Lab Course: Visualization and Medical Image Analysis

Course

  • Lecturer(s):
  • Start: 03.11.2020 (online)
  • Dates: by arrangement
  • Course number: MA-INF 2220 - Lab Visualization and Medical Image Analysis
  • Curriculum: Master , B-IT Master Media Informatics
  • Requirements: Knowledge from one of our classes is recommended.
  • Exams: TBA

Description

In this lab, you will learn to implement state-of-the-art methods from the field of visualization or medical image analysis. If you are interested in one of the topics below, please contact the corresponding direct advisor. If you would like to propose your own topic within the scope of visualization and medical image analysis, please contact Prof. Schultz.

Please use our LaTeX template or a similar layout for your lab report.

Available Topics

A web interface for CoBundleMAP
CoBundleMAP is a manifold learning based method for jointly parameterizing streamlines from diffusion MRI tractography that was recently developed in our group. It enables the comparison of imaging biomarkers between subjects and brain hemispheres along white matter tracts. Your task within this lab would be to design and implement a web interface for visualizing and exploring the results, similar to the AFQ-Browser that was developed at the University of Washington. Ideally, this would be a group effort of multiple students.
Main reference: Khatami et al., CoBundleMAP: Consistent 2D Parameterization Of Fiber Bundles Across Subjects and Hemispheres. Proc. IEEE Int'l Symposium on Biomedical Imaging, pp. 1475-1478, 2019
Direct advisor: Mohammad Khatami

Graph-based modeling of resting-state functional MRI for supervised classification
Connectivity matrices have been derived from resting-state functional MRI data to model the functional connectivity between cortical regions. It has been demonstrated that, based on such models, it is possible to extract features for supervised classification tasks, e.g., to detect diseases. In this lab, you will pursue such an approach based on data from the Human Connectome Project.
Main reference: Khazaee et al., Identifying patients with Alzheimer’s disease using resting-state fMRI and graph theory. Clinical Neurophysiology 126:2131-2141, 2015
Direct advisor: Mohammad Khatami

Meta Learning for Domain Generalization
Imaging the same anatomical structures with different medical scanners often leads to a different appearance, or different image characteristics. This domain shift can lead to a drastic reduction of accuracy when applying neural networks that have been trained on images from one scanner to those from other devices, or with slightly different settings. It is the goal of this lab to implement and evaluate a meta learning approach for better generalization from a limited number of scanners to images from a new scanner.
Main reference: Khandelwal et al., Domain Generalizer: A Few-Shot Meta Learning Framework for Domain Generalization in Medical Imaging. In: Domain Adaptation and Representation Transfer. LNCS 12444, pp. 73-84, 2020 (Paper on arXiv)
Direct advisor: Rasha Sheikh

Visualizing CNNs for Semantic Segmentation
For image classification with convolutional neural networks, many approaches have been developed to visualize which regions in the input image are most relevant for determining the output label. The task of this lab is to explore strategies for visualizing how the input determines the output in case of semantic segmentation. This is a much more complex problem, since the output is one label per pixel. If successful, this lab could be followed up with a M.Sc. thesis project.
Main reference: Wickstrøm et al., Uncertainty and interpretability in convolutional neural networks for semantic segmentation of colorectal polyps. Medical Image Analysis 60:101619, 2020
Direct advisor: Rasha Sheikh

Image Compression with Partial Differential Equations
Partial differential equations have an impressive ability to reconstruct images from a sparse subset of their pixels. When carefully selecting which pixels to preserve, this can be exploited for image compression. It is the goal of this lab to re-implement and extend such a compression codec that is based on rectangular subdivision and edge-enhancing diffusion (R-EED).
Main reference: Schmaltz et al., Beating the Quality of JPEG 2000 with Anisotropic Diffusion. Proc. DAGM, pp. 452-461, 2009
Direct advisor: Ikram Jumakulyyev

4D Imputation of Signal Dropout in Diffusion MRI
Diffusion MRI enables the non-invasive investigation of tissue micro-structure. However, it suffers from acquisition artifacts such as the signal dropout caused by bulk motion during the diffusion encoding part of the imaging. A prior publication from our group has detected and imputed signal loss using sparse signal modeling. A prior lab has explored the use of spatial inpainting in this context. The goal of this lab will be to fuse both approaches, to exploit all available information via four dimensional imputation.
Main references: Koch et al., SHORE‐based detection and imputation of dropout in diffusion MRI. Magnetic Resonance in Medicine 82(6):2286-2298, 2019
Galic et al., Image Compression with Anisotropic Diffusion. Journal of Mathematical Imaging and Vision 31:255-269, 2008
Direct advisor:
Ikram Jumakulyyev

Guidelines for a strong lab

To successfully finish a lab in our group, you have to produce three main results:

  • practical result. In most cases, this will be a software implementation of a method described in the scientific literature or by your advisor. You also need to conduct experiments to demonstrate the extent to which the implementation fulfills its purpose, and to investigate its limitations. Alternatively, in specific cases, the practical result might consist of getting one or several existing software packages to work, and performing a systematic evaluation. In any case, this is the most important part of the lab, and we will grade it based on how challenging it was to achieve the result, on its quality (correctness, efficiency, readability of your code), on how independently you worked and contributed your own ideas, as well as on how carefully you designed, conducted, and interpreted your experiments.

  • written report. The report should describe the overall design of the software, provide the information that is required to use it, including dependencies (such as external libraries), and highlight the steps that took most of your time and effort. It should also present and interpret your experiments, and conclude with a clear statement what you achieved in the lab, and what might be left for potential future work. We recommend writing the report in LaTeX, since that is a useful skill for the MSc thesis; however, this will not affect your grade. Compared to a seminar report, the length is more flexible: The typical range is 5-10 pages per person, but it will depend on your specific task and needs.

  • An oral presentation that should take no more than 30 minutes per person, followed by answering questions about your lab. Your answers should demonstrate that you have a firm understanding of the tools you have been using, and are able to defend your design choices and experimental setups. The presentation should contain an introduction that clearly explains what your topic is about and why it is important. Your presentation should be supported by suitable media, such as projected slides. It is up to you how you create them (LaTeX, PowerPoint, other means). If it fits your topic, feel free to use other media also, such as the whiteboard or a live demo.

The practical result and written report have to be submitted before the presentation; we will tell you the exact dates well in advance. It is strongly recommended that you discuss your results and report with your direct advisor – who is usually one of our PhD students – and perform a practice presentation. Please contact them about this well in advance. If you would like to get direct feedback from Prof. Schultz, please submit a draft at least two weeks in advance.

Write the report in your own words. In the “References” section, list all papers and resources that have helped you to complete the lab, and explain in the text what role they played. If you would like to use images, plots, or diagrams produced by others, or if you would like to cite one of your sources verbatim, you have to clearly mark which material has been copied, and from which source. The same applies to source code you might have used from open-source packages or other sources, including fellow students. If you copy source code, text, or any other material without clearly marking it, this will be considered as plagiarism and can lead to failing the module!

A lab amounts to around one third of the overall workload of a full-time semester, which means that you will not be able to complete a strong lab unless you work on it continuously throughout the semester – you cannot expect to do everything at the last minute. Finally, please keep in mind that a nice report and presentation will not make up for missing or poor results, but that a report that is extremely short or difficult to understand, or a chaotic presentation, can make it hard for us to fully appreciate the practical work that you have done.