Gang G J, Siewerdsen J H, Stayman J W
Department of Biomedical Engineering, Johns Hopkins University, Baltimore MD, USA 21205.
Proc SPIE Int Soc Opt Eng. 2017 Feb 11;10132. doi: 10.1117/12.2255517. Epub 2017 Mar 9.
This work presents a task-driven joint optimization of fluence field modulation (FFM) and regularization in quadratic penalized-likelihood (PL) reconstruction. Conventional FFM strategies proposed for filtered-backprojection (FBP) are evaluated in the context of PL reconstruction for comparison.
We present a task-driven framework that leverages prior knowledge of the patient anatomy and imaging task to identify FFM and regularization. We adopted a maxi-min objective that ensures a minimum level of detectability index () across sample locations in the image volume. The FFM designs were parameterized by 2D Gaussian basis functions to reduce dimensionality of the optimization and basis function coefficients were estimated using the covariance matrix adaptation evolutionary strategy (CMA-ES) algorithm. The FFM was jointly optimized with both space-invariant and spatially-varying regularization strength () - the former via an exhaustive search through discrete values and the latter using an alternating optimization where was exhaustively optimized locally and interpolated to form a spatially-varying map.
The optimal FFM inverts as increases, demonstrating the importance of a joint optimization. For the task and object investigated, the optimal FFM assigns more fluence through less attenuating views, counter to conventional FFM schemes proposed for FBP. The maxi-min objective homogenizes detectability throughout the image and achieves a higher minimum detectability than conventional FFM strategies.
The task-driven FFM designs found in this work are counter to conventional patterns for FBP and yield better performance in terms of the maxi-min objective, suggesting opportunities for improved image quality and/or dose reduction when model-based reconstructions are applied in conjunction with FFM.
本研究提出了一种在二次惩罚似然（PL）重建中对注量场调制（FFM）和正则化进行任务驱动的联合优化方法。针对滤波反投影（FBP）提出的传统FFM策略在PL重建的背景下进行评估以作比较。
我们提出了一个任务驱动框架，该框架利用患者解剖结构和成像任务的先验知识来确定FFM和正则化。我们采用了一个极大极小目标，以确保图像体积中各个样本位置的可检测性指数（）达到最低水平。FFM设计由二维高斯基函数参数化，以降低优化的维度，并使用协方差矩阵自适应进化策略（CMA-ES）算法估计基函数系数。FFM与空间不变和空间变化的正则化强度（）联合优化——前者通过对离散值的穷举搜索，后者使用交替优化，其中在局部进行穷举优化并进行插值以形成空间变化图。
随着的增加，最优FFM会反转，这表明联合优化的重要性。对于所研究的任务和对象，最优FFM通过衰减较小的视图分配更多的注量，这与针对FBP提出的传统FFM方案相反。极大极小目标使整个图像的可检测性均匀化，并实现比传统FFM策略更高的最低可检测性。
本研究中发现的任务驱动FFM设计与FBP的传统模式相反，并且在极大极小目标方面具有更好的性能，这表明当基于模型的重建与FFM结合应用时，有机会提高图像质量和/或降低剂量。
