Abstract
The goal of this project is to reconstruct a 3D image of various non-overlapping shapes from a limited number of projections (views of the shapes, see fig. 1). Specifically, we look at 2D tomography and try to recreate a segmentation mask of the image that would separate out the areas of the image with shapes and those without. To begin, we generate the shapes in a box and form binary segmentation masks (arrays that showed only the location of the shapes, not any details about their values) from the images for use in the training. In addition to the binary segmentation masks, we generate the image’s few view (4 and 40) filtered backprojection reconstructions. Using these, we implement a UNet architecture with 4 convolutional layers (fig. 2) to try and find the segmentation mask from the input FBP. The output of the network is a probability map that shows the likelihood of each pixel in the image corresponding to a 1 or 0 in the segmentation mask of the FBP (fig. 3).
This was a short summer project worked with Los Alamos National Labaratory. This problem has applications in baggage screening scenarios.
