AI 中文总结
研究针对同步辐射微纳断层扫描中准确确定旋转轴位置的问题，提出基于学习的方法，将中心选择作为二元分类问题，结合视觉Transformer与多实例学习，经测试该方法估计准确、鲁棒性好且具可解释性，已集成到软件包助力实验。
AI 中文摘要
准确确定旋转轴位置是平行束同步辐射微断层扫描中无伪影重建的前提。传统方法如Vo方法依赖于正弦图特征，对于低对比度或弱吸收样本可能失效。我们提出一种基于学习的方法，将中心选择视为二元分类问题，使用预训练的DINOv2视觉Transformer与基于注意力的多实例学习相结合，并在断层图像上进行端到端微调。推理时，该算法应用于在一系列候选中心重建的断层图像堆栈，以选择最佳重建中心。我们在两个独立数据源上测试了该方法的估计准确性，平均绝对误差始终低于1像素。我们还测试了该方法对稀疏或噪声采集的鲁棒性，当投影数量减少到十分之一或泊松噪声的空白扫描因子增加到10时，性能保持一致。我们还通过映射连续空间特征对整体分类任务的相对贡献来说明该方法的可解释性。该方法以开源命令行工具tomo-center的形式提供，已集成到多个断层扫描软件包中，以协助日常束线操作中的实验。
英文摘要
Accurate determination of the rotation-axis position is a prerequisite for artifact-free reconstruction in parallel-beam synchrotron micro-tomography. Traditional approaches such as Vo's method rely on sinogram features that can fail for low-contrast or weakly absorbing specimens. We present a learning-based method that treats center selection as a binary classification problem, using a DINOv2-pretrained vision transformer aggregated with attention-based multiple-instance learning, fine-tuned end-to-end on tomographic images. At inference time, the proposed algorithm was applied to a stack of tomograms reconstructed at a sweep of candidate centers to select the optimal center for reconstruction. We tested the estimation accuracy of the proposed method on two independent data sources and consistently achieved a mean absolute error of below 1 pixel. We also tested the method robustness to sparse or noisy acquisitions with the same datasets and demonstrated consistent performance when the number of projections was reduced by a factor of up to 10 or the blank scan factor of the underlying Poisson's noise was increased to 10. We also illustrated the interpretability of the proposed method by mapping out the relative contributions of continuous spatial features to the overall classification task. This method, delivered as tomo-center, an open-source command-line tool, has been integrated into several tomography software packages to assist experiments during the routine beamline operations.
