分享:
分享到微信朋友圈
X
临床研究
基于双参数MRI评估直肠磁敏感伪影对前列腺癌诊断影响的研究
汪征 胡磊 陆蓬 刘松 付成志 余松 余成新

Cite this article as WANG Z, HU L, LU P, et al. Influence of rectal susceptibility artifacts on diagnosis of prostate cancer based on biparametric magnetic resonance imaging[J]. Chin J Magn Reson Imaging, 2024, 15(5): 134-140, 147.本文引用格式汪征, 胡磊, 陆蓬, 等. 基于双参数MRI评估直肠磁敏感伪影对前列腺癌诊断影响的研究[J]. 磁共振成像, 2024, 15(5): 134-140, 147. DOI:10.12015/issn.1674-8034.2024.05.021.


[摘要] 目的 探讨MRI直肠磁敏感伪影对于前列腺癌主观评价和深度学习计算机辅助诊断(deep learning-based computer aided diagnosis, DL-CAD)的影响。材料与方法 回顾性分析685例行双参数MRI检查患者的影像资料,所有病例均通过穿刺活检或手术切片获得病理结果,由三组不同年资放射科阅片医师(阅片者1~6)分别依据前列腺成像报告和数据系统(prostate imaging report and data system, PI-RADS)v2.1版对前列腺MRI上的可疑病灶进行独立评审。另外两位阅片医师(阅片者甲、乙)对MRI上是否存在直肠伪影以及伪影程度进行评分。构建基于前列腺MRI的DL-CAD诊断模型评估直肠伪影对于深度学习诊断模型的影响。采用加权Kappa系数进行直肠伪影评估的一致性检验。采用χ2检验比较不同年资阅片医师PI-RADS评分、直肠伪影评分差异。采用多读者多病例受试者工作特征曲线(multi-reader multi-case receiver operating characteristic curve, MRMC-ROC)比较不同阅片者的诊断差异。采用受试者工作特征(receiver operating characteristic, ROC)曲线下面积(area under the curve, AUC)评估DL-CAD的诊断效能。采用DeLong检验比较AUC值差异。P<0.05为差异具有统计学意义。结果 本研究共纳入685例患者,其中前列腺癌组共199例,良性病变组共486例。在主观评价方面,低年资阅片者1的AUC无伪影为0.772, AUC有伪影为0.644,差异具有统计学意义(P=0.023)。低年资阅片者2的AUC无伪影为0.809,AUC有伪影为0.682,差异具有统计学意义(P=0.007)。而中高年资阅片者诊断效能差异均无统计学意义(P>0.05)。在不同程度的直肠伪影评估方面,所有医师的诊断效能AUC差异均无统计学意义(0.071≤P<0.973)。基于主观评分标准,两位医师对直肠伪影评分一致性为0.851。在直肠伪影亚组分析方面,外周带无伪影区域AUC高于有伪影区域(阅片者1:0.754 vs. 0.532;阅片者2:0.771 vs. 0.580),且差异存在统计学意义(P<0.05),剩余亚组比较差异均无统计学意义(P>0.05)。在深度学习方面,DL-CAD的AUC无伪影为0.794,AUC有伪影为0.538,差异具有统计学意义(P<0.05)。DL-CAD的AUC轻度伪影为0.546,AUC中度伪影为0.590,AUC重度伪影为0.481,轻、中、重度伪影对DL-CAD诊断效能差异均无统计学意义(P>0.05)。结论 直肠磁敏感伪影对于主观视觉评价及DL-CAD评估均有显著性负面影响,对于主观视觉评价和DL-CAD评估影响方式存在差异。
[Abstract] Objective To explore the impact of rectal susceptibility artifacts on the subjective evaluation and deep learning-based computer aided diagnosis (DL-CAD) in MRI-based prostate cancer diagnosis.Materials and Methods A retrospective analysis was conducted on 685 patients who underwent biparametric magnetic resonance imaging (bpMRI). All patients have confirmed pathological results via either biopsy or surgical resection. Three groups of radiologists (Reader 1-6) with varying years of experience independently reviewed suspicious lesions on prostate MRI according to the Prostate Imaging Reporting and Data System (PI-RADS) version 2.1. The other two readers scored whether there were rectal artifacts on MRI and the degree of artifacts. A DL-CAD model based on prostate MRI was constructed to evaluate the impact of rectal artifacts on the deep learning-based diagnostic model. The weighted Kappa coefficient was used for the consistency test of rectal artifact assessment. Differences in PI-RADS scores and rectal artifact scores among radiologists with different years of experience were compared using the chi-square test. The diagnostic differences among readers were compared using the multi-reader multi-case receiver operating characteristic curve (MRMC-ROC). The area under the curve (AUC) was used to evaluate the diagnostic performance of DL-CAD. The DeLong test was used to compare the differences in AUC values. A significance level of P<0.05 was considered statistically significant.Results This study included a total of 685 patients, comprising 199 cases of prostate cancer and 486 cases of benign lesions. In subjective evaluation, the AUC for junior Reader 1 was 0.772 without artifacts and 0.644 with artifacts, a statistically significant difference (P=0.023), while the AUC for junior Reader 2 was 0.809 without artifacts and 0.682 with artifacts, a statistically significant difference (P=0.007). The difference was not statistically significant (P>0.05) between the diagnostic performance of the middle and senior readers. Regarding the assessment of different degrees of rectal artifacts, there were no statistically significant differences in the diagnostic performance AUC among all readers (0.071≤P<0.973). Based on subjective scoring criteria, the other two readers rated the rectal artifact with a consistency of 0.851. In rectal artifact subgroup analysis, the AUC in the area without artifacts was higher than that in the area with artifacts in peripheral zone (Reader 1: 0.754 vs. 0.532; Reader 2: 0.771 vs. 0.580), and these differences were statistically significant (P<0.05). However, no statistically significant differences were observed in the remaining subgroups (P>0.05). In deep learning, the AUC without artifacts was 0.794 and the AUC with artifacts 0.538 for DL-CAD, and the difference was statistically significant(P<0.05). The AUC with mild artifacts were 0.546, the AUC with moderate artifacts were 0.590, and the AUC with severe artifacts were 0.481, and there was no significant difference in the diagnostic performance of DL-CAD (P>0.05).Conclusions Rectal susceptibility artifacts have significant negative effects on subjective visual assessment and DL-CAD assessment. There are differences in the impact of rectal artifacts on subjective visual assessment and DL-CAD assessment.
[关键词] 直肠磁敏感伪影;前列腺癌;磁共振成像;前列腺成像报告和数据系统;深度学习
[Keywords] rectal susceptibility artifacts;prostate cancer;magnetic resonance imaging;prostate imaging report and data system;deep learning

汪征 1, 2   胡磊 3   陆蓬 1, 2   刘松 1, 2   付成志 1, 2   余松 1, 2   余成新 1, 2*  

1 三峡大学第一临床医学院,宜昌 443000

2 宜昌市中心人民医院放射科,宜昌 443000

3 广东省人民医院放射科,广州 519041

通信作者:余成新,E-mail:ycyucx@163.com

作者贡献声明::余成新设计本研究的方案,对稿件重要的智力内容进行了修改;汪征起草和撰写稿件,获取、分析并解释本研究的数据;胡磊、陆蓬、刘松、付成志、余松获取、分析或解释本研究的数据,对稿件重要的内容进行了修改;胡磊获得了国家自然科学基金项目资助;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: 国家自然科学基金项目 82302130
收稿日期:2023-12-21
接受日期:2024-04-17
中图分类号:R445.2  R737.25 
文献标识码:A
DOI: 10.12015/issn.1674-8034.2024.05.021
本文引用格式汪征, 胡磊, 陆蓬, 等. 基于双参数MRI评估直肠磁敏感伪影对前列腺癌诊断影响的研究[J]. 磁共振成像, 2024, 15(5): 134-140, 147. DOI:10.12015/issn.1674-8034.2024.05.021.

0 引言

       前列腺癌(prostate cancer, PCa)是男性泌尿生殖系统中最常见的恶性肿瘤,其发病率在世界范围内位居男性恶性肿瘤第2位[1]。精准诊断PCa有助于减少低风险患者的过度诊疗,延长高风险患者生存时间,改善其生活质量[2]。前列腺MRI是一种有效的无创性方法,用于PCa的早期诊断、鉴别诊断、分期、治疗决策、预后评估及随访[3, 4]。前列腺MRI在诊断前列腺良恶性病变方面具有很高的诊断价值[5],然而受扫描设备、扫描序列、场强、患者自身条件等诸多因素的影响,前列腺MRI好发各种伪影,而直肠磁敏感伪影是前列腺最常见的伪影[6]。直肠伪影主要由直肠扩张以及直肠内容物所导致的局部磁场不均衡所引起。直肠磁敏感伪影可以导致前列腺组织局部信号丢失或信号强度下降,使正常组织和病变区域难以区分,从而引起医生对前列腺疾病的误诊或漏诊[7]。此外,由于直肠磁敏感伪影影响前列腺MRI的影像质量,患者可能需要进行多次复查,以获得更加准确的诊断结果,这不仅增加了患者的经济负担和心理压力,还可能导致诊断和治疗的延误,影响治疗决策。

       相关研究显示,与有经验的放射科医生相比,基于深度学习的计算机辅助诊断(deep learning-based computer aided diagnosis, DL-CAD)系统在MRI诊断PCa方面具有同等或更高的诊断性能和评价重复性,所需诊断时间和劳动力更少[8, 9, 10, 11, 12, 13]。直肠伪影对于DL-CAD在PCa检测中的影响也是一个重要的研究领域。作为深度学习的核心,深度神经网络(deep neural networks, DNNs)已经被广泛应用于PCa的诊断。近年来,许多基于DNNs的分类模型被提出,并在PCa分类任务上取得了显著效果。但这些基于DNNs的分类模型性能严重依赖于输入前列腺MRI图像质量[14, 15, 16]。值得注意的是,前列腺MRI图像通常包含各种伪影,这可能会误导基于DNNs的分类模型,从而导致错误的预测[17]。然而,现有的基于DNNs的分类方法通常不考虑直肠伪影的影响,直肠伪影对于DL-CAD是否存在影响目前尚不明确。因此,本研究旨在探讨MRI直肠磁敏感伪影对于PCa主观评价和DL-CAD的影响。

1 材料与方法

1.1 研究对象

       本研究遵守《赫尔辛基宣言》,并经宜昌市中心人民医院伦理委员会批准,免除受试者知情同意,批准文号:2023-001-01。回顾性分析2019年1月至2022年7月在宜昌市中心人民医院进行双参数MRI检查患者的临床和影像资料。纳入标准:(1)在双参数MRI图像上有明确的前列腺病变;(2)完善的临床信息和MRI报告,包括前列腺特异性抗原(prostate specific antigen, PSA)、前列腺影像报告和数据系统(Prostate Imaging Report and Data System, PI-RADS)评分等。排除标准:(1)检查过程中患者配合欠佳,成像质量不佳,序列不完整;(2)无法获得最终的组织病理诊断。起初共有711例病例,根据纳入和排除标准,其中26例被排除在外,最终研究纳入的参与者为685例,其中前列腺癌患者为199例,良性病变患者为486例。

       患者临床信息包括:年龄、PSA、恶性结节位置和直径、格里森(gleason, GS)评分、病理分期(表1)。GS评分是目前应用最广泛的组织学评价前列腺癌的评分系统。2014年国际泌尿病理协会专家共识会议对前列腺癌GS分级标准进行修订,更为详细和明确界定了前列腺癌GS各级别的形态学标准[18],分别如下:1级为单个分化良好的腺体紧密排列,形成界限清楚的结节;2级为单个分化良好的腺体较疏松排列,形成界限较清楚的结节;3级为分散、独立的分化良好的腺体;4级为分化不良、融合的或筛状的腺体;5级为缺乏腺性分化和/或坏死。

表1  纳入患者的基本临床信息及病理特征
Tab. 1  Basic clinical information and pathological characteristics of the included patients

1.2 MRI采集

       采用3.0 T MRI扫描仪(Ingenia CX,飞利浦,荷兰)和外部表面相控阵体部线圈进行采集,包括横断位T2加权成像(T2 weighted imaging, T2WI)和扩散加权成像(diffusion-weighted imaging, DWI),并通过DWI不同b值计算出相应的表观扩散系数(apparent diffusion coefficient, ADC),扫描参数详见表2

表2  磁共振序列参数
Tab. 2  Magnetic resonance sequence parameters

1.3 MRI前列腺诊断评估

       六名具有不同年限诊断经验的放射科医师(阅片者1和阅片者2为1~3年诊断经验的低年资诊断医师,阅片者3和阅片者4为5~10年诊断经验的中年资诊断医师,阅片者5和阅片者6为大于10年诊断经验的高年资诊断医师)在不知晓患者临床信息及病理结果的情况下分别独自对前列腺MRI上的可疑病灶进行独立阅片,并依据PI-RADS v2.1版本给出患者层面的PI-RADS评分。

1.4 MRI直肠伪影评估

       两名具有5年诊断经验的放射科医师(阅片者甲、阅片者乙)对MRI图像上是否存在直肠伪影以及伪影程度进行评分。当两人意见不统一时,由第三名具有15年MRI诊断经验高年资医师(阅片者丙)进行复核并决定最终评分结果。直肠伪影评分参考既往文献[19]:1分:无伪影,图像质量优良;2分:轻度伪影,轻度影响诊断;3分:中度伪影,中度影响诊断;4分:重度伪影,重度影响诊断(图1)。

图1  直肠伪影评分示意图。1A~1C:当伪影累及直肠旁外周带且受累<50%时,显示轻度伪影;1D~1F:当伪影累及直肠旁外周带且受累51%~100%时,显示中度伪影;1G~1I,当伪影累及到移行带时,显示重度伪影。
Fig. 1  Schematic diagram of rectal artifact scores. 1A-1C: Selected images used to show mild artifact, when < 50% of the peripheral zone (PZ) next to the rectum is involved; 1D-1F: Moderate artifact, when 51%-100% of the PZ is affected without involving the transition zone (TZ); 1G-1I: Severe artifact, when the artifact extends into the TZ.

1.5 前列腺MRI深度学习诊断模型

       为了评估MRI直肠伪影对于深度学习诊断模型的影响。在本研究中,我们以3D-nnUNet[20]为网络框架,构建了前列腺MRI深度学习诊断模型。该模型使用经过标注的公共数据集“Prostate Imaging: Cancer AI (Version 1.1)”进行训练(详情参见https://zenodo.org/record/6667655)。该数据集包括了来自三个中心(Radboud大学医学中心、University Medical Center Groningen和Ziekenhuis Groep Twente)1476位临床怀疑PCa(PSA水平升高、直肠指诊发现异常等)男性患者的1500份匿名前列腺MRI扫描数据。患者信息包括:基本临床信息(年龄、前列腺体积、PSA、前列腺特异性抗原密度)、基本采集信息(扫描仪制造商、扫描仪型号名称、扩散b值)和bpMRI扫描信息。该数据集数据标注由1名训经过训练的研究者或放射科住院医师在3名放射科专家的监督下完成。标注者参照病理诊断报告以及完整的前列腺切除标本对体素水平PCa病变进行标注,并记录患者水平PCa预后。

       前列腺MRI深度学习诊断模型包含前列腺腺体分割模型及前列腺病变检测模型。首先,采用前列腺腺体分割模型利用T2WI图像生成中央和周围腺体的分割掩模。随后,前列腺病变检测模型利用包括分割的中央和周围腺体掩模,以及T2WI、DWI和ADC图像作为输入。该模型生成一个代表不同前列腺区域存在病变的置信水平病变置信度图。随后,利用病变置信度图,生成相应的病变候选区域及其PCa检测概率。

1.6 统计学分析

       统计学分析采用R v4.10进行统计(R统计计算基金会,维也纳,奥地利,https://www.R-project.org/)。使用单样本Kolmogorov-Smirnov检验检测数据是否服从正态分布。服从正态分布的连续变量差异分析采用独立样本t检验,非正态分布的连续变量差异比较采用Mann-Whitney U检验。分类变量差异分析采用χ2检验。采用多读者多病例受试者工作特征曲线(multi-reader multi-Case receiver operating characteristic curve, MRMC-ROC)比较不同年资医师对直肠伪影的诊断差异。直肠伪影评分一致性检验采用Kappa系数,κ系数评定标准如下:0.01~0.20为一致性较差;0.21~0.40为一致性一般;0.41~0.60为一致性中等;0.61~0.80为一致性较好;0.81~0.99为一致性很好。深度学习模型诊断表现通过受试者工作特征(receiver operating characteristic, ROC)曲线下面积(area under the curve,AUC)进行评估。AUC值差异比较使用DeLong检验。P<0.05认为差异具有统计学意义。

2 结果

2.1 一般资料

       共纳入患者685例,其中前列腺癌组共199例,年龄(72±8)岁,总前列腺特异性抗原41.66(13.33,100.00)ng/mL;良性病变组共486例,年龄(69±8)岁,总前列腺特异性抗原6.50(3.29,12.16)ng/mL。患者基本临床信息详见表1。前列腺癌组、良性病变组分别在年龄、总PSA、游离PSA、游离PSA/总PSA方面差异均有统计学意义(P<0.05)。在PCa各伪影组癌灶长径方面,无伪影组癌灶长径约32.9(20.6,44.8)mm,有伪影组癌灶长径约30.5(19.0,41.6)mm,轻度、中度、重度伪影组癌灶长径分别约30.2(18.7,41.4)mm、33.0(23.5,42.2)mm、32.7(21.8,43.3)mm,各组在癌灶长径方面差异均无统计学意义(P>0.05)。

2.2 直肠伪影评分

       直肠伪影评分一致性检验采用Kappa系数,两名医师(阅片者甲、阅片者乙)对直肠伪影评分一致性κ系数为0.851,P<0.05,说明两位医师对直肠伪影评估具有很好的一致性,详情见表3

       前列腺癌组、良性病变组在不同程度伪影评分方面差异均具有统计学意义(P<0.05),其中,无直肠伪影病例占总病例63.8%(437/685)。轻度、中度、重度伪影占总病例分别为10.5%(72/685)、14.9%(102/685)、10.8%(74/685),详情见表4

表3  直肠伪影评分一致性检验
Tab. 3  Consistency test of rectal artifact assessment
表4  伪影评分表
Tab. 4  artifact score table

2.3 诊断准确性评估

       在主观评价方面,各年资医师(阅片者1~6)的PI-RADS评分详见表5。通过不同年资医师诊断效能ROC曲线分析发现,两位高年资医师获得最高的诊断效能,AUC值分别为0.821、0.838;两位中年资医师诊断效能次之,AUC值分别为0.785、0.802;两位低年资医师诊断效能相对最低,AUC值分别为0.740、0.776。六位不同年资医师的诊断效能AUC平均值为0.793,说明中高年资医师诊断效能均高于低年资医师,详情见表6图2

图2  不同年资阅片者对685名患者前列腺PI-RADS评分的受试者工作特征曲线。
Fig. 2  Receiver operating characteristic curve of PI-RADS scores for 685 patients by different seniority readers.
表5  不同年资医师的PI-RADS评分
Tab. 5  PI-RADS scores of radiologists with different seniority
表6  不同年资阅片医师的AUC和95% CI
Tab. 6  AUC and 95% CI of radiologist with different seniority

2.4 直肠伪影对主观评价影响的亚组分析

       通过MRMC-ROC曲线分析发现,低年资阅片者1的AUC无伪影为0.772,AUC有伪影为0.644,差异具有统计学意义(P=0.023);低年资阅片者2的AUC无伪影为0.809,AUC有伪影为0.682,差异具有统计学意义(P=0.007),而中高年资阅片者诊断效能差异均无统计学意义(P>0.05),说明直肠伪影对于低年资医师诊断效能影响较大,而对于中高年资医师影响不大,详情见表7图3

       根据对不同程度的伪影进行分析发现,所有医师的诊断效能AUC在不同程度的直肠伪影干扰下诊断差异均无统计学意义(0.071≤P<0.973)。根据结节位置和有无伪影对不同医师诊断AUC进行进一步亚组分析发现,伪影对于低年资医师在外周带PCa诊断准确性方面存在影响,无伪影区域AUC高于有伪影区域(阅片者1:0.754 vs. 0.532;阅片者2:0.771 vs. 0.580),且差异具有统计学意义(P<0.05)。剩余亚组比较差异均无统计学意义(P>0.05)。

图3  阅片者1~6对有无伪影比较的受试者工作特征(ROC)曲线图。蓝线代表伪影组,黄线代表无伪影组。
Fig. 3  Receiver operating characteristic curves of the Reader 1-6 in no artifacts group and artifacts group. Blue lines represent artifacts group, and yellow lines represent no artifacts group.
表7  不同年资阅片医师对有无伪影的诊断差异比较
Tab. 7  Comparison of diagnostic differences between radiologists with or without artifacts in different seniority

2.5 直肠伪影对DL-CAD影响的分析

       通过DL-CAD对有无直肠伪影及伪影严重程度的ROC曲线及AUC值分析可以发现,DL-CAD的AUC无伪影为0.794,AUC有伪影为0.538,有无直肠伪影对DL-CAD诊断效能差异具有统计学意义(P<0.05),说明DL-CAD在没有直肠伪影的前提下诊断效能更好,详情见表8图4。DL-CAD的AUC轻度伪影为0.546,AUC中度伪影为0.590,AUC重度伪影为0.481,轻、中、重度伪影对DL-CAD诊断效能差异均无统计学意义(P>0.05)详情见表9

图4  DL-CAD对有无伪影(4A)及不同程度伪影(4B)的受试者工作特征曲线图。DL-CAD:基于深度学习的计算机辅助诊断。
Fig. 4  Receiver operating characteristic curve of DL-CAD for the presence and absence of artifacts (4A) and the severity of artifacts (4B). DL-CAD: deep learning-based computer aided diagnosis.
表8  DL-CAD对有无伪影及不同程度伪影的AUC和95% CI
Tab. 8  AUC and 95% CI of DL-CAD for the presence and absence of artifacts and the severity of artifacts
表9  DL-CAD对有无伪影及不同程度伪影的诊断效能对比
Tab. 9  Comparison of the diagnostic performance of DL-CAD for the presence and absence of artifacts and the severity of artifacts

3 讨论

       本研究探讨了直肠磁敏感伪影在PCa磁共振诊断方面对于不同年资放射科医生的主观视觉评价及DL-CAD评估的影响,结果发现,直肠伪影主要影响具有相对较少诊断经验的低年资放射科医生的诊断效能,而经验丰富的中高年资放射科医生则可以有效鉴别伪影和病灶,免受干扰。同时,直肠伪影确实会干扰DL-CAD评估并降低其诊断效能,并且DL-CAD诊断效能并没有和伪影严重程度呈负相关关系,说明直肠伪影对于DL-CAD影响更为复杂。这为理解直肠伪影对于主观评价和DL-CAD诊断的干扰提供了有力证据,对后续制订特定抵御直肠伪影干扰的MRI扫描策略以及诊断策略具有指导意义。

3.1 直肠磁敏感伪影对主观评价的结果分析

       尽管在直肠伪影方面已经进行了相关研究[21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31],但是根据改良版的PI-RADS v2.1,有关患者肠道准备问题并没有达成共识。同时,在既往研究中,ANTUNES等[19]和CAGLIC等[32]的研究探讨了直肠伪影对于基于PI-RADS v2.0的PCa主观评价的影响。ANTUNES等的研究表明,由于直肠伪影的存在,36.8%的恶性病变被低估。同样,CAGLIC等发现,直肠扩张对T2WI和DWI图像的质量均有显著的负面影响。虽然在2019年,为了提升阅片者诊断的异质性,美国放射学会推出了改良版的PI-RADS v2.1,但相关研究[33, 34, 35]依然表明,减少直肠伪影可以显著提升诊断者的诊断准确度。然而,上述研究在分析直肠伪影对于阅片者诊断效能的影响时,均未将阅片者诊断经验纳入分析因素。本研究发现直肠伪影对于不同年资医师诊断影响存在差异,可能是由于高年资医生具有较丰富的临床经验,熟悉各种伪影的特点,能够区分真实的病变和伪影,进而更好地解释前列腺MRI图像,而相比之下,年轻医生相对缺乏此类经验,在一定程度上受到直肠伪影的影响,从而影响他们的诊断准确性。

3.2 直肠磁敏感伪影亚组结果分析

       通过对直肠伪影进行亚组分析发现,所有医师在不同程度直肠伪影中的诊断AUC差异并无统计学意义,而直肠伪影影响主观视觉评价的差异主要来源于外周带区域结节,这可能与PCa自身发病特点以及前列腺解剖特点有关。70%~75%的前列腺癌发生在外周带[36],而直肠位于前列腺后方,直肠伪影最常累计的位置也在外周带[32, 37]。在本研究中,为了提升直肠伪影评估的一致性,参考既往方法[19],根据直肠伪影的累及范围对直肠伪影进行评分。一方面,2~4级直肠伪影评分均累计外周带,另一方面,直肠伪影由直肠区域发出,呈放射状消散,主要累及范围在外周带,中央带受直肠伪影累计较轻,这可能是不同评分的直肠伪影组AUC差异无统计学意义的潜在原因。

3.3 直肠磁敏感伪影对DL-CAD的结果分析

       通过对DL-CAD诊断效能进行分析发现,直肠伪影确实会干扰DL-CAD评估并降低其诊断效能,结果显示,DL-CAD在没有直肠伪影的患者中诊断效能更好,但是,不同伪影严重程度患者之间DL-CAD诊断效能无明显差异。这说明,直肠伪影对于深度学习影响比较复杂。既往研究表明[38],相对于人眼评估,深度学习对于图像微小扰动的变化更加敏感,轻微角度的改变,局部噪声的添加,甚至一个像素点的改变都可能导致训练良好的模型出现误诊或者漏诊。本研究发现直肠伪影对于主观评价和深度学习模型的诊断干扰并不相同,即使不会导致医生出现误诊的轻微伪影也可以导致模型效能的显著下降,这提示我们在临床实践中,对于深度学习基于图像质量较差的患者给出的诊断结果解释应保持慎重的态度。

3.4 抵御直肠磁敏感伪影干扰的临床策略

       目前,临床上减轻前列腺直肠伪影主要从临床策略和成像手段两方面着手。例如,在MRI检查前进行肠道准备可以有效减少直肠伪影的发生概率[22];而更加先进的成像技术,例如,缩小视野成像[39]、平行成像、分段读出平面回波成像(echo planar imaging,EPI)序列等也被证明可以一定程度地减少宏观直肠伪影的严重程度。然而,目前尚没有一种策略可以完全消除直肠伪影。在我们参与的一项多中心前列腺MRI合作项目中,我们尝试在模型算法层面上抵御直肠伪影干扰[40],让模型主动适应直肠伪影噪声,为直肠伪影干扰PCa诊断提供了新的思路。

3.5 本研究的局限性

       本研究存在以下不足:(1)本研究为单中心的回顾性研究设计,可能存在选择偏差;(2)阅片者评分也是主观诊断,存在主观偏倚;(3)不同b值的DWI图像对直肠伪影的显示可能存在一些偏差;(4)未探讨直肠伪影对T2WI图像、DWI图像及ADC图的具体影响。上述不足之处将在后续多中心、前瞻性的临床研究中不断完善和改进。

4 结论

       综上所述,直肠磁敏感伪影对于主观视觉评价及DL-CAD评估均有显著性负面影响,对于主观视觉评价和DL-CAD评估影响方式存在差异。

[1]
SUNG H, FERLAY J, SIEGEL R L, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA Cancer J Clin, 2021, 71(3): 209-249. DOI: 10.3322/caac.21660.
[2]
HECTORS S J, CHERNY M, YADAV K K, et al. Radiomics features measured with multiparametric magnetic resonance imaging predict prostate cancer aggressiveness[J]. J Urol, 2019, 202(3): 498-505. DOI: 10.1097/JU.0000000000000272.
[3]
PECORARO M, TURKBEY B, PURYSKO A S, et al. Diagnostic accuracy and observer agreement of the MRI prostate imaging for recurrence reporting assessment score[J]. Radiology, 2022, 304(2): 342-350. DOI: 10.1148/radiol.212252.
[4]
ABREU-GOMEZ J, DIAS A B, GHAI S. PI-RR: the prostate imaging for recurrence reporting system for MRI assessment of local prostate cancer recurrence after radiation therapy or radical prostatectomy-a review[J]. AJR Am J Roentgenol, 2023, 220(6): 852-861. DOI: 10.2214/AJR.22.28665.
[5]
BREMBILLA G, GIGANTI F, SIDHU H, et al. Diagnostic accuracy of abbreviated Bi-parametric MRI (a-bpMRI) for prostate cancer detection and screening: a multi-reader study[J/OL]. Diagnostics, 2022, 12(2): 231 [2023-10-15]. https://pubmed.ncbi.nlm.nih.gov/35204322/. DOI: 10.3390/diagnostics12020231.
[6]
KYRIAZI S, BLACKLEDGE M, COLLINS D J, et al. Optimising diffusion-weighted imaging in the abdomen and pelvis: comparison of image quality between monopolar and bipolar single-shot spin-echo echo-planar sequences[J]. Eur Radiol, 2010, 20(10): 2422-2431. DOI: 10.1007/s00330-010-1826-3.
[7]
ROSENKRANTZ A B, TANEJA S S. Radiologist, be aware: ten pitfalls that confound the interpretation of multiparametric prostate MRI[J]. AJR Am J Roentgenol, 2014, 202(1): 109-120. DOI: 10.2214/AJR.13.10699.
[8]
HIREMATH A, SHIRADKAR R, MERISAARI H, et al. Test-retest repeatability of a deep learning architecture in detecting and segmenting clinically significant prostate cancer on apparent diffusion coefficient (ADC) maps[J]. Eur Radiol, 2021, 31(1): 379-391. DOI: 10.1007/s00330-020-07065-4.
[9]
WINKEL D J, TONG A, LOU B, et al. A novel deep learning based computer-aided diagnosis system improves the accuracy and efficiency of radiologists in reading biparametric magnetic resonance images of the prostate: results of a multireader, multicase study[J]. Invest Radiol, 2021, 56(10): 605-613. DOI: 10.1097/RLI.0000000000000780.
[10]
HU L, WEI L M, WANG S H, et al. Better lesion conspicuity translates into improved prostate cancer detection: comparison of non-parallel-transmission-zoomed-DWI with conventional-DWI[J]. Abdom Radiol, 2021, 46(12): 5659-5668. DOI: 10.1007/s00261-021-03268-5.
[11]
BULTEN W, PINCKAERS H, VAN BOVEN H, et al. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study[J]. Lancet Oncol, 2020, 21(2): 233-241. DOI: 10.1016/S1470-2045(19)30739-9.
[12]
王绎忱, 张馨心, 胡满仓, 等. 深度学习重建技术在优化前列腺磁共振T2加权成像扫描时间和图像质量中的应用价值[J]. 磁共振成像, 2023, 14(5): 48-52, 59. DOI: 10.12015/issn.1674-8034.2023.05.010.
WANG Y C, ZHANG X X, HU M C, et al. Value of deep learning reconstruction in optimizing prostate MR T2-weighted imaging scanning time and imaging quality[J]. Chin J Magn Reson Imag, 2023, 14(5): 48-52, 59. DOI: 10.12015/issn.1674-8034.2023.05.010.
[13]
胡尘翰, 乔晓梦, 胡冀苏, 等. 基于双参数MRI的深度学习-临床混合模型对临床显著性前列腺癌诊断价值的研究[J]. 磁共振成像, 2024, 15(2): 90-96. DOI: 10.12015/issn.1674-8034.2024.02.013.
HU C H, QIAO X M, HU J S, et al. The utility of deep learning-clinical combined model based on bi-parametric MRI for diagnosis of clinically significant prostate cancer[J]. Chin J Magn Reson Imag, 2024, 15(2): 90-96. DOI: 10.12015/issn.1674-8034.2024.02.013.
[14]
GIGANTI F, ROSENKRANTZ A B, VILLEIRS G, et al. The evolution of MRI of the prostate: the past, the present, and the future[J]. AJR Am J Roentgenol, 2019, 213(2): 384-396. DOI: 10.2214/AJR.18.20796.
[15]
ZHU Q K, DU B, TURKBEY B, et al. Exploiting interslice correlation for MRI prostate image segmentation, from recursive neural networks aspect[J/OL]. Complexity, 2018, 2018: 4185279 [2023-10-15]. https://downloads.hindawi.com/journals/complexity/2018/4185279.pdf. DOI: 10.1155/2018/4185279.
[16]
可赞, 李亮, 宋鑫洋, 等. 基于深度学习重建技术改善前列腺T2WI图像质量的研究[J]. 磁共振成像, 2023, 14(5): 41-47. DOI: 10.12015/issn.1674-8034.2023.05.009.
KE Z, LI L, SONG X Y, et al. Novel deep learning-based T2-weighted imaging of the prostate provides superior image quality[J]. Chin J Magn Reson Imag, 2023, 14(5): 41-47. DOI: 10.12015/issn.1674-8034.2023.05.009.
[17]
CUOCOLO R, CIPULLO M B, STANZIONE A, et al. Machine learning applications in prostate cancer magnetic resonance imaging[J/OL]. Eur Radiol Exp, 2019, 3(1): 35 [2023-10-15]. https://pubmed.ncbi.nlm.nih.gov/31392526/. DOI: 10.1186/s41747-019-0109-2.
[18]
EPSTEIN J I, EGEVAD L, AMIN M B, et al. The 2014 international society of urological pathology (ISUP) consensus conference on gleason grading of prostatic carcinoma: definition of grading patterns and proposal for a new grading system[J]. Am J Surg Pathol, 2016, 40(2): 244-252. DOI: 10.1097/PAS.0000000000000530.
[19]
ANTUNES N, VAS D, SEBASTIA C, et al. Susceptibility artifacts and PIRADS 3 lesions in prostatic MRI: how often is the dynamic contrast-enhance sequence necessary?[J]. Abdom Radiol, 2021, 46(7): 3401-3409. DOI: 10.1007/s00261-021-03011-0.
[20]
ISENSEE F, JAEGER P F, KOHL S A A, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation[J]. Nat Methods, 2021, 18(2): 203-211. DOI: 10.1038/s41592-020-01008-z.
[21]
VAN GRIETHUYSEN J J M, BUS E M, HAUPTMANN M, et al. Gas-induced susceptibility artefacts on diffusion-weighted MRI of the rectum at 1.5 T-Effect of applying a micro-enema to improve image quality[J]. Eur J Radiol, 2018, 99: 131-137. DOI: 10.1016/j.ejrad.2017.12.020.
[22]
PLODECK V, RADOSA C G, HÜBNER H M, et al. Rectal gas-induced susceptibility artefacts on prostate diffusion-weighted MRI with epi read-out at 3.0T: does a preparatory micro-enema improve image quality?[J]. Abdom Radiol, 2020, 45(12): 4244-4251. DOI: 10.1007/s00261-020-02600-9.
[23]
COSKUN M, MEHRALIVAND S, SHIH J H, et al. Impact of bowel preparation with Fleet's™ enema on prostate MRI quality[J]. Abdom Radiol, 2020, 45(12): 4252-4259. DOI: 10.1007/s00261-020-02487-6.
[24]
LIM C, QUON J, MCINNES M, et al. Does a cleansing enema improve image quality of 3T surface coil multiparametric prostate MRI?[J]. J Magn Reson Imaging, 2015, 42(3): 689-697. DOI: 10.1002/jmri.24833.
[25]
SLOUGH R A, CAGLIC I, HANSEN N L, et al. Effect of hyoscine butylbromide on prostate multiparametric MRI anatomical and functional image quality[J/OL]. Clin Radiol, 2018, 73(2): 216.e9-216.e14 [2023-10-15]. https://pubmed.ncbi.nlm.nih.gov/28803622/. DOI: 10.1016/j.crad.2017.07.013.
[26]
ULLRICH T, QUENTIN M, SCHMALTZ A K, et al. Hyoscine butylbromide significantly decreases motion artefacts and allows better delineation of anatomic structures in mp-MRI of the prostate[J]. Eur Radiol, 2018, 28(1): 17-23. DOI: 10.1007/s00330-017-4940-7.
[27]
WAGNER M, RIEF M, BUSCH J, et al. Effect of butylscopolamine on image quality in MRI of the prostate[J]. Clin Radiol, 2010, 65(6): 460-464. DOI: 10.1016/j.crad.2010.02.007.
[28]
ROETHKE M C, KURU T H, RADBRUCH A, et al. Prostate magnetic resonance imaging at 3 Tesla: is administration of hyoscine-N-butyl-bromide mandatory?[J]. World J Radiol, 2013, 5(7): 259-263. DOI: 10.4329/wjr.v5.i7.259.
[29]
PURYSKO A S, MIELKE N, BULLEN J, et al. Influence of Enema and dietary restrictions on prostate MR image quality: a multireader study[J]. Acad Radiol, 2022, 29(1): 4-14. DOI: 10.1016/j.acra.2020.10.019.
[30]
SCHMIDT C, HÖTKER A M, MUEHLEMATTER U J, et al. Value of bowel preparation techniques for prostate MRI: a preliminary study[J]. Abdom Radiol, 2021, 46(8): 4002-4013. DOI: 10.1007/s00261-021-03046-3.
[31]
BRENNAN D L, LAZARAKIS S, LEE A, et al. Do antispasmodics or rectal enemas improve image quality on multiparametric prostate MRI? An 'Evidence-Based Practice' review of the literature[J]. Abdom Radiol, 2021, 46(6): 2770-2778. DOI: 10.1007/s00261-020-02916-6.
[32]
CAGLIC I, HANSEN N L, SLOUGH R A, et al. Evaluating the effect of rectal distension on prostate multiparametric MRI image quality[J]. Eur J Radiol, 2017, 90: 174-180. DOI: 10.1016/j.ejrad.2017.02.029.
[33]
ARNOLDNER M A, POLANEC S H, LAZAR M, et al. Rectal preparation significantly improves prostate imaging quality: assessment of the PI-QUAL score with visual grading characteristics[J/OL]. Eur J Radiol, 2022, 147: 110145 [2023-10-16]. https://pubmed.ncbi.nlm.nih.gov/35007983. DOI: 10.1016/j.ejrad.2021.110145.
[34]
HUANG Y H, ÖZÜTEMIZ C, RUBIN N, et al. Impact of 18-french rectal tube placement on image quality of multiparametric prostate MRI[J]. AJR Am J Roentgenol, 2021, 217(4): 919-920. DOI: 10.2214/AJR.21.25732.
[35]
REISCHAUER C, CANCELLI T, MALEKZADEH S, et al. How to improve image quality of DWI of the prostate-enema or catheter preparation?[J]. Eur Radiol, 2021, 31(9): 6708-6716. DOI: 10.1007/s00330-021-07842-9.
[36]
TURKBEY B, ROSENKRANTZ A B, HAIDER M A, et al. Prostate imaging reporting and data system version 2.1: 2019 update of prostate imaging reporting and data system version 2[J/OL]. Eur Urol, 2019, 76(3): 340-351. DOI: 10.1016/j.eururo.2019.02.033.
[37]
DIETRICH O, REISER M F, SCHOENBERG S O. Artifacts in 3-T MRI: physical background and reduction strategies[J]. Eur J Radiol, 2008, 65(1): 29-35. DOI: 10.1016/j.ejrad.2007.11.005.
[38]
HIRANO H, MINAGI A, TAKEMOTO K. Universal adversarial attacks on deep neural networks for medical image classification[J/OL]. BMC Med Imaging, 2021, 21(1): 9 [2023-10-16]. https://pubmed.ncbi.nlm.nih.gov/33413181/. DOI: 10.1186/s12880-020-00530-y.
[39]
HU L, FU C X, SONG X Y, et al. Automated deep-learning system in the assessment of MRI-visible prostate cancer: comparison of advanced zoomed diffusion-weighted imaging and conventional technique[J/OL]. Cancer Imaging, 2023, 23(1): 6 [2023-10-16]. https://pubmed.ncbi.nlm.nih.gov/36647150/. DOI: 10.1186/s40644-023-00527-0.
[40]
HU L, GUO X Y, ZHOU D W, et al. Development and validation of a deep learning model to reduce the interference of rectal artifacts in MRI-based prostate cancer diagnosis[J/OL]. Radiol Artif Intell, 2024, 6(2): e230362 [2024-03-06]. https://pubmed.ncbi.nlm.nih.gov/38446042. DOI: 10.1148/ryai.230362.

上一篇 多参数心脏磁共振在低龄儿童急性心肌炎中的诊断和短期随访价值
下一篇 DCE-MRI联合IVIM-DWI预测早期宫颈癌盆腔淋巴结转移的价值
  
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