End-To-End Deep Learning Material Discrimination Using Dual-Energy LINAC-CT

Material discrimination, or more precise material identification, has a long history in computed tomography (CT). However, the focus of research has generally been on clinical CT applications, rather than industry-relevant samples. Yet the areas of application differ from one another in key characteristics. In industrial applications, the tube voltages applied are significantly higher than those typically encountered in clinical scan setups, especially when dealing with large objects and highly attenuating materials like iron or nickel alloys. This has the consequence that images contain no detectable fingerprint attributable to absorption at characteristic K-edges, which has been the key to material identification for most clinical applications. In this work, we propose a Z − ρ (atomic number-density) decomposition algorithm, which is driven by a newly developed deep learning architecture inspired by the U-Net infused with transformer mechanics. Using this approach, a wide range of materials can be distinguished by a single model. Especially for quality control, our model can be fine-tuned using prior information or domain knowledge to ensure the best model precision, which is a subject of current research. The model is trained exclusively on simulated data but can compute predictions from experimental test scans. Since deep learning models must be trained on large datasets, it is advantageous to generate synthetic CT datasets using a simulation. This approach helps avoid costly machine time, eliminates the need for manual labeling (itself susceptible to human error), and enables the creation of scenarios that may be unfeasible in a specific experimental setup. We are using an in-house simulation pipeline which we describe below. Our trained model predicts aluminum correctly. Copper, brass and stainless steel show a systematic offset, which will be studied in a future work. © 2024-by the Authors.