or at the regional level. In particular, we will seek to minimise the margin in an SVM-like formulation, in order to derive an importance map of voxels or regions contributing to the confusion. Based on this information, we will compute projection matrices that can allow a classifier trained on a site to be used on data from other sites in a principled manner.

[Spring 2011] Learning about the brain using the connectomes database

Student: Veronica Andrade
Responsible: Jonas Richiardi

Recent neuroimaging data-sharing initiatives can provide new insights about the human brain, in particular, about how it is connected in the functional (how do brain regions "talk" to each other) and structural (how are brain regions connected by white matter tracts) sense. The purpose of this project is twofold. First, we want to build a structural connectivity matrix from MR diffusion-tensor-imaging data using the FSL probabilistic tractography method. Second, we want to compare the structural against the functional connectivity.

[Spring 2011] Multidimensional random walk embedding of EEG microstate sequences

Student: Gabriel Cuendet
Responsible: Isik Karahanoglu

Traditional analysis of spontaneous EEG is based on the well-known separation in various frequency bands (eg, power in the alpha band). Another approach, much less known, is to consider the instantaneous topography of the scalp potential, named microstate. These microstate are stable during a time window of 80 ms. Another striking observation is that most of the power in spontaneous EEG can be explained by four microstate topographies only (see Figure). An outstanding question is to understand the syntax of the microstate language; i.e., how do these microstates vary over time. In this project, the aim is to build a realistic simulator of microstate sequences. The implementation will be done in Matlab.
This project is in close collaboration with Dr. Juliane Britz of the EEG Lab headed by Prof. Christoph Michel at the University of Geneva.

[Spring 2011] Graph wavelets for adapted domains in image processing

Reserved: Elena Najdenovska
Responsible: Nora Leonardi

The wavelet transform is an exquisite tool to analyze piecewise polynomial signals; i.e., the smooth part goes into the coarse scales, while the transients (spikes, edges, ...) are captured by coefficients at the fine scales. Recent advances in wavelet design proposed extensions for graphs. The graph is defined by its vertices and edges; the wavelets follow the neighborhood and scaling relationships that are defined in this way. Here, we propose to use the graph wavelets for decomposition of images where the domain of interest is irregular. For example, we want the wavelet transform to be adapted for a particular domain such as the cortical layer in brain images. The implementation will be done in Matlab.

[Fall 2010] Preprocessing of elderly handwriting for early diagnostic support in neurodegenerative diseases

Student: Dat Nguyen
Responsible: Jonas Richiardi

It is well known that fine motor skills, including drawing and handwriting, are affected by dementia and neurodegenerative diseases. Quantitative analysis of these skills via pattern recognition methods may yield complementary markers of disease that can be used in clinical practice to refine diagnosis and prognosis.
The existing body of work on handwriting and drawing processing has so far not focused on processing on-line pen data from an elderly population. The extreme variability found in the population of interest, further compounded by potential disease impact, renders the task challenging. A novel dataset jointly designed with Dr. Armin von Gunten of the Psychiatry department at CHUV, and collected in a real-world medical setting, opens new research avenues and ensures the clinical relevancy of the project.
The goal of this project is to develop a front-end toolkit for signal preprocessing and feature extraction on elderly handwriting and drawing, which will interface with a machine learning backend.
This project is in collaboration with Dr. Andrzej Drygajlo of the speech processing and biometrics group, Laboratory of IDIAP. Matlab and signal processing skills are essential.

[Fall 2010] Benchmark for brain atlases

Student: Manuel Wuetrich
Responsible: Nora Leonardi and Jonas Richiardi

Magnetic resonance imaging provides us with a (non-invasive) window on the brain, both for structural and functional data. Brain atlases can help to reduce the massive amount of voxels available (50'000) to a limited number of brain regions (about 100). In our lab, using atlasing is often a critical step prior to data processing such as brain decoding (i.e., inferring function from the imaging data). However, very different methods of atlasing are available and their comparison is not straigthforward. The aim of this project is to compare the state-of-the-art atlasing methods and to evaluate their quality (on structural data) and implications (for functional data), both for healthy and clinical populations. We will also investigate and develop objective, automated measures of atlasing quality. Matlab and command-line Linux skills (bash/python/PERL scripting) are essential.

[Fall 2010] Analytic sensing for systems governed by the wave equation

Student: Vagia Tsiminaki
Responsible: Djano Kandaswamy and Ivana Jovanovic

Analytic sensing is a new source imaging method that allows multi-source/multi-dipole localization from boundary measurements. First, the sensing principle relies on the divergence theorem to link the boundary measurements to volume information using a test function, the so-called analytic sensor. Second, the annihilation principle allows to find the sources' positions in a non-iterative way. The basic method has been developed for electroencephalography (EEG). In this project, we want to investigate the possible extension of analytic sensing to systems governed by the wave equation.