Application of Deep Learning Segmentation to Similar Contrast Materials in X-Ray CT Imaging

Author(s):

Keaton Klaff and Jennifer Sietins

Abstract:

Accurate segmentation of X-ray computed tomography (CT) images is critical for quantitative analysis of composite materials. Segmenting samples with low-contrast material phases is often challenging and time-consuming. This study compares two segmentation approaches applied to CT data from an aluminum–silicon carbide (Al-SiC) metal matrix composite specimen. Dual-energy CT images were segmented using Zeiss’ Dual Scan Contrast Visualizer (DSCoVr) software, and single-energy CT images were segmented using a 2D U-Net model implemented in Dragonfly Pro. The U-Net method was more precise in capturing phase boundaries and less prone to over-segmentation due to noise, at the expense of greater computational time and user effort. These results highlight a trade-off between accuracy and efficiency and offer guidance for selecting segmentation methods for low-contrast material phases in CT images.