分享:
分享到微信朋友圈
X
特别关注
基于深度学习重建扩散加权成像图像评估甲状腺相关性眼病的活动性
王云梦 崔园园 倪双爽 代建昆 万欣怡 陈歆 姜沁伶 程宇欣 张天然 马宜传 萧毅

Cite this article as: WANG Y M, CUI Y Y, NI S S, et al. Deep learning-based reconstruction of diffusion-weighted imaging images to assess the activity of thyroid-associated ophthalmopathy[J]. Chin J Magn Reson Imaging, 2024, 15(10): 36-42, 68.本文引用格式:王云梦, 崔园园, 倪双爽, 等. 基于深度学习重建扩散加权成像图像评估甲状腺相关性眼病的活动性[J]. 磁共振成像, 2024, 15(10): 36-42, 68. DOI:10.12015/issn.1674-8034.2024.10.007.


[摘要] 目的 探讨深度学习重建(deep learning reconstruction, DLR)扩散加权成像(diffusion weighted imaging, DWI)图像在甲状腺相关性眼病(thyroid-associated ophthalmopathy, TAO)活动期与非活动期评估中的价值。材料与方法 2023年4月至9月前瞻性纳入73例临床诊断为TAO的患者(活动期46例,非活动期27例)及26例健康对照组。所有受试者均使用3.0 T磁共振扫描仪和21通道头颈联合线圈进行眼眶MRI扫描。对视野优化无失真的单次激发成像和多路复用敏感度编码技术(field of view optimized and constrained undistorted single-shot imagingand multiplexed sensitivity encoding, FOCUS MUSE)的DWI序列进行传统重建(conventional reconstruction, ConR)和DLR。两名医师采用四分制评分量表独立对两种序列的图像质量进行主观评价。通过测量眼外肌的信噪比(signal-to-noise ratio, SNR)和对比噪声比(contrast-to-noise ratio, CNR)对图像质量进行客观评估。使用眼外肌的DWI表观扩散系数(apparent diffusion coefficient, ADC)值评估TAO的活动性。使用Wilcoxon检验比较ConR和DLR DWI间SNR、CNR和ADC值的差异。以临床活动度评分(Clinical Activity Score, CAS)作为金标准,采用Kruskal-Wallis检验比较健康对照组、活动期组和非活动期组TAO患者的ADC值差异。使用受试者工作特征(receiver operating characteristic, ROC)曲线评价DLR DWI鉴别活动性TAO的性能,Spearman评估眼外肌的ADC值与CAS之间的相关性。结果 DLR DWI在边界清晰度和整体图像质量上的主观评分显著高于ConR DWI,两序列的组内和组间一致性均达良好(Kappa>0.650)。和ConR相比,眼外肌DLR DWI的SNR和CNR显著提高(P<0.001)。ConR和DLR DWI的眼外肌ADC值差异无统计学意义(P>0.05)。在两种序列中,活动期组TAO患者眼外肌的ADC值均显著高于非活动期组TAO患者和健康对照组(P<0.001)。非活动期组TAO患者与健康对照组的ADC值差异均无统计学意义(P>0.05)。ConR DWI(r=0.637,P<0.001)和DLR DWI(r=0.662,P<0.001)的眼外肌ADC值均与CAS显著正相关。在活动性评估中,DLR DWI的诊断效能高于ConR DWI(曲线下面积:0.959 vs. 0.939,P=0.020)。结论 DLR在不增加扫描时间的同时提高了眼眶的图像质量。相较于ConR,基于DLR DWI得到的ADC值在鉴别TAO的活动性及与CAS的相关性方面均有所提升。
[Abstract] Objective To investigate the value of deep learning reconstruction (DLR) orbital diffusion weighted imaging (DWI) images in the assessment of active and inactive stages of thyroid-associated ophthalmopathy (TAO).Materials and Methods This prospectively study included 73 clinically diagnosed TAO patients (46 active TAO, 27 inactive TAO) and 26 healthy controls from April to September 2023. All participants underwent orbital MRI scans using a 3.0 T MRI scanner and a 21ch head-and-neck combined coil. DWI sequences with field of view optimized and constrained undistorted single-shot imaging and multiplexed sensitivity encoding (FOCUS MUSE) were reconstructed by conventional reconstruction (ConR) and DLR. Two diagnostic radiologists independently subjectively evaluated the image quality of the two sequences using a four-point Likert scale. The image quality was objectively evaluated by measuring the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of extraocular muscle (EOM). The DWI apparent diffusion coefficient (ADC) of EOM was used to distinguishing active from inactive TAO. The Wilcoxon was applied to test the difference of SNR, CNR, and ADC between ConR and DLR DWI, separately. Using the Clinical Activity Score (CAS) as the gold standard. The Kruskal-Wallis test was used to compare the difference of ADC between healthy controls, active and inactive TAO patients. Receiver operating characteristics (ROC) curves were used to compare the diagnostic performance of EOM ADC for differentiating active from inactive TAO patients between ConR and DLR DWI. The correlation between the EOM ADC and CAS of TAO patients was analyzed using Spearman's rank correlation coefficient.Results DLR DWI had significantly higher subjective scores than ConR DWI for Sharpness of boundaries and overall image quality. The intra- and inter-reader agreement for both sequences was good (Kappa>0.650). Significantly higher SNR and CNR in EOM DLR DWI compared to ConR (all P<0.001). No significant difference of EOM ADC was observed between ConR and DLR DWI (P>0.05). In both sequences, the EOM ADC obtained was significantly higher in the active TAO than in both inactive TAO and healthy controls, respectively (all P<0.001). There was no significant difference of EOM ADC between inactive TAO and healthy controls (P>0.05). The EOM ADC extracted from both ConR DWI (r=0.637, P<0.001) and DLR DWI (r=0.662, P<0.001) was significantly positively correlated with the CAS. Compared with ConR DWI, DLR DWI presented better performance for discriminating active from inactive TAO patients (area under the curve: 0.959 vs. 0.939, P=0.020).Conclusions DLR improved the image quality of orbital DWI without increasing scan time. Compared to ConR, ADC values obtained based on DLR DWI were improved in identifying the activity of TAO and correlation with CAS.
[关键词] 甲状腺相关性眼病;眼外肌;扩散加权成像;深度学习重建;磁共振成像
[Keywords] thyroid-associated ophthalmopathy;extraocular muscle;diffusion weighted imaging;deep learning reconstruction;magnetic resonance imaging

王云梦 1, 2   崔园园 2   倪双爽 2   代建昆 3   万欣怡 2   陈歆 2   姜沁伶 2   程宇欣 2   张天然 2   马宜传 4*   萧毅 2*  

1 蚌埠医科大学研究生院,蚌埠 233000

2 海军军医大学第二附属医院放射科,上海 200003

3 通用电气医疗系统贸易发展(上海)有限公司,上海 200120

4 蚌埠医科大学第一附属医院放射科,蚌埠 233000

通信作者:萧毅,E-mail: xiaoyi@188.com 马宜传,E-mail: myc57688754@163.com

作者贡献声明:萧毅、马宜传设计本研究的方案,对稿件的重要内容进行了修改;王云梦起草和撰写稿件,获取、分析并解释本研究的数据;崔园园、倪双爽、代建昆、万欣怡、陈歆、姜沁伶、程宇欣、张天然获取、分析和解释本研究的数据,对稿件重要内容进行了修改;萧毅获得了国家重点研发计划项目、国家自然科学基金项目、军队后勤保健课题项目资助;全体作者都同意发表最后的修改稿,同意对本研究的所有方面负责,确保本研究的准确性和诚信。


基金项目: 国家重点研发计划项目 2022YFC2410002 国家自然科学基金项目 82271994 军队后勤保健课题项目 22BJZ07
收稿日期:2024-01-10
接受日期:2024-05-13
中图分类号:R445.2  R581.1  R771.3 
文献标识码:A
DOI: 10.12015/issn.1674-8034.2024.10.007
本文引用格式:王云梦, 崔园园, 倪双爽, 等. 基于深度学习重建扩散加权成像图像评估甲状腺相关性眼病的活动性[J]. 磁共振成像, 2024, 15(10): 36-42, 68. DOI:10.12015/issn.1674-8034.2024.10.007.

0 引言

       甲状腺相关性眼病(thyroid-associated ophthalmopathy, TAO)是一种自身免疫性疾病,居成人眼眶疾病发病率首位[1, 2]。TAO的病理变化以活动期的炎症、水肿和非活动期的纤维化、脂肪浸润为主[1]。不同分期的患者其治疗方式不同,准确分期有利于制订精准的治疗方案[1]。临床上依据临床活动度评分(Clinical Activity Score, CAS)对TAO进行分期[3]。然而,CAS对于亚临床患者和治疗期间的疾病进展不敏感且无法评估眼眶深部结构[4, 5]。因此,非侵入的定量技术可能更客观、准确地评估TAO。

       既往研究显示,扩散加权成像(diffusion weighted imaging, DWI)有助于评估TAO的活动性[6]。但眼眶深部解剖构造复杂,常规DWI易产生伪影和失真,影响TAO受累结构(如眼外肌)的显示及量化,从而影响TAO活动性的评估。前期研究报道了快速梯度与自旋回波杂交的刀锋技术提高了眼眶DWI图像质量,但此技术扫描时间长且吸收比率高[7, 8]。视野优化无失真单次激发成像(field of view optimized and constrained undistorted single-shot imaging, FOCUS)和多路复用敏感度编码技术(multiplexed sensitivity encoding, MUSE)可同时实现小视野成像和多次激发,从而减小磁敏感伪影和几何失真[9, 10]。另外,深度学习已应用在放射学领域,商用版的深度学习重建(deep learning reconstruction, DLR)算法(AIRTM ReconDL, GE Healthcare)在不增加扫描时间的同时降低噪声、减轻Gibbs伪影,可获取高信噪比(signal-to-noise ratio, SNR)且边缘锐利的图像[11]。前期研究已将其应用于前列腺[12, 13]、膝关节[14]和脊柱[15]等部位,展示了较好的图像质量及病灶显示能力。

       目前尚未有FOCUS MUSE技术应用于TAO。因此,本研究拟使用FOCUS MSUE DWI对眼眶进行成像,探究DLR在提升眼眶DWI图像质量中的作用,以及评估其在TAO活动性评价中的价值,为TAO患者的MRI检查提供更可靠的定量技术,有助于制订更精准的临床治疗方案。

1 材料与方法

1.1 研究对象

       本前瞻性研究遵守《赫尔辛基宣言》,并通过上海长征医院医学伦理委员会批准(批准文号:82170858),受试者均已签署知情同意书。

       于2023年4月至9月期间,共纳入来上海长征医院内分泌科就诊并于放射科进行MRI检查的73例TAO患者。与此同时招募26名年龄和性别匹配的健康对照者入健康对照组。TAO患者纳入标准:(1)符合TAO的Bartley诊断标准[16];(2)年龄>18岁;(3)无眼眶其他疾病及外伤史。排除标准:(1)存在MRI禁忌证者;(2)图像质量较差,如假牙、金属植入物及运动伪影等影响影像诊断和数据分析;(3)MRI显示眼眶内存在肿瘤等其他疾病。健康对照组纳入标准:(1)无眼眶疾病史如外伤、肿瘤、视神经病变及其他原因不明的炎症性疾病;(2)眼眶MRI检查无异常。排除标准:(1)存在MRI禁忌证者;(2)图像质量较差,如假牙、金属植入物及运动伪影等影响影像诊断和数据分析。

1.2 临床评估

       临床内分泌科医生根据CAS对每位TAO患者进行评分。CAS采取7分制评分标准[16]:(1)自发性球后疼痛;(2)眼球运动后疼痛;(3)眼睑充血;(4)眼睑水肿;(5)结膜充血;(6)结膜水肿;(7)泪阜肿胀。每项1分,共7分。评分≥3分为活动期;评分<3分为非活动期。

1.3 眼眶MRI检查

       所有患者均采用3.0 T磁共振扫描仪(SIGNA Premier; GE Healthcare, Milwaukee, USA)和21通道头颈联合线圈进行眼眶MRI扫描。患者均采取仰卧位,扫描范围包括整个眼眶和视神经。所有患者均进行眼眶常规T2WI的反转恢复序列(short tau inversion recovery imaging, STIR)和FOCUS MUSE DWI序列的扫描。T2WI-STIR:重复时间4427 ms,回波时间85 ms,视野20 cm×20 cm,矩阵320×240,层厚2 mm,层数19,扫描时间1 min 15 s;FOCUS MUSE DWI:重复时间4500 ms,回波时间58.7 ms,视野20 cm×12 cm,矩阵140×140,层厚2 mm,层数19,b值=0、800 s/mm2,扫描时间2 min 15 s。FOCUS MUSE DWI图像分别使用ConR和DLR进行重建。DLR技术采用GE公司的AIRTM Recon DL技术,它是基于超过10 000对无伪影、高SNR、高空间分辨率图像及其对应的低SNR、低空间分辨率图像深度卷积网络的训练数据库的训练算法。

1.4 图像分析

1.4.1 主观评价

       两名具有10年(阅片者1)和15年(阅片者2)临床经验的放射诊断科副主任医师独立评估所有受试者的图像质量。他们对所有受试者的临床信息均未知。采用四分制评分量表,从磁敏感伪影、边界清晰度、几何失真和整体图像质量四个方面对眼外肌进行评分。一个月后阅片者1再次进行独立评估。评分标准详见表1

表1  主观评分标准
Tab. 1  Subjective scoring criteria

1.4.2 客观评价

       使用SNR和对比噪声比(contrast-to-noise ratio, CNR)对所有DWI进行定量评估。参考T2WI-STIR序列,谨慎避开周围结构,在DWI图像中,于两侧眼外肌肌腹显示最大层面进行勾画感兴趣区(region of interest, ROI)。为避开容积效应,ROI的面积占据眼外肌肌腹显示最大层面的2/3(图1A)。在相应层面的双侧颞肌处各勾画一个ROI,分别取其信号强度和标准差的平均值代表颞肌的信号强度和标准差用以计算SNR和CNR[17, 18]。SNR定义为眼外肌与颞肌信号强度的比值。CNR定义为眼外肌和颞肌的信号强度差与眼外肌和颞肌平方和的平方根的比值。将眼外肌ROI复制到对应的表观扩散系数(apparent diffusion coefficient, ADC)图,获取每条眼外肌的ADC值(图1B)。

图1  眼外肌的ROI勾画过程。1A:DWI图像;1B:DWI ADC图像。在双侧眼外肌显示最大层面及双侧颞肌上进行手动勾画ROI,并将眼外肌的ROI复制到对应ADC图上用以记录ADC值。图中蓝色圆圈分别代表勾画双侧眼外肌和颞肌的ROI。DWI:扩散加权成像;ADC:表观扩散系数;ROI:感兴趣区。
Fig. 1  ROI delineation process of extraocular muscle. 1A: DWI image; 1B: DWI ADC image. The ROIs of bilateral EOM and temporal muscle are drawn at the muscle belly's maximum cross section. The ROIs of EOMs defined above are copied to the ADC map and the ADC within each ROI is used for further analysis. The blue circles in the figure represent ROIs for drawing the bilateral extraocular muscles and temporal muscles, respectively. DWI: diffusion weighted imaging; ADC: apparent diffusison coefficient; ROI: region of interest.

1.5 统计学分析

       所有数据均采用SPSS(version 25.0; IBM, NY, USA)和MedCalc(version 22.0; Mariakerke, Belgium)软件进行统计学分析。采用Kolmogorov-Smirnov检验对定量资料进行正态分布和方差齐性检验。正态分布的定量资料用平均值±标准差描述,否则采用中位数(四分位间距)表示。使用t检验、卡方检验比较健康对照组和TAO患者组间年龄和性别差异。采用Kappa分析评估两种序列的组内和组间一致性(Kappa≤0.400,差;0.400<Kappa≤0.600,中等;0.600<Kappa<0.800,良好;Kappa≥0.800,优秀)。采用Wilcoxon符号秩和检验比较ConR和DLR的FOCUS MUSE DWI在眼外肌上的主观评分、SNR、CNR和ADC值。根据CAS将TAO患者分为活动期和非活动期,比较健康对照组、活动期组和非活动期组TAO患者的ADC值差异,如数据满足正态分布和方差齐性检验使用单因素方差分析,如不满足则使用Kruskal-Wallis检验。采用Dunn-Bonferroni事后检验对两两比较进行显著性分析。Spearman相关系数评估眼外肌ADC值与CAS的相关性。采用受试者工作特征(receiver operating characteristics, ROC)曲线评价ADC值区分活动期组和非活动期组TAO患者的诊断性能。DeLong检验比较曲线下面积(area under the curve, AUC)的差异。P<0.05为差异有统计学意义。

2 结果

2.1 一般资料

       该研究纳入73例TAO患者(活动期46例,非活动期27例)和26例健康对照组。基线资料见表2。TAO患者组与健康对照组的年龄(P=0.228)、性别(P=0.815)差异均无统计学意义。活动期组和非活动期组TAO患者间的CAS差异具有统计学意义(P<0.001)。

表2  TAO患者组与健康对照组的基线资料
Tab. 2  The clinical characteristics of patients with TAO and healthy controls

2.2 主观图像评估

       在眼外肌的磁敏感伪影和几何失真方面,两种序列的评分结果差异无统计学意义(P>0.050)(表3表4图2)。在眼外肌的边界清晰度和整体图像质量方面,FOCUS MUSE-DLR的评分均优于FOCUS MUSE-ConR(P<0.001)(表5表6图2)。两种序列的组内和组间一致性均为良好(Kappa>0.650)。

图2  活动期组与非活动期组TAO患者各一例影像图像。2A~2D:男,73岁,活动期TAO患者,CAS=3,影像学表现为双侧内直肌和右侧外直肌明显增粗、肿胀。2E~2H:女,61岁,非活动期TAO患者,CAS=1,影像学表现为双侧内直肌和外直肌未见明显肿胀。2A、2E:FOCUS MUSE-DLR图像;2B、2F:FOCUS MUSE-ConR图像;2C、2G:FOCUS MUSE-DLR ADC图像。2D、2H:FOCUS MUSE-ConR ADC图像。双侧内直肌和外直肌在FOCUS MUSE-DLR(2A、2E)上信号更均匀,且边界较锐利,而在FOCUS MUSE-ConR(2B、2F)上噪声较明显。基于DLR的ADC图上可清晰分辨视神经双侧的脑脊液(2C,红箭),而在基于ConR的ADC图上视神经与双侧脑脊液的界限较模糊(2D,红箭)。在DWI序列和ADC图中,红色矩形范围内的各眼眶结构被放大。TAO:甲状腺相关性眼病;CAS:临床活动度评分;FOCUS MUSE:视野优化约束无失真和多路复用敏感度编码技术;ADC:表观扩散系数;DLR:深度学习重建;ConR:传统重建。
Fig. 2  MRI images of one active and one inactive TAO patient, respectively. 2A-2D represent a 73-year-old female active TAO with CAS=3. Imaging findings show marked thickening and swelling of the medial rectus muscles bilaterally and the right lateral rectus muscle. 2E-2H represent a 61-year-old male inactive TAO with CAS=1. Imaging findings show no significant swelling of the medial rectus muscles and lateral rectus muscles bilaterally. 2A, 2E: FOCUS MUSE-DLR images; 2B, 2F: FOCUS MUSE-ConR images; 2C, 2G: FOCUS MUSE-DLR ADC map; 2D, 2H: FOCUS MUSE-ConR ADC map. Bilateral medial rectus muscles and lateral rectus muscle have more homogeneous signals and sharper borders on FOCUS MUSE-DLR (2A, 2E), whereas noise is more pronounced on FOCUS MUSE-ConR (2B, 2F). The cerebrospinal fluid (2C, red arrow) located bilateral to the optic nerve is more clearly presented in the FOCUS MUSE DWI-DLR ADC map because the improvement of DWI image quality. In contrast, the boundaries between the optic nerve and bilateral cerebrospinal fluid are blurred in the FOCUS MUSE-ConR ADC map (2D, red arrows). In the DWI sequences and ADC maps, the orbital structures within the red rectangle are magnified. TAO: thyroid-associated ophthalmopathy; FOCUS MUSE: field of view optimized and constrained undistorted single-shot imaging and mulitiplexed sensitivity encoding; ADC: apparent diffusion coefficient; DLR: deep learning reconstruction; ConR: conventional reconstruction.
表3  DLR前后的FOCUS MUSE序列在磁敏感伪影中的一致性分析
Tab. 3  Consistency analysis of FOCUS MUSE with and without deep learning reconstruction in terms of susceptibility artifacts
表4  DLR前后的FOCUS MUSE序列在几何失真中的一致性分析
Tab. 4  Consistency analysis of FOCUS MUSE with and without deep learning reconstruction in terms of geometric distortions
表5  DLR前后的FOCUS MUSE序列在边界清晰度中的一致性分析
Tab. 5  Consistency analysis of FOCUS MUSE with and without deep learning reconstruction in terms of sharpness of boundaries
表6  DLR前后的FOCUS MUSE序列在整体图像质量中的一致性分析
Tab. 6  Consistency analysis of FOCUS MUSE with and without deep learning reconstruction in terms of overall image quality

2.3 DLR和ConR DWI眼外肌SNR、CNR和ADC值比较

       如表7所示,相较于ConR DWI,DLR DWI显著提升了外直肌的SNR(P<0.001)和CNR(P<0.001)。同样对内直肌而言,DLR DWI的SNR(P<0.001)和CNR(P<0.001)也显著高于ConR DWI。外直肌和内直肌的ADC值分别在两种重建方法之间的差异均无统计学意义(P>0.05)。

表7  ConR和DLR FOCUS MUSE DWI间眼外肌SNR、CNR和ADC的比较
Tab. 7  Comparison of extraocular muscles SNR, CNR and ADC between ConR and DLR FOCUS MUSE DWI

2.4 健康对照组、活动期组和非活动期组TAO患者间ADC值比较

       如图3所示,ConR DWI揭示活动期组的眼外肌ADC值[1.59(1.54, 1.67)×10-3 mm2/s]显著高于非活动期组[1.40(1.34, 1.46)×10-3 mm2/s;P<0.001]和健康对照组[1.38(1.29, 1.43)×10-3 mm2/s;P<0.001]。DLR DWI也表明活动期组的眼外肌ADC值[1.64(1.55, 1.71)×10-3 mm2/s]显著高于非活动期组[1.40(1.35, 1.45)×10-3 mm2/s;P<0.001]和健康对照组[1.44(1.38, 1.51)×10-3 mm2/s; P<0.001]。ConR和DLR DWI的结果均显示非活动期组和健康对照组的眼外肌ADC值差异无统计学意义(P>0.05)。

图3  健康对照组、活动期组和非活动期组TAO患者间眼外肌ADC值比较。3A:DLR图像中三组人群的比较;3B:ConR图像中三组人群的比较。TAO:甲状腺相关性眼病;FOCUS MUSE:视野优化约束无失真和多路复用敏感度编码技术;DLR:深度学习重建;ConR:传统重建;ADC:表观扩散系数。
Fig. 3  Comparison of extraocular muscles ADC among healthy controls, active and inactive TAO patients. 3A: Comparison of three groups in DLR images; 3B: Comparison of three groups in ConR images. TAO: thyroid-associated ophthalmopathy; FOCUS MUSE: field of view optimized and constrained undistorted single-shot imaging and mulitiplexed sensitivity encoding; DLR: deep learning reconstruction; ConR: conventional reconstruction; ADC: apparent diffusion coefficient.

2.5 眼外肌ADC值与CAS相关性

       如图4所示,ConR DWI(r=0.637,P<0.001)和DLR DWI(r=0.662,P<0.001)获取的眼外肌ADC值与CAS均显著正相关。

图4  眼外肌ADC值与CAS的相关性。4A:DLR图像中眼外肌的ADC值与CAS的相关性;4B:ConR图像中眼外肌的ADC值与CAS的相关性。ADC:表观扩散系数;CAS:临床活动度评分;DLR:深度学习重建;ConR:传统重建。
Fig. 4  The correlation between extraocular muscle ADC value and CAS. 4A: Correlation between ADC values and CAS of extraocular muscles in DLR images; 4B: Correlation between ADC values and CAS of extraocular muscles in ConR images. ADC: apparent diffusion coefficient; CAS: Clinical Activity Score; DLR: deep learning reconstruction; ConR: conventional reconstruction.

2.6 活动期组与非活动期组TAO鉴别

       如图5所示,DLR和ConR DWI获取的眼外肌ADC值均能显著鉴别活动期组和非活动期组TAO患者(P<0.001)。DLR DWI的鉴别效能高于ConR DWI(AUC:0.959 vs. 0.939,P=0.020)。ConR DWI的眼外肌ADC阈值为1.485×10-3 mm2/s时对活动性TAO患者具有最佳诊断性能,敏感度为87.04%,特异度为94.57%,准确度为91.78%。而DLR DWI的眼外肌ADC阈值为1.505×10-3 mm2/s时具有最佳诊断性能,敏感度为94.44%,特异度为95.65%,准确度为95.21%。

图5  眼外肌ADC鉴别活动期组与非活动期组TAO患者ROC曲线。ADC:表观扩散系数;TAO:甲状腺相关性眼病;ROC:受试者工作特征;FOCUS MUSE:视野优化约束无失真和多路复用敏感度编码技术;DLR:深度学习重建;ConR:传统重建。
Fig. 5  ROC curve of the differentiation between active and inactive TAO using the extraocular muscles ADC value. ADC: apparent diffusion coefficient; TAO: thyroid-associated ophthalmopathy; ROC: receiver operating characteristic; FOCUS MUSE: field of view optimized and constrained undistorted single-shot imaging and mulitiplexed sensitivity encoding; DLR: deep learning reconstruction; ConR:conventional reconstruction.

3 讨论

       本研究使用FOCUS MUSE技术获取高分辨率和极大抑制磁敏感伪影的眼眶DWI图像,探索DLR对FOCUS MUSE DWI图像质量和其对TAO活动性鉴别效能的影响。结果显示,通过ConR和DLR DWI获取的眼外肌的ADC值均与CAS显著相关;相对于ConR,DLR能在不额外增加扫描时间的情况下,显著提升眼眶DWI图像的SNR和CNR,图像质量的提升使得眼外肌ADC值鉴别活动期与非活动期TAO的效能有所提高。

3.1 DLR在眼眶MRI图像质量中的优势

       前期研究已表明,DWI有助于评估TAO的活动性[6, 19]。目前临床上常使用单激发平面回波成像技术(echo planar imaging, EPI)采集DWI图像,但产生的磁敏感伪影和几何失真影响TAO的准确评估[20]。随着成像技术发展,新的基于EPI的DWI成像方法对磁场均匀性引起的图像伪影得到很好的控制。比如,FOCUS技术通过非共面射频脉冲选择性激发局部区域以实现小视野成像的目的,从而减少成像所需信号的读出时间,减少EPI DWI磁敏感伪影[9]。MUSE技术在相位编码方向进行多次激发、交错式填充以实现高分辨率和低磁敏感伪影的EPI DWI图像[10]。MUSE整合FOCUS技术(即FOCUS MUSE)进一步减轻伪影和失真[21]。尽管新的基于EPI的DWI技术极大提高了图像质量,但扩散梯度的施加会引起磁共振信号的降低,使得图像SNR降低[2]。在临床可接受的扫描时间内获取高分辨率、高SNR眼眶DWI图像仍具有一定挑战性。并行采集成像、压缩感知和高通道表面线圈等方法可以获取高SNR的图像质量,但在加速倍数过大时,会由于采样不足导致图像细节模糊[22, 23, 24]。近年来,深度学习在放射学领域的应用不断取得进展,特别是在疾病和病变的检测、分期、减少噪声和伪影等方面[25, 26, 27, 28]

       本研究中使用的DLR算法是一种基于1000对高空间分辨率和高SNR图像及对应的低空间分辨率和低SNR图像深度卷积网络的训练数据库的训练算法[29, 30]。直接作用于MRI的原始K空间,以提高图像SNR和锐利度,从而提高病变的检出率[31]。该方法可以弥补SNR的损失,且不延长扫描时间,从而降低了由于眼球运动产生伪影的概率,增加了患者的依从性。对比研究结果显示,DLR DWI图像中眼外肌的SNR和CNR均优于ConR,这对于准确评估眼外肌的受累情况至关重要。本研究使用的DLR算法可以在不改变K空间范围的条件下,直接从完全或欠采样的K空间数据中重建图像,去除噪声和Gibbs伪影,以实现低噪声、高SNR的图像质量[32]。基于DLR算法得到的高SNR图像有利于显示眼眶后部气-骨-软组织交界面的眼外肌,从而精准识别其受累程度。本研究将DLR算法应用在解剖结构相对较小的眼眶中证明了此算法的可行性。

3.2 DLR DWI在评估活动性TAO中的优势

       临床上,TAO另一个重要的关键点是准确区分活动期与非活动期,这是TAO患者选择治疗方案的关键。目前国内外指南中CAS被广泛用于判断TAO患者的活动性并指导临床治疗方案[33],但实践证明,CAS较为主观,在指导治疗中存在一定偏差。影像学可无创显示球后软组织微观结构的变化[34]。DWI通过检测水分子的扩散运动,可定量评估TAO,反映疾病的炎症过程[33]。ADC值的定量监测在TAO的活动性和治疗反应性方面均具有良好相关性[35]。POLITI等[36]报道TAO患者的眼外肌ADC值与CAS呈正相关。WU等[37]分析TAO患者治疗前后的ADC值,发现均与CAS具有相关性。LIU等[38]的研究结果显示,活动期组的ADC值明显高于非活动期组,且经激素治疗后其ADC值有所下降。HU等[39]也发现,两组人群间的ADC值存在明显差异。本研究结果显示DLR DWI获取的眼外肌ADC值与CAS也显著正相关,活动期组的ADC值均高于非活动期组,与既往研究结论相同,再次证明眼外肌ADC值可作为定量评估TAO活动性的指标,并监测眼外肌的炎症情况。

       既往研究都聚焦在传统重建的DWI序列上,本研究首次比较了两种重建方法在评估TAO活动性的差异,结果显示,DLR DWI获取的ADC值在鉴别活动期与非活动期TAO患者中的效能更高,并且提高了鉴别的特异度和准确度。考虑是由于DLR显著提升了DWI图像的SNR、减轻了Gibbs伪影,使得图像质量有所提升,提高眼外肌ADC值的鲁棒性。

3.3 局限性

       本研究存在一定局限性。第一,本研究属于前瞻性研究,纳入的样本量小,随着后续更多病例的入组,可以对研究结果进一步验证;第二,本研究中仅评估了DLR DWI在鉴别活动期TAO患者中的作用及与常规方法的比较,DLR DWI能否用于预测TAO的疗效需要进一步研究总结;第三,TAO的一个重要特征是胶原蛋白的产生和沉积,化学交换饱和转移成像(chemical exchange saturation transfer imaging, CEST)可能检测到此变化,基于DLR的CEST能否进一步提高活动期与非活动期的鉴别效能,将在后续研究中进行探讨。

4 结论

       与传统重建方法相比,DLR能在不增加额外扫描时间的情况下,显著提高眼眶DWI图像的SNR和CNR,且ADC值的定量不受影响。图像质量的提升提高了活动期与非活动期TAO患者的鉴别效能,DLR DWI的使用有利于更准确地对TAO活动性进行诊断。

[1]
BARRIO-BARRIO J, SABATER A L, BONET-FARRIOL E, et al. Graves' ophthalmopathy: visa versus EUGOGO classification, assessment, and management[J/OL]. J Ophthalmol, 2015, 2015: 249125 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/26351570/. DOI: 10.1155/2015/249125.
[2]
ČIVRNÝ J, KARHANOVÁ M, HÜBNEROVÁ P, et al. MRI in the assessment of thyroid-associated orbitopathy activity[J]. Clin Radiol, 2022, 77(12): 925-934. DOI: 10.1016/j.crad.2022.08.124.
[3]
BURCH H B, PERROS P, BEDNARCZUK T, et al. Management of thyroid eye disease: a consensus statement by the American thyroid association and the European thyroid association[J]. Thyroid, 2022, 32(12): 1439-1470. DOI: 10.1089/thy.2022.0251.
[4]
DOLMAN P J. Grading severity and activity in thyroid eye disease[J/OL]. Ophthalmic Plast Reconstr Surg, 2018, 34(4S Suppl 1): S34-S40 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/29952931/. DOI: 10.1097/IOP.0000000000001150.
[5]
TORTORA F, CIRILLO M, FERRARA M, et al. Disease activity in Graves' ophthalmopathy: diagnosis with orbital MR imaging and correlation with clinical score[J]. Neuroradiol J, 2013, 26(5): 555-564. DOI: 10.1177/197140091302600509.
[6]
KILICARSLAN R, ALKAN A, ILHAN M M, et al. Graves' ophthalmopathy: the role of diffusion-weighted imaging in detecting involvement of extraocular muscles in early period of disease[J/OL]. Br J Radiol, 2015, 88(1047): 20140677 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/25525866/. DOI: 10.1259/bjr.20140677.
[7]
FU Q, LIU D X, MA H, et al. Turbo gradient and spin-echo BLADE-DWI for extraocular muscles in thyroid-associated ophthalmopathy[J/OL]. J Clin Med, 2023, 12(1): 344 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/36615144/. DOI: 10.3390/jcm12010344.
[8]
LI Z Q, PIPE J G, LEE C Y, et al. X-PROP: a fast and robust diffusion-weighted propeller technique[J]. Magn Reson Med, 2011, 66(2): 341-347. DOI: 10.1002/mrm.23033.
[9]
TANABE M, HIGASHI M, BENKERT T, et al. Reduced field-of-view diffusion-weighted magnetic resonance imaging of the pancreas with tilted excitation plane: a preliminary study[J]. J Magn Reson Imaging, 2021, 54(3): 715-720. DOI: 10.1002/jmri.27590.
[10]
CHEN N K, GUIDON A, CHANG H C, et al. A robust multi-shot scan strategy for high-resolution diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE)[J/OL]. Neuroimage, 2013, 72: 41-47 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/23370063/. DOI: 10.1016/j.neuroimage.2013.01.038.
[11]
LEE K L, KESSLER D A, DEZONIE S, et al. Assessment of deep learning-based reconstruction on T2-weighted and diffusion-weighted prostate MRI image quality[J/OL]. Eur J Radiol, 2023, 166: 111017 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/37541181/. DOI: 10.1016/j.ejrad.2023.111017.
[12]
UEDA T, OHNO Y, YAMAMOTO K, et al. Deep learning reconstruction of diffusion-weighted MRI improves image quality for prostatic imaging[J]. Radiology, 2022, 303(2): 373-381. DOI: 10.1148/radiol.204097.
[13]
TAJIMA T, AKAI H, SUGAWARA H, et al. Feasibility of accelerated whole-body diffusion-weighted imaging using a deep learning-based noise-reduction technique in patients with prostate cancer[J/OL]. Magn Reson Imaging, 2022, 92: 169-179 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/35772583/. DOI: 10.1016/j.mri.2022.06.014.
[14]
JOHNSON P M, LIN D J, ZBONTAR J, et al. Deep learning reconstruction enables prospectively accelerated clinical knee MRI[J/OL]. Radiology, 2023, 307(2): e220425 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/36648347/. DOI: 10.1148/radiol.220425.
[15]
ALMANSOUR H, HERRMANN J, GASSENMAIER S, et al. Deep learning reconstruction for accelerated spine MRI: prospective analysis of interchangeability[J/OL]. Radiology, 2023, 306(3): e212922 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/36318032/. DOI: 10.1148/radiol.212922.
[16]
MOURITS M P, PRUMMEL M F, WIERSINGA W M, et al. Clinical activity score as a guide in the management of patients with Graves' ophthalmopathy[J]. Clin Endocrinol, 1997, 47(1): 9-14. DOI: 10.1046/j.1365-2265.1997.2331047.x.
[17]
TORTORA F, PRUDENTE M, CIRILLO M, et al. Diagnostic accuracy of short-time inversion recovery sequence in Graves' Ophthalmopathy before and after prednisone treatment[J]. Neuroradiology, 2014, 56(5): 353-361. DOI: 10.1007/s00234-014-1332-4.
[18]
MAYER E, HERDMAN G, BURNETT C, et al. Serial STIR magnetic resonance imaging correlates with clinical score of activity in thyroid disease[J]. Eye, 2001, 15(Pt 3): 313-318. DOI: 10.1038/eye.2001.102.
[19]
ABDEL RAZEK A A, EL-HADIDY M, MOAWAD M E, et al. Performance of apparent diffusion coefficient of medial and lateral rectus muscles in Graves' orbitopathy[J]. Neuroradiol J, 2017, 30(3): 230-234. DOI: 10.1177/1971400917691993.
[20]
AN H, MA X D, PAN Z Y, et al. Qualitative and quantitative comparison of image quality between single-shot echo-planar and interleaved multi-shot echo-planar diffusion-weighted imaging in female pelvis[J]. Eur Radiol, 2020, 30(4): 1876-1884. DOI: 10.1007/s00330-019-06491-3.
[21]
BAI Y, PEI Y G, LIU W V, et al. MRI: evaluating the application of FOCUS-MUSE diffusion-weighted imaging in the pancreas in comparison with FOCUS, MUSE, and single-shot DWIs[J]. J Magn Reson Imaging, 2023, 57(4): 1156-1171. DOI: 10.1002/jmri.28382.
[22]
LUSTIG M, DONOHO D, PAULY J M. Sparse MRI: the application of compressed sensing for rapid MR imaging[J]. Magn Reson Med, 2007, 58(6): 1182-1195. DOI: 10.1002/mrm.21391.
[23]
PRUESSMANN K P, WEIGER M, SCHEIDEGGER M B, et al. SENSE: sensitivity encoding for fast MRI[J]. Magn Reson Med, 1999, 42(5): 952-962.
[24]
SODICKSON D K, GRISWOLD M A, JAKOB P M, et al. Signal-to-noise ratio and signal-to-noise efficiency in SMASH imaging[J]. Magn Reson Med, 1999, 41(5): 1009-1022. DOI: 10.1002/(sici)1522-2594(199905)41:5<1009:aid-mrm21>3.0.co;2-4.
[25]
YASAKA K, AKAI H, KUNIMATSU A, et al. Liver fibrosis: deep convolutional neural network for staging by using gadoxetic acid-enhanced hepatobiliary phase MR images[J]. Radiology, 2018, 287(1): 146-155. DOI: 10.1148/radiol.2017171928.
[26]
HIGAKI T, NAKAMURA Y, TATSUGAMI F, et al. Improvement of image quality at CT and MRI using deep learning[J]. Jpn J Radiol, 2019, 37(1): 73-80. DOI: 10.1007/s11604-018-0796-2.
[27]
CHEN Q, FANG S, YUCHEN Y, et al. Clinical feasibility of deep learning reconstruction in liver diffusion-weighted imaging: improvement of image quality and impact on apparent diffusion coefficient value[J/OL]. Eur J Radiol, 2023, 168: 111149 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/37862927/. DOI: 10.1016/j.ejrad.2023.111149.
[28]
VAN DER VELDE N, HASSING H C, BAKKER B J, et al. Improvement of late gadolinium enhancement image quality using a deep learning-based reconstruction algorithm and its influence on myocardial scar quantification[J]. Eur Radiol, 2021, 31(6): 3846-3855. DOI: 10.1007/s00330-020-07461-w.
[29]
WANG X Z, MA J F, BHOSALE P, et al. Novel deep learning-based noise reduction technique for prostate magnetic resonance imaging[J]. Abdom Radiol, 2021, 46(7): 3378-3386. DOI: 10.1007/s00261-021-02964-6.
[30]
SHU Z, XINZENG W, WILLIAM S, et al. Deep Learning-based image reconstruction improves CEST MRI [J/OL]. Proc. Intl. Soc. Mag. Reson. Med. 29 (2021) [2024-01-09]. https://index.mirasmart.com/ISMRM2021/PDFfiles/1457.html.
[31]
KIM M, KIM H S, KIM H J, et al. Thin-slice pituitary MRI with deep learning-based reconstruction: diagnostic performance in a postoperative setting[J]. Radiology, 2021, 298(1): 114-122. DOI: 10.1148/radiol.2020200723.
[32]
ZOCHOWSKI K C, TAN E T, ARGENTIERI E C, et al. Improvement of peripheral nerve visualization using a deep learning-based MR reconstruction algorithm[J/OL]. Magn Reson Imaging, 2022, 85: 186-192 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/34715288/. DOI: 10.1016/j.mri.2021.10.038.
[33]
FEENEY C, LINGAM R K, LEE V, et al. Non-EPI-DWI for detection, disease monitoring, and clinical decision-making in thyroid eye disease[J]. AJNR Am J Neuroradiol, 2020, 41(8): 1466-1472. DOI: 10.3174/ajnr.A6664.
[34]
BARTALENA L, KAHALY G J, BALDESCHI L, et al. The 2021 European Group on Graves' orbitopathy (EUGOGO) clinical practice guidelines for the medical management of Graves' orbitopathy[J]. Eur J Endocrinol, 2021, 185(4): G43-G67. DOI: 10.1530/EJE-21-0479.
[35]
LINGAM R K, MUNDADA P, LEE V. Novel use of non-echo-planar diffusion weighted MRI in monitoring disease activity and treatment response in active Grave's orbitopathy: an initial observational cohort study[J]. Orbit, 2018, 37(5): 325-330. DOI: 10.1080/01676830.2017.1423343.
[36]
POLITI L S, GODI C, CAMMARATA G, et al. Magnetic resonance imaging with diffusion-weighted imaging in the evaluation of thyroid-associated orbitopathy: getting below the tip of the iceberg[J]. Eur Radiol, 2014, 24(5): 1118-1126. DOI: 10.1007/s00330-014-3103-3.
[37]
WU Y, LAO Z, ZHANG S, et al. Evaluation of extraocular muscles in patients with thyroid associated ophthalmopathy using apparent diffusion coefficient measured by magnetic resonance imaging before and after radiation therapy[J]. Acta Radiol, 2022, 63(9): 1180-1186. DOI: 10.1177/02841851211034042.
[38]
LIU X T, SU Y, JIANG M D, et al. Application of magnetic resonance imaging in the evaluation of disease activity in Graves' ophthalmopathy[J]. Endocr Pract, 2021, 27(3): 198-205. DOI: 10.1016/j.eprac.2020.09.008.
[39]
HU H, CHEN L, ZHOU J, et al. Multiparametric magnetic resonance imaging for differentiating active from inactive thyroid-associated ophthalmopathy: added value from magnetization transfer imaging[J/OL]. Eur J Radiol, 2022, 151: 110295 [2024-01-09]. https://pubmed.ncbi.nlm.nih.gov/35405579/. DOI: 10.1016/j.ejrad.2022.110295.

上一篇 基于深度学习重建和传统TSE序列在直肠癌磁共振检查的对比研究
下一篇 深度学习重建技术在乳腺MRI图像质量优化中的应用研究
  
诚聘英才 | 广告合作 | 免责声明 | 版权声明
联系电话:010-67113815
京ICP备19028836号-2