AI Platform for Learning 3D Cellular Objects in High-Resolution Holo-Tomographic Volumes

A unique holo-tomographic microscope allows recording cellular 3D volumetric images and detecting unlabeled structures at an unprecedented subcellular resolution in both space and time (nanolive.ch). Because of the lack of labels, powerful Artificial Intelligence (AI) is required to help experimenters to label 3D structures and pursue with quantitative analysis of their dataset. Thus, the AI-driven analysis of holo-tomographic cellular volumes will revolutionize the field of live cell imaging.

 

Goals

  • Develop an AI-framework with deep learning algorithms for learning 3D high-resolution cellular and subcellular details within holo-tomographic volumes
  • Benchmark both predictive and computational performance of developed algorithms
  • Based on “Best Software Practices” industrialize most promising AI-algorithm including graphical user interface (GUI) and System Tests
Tasks

  • Familiarize with holo-tomographic microscopy – including practical experiments
  • Explore current Machine Learning (ML) and Deep Learning (DL) algorithms to learn subcellular 3D structures in holo-tomographic volumes
  • Setup AI-framework (workflow) with functionality:
    • Load and preprocess holo-tomographic data (provided by Nanolive)
    • Train-validate-test
    • Visualize and store
  • Implement Prototype including:
    • The currently existing ML-algorithms as starting point (base line)
    • The most promising ML – / DL-algorithm evaluated above
    • The possible extension of the initial ground truth data
    • Benchmark and iteratively improve implemented ML- / DL-algorithms
  • Industrialize most promising ML- / DL-algorithm using best Software Practices including:
    • Test Driven Development
    • Modularity
    • Extensibility: e.g. new features via on-line improving labelling quality
    • Ergonomic Graphical User Interfaces (GUI)
  • Document and present Master Thesis
    • Final document
    • One-Minute Video Summary
    • 30 minute presentation
Possible Technologies

  • Machine/Deep Learning and Software Engineering
  • Python, TensorFlow, mxNet
  • Optional: scikit-learn, ImageJ, MATLAB
Literature

  • Choi, W. et al. Tomographic phase microscopy. Nat. Methods 4, 717-719 (2007).
  • Cotte, Y. et al. Marker-free phase nanoscopy. Nat. Photonics 7, 113-117 (2013).
  • Chen, L,C. et al. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs IEEE Transactions on Pattern Analysis and Machine Intelligence

Kind of Work
40% Research, 40% Software Development, 20% Benchmarking

Requirements
Machine Learning & Deep Learning Basics, CV, Python, Matlab

Time & Effort
Master’s Thesis, 1 Person

Contact 
fabian.schenkel@scs.ch

SCS - Studienarbeiten
SCS - Studienarbeiten

Supercomputing Systems AG  •  Technoparkstrasse 1  •  CH-8005 Zürich  •  Phone +41 43 456 16 00  •  Fax +41 43 456 16 10  •  info@scs.ch  •  Impressum  •  Datenschutzerklärung