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:9

Benchmarking of numerical methods for solving inverse problems

*Zhiwei Huang*

Master Semester Project: Summer 2019

Inverse problems are at the heart of many microscopy and medical imaging modalities where one aims at recovering an unknown object from given measurements. Such a problem is generally addressed through the minimization of a given functional composed of a data-fidelity term plus a regularization term. Within the Biomedical Imaging Group, we are currently developing a Matlab library (http://bigwww.epfl.ch/algorithms/globalbioim/) unifying the resolution of inverse problems. This library is based on several blocks (forward models, data-fidelity terms, regularizers, algorithms) that can be combined to solve any inverse problem. Given an imaging modality, one can thus easily compare methods that use different data terms, regularizers or algorithms. The goal of this project is to develop a Matlab code which, for a given modality, outputs in an elegant way different metrics showing the performances obtained using all the combinations of blocks (forward models, data-fidelity terms, regularizers, algorithms) that are available within the Library.

Supervision:

- Pakshal Bohra, pakshal.bohra@epfl.ch, BM 4.140
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Emmanuel Soubies

Reconstruction of autofluorescence optical projection tomography

*Thomas Ramseier*

Master Semester Project: Summer 2019

In this project, the student will develop an image reconstruction method for a novel autofluorescence optical projection tomography dataset prepared by our collaborators in the Laboratory of Nanoscale Biology at EPFL. Optical projection tomography is a method for creating a 3D image of a sample. It does this by computationally combining many projection images (similar to the X-ray image a doctor might take of a broken bone), taken from different angles. As the first goal of the project, we would like to apply tomography reconstruction software that our group has already developed to this novel data, compare different methods, and tune reconstruction parameters. As a second goal, we would like to explore methods of performing high-resolution, 3D reconstructions (as large as 1024x1024x1024), where RAM becomes a bottleneck; this will require carefully splitting the reconstruction volume into subvolumes.

Supervision:

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

An off-the-grid algorithm in ImageJ for 3D single-molecule localization microscopy

*Amandine Evard*

Master Semester Project: Summer 2019

A new class of algorithms has recently emerged in the literature for the recovery of point source signals from altered and noisy measurements. These methods are able to perform the reconstruction, without requiring any discretization of the domain, by solving an infinite dimensional optimization problem. They interleave convex optimization updates (corresponding to adding a new point source) with non-convex optimization steps (corresponding to changing the intensities and positions of the point sources). Single-molecule localization microscopy is an imaging technique in fluorescence microscopy that is able to bypass the diffraction limit and reach nanoscale resolution for the imaging of sub-cellular structures in cells (e.g., microtubules). High performance numerical solvers are needed to locate precisely the positions of the fluorescent molecules. The previously mentioned class of algorithms are currently the one that obtain the state-of-the-art results in this application. The goal of this project is to implement one of these methods, called the Sliding Frank-Wolf algorithm, in Java as an ImageJ/Fiji plugin so that it can be usable by biologists. It will include a user interface and permit automatic processing of large datasets and output super-resolved images.

Supervision:

- Thanh-An Pham, thanh-an.pham@epfl.ch, BM 4.140
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Quentin Denoyelle

Learning Spline-based activations for very deep learning

*Joaquim Campos*

Master Semester Project: Summer 2019

In neural networks, typically, only the weights of the linear layers are learnt, and the non-linear activation functions of the neurons are kept unchanged. However, learning these activation functions has gained popularity with promising results. Generally these activations are assumed to lie in a family of functions which are parameterized by small number of learnable parameters. Recently, in [1], it was shown that within a large family of activation functions the optimal ones are composed of multiple ReLUs (Rectified Linear Units) whose weights and locations are unknown apriori. However, learning the location and the coefficients of these ReLUs can be computationally challenging. Therefore, we use a spline-based parameterization to learn these coefficients which decreases the computational complexity. The goal of this project is to learn the spline-based activation functions in very deep neural networks. The task will be to build and evaluate the performance of very deep neural networks with learnable spline-based activations on classification problems. The project will be in Pytorch and the learnable activation module are already available. Pre-requisites: Either already having knowledge of Deep learning, Pytorch, and Signal Processing or going to take courses on these subjects in parallel to the project. [1] M. Unser. `A representer theorem for deep neural networks', arXiV, 2018.

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

Implementing Deep-learning-based iterative algorithm to solve inverse problem of MRI

*Huy Thong Phan*

Master Semester Project: Summer 2019

MRI or CT imaging requires to solve inverse problem of reconstructing the image of internal organs from the acquired measurements. Recently, neural networks have been used to reconstruct these images. These networks are trained to directly output the image when the measurements (or a primitive image obtained from measurements) are given as an input. However, these CNN-based approaches usually lack a feedback mechanism to enforce that the reconstructed image is consistent with the measurements. In [1], we use an iterative framework for CT imaging, wherein measurement consistency is enforced, while a neural network recursively improves these images. This approach outperforms the previous approaches and is more robust. The goal of the project would be to extend this algorithm for MRI modality. The task will be to implement this algorithm in Matlab (using GlobalBioIm Library), with learning framework in Pytorch. The framework for the CT case is already available in MatconvNet (in Matlab) [2]. Pre-requisites: Either already having knowledge of Deep learning, Pytorch, and Image Processing or going to take courses on these subjects in parallel to the project. [1] H. Gupta, et al. `CNN-Based Projected Gradient Descent for Consistent CT Image Reconstruction'. IEEE transactions on medical imaging, 2018. [2] https://github.com/harshit-gupta-epfl/CNN-RPGD . [3] http://bigwww.epfl.ch/globalbioim/ .

Supervision:

- Harshit Gupta, harshit.gupta@epfl.ch, BM 4.140
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Than-an Pham

Simulating realistic synthetic data sets for developing a self-driving microscope

*Robin Lang*

Master Semester Project: Winter 2019

**Motivation**: Many of the breakthroughs made in the fields of computer science and image processing have not yet been applied to the main tools driving biological discovery. The project proposed here would contribute to developing a self-driving, or "smart" super-resolution microscope in single molecule localization microscopy (SMLM). Super-resolution microscopy circumvents the diffraction limit of light, allowing nanoscale spatial resolution while retaining the advantages of fluorescence microscopy

**Goal**: The goal is to build a Java interactive software ("Flight Simulator") for generating simulated super-resolution datasets. It should also enable on-the-fly modification of the acquisition parameters in an attempt to mimic, for the first time, what would actually happen on the microscope when an expert does the acquisition. Such an interface is a requirement for the development, validation and comparison of future automated, machine learning driven, routines dedicated to super-resolution microscopy.

The Master project will be shared between two labs (the Laboratory of Experimental Biophysics led by Prof. S. Manley & the Biomedical Imaging Group led by Prof. M. Unser) and under the joint supervision of Dr. Sage and Dr. Griffié, allowing the student to work in a pluri-disciplinary environment

Supervision:

- Daniel Sage, daniel.sage@epfl.ch, BM 4.135, Tel: 021 693 51 89
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Juliette Griffié and Suliana Manley LEB EPFL

Deep Learning for Angle Estimation in Cryo-EM

*Jelena Bancac*

Master Diploma Project: Winter 2019

Single-particle cryo-electron microscopy (cryo-EM) is a Nobel-prized technology that aims to characterize the 3D structure of proteins at the atomic level. The electron microscope first images numerous (~100k) replicates of a protein, positioned at various orientations. Algorithms then reconstruct a high-resolution 3D structure from the acquired images. The main challenge in cryo-EM reconstruction, compared to traditional tomographic set-ups, is that the angles at which the images were taken are unknown. Another challenge is that the images are extremely noisy and blurred. The sheer amount of images per protein (~100k), as well as the number of imaged proteins (~4k), should however enable a data-driven approach to overcome those challenges. Project goal: Design a neural network to estimate the angular relation between images of a protein. The developed neural network will be trained and tested on simulated and real data. Prerequisites: Experience with Python programming. Experience with (Deep) Machine Learning (with any framework) is desirable. No experience in biology is required. Experience in imaging is a plus.

Supervision:

- Laurène Donati, laurene.donati@epfl.ch, BM 4.139
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Michaël Defferrard (LTS2)

Phase Unwrapping with Deep Learning

*Jérome Savary*

Master Semester Project: Summer 2018

Phase unwrapping is a seminal imaging problem encountered in numerous applications, such as interferometric synthetic aperture radar, magnetic resonance imaging, digital holography microscopy (DHM), or optical diffraction tomography (ODT). In all these applications, the measurements of the true phase values lie between 0 and 2π (i.e., the true phase values are observed modulo 2π). Phase unwrapping consists in recovering the original phase from these wrapped and possibly noisy measurements. In DHM / ODT, optically thick samples yield 2D images difficult to unwrap. Conventional techniques can fail to recover the true phase because of steep gradient. In this project, the student will implement a convolutional neural network to solve the phase unwrapping problem in DHM / ODT using Python (TensorFlow / PyTorch), and compare its performance to other existing approaches.

Supervision:

- Thanh-An Pham, thanh-an.pham@epfl.ch, BM 4.140
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75
- Kyong Hwan Jin

(Theoretical Project) Generating Sparse Stochastic Processes

*Leello Dadi*

Master Diploma Project: Summer 2017

Sparse stochastic processes are continuous time stochastic models for sparse signals. It has been shown that such processes are limit points (in distribution) of a sequence of compound Poisson processes, which are roughly a random sum of Dirac impulses that each have independent random heights. The goal of this project is to use this theoretical observation in order to generate sparse stochastic processes with arbitrary accuracy. It requires a theoretical understanding of the convergence with some bounds guaranteeing the accuracy of the generated process. This project will also include an implementation that can be used to model natural signals and images. The student should have a solid understanding of functional analysis and probability theory, and basic knowledge of programming.

Supervision:

- Shayan Aziznejad, shayan. aziznejad@epfl.ch, BM 4.138
- Michael Unser, michael.unser@epfl.ch, BM 4.136, Tel: 021 693 51 75