U-net-based deformation vector field estimation for motion-compensated 4D-CBCT reconstruction
U-net-based deformation vector field estimation for motion-compensated 4D-CBCT reconstruction
复制标题
DOI：
10.1002/mp.14150
复制
发表时间：
2020-04-27
期刊：
影响因子：
3.8
通讯作者：
Wang, Jing
中科院分区：
文献类型：
--
作者：
Huang, Xiaokun;Zhang, You;Wang, Jing
目的 对于四维锥形束计算机断层扫描（4D-CBCT），其图像质量通常会因相位排序后每个呼吸相位的投影不足而降低。最近，我们开发了一种同步运动估计和图像重建 (SMEIR) 技术，该技术可以通过合并作为变形矢量场 (DVF) 生成的相间运动模型来提高肺部 4D-CBCT 重建质量。同时运动估计和图像重建使用强度驱动的二维 (2D)-三维 (3D) 变形技术，通过强度匹配的 2D 投影来估计这些 DVF。然而，2D-3D 变形可能无法生成准确的肺内 DVF，因为复杂的小型肺结构的运动只会导致 2D 投影上的细微强度变化，而不足以驱动准确的 DVF 优化。本研究旨在开发基于卷积神经网络（CNN）的方法来微调 2D-3D 变形 DVF，以提高 4D-CBCT 重建的效率和准确性。 方法 我们为本研究构建了两种基于 U-net 的架构。第一个架构 (U-net-3C) 使用 2D-3D 变形估计 DVF（在三个基本方向）作为具有三个通道 (3C) 的输入，输出是微调的 DVF。对于第二种架构 (U-net-4C)，添加了由 SMEIR 重建的参考相 CBCT 图像作为附加输入通道 (4C)，以表示患者特定的肺部异质特性。两种模型的输出经过微调的高质量 DVF 再次输入到 SMEIR 工作流程中，作为优化的运动模型，以生成最终的 4D-CBCT。使用五重交叉验证对 11 个肺部患者病例对两种方法进行了评估。我们还通过原始SMEIR和SMEIR-Bio（具有生物力学建模的SMEIR）算法重建了4D-CBCT以进行比较。通过均方根误差 (RMSE)、通用质量指数 (UQI) 和归一化互相关 (NCC) 等指标定量评估 4D-CBCT 准确性。通过手动跟踪肺部标志来评估 DVF 准确性。我们还使用上述指标基于重建的 4D-CBCT 质量在 SPARE 挑战数据集上评估了我们提出的方法。结果 SMEIR-U-net-3C、SMEIR-U-net-4C、SMEIR-Bio 和 SMEIR 的平均（+/- 标准偏差）残余 DVF 误差分别为 3.88 +/- 3.12 mm、3.71 +/- 2.90 mm、3.75 +/- 3.40分别为 5.73 +/- 4.61 毫米和 5.73 +/- 4.61 毫米。与其他方法相比，SMEIR-U-net-3C 和 SMEIR-U-net-4C 生成的 RMSE、UQI 和 NCC 普遍得到改善的图像。与SMERI-U-net-3C相比，SMEIR-U-net-4C具有略高的4D-CBCT重建和DVF估计精度。对于SPARE数据集，SMEIR-U-net-3C、SMEIR-U-net-4C、SMEIR-Bio和SMEIR的UQI分别为0.96、0.97、0.96和0.94。结论基于CNN的模型可以实现快速（10 s）和准确的DVF微调，从而提高4D-CBCT重建的效率和准确性。
Purpose For four-dimensional cone-beam computed tomography (4D-CBCT), its image quality is usually degraded by insufficient projections at each respiratory phase after phase-sorting. Recently, we developed a simultaneous motion estimation and image reconstruction (SMEIR) technique, which can improve lung 4D-CBCT reconstruction quality by incorporating an interphase motion model generated as deformation vector fields (DVFs). Simultaneous motion estimation and image reconstruction uses an intensity-driven two-dimensional (2D)-three-dimensional (3D) deformation technique to estimate these DVFs by intensity-matching 2D projections. However, 2D-3D deformation may fail to generate accurate intra-lung DVFs, since the motion of intricate, small lung structures only leads to subtle intensity variations on 2D projections that are insufficient to drive accurate DVF optimization. This study is to develop convolutional neural network (CNN)-based methods to fine-tune the 2D-3D deformation DVFs to improve the efficiency and accuracy of 4D-CBCT reconstruction.Methods We built two U-net-based architectures for this study. The first architecture (U-net-3C) uses 2D-3D deformation-estimated DVFs (in three cardinal directions) as the input with three channels (3C), and outputs are fine-tuned DVFs. For the second architecture (U-net-4C), the reference phase CBCT image reconstructed by SMEIR was added as an additional input channel (4C) to represent patient-specific heterogeneous properties of the lung. The output fine-tuned high-quality DVFs of both models were input again into the SMEIR workflow, as an optimized motion model, to generate the final 4D-CBCT.Both methods were evaluated on 11 lung patient cases, using fivefold cross-validation. We also reconstructed 4D-CBCTs by the original SMEIR and the SMEIR-Bio (SMEIR with biomechanical modeling) algorithms for comparison. The 4D-CBCT accuracy was quantitatively assessed through metrics including root-mean-square-error (RMSE), universal quality index (UQI), and normalized cross-correlation (NCC). The DVF accuracy was evaluated by manually tracked lung landmarks. We also evaluated our proposed methods on the SPARE challenge dataset based on reconstructed 4D-CBCT quality using the above metrics.Results The average (+/- standard deviation) residual DVF errors of SMEIR-U-net-3C, SMEIR-U-net-4C, SMEIR-Bio, and SMEIR were 3.88 +/- 3.12 mm, 3.71 +/- 2.90 mm, 3.75 +/- 3.40 mm, and 5.73 +/- 4.61 mm, respectively. The SMEIR-U-net-3C and SMEIR-U-net-4C generated images of generally improved RMSE, UQI, and NCC as compared to the other methods. Compared with SMERI-U-net-3C, SMEIR-U-net-4C has slightly higher 4D-CBCT reconstruction and DVF estimation accuracy. For the SPARE dataset, the UQI for SMEIR-U-net-3C, SMEIR-U-net-4C, SMEIR-Bio, and SMEIR were 0.96, 0.97, 0.96, and 0.94.Conclusion The CNN-based models can achieve fast (10 s) and accurate DVF fine-tuning to improve the efficiency and accuracy of 4D-CBCT reconstruction.