Biological analysis at extremely small scales is becoming increasingly more important in life sciences. To obtain in-depth understanding of biological tissue and cell structure FEI tools like a Scanning Electron Microscope (SEM) are used to create nanometer scale images. These 2D images are created from slices of a 3D tissue volume. Obtaining such a volume is a very time consuming and therefore expensive task; often it requires months to cut and scan the volume on such expensive tools. In total a representative 3D tissue volumes contains over 1 Tera voxels, which is a huge dataset. To speed-up the slow and expensive acquisition process we would like to exploit the spatial structure in these volumes. A very sparse but fast acquisition and the use of advanced interpolation algorithms can reconstruct the missing parts.
Develop and evaluate multiple reconstruction algorithms. The starting point is a very recent publication based upon a Very Deep Super Resolution (VDSR) Convolutional Network. This network uses 20 layers of non-linear transformations to reconstruct high resolution features in images. The project should result in a Deep Learning based technology that performs correct reconstructions for multiple sparsity settings. In addition, the reconstruction phase should be fast large images of 4k resolution should be reconstructed in a few seconds.
Programming, algorithm development.