Biomedical Imaging Group

Student Projects

EPFL > BIG > Teaching > Student Projects > On going projects |

CONTENTS |

Student Projects |

STUDENT PROJECTS |

Conditions:The projects are reserved for EPFL students or students of mobility program. |

Number of projects:6

A study of the tip detection problem

*Alexandre Cherqui*

Master Semester Project: Winter 2017

More or less rounded extremities (or "tips") are extremely common features in biological images. They are very important features in many image analysis situations, e.g. when trying to measure the length of bacteria or worms body, or when segmenting legs, whiskers or antennas of model organisms. An extensive amount of methods are available for the detection of other features (such as blobs or corners), but how to solve the tip detection problem remains much less clear. The goal of this project is to study available solutions for detecting tips. This implies identifying, understanding (possibly reimplementing), analyzing, comparing and characterizing the different approaches including e.g. filter design, keypoint detection methods and machine learning. The final goal is to build a catalogue of the different methods for tip detection specifying their advantages, limitations, and how to use them. From this, the interested student could in addition identify recurring problems in existing approaches and propose novel algorithms for tip detections.

Supervision:

- Virginie Uhlmann, virginie.uhlmann@epfl.ch, BM 4.142, Tel: 021 693 1136
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75

B-spline implementation to find the solution of continuous domain total-variation minimization problem

*Thomas Debarre*

Master Semester Project: Winter 2017

In MRI and other real world applications, the measurements are generally obtained through a continuous-domain transformation of a continuous-domain signal (Fourier samples for MRI case). Yet, for computational feasibility, the inverse problem formulated to numerically reconstruct the signal from these measurements, are often formulated in discrete-domain. Continuous-domain formulation of inverse problems therefore can be advantageous in this sense, provided there is a way to tackle the computational complexity of the reconstruction task. Recently in [1], the solution for inverse problems in continuous domain with Total variation regularization is found out to be non-uniform spline. To perform the reconstruction we use Green's function of the operator used in regularization, as the dictionary basis. However, this often results in ill-conditioned system matrices leading to poor convergence rate. We propose a student project to use the corresponding B-splines as the dictionary basis for the TV-regularized solution. The resultant system matrix in this case is expected to be better conditioned and an appropriate algorithm can result in faster convergence to a solution. The task will be to effectively implement this formulation and contrast it with the results of the previous formulation. The student will have to understand the theoretical background of the problem and convex optimization techniques, and then implement the formulation in MATLAB. Prerequisites: Convex optimization [1] M. Unser, J. Fageot, and J. P. Ward, “Splines are universal solutions of linear inverse problems with generalized-TV regularization,” arXiv preprint arXiv:1603.01427, 2016.

Supervision:

- Harshit Gupta, harshit.gupta@epfl.ch, BM 4140, Tel: +41 21 69 35142
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75

3D Steerable Filter Learning for Efficient Volumetric Image Analysis

*Camille Boymond*

Master Diploma Project: Winter 2017

The use of deep convolutional neural networks (CNN) for object recognition in computer vision has shown to provide excellent results in many applications. Deep CNNs learn multiple filters in each convolutional layer of a deep neural network architecture using backpropagation weight updates. A major drawback of the latter is the requirement of large amounts of training data and computational time to learn all pixel weights (i.e., free parameters) of the filters. Moreover, CNNs are not rotation-invariant and require extensive re-training with augmented data (e.g., rotated versions of the training images), which degrades the specificity of the learned filters. Steerable filters are used on image analysis as efficient and accurate rotation-invariant object detectors. They are excellent candidates to overcome these drawbacks. The 2D theory has been recently adapted to classification problems and applied to texture analysis. The goal of this project is to extend the framework to the 3D setting, where rotation-invariance is even more important. This presents both mathematical and implementation challenges.

Supervision:

- Adrien Depeursinge, adrien.depeursinge@epfl.ch, BM 4141, Tel: 021 693 5115
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Julien Fageot, julien.fageot@epfl.ch, BM 4.139, Tel: 021 693 3701

Optimal Configurations for Parallel-Beam Computed Tomography

*Matthieu Broisin*

Master Semester Project: Winter 2017

Parallel-beam computed tomography aims at reconstructing the 3D volume of an object from its 2D projection measurements. As most rays used in CT imaging (x-rays, electron-rays, etc.) are harmful for the object being imaged, an important issue is to maximally reduce the radiation dose necessary for high-quality 3D reconstructions. In this project, we will study, through 2D and 3D simulations, the impact of two dose-reduction approaches (i.e., reducing the number of tilt views Vs reducing the dose per view) on the quality of the reconstructed image, at different levels of gaussian noise. This shall be done for two distinct types of reconstruction algorithms: filter-back projection (FBP) algorithms and more advanced iterative algorithms.

Supervision:

- Laurène Donati, laurene.donati@epfl.ch, BM 4.139
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Daniel sage, daniel.sage@epfl.ch

Active Contour for Jointly Segmentation of Multiple Cells

*Aymeric Galan*

Master Semester Project or Master Diploma Project: Winter 2017

Active contours are powerful methods for the segmentation of biomedical images. However, they usually segment only well cell at a time, which makes fastidious the segmentation of microscopic images that often contain hundreds of cells. The goal of this project is to develop an active contour that is able to jointly segment multiple cells. We will work on fluorescence images of the embryo of c-Elegans. In fact, the study of early cell division is an active field of research in developmental biology. Good notions on image processing and optimization are strongly advised, and experience with java is recommended.

Supervision:

- Anaïs Badoual, anais.badoual@epfl.ch, BM 4142, Tel: 31136
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Daniel sage, daniel.sage@epfl.ch

Rotation axis estimation for parallel-beam X-ray CT

*claire.stoffel@epfl.ch*

Bachelor Semester Project or Master Semester Project: Winter 2017

In parallel-beam X-ray CT, a 3D image of an object is reconstructed from a collection of its 2D X-ray projections taken from different angles around a fixed rotation axis. The location of the axis must be estimated accurately or serious artifacts arise during reconstruction. In this project, the student will implement a method (already existing in the literature) for estimating the rotation axis directly from the data and compare its performance to a baseline method on simulated and real data.

Supervision:

- Mike McCann, michael.mccann@epfl.ch, BM 4141
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75