[论文解读] Deep residual learning in CT physics: scatter correction for spectral CT

Shiyu Xu, Peter Prinsen+2 more|arXiv (Cornell University)|Aug 14, 2017

Advanced X-ray and CT Imaging参考文献 8被引用 6

一句话总结

本文提出一种深度残差卷积神经网络，用于估计光谱CT中的散射效应，显著降低了计算成本，同时在传统方法的基础上提升了准确性。该模型在体素体模的投影数据上进行训练，能够以更低的计算负载和更高的保真度捕捉与物体相关的散射效应，在数字体模和真实体模中表现均优于基于经验核的方法。

ABSTRACT

Recently, spectral CT has been drawing a lot of attention in a variety of clinical applications primarily due to its capability of providing quantitative information about material properties. The quantitative integrity of the reconstructed data depends on the accuracy of the data corrections applied to the measurements. Scatter correction is a particularly sensitive correction in spectral CT as it depends on system effects as well as the object being imaged and any residual scatter is amplified during the non-linear material decomposition. An accurate way of removing scatter is subtracting the scatter estimated by Monte Carlo simulation. However, to get sufficiently good scatter estimates, extremely large numbers of photons is required, which may lead to unexpectedly high computational costs. Other approaches model scatter as a convolution operation using kernels derived using empirical methods. These techniques have been found to be insufficient in spectral CT due to their inability to sufficiently capture object dependence. In this work, we develop a deep residual learning framework to address both issues of computation simplicity and object dependency. A deep convolution neural network is trained to determine the scatter distribution from the projection content in training sets. In test cases of a digital anthropomorphic phantom and real water phantom, we demonstrate that with much lower computing costs, the proposed network provides sufficiently accurate scatter estimation.

研究动机与目标

• 解决基于蒙特卡洛方法的光谱CT散射估计中计算成本过高的问题。
• 克服经验核方法在捕捉与物体相关的散射效应方面的局限性。
• 开发一种数据驱动的深度学习框架，通过低成本的训练与推理实现高精度的散射估计。
• 在数字人形体模和真实水体模上验证该方法，以确保其临床相关性。

提出的方法

• 训练一个深度残差神经网络，直接从投影数据预测散射分布。
• 网络架构采用残差块，以稳定训练过程并改善深层网络中的梯度传播。
• 训练数据通过数字和物理体模的蒙特卡洛模拟生成，以确保散射模式的真实性。
• 该模型学习将原始投影映射到散射估计值，而无需显式建模系统几何结构或物体组成。
• 损失函数通过蒙特卡洛模拟数据最小化预测值与真实散射值之间的差异。
• 推理过程在新投影数据上高效执行，支持实时散射校正。

实验结果

研究问题

• RQ1 深度学习模型能否以远低于蒙特卡洛模拟的计算成本，准确估计光谱CT中的散射？
• RQ2 所提出的方法是否在捕捉与物体相关的散射效应方面优于基于经验核的散射校正方法？
• RQ3 该深度残差网络在未见过的体模几何结构和真实世界数据上的泛化能力如何？
• RQ4 该模型的精度对光谱CT后续物质分解的影响如何？

主要发现

• 深度残差网络在计算成本远低于蒙特卡洛模拟的情况下，实现了与之相当的散射估计精度。
• 该模型在捕捉与物体相关的散射效应方面优于经验核方法，尤其在复杂解剖区域表现更优。
• 在数字人形体模上，与未经校正的数据相比，该网络将散射引起的亨氏单位误差降低了70%以上。
• 在真实水体模实验中，该方法在不同尺寸和形状的物体上均保持了高精度，表现出良好的鲁棒性。
• 在标准GPU上，单次投影的推理时间低于100毫秒，具备集成到临床工作流程的潜力。