U-Net Segmentation Methods for Variable-Contrast XCT Images of Methane-Bearing Sand Using Small Training Datasets

Volume Segmantics: A Python Package for Semantic Segmentation of Volumetric Data Using Pre-trained PyTorch Deep Learning Models

“…Even though these 2D models are quicker to train and require fewer computational resources than their 3D counterparts (Alvarez-Borges et al, 2022), when predicting a segmentation for a volume, the lack of 3D context available to these models can lead to striping artifacts in the 3D output, especially when viewed in planes other than the one used for prediction. To overcome this, a multi-axis prediction method is used, and the multiple predictions are merged by using maximum probability voting.…”

“…In a completely different context, SXCT datasets were collected on a soil system in which methane bubbles were forming in brine amongst sand particles. The utility of a pre-trained 2D U-Net was investigated to segment these variable-contrast image volumes in comparison to a 3D U-Net with no prior training (Alvarez-Borges et al, 2022). In this case, the training data ranged in size from 384 3 pixels to 572 3 .…”