Deep-learning based porosities segmentation applied to twin robotic CT images

This work stands in the framework of twin robotic Computed Tomography. Such type of acquisition allows the possibility of new viewing angles included in the set of reconstructed projections, compared to classical circular acquisition. However, due to the size and shape of the measured objects as well as the kinematic constraints of the robots, not all viewing angles can be accessed. For many applications, generating sufficient information for a mathematically correct and reliable reconstruction is therefore not possible. Reconstruction with classical algorithms then leads to artifacts corresponding to Sparse-View acquisition (streak artifacts) and limited angle acquisition. The aim of this work is to evaluate the benefit of a Deep-Learning based algorithm for 3D porosities segmentation on CT data acquired with a twin robotic CT system, compared to classical image segmentation. Classical SART algorithm is chosen for the reconstruction. In order to use supervised approaches for the segmentation task, we build a database using CIVA software. The database is used to train a 3D UNet based algorithm for segmentation.