Abstract. X-ray computed tomography has established itself as a crucial tool in the analysis of rock materials, providing the ability to visualise intricate 3D microstructures and capture quantitative information about internal phenomena such as structural damage, mineral reactions, and fluid–rock interactions. The efficacy of this tool, however, depends significantly on the precision of image segmentation, a process that has seen varied results across different methodologies, ranging from simple histogram thresholding to more complex machine learning and deep-learning strategies. The irregularity in these segmentation outcomes raises concerns about the reproducibility of the results, a challenge that we aim to address in this work. In our study, we employ the mass balance of a metamorphic reaction as an internal standard to verify segmentation accuracy and shed light on the advantages of deep-learning approaches, particularly their capacity to efficiently process expansive datasets. Our methodology utilises deep learning to achieve accurate segmentation of time-resolved volumetric images of the gypsum dehydration reaction, a process that traditional segmentation techniques have struggled with due to poor contrast between reactants and products. We utilise a 2D U-net architecture for segmentation and introduce machine-learning-obtained labelled data (specifically, from random forest classification) as an innovative solution to the limitations of training data obtained from imaging. The deep-learning algorithm we developed has demonstrated remarkable resilience, consistently segmenting volume phases across all experiments. Furthermore, our trained neural network exhibits impressively short run times on a standard workstation equipped with a graphic processing unit (GPU). To evaluate the precision of our workflow, we compared the theoretical and measured molar evolution of gypsum to bassanite during dehydration. The errors between the predicted and segmented volumes in all time series experiments fell within the 2 % confidence intervals of the theoretical curves, affirming the accuracy of our methodology. We also compared the results obtained by the proposed method with standard segmentation methods and found a significant improvement in precision and accuracy of segmented volumes. This makes the segmented computed tomography images suited for extracting quantitative data, such as variations in mineral growth rate and pore size during the reaction. In this work, we introduce a distinctive approach by using an internal standard to validate the accuracy of a segmentation model, demonstrating its potential as a robust and reliable method for image segmentation in this field. This ability to measure the volumetric evolution during a reaction with precision paves the way for advanced modelling and verification of the physical properties of rock materials, particularly those involved in tectono-metamorphic processes. Our work underscores the promise of deep-learning approaches in elevating the quality and reproducibility of research in the geosciences.

中文翻译：

---

使用时间分辨 X 射线微计算机断层扫描中的内标来量化固态矿物反应中的晶粒级发展

摘要。 X 射线计算机断层扫描已成为岩石材料分析的重要工具，能够可视化复杂的 3D 微观结构并捕获有关结构损伤、矿物反应和流体-岩石相互作用等内部现象的定量信息。然而，该工具的功效在很大程度上取决于图像分割的精度，这一过程在不同的方法中出现了不同的结果，从简单的直方图阈值到更复杂的机器学习和深度学习策略。这些分割结果的不规则性引起了人们对结果可重复性的担忧，这是我们在这项工作中旨在解决的挑战。在我们的研究中，我们采用变质反应的质量平衡作为内部标准来验证分割准确性并揭示深度学习方法的优势，特别是它们有效处理广泛数据集的能力。我们的方法利用深度学习来实现石膏脱水反应的时间分辨体积图像的精确分割，由于反应物和产物之间的对比度较差，传统分割技术一直难以实现这一过程。我们利用 2D U-net 架构进行分割，并引入机器学习获得的标记数据（特别是随机森林分类），作为解决从成像获得的训练数据的局限性的创新解决方案。我们开发的深度学习算法表现出了非凡的弹性，能够在所有实验中一致地分割体积阶段。此外，我们训练有素的神经网络在配备图形处理单元（GPU）的标准工作站上表现出令人印象深刻的短运行时间。为了评估我们工作流程的精度，我们比较了脱水过程中石膏与烧石膏的理论和测量摩尔演化。所有时间序列实验中预测体积和分段体积之间的误差均落在理论曲线的 2% 置信区间内，证实了我们方法的准确性。我们还将所提出的方法获得的结果与标准分割方法进行了比较，发现分割体积的精度和准确度有了显着的提高。这使得分段计算机断层扫描图像适合提取定量数据，例如反应过程中矿物生长速率和孔径的变化。在这项工作中，我们引入了一种独特的方法，通过使用内部标准来验证分割模型的准确性，展示了其作为该领域强大且可靠的图像分割方法的潜力。这种精确测量反应过程中体积演化的能力为岩石材料物理特性的高级建模和验证铺平了道路，特别是那些涉及构造变质过程的岩石材料。我们的工作强调了深度学习方法在提高地球科学研究的质量和可重复性方面的前景。