背景和目的：

磁共振（MR）模拟定位技术是最新的放疗模拟定位技术之一，与常用的CT模拟定位技术相比，磁共振图像（MRI）具有软组织分辨率高、无辐射剂量等优点。但目前还存在如下关键问题需要解决：1）缺乏MR模拟定位图像验收测试的总体策略和具体方案 ；2）单次激发扩散加权回波平面成像（DWEPI）是监测肿瘤生物学特性的重要序列，其几何精度亟待改进；3）无法提供电子密度信息用于放疗计划的剂量计算。针对上述关键问题，本课题拟建立完整的MR模拟定位图像验收测试流程，编写半自动图像分析程序，对模体扫描图像各项指标进行分析，完成对扫描仪和模拟定位射频线圈等硬件设备以及放疗模拟定位序列的验收测试。并对单次激发扩散加权EPI几何失真情况进行测量，研究满足放疗模拟定位精度的脑部肿瘤EPI扫描序列，将基于多路敏感度编码（MUSE）的多次激发EPI序列应用于放疗模拟定位，以减少EPDWI序列的几何畸变。拟基于鼻咽癌MR模拟定位图像，建立卷积神经网络模型，从MR图像合成CT图像，并验证模型的准确性。

材料和方法：

所有图像均在3.0T大孔径MR模拟定位机（Discovery MR750W，GE Healthcare）采集。使用ACR模体进行图像质量分析，编程实现自动分析几何畸变百分比（%GD）、层位置（SP）、层厚（ST）、百分图像均匀性（PIU）、伪影比（GR）和信噪比（SNR）6个定量指标和半自动分析高对比空间分辨率(HCSR)及低对比分辨率（LCD)2个指标。使用ACR标准序列和正交头线圈检查扫描仪的基本性能。使用ACR标准序列测试6通道分离式头部放疗模拟定位线圈和体部相控阵线圈，并与诊断线圈进行比较。使用放疗临床结构像脉冲序列配合模拟定位射频线圈，对序列参数设置和扫描仪的高级性能进行验收测试。并通过模体偏中心摆位，完成大孔径扫描仪模拟定位序列几何畸变测量。

建立间隔采集的多次激发EPI序列和基于MUSE的无导航脉冲重建算法。采集T2 PORPELLE序列作为参考图像，与SS-DWEPI和MUSE-DWEPI的b0和b1000图像（b = 0s / mm²和1000s / mm²）进行比较。开发内部程序进行自动交互式几何畸变测量和感兴趣区域自动ADC值计算。在模体研究中，在近中心层面和远离中心层面进行扫描，自动计算MUSE-DWEPI和SS-DWEPI控制点的相对位移矢量，及模体直径的％GD。在患者研究中，入组需行MR模拟定位的10名脑肿瘤患者（其中6名患者为术后脑胶质瘤，4名患者为脑转移瘤）。基于T2 PROPELLER，SS-DWEPI和MUSE-DWEPI分别勾画肿瘤靶区和四个脑区，使用Dice相似性系数（DSC）和Hausdorff距离（HD）来计算几何畸变的水平。

回顾性分析在我院行放射治疗的早期鼻咽癌患者20例。每位患者分别完成MRI和CT模拟定位扫描。对MRI T1 FSE序列和CT图像进行配准及预处理。建立端到端的全卷积神经网络，采用10折交叉验证方法，训练伪CT（pCT）生成模型。基于体素点计算平均绝对误差（MAE）、平均误差（ME）和均方根误差（RMSE），验证pCT的准确性。

结果:

ACR模体图像自动分析结果与人工分析结果，各项指标差别均可控制在1%或0.05 mm以内，平均每套图像的分析时间从人工分析的16.27分钟，下降到1.73分钟。使用ACR标准序列和正交头线圈，图像质量各项指标均可达到ACR推荐的标准。6通道头部放疗模拟定位线圈与8通道诊断头部线圈相比，PIU降低约34％，其它7个图像质量参数均在可接受范围。通过强度校准方法，PIU可以提高到85％以上。模拟定位序列包括T1加权和T2加权图像，几何畸变在等中心和偏离中心10 cm处%GD均可控制±1％以内。

EPI序列模体研究，在近中心层面和在z方向远离中心层面采用不同的b值测量，SS-DWEPI序列的％GD 均在7％-8％的范围内波动，而 MUSE-DWEPI序列％GD 的波动范围均可控制在2％-3％，图像畸变主要沿相位编码方向。在近中心层面，SS-DWEPI序列 b0和b1000所有控制点平均相对位移为4.45 ± 3.44 mm，MUSE-DWEPI 序列b0和b1000的平均相对位移为2.17 ± 1.9 mm。从距离中心层面0.8 cm增加到4.8 cm时，SS-DWEPI序列控制点相对位移的平均值增加8.3 mm，较MUSE-DWEPI明显。对于入组所有病人勾画的脑区和靶区，使用MUSE-DWEPI较SS-DWEPI序列，DSC值平均提高0.32 ± 0.31（p <0.01），HD值平均降低3.69 ±1.36 mm（p <0.01）。

对1433套配对的MRI和CT模拟定位图片，训练pCT生成模型，进行10折的交叉验证。平均ME为 - 9.3 ± 16.9 HU，平均MAE为102.6 ± 11.4 HU，平均RMSE为209.8 ± 22.6 HU。模型训练的平均时间为2.22 ± 0.04小时，每例病人pCT生成的平均时间为7.90 ± 0.47秒。

结论：

本课题建立了一套半自动MR模拟定位图像验收测试方法，保证了MR模拟定位系统投入临床使用的精度和图像质量，及图像分析的效率和精度；首次提出将相位编码方向分段的多次激发EPI序列和基于MUSE的无导航脉冲重建算法应用于放疗模拟定位，通过模体研究和患者研究，证明与SS-DWEPI相比，MUSE-DWEPI序列显著减少图像的几何畸变，可以应用于脑部放疗模拟定位；建立端到端的卷积神经网络，实现从鼻咽癌MRI到pCT的转换，且具有较高的转换精度和效率，在基于单独MRI的放疗流程中，可减少患者额外CT的使用，降低辐射剂量，保证勾画的精度。

background and ives:

magnetic resonance (mr) simulation is one of the latest radiotherapy simulation technologies. compared with ct simulation, mr simulation has the advantages of nonionizing radiation and superior soft-tissue. however, the following key problems are still need to be solved: 1) lack of the overall strategy and method for acceptance and commissioning of an mr simulator; 2) as an commonly used  sequence for monitoring the biological characteristics of tumors, the geometric accuracy of single-shot diffusion-weighted echo planar imaging (dwepi) needs to be improved; 3) the electron density information cannot be derived from mri for dose calculation of the radiotherapy plan. we proposed a semi-automatic method for image acceptance and commissioning for the scanner, the radiofrequency coils, and pulse sequences for an mr simulator. and we aimed to investigate the use of multiplexed sensitivity-encoding dwepi (muse-dwepi) to reduce the geometric distortions to allow its use in brain radiotherapy. lastly, we aimed to generate pseudo-ct(pct) from mri using convolutional neural networks(cnns) for nasopharyngeal carcinoma, and the accuracy of the model was verified

materials and methods:

all the mr scans in this study were performed on a 3.0 t scanner (discovery mr750w，ge healthcare). the acr mri accreditation large phantom was used for image quality analysis with eight parameters. codes and procedures were developed for automatic analysis of percentage geometric distortion(%gd), slice position(sp), slice thickness(st), percent image uniformity(piu), ghosting ratio(gr) and signal-to-noise ratio(snr), and semi-automatic analysis of high-contrast spatial resolution (hcsr) and low contrast detectability(lcd). standard acr sequences with a split head coil were adopted to examine the scanner’s basic performance. the performance of simulation radiofrequency coils were measured and compared using the standard sequence with different clinical diagnostic coils. we used simulation sequences with simulation coils to test the quality of image and advanced performance of the scanner. for testing the geometric accuracy of whole fov, the acr phantom was installed in both the isocenter and 10 cm off-center.

muse algorithm was applied for interleaved dwepi technique of multi-shot acquisition without the need for navigator echoes. the image distortion levels in ss-dwepi and muse-dwepi were compared using t2 propeller images as the reference. in-house programs were developed for automatic interactive geometric distortion measurements and adc value extraction for regions of interest. in a phantom study, phantom diameters measured using muse-dwepi and ss-dwepi were compared and the percentages of geometric distortion (%gd) were calculated. the shifting vectors of control points were also plotted and calculated. in a patient study, ten patients (six with post-surgery glioma, four with brain metastases) requiring mri were enrolled. the tumor targets and four brain regions were delineated based on t2 propeller, ss-dwepi, and muse-dwepi, using the dice similarity coefficient (dsc) and the hausdorff distance (hd) to quantify the level of geometric distortion.

20 patients with early nasopharyngeal carcinoma who accepted radiotherapy in our department were enrolled retrospectively. each patient had received ct and mri simulation before treatment. paired mri t1 fse and ct images were registered and preprocessed.  an end-to-end full convolutional neural network was established, and a 10-fold cross-validation method was used to train a pseudo-ct (pct) generation model. the mean absolute error (mae), mean error (me), and root mean square error (rmse) were calculate voxel by voxel to verify the accuracy of pct.

results:

the semi-automatic image analysis program developed by this subject can automatically analyze six image quality parameter indicators. the difference between the automatic analysis result and the manual analysis result was within 1% and 0.05 mm, and the average analysis time of each set of images decreased from 16.27 min to 1.73 min. when using standard acr sequences with a split head coil, image quality passed all acr recommended criteria. the image intensity uniformity with a simulation radiofrequency coil decreased about 34% compared with the 8-channel diagnostic head coil, while the other six image quality parameters were acceptable. those two image quality parameters could be improved to more than 85% by built-in intensity calibration methods. the geometric distortion of simulation sequences such as t1-weighted and t2-weighted images was well-controlled in the isocenter and 10 cm off-center within a range of ± 1%.

in the phantom study for epi, the most prominent image distortion was along the phase encoding direction in terms of %gd fluctuated from 7%–8% for ss-dwepi (b = 0 s/mm2 and 1000 s/mm2) and fluctuated from 2%–3% for muse-dwepi (b = 0 s/mm2 and 1000 s/mm2)  with different positions and b values. the mean relative displacement of all control points was 4.45 ± 3.44 mm for ss-dwepi b0 and 2.17 ± 1.9 mm for muse-dwepi b0 close to the isocenter. increasing the distance away from the isocenter in the z direction (from 0.8 cm to 4.8 cm), the distortion increased 8.3 mm in ss-dwepi which is much more than muse-dwepi. for all brain regions and targets involved, the mean improvement when switching from ss-dwepi to muse-dwepi was 0.32 ± 0.31 in dsc (p <0.01) and 3.69 ±1.36 mm in hd (p <0.01）.

the pct generating cnn model was trained using 1433 paired mri and ct simulation images with 10-fold cross-validation. the average me was - 9.3 ± 16.9 hu, the average mae was 102.6 ± 11.4 hu, and the average rmse was 209.8 ± 22.6 hu. the average time for model training was 2.22 ± 0.04 hours, and the average time for pct generation per patient was 7.90 ± 0.47 seconds.

conclusions:

we developed a semi-automatic image quality analysis method for quantitative evaluation of images and commissioning of an mri simulator. the accuracy and safety of mr simulation is ensured when applied in clinic, and the consistency and reproducibility of image analysis is improved. we first investigated that the muse technique produces multi-shot dwi data in phase direction with less geometric distortion compared with ss-dwepi. muse-dwepi might be a promising application strategy for dwi in radiotherapy. our established end-to-end convolutional neural networks have the good performance in generating pct from mri evaluated with accuracy and efficiency for nasopharyngeal carcinoma. it might be a potential tool for mri-only radiotherapy to reduce ionizing radiation by ct simulation and ensure delineating precisely.