Resolution enhancement and segmentation of polar bubbly ice micro CT scans via 3D convolutional neural network

Accurate segmentation of 3D micro CT scans is a key step in the process of analysis of the microstructure of porous materials. In polar ice core studies, the environmental effects on the firn column could be detected if the microstructure is digitized accurately. The most challenging task is to obtain the microstructure parameters of the bubbly ice section. To identify the minimum, necessary resolution, the bubbly ice micro CT scans with different resolutions (120, 60, 30, 12 $\mu m$) were compared object-wise via a region pairing algorithm. When the minimum resolution was found to be 60 $\mu m$, for generating the training/validation dataset, 4 ice core samples from a depth range of 96 to 108 meters (bubbly ice) were scanned with 120 $\mu m$ (input images) and another time with 4 times higher resolution (30 $\mu m$) to build ground truth. A specific pipeline was designed with non-rigid image registration to create an accurate ground truth from 4 times higher resolution scans. Then, two SOTA deep learning models (3D-Unet and FCN) were trained and later validated to perform super-resolution segmentation by taking input of $120\mu m$ resolution data and giving the output of binary segmented with two times higher resolution ($60\mu m$). Finally, the outputs of CNN models were compared with traditional rule-based and unsupervised methods on blind test data. It is observed the 3D-Unet can segment low-resolution scans with an accuracy of 96% and an f1-score of 80.8% while preserving microstructure having less than 2% error in porosity and SSA.