分享:
分享到微信朋友圈
X
临床研究
基于不同扩散模型参数图的影像组学分析磁共振早期诊断临床显著性前列腺癌的价值
杜兵 戚轩 杨宏楷 齐东 何永胜

Cite this article as: DU B, QI X, YANG H K, et al. To analyze the value of radiomics based on different diffusion model parameter maps in the early diagnosis of clinically significant prostate cancer by magnetic resonance imaging[J]. Chin J Magn Reson Imaging, 2024, 15(2): 83-89.本文引用格式杜兵, 戚轩, 杨宏楷, 等. 基于不同扩散模型参数图的影像组学分析磁共振早期诊断临床显著性前列腺癌的价值[J]. 磁共振成像, 2024, 15(2): 83-89. DOI:10.12015/issn.1674-8034.2024.02.012.


[摘要] 目的 旨在探讨基于磁共振单指数和扩散峰度模型功能参数图的影像组学分析早期诊断临床显著性前列腺癌(clinically significant prostate cancer, csPCa)的价值。材料与方法 回顾性地分析2022年4月至2023年7月就诊于马鞍山市人民医院的前列腺疾患病例238例,经超声下引导穿刺或手术病理证实,其中csPCa 96例、非临床显著性前列腺癌(non-clinically significant prostate cancer, ncsPca)142例,年龄56~84(62.34±7.62)岁。将238例患者按照7∶3的比例进行随机分组为训练集和测试集。所有患者均行MRI多参数扫描,通过后处理生成表观扩散系数(apparent diffusion coefficient, ADC)伪彩图,并得到扩散峰度模型中的平均扩散峰度(mean kurtosis, MK)和平均扩散系数(mean diffusivty, MD)伪彩图,图像预处理后,提取各个功能参数图的共计1 056个组学特征,对ADC、MD和MK模型的数据采用最大相关最小冗余(maximum relevance minimum redundancy, MRMR)算法和最小绝对收缩和选择算子(least absolute shrinkage and selection operator, LASSO)消除冗余、进行特征降维,保留与标签高相关的特征,应用10倍交叉验证后得到特征子集。最终ADC模型筛选出5个组学特征,MD模型筛选出6个组学特征,MK模型筛选出6个组学特征,建立逻辑回归模型,分别计算临床模型、影像学模型和临床-影像学联合模型的阈值、准确度、敏感度、特异度,绘制受试者工作特征(receiver operating characteristic, ROC)曲线并计算曲线下面积(area under the curve, AUC)及95%置信区间(confidence interval, CI),利用DeLong检验对各个模型进行两两组合,比较两组间的AUC值是否具有统计学意义,进一步使用决策曲线分析(decision curve analysis, DCA)评估模型的净获益。结果 临床模型在训练集中的AUC、特异度和敏感度分别为0.840(95% CI:0.778~0.901)、78.7%、76.8%,在测试集中分别为0.675(95% CI:0.539~0.812)、79.0%、59.2%。影像学模型中ADC模型在训练集中的AUC、特异度和敏感度分别为0.927(95% CI:0.890~0.964)、81.9%、86.9%,在测试集中分别为0.909(95% CI:0.835-0.983)、90.6%、84.1%;MD模型在训练集中的AUC、特异度和敏感度分别为0.934(95% CI:0.899~0.969)、85.1%、84.0%,在测试集中分别为0.960(95% CI:0.910~1.000)、93.0%、85.1%;MK模型在训练集中的AUC、特异度和敏感度分别为0.935(95% CI:0.900~0.971)、90.4%、84.0%,在测试集中分别为0.856(95% CI:0.770~0.941)、81.3%、66.6%。临床-影像学联合模型在训练集中的AUC、特异度和敏感度分别为0.946(95% CI:0.912~0.980)、88.2%、89.8%,在测试集中分别为0.963(95% CI:0.925~1.000)、93.0%、85.1%。DeLong检验结果显示影像学模型和临床-影像学联合模型两两比较差异均无统计学意义(P>0.05),临床模型与其他两个模型的AUC值差异具有统计学意义(Z=2.836,P=0.004)。DCA显示各个模型的阈值概率在0.1~1.0范围内,对临床有净获益,不同模型对csPCa的诊断均具有较高的诊断效能,以临床-影像学联合模型的诊断效能最高。结论 MRI单指数、扩散峰度模型功能参数图的影像组学分析技术是csPCa的有效检出方法,构建的临床-影像学联合模型对csPCa具有较高的诊断价值,能够为临床早期诊断和治疗提供相关技术支持。
[Abstract] Objective To explore the predictive value of radiomics analysis basedon magnetic resonance single-index and diffusion kurtosis model functional parameter maps for clinically significant prostate cancer (csPCa).Materials and Methods A retrospective analysis was conducted on 238 prostate patients who visited Ma'anshan People's Hospital from April 2022 to July 2023. They were confirmed by ultrasound-guided puncture or surgical pathology, including 96 csPCa patients and 142 non-csPCa patients. The age of the patients 56-84 (62.34±7.62) years old. The Clinical data within and between the groups were compared. All patients underwent magnetic resonance multi-parameter scanning, after post-processing, the apparent diffusion coefficient (ADC) pseudo-color plots were generated, and the mean kurtosis (MK) and mean diffusivty (MD) pseudo-color plots in the diffusion kurtosis model were obtained. After image preprocessing, the image features of eachfunctional parameter map are extracted. There are a total of 1 056 radiomics features. The maximum correlation minimum redundancy (MRMR) algorithm and least absolute shrinkage and selection operator (LASSO) are used to eliminateredundancy, perform feature dimensionality reduction, and retain high-quality labels for the data of ADC, MD, and MK models. For relevant features, 10-foldcross-validation was applied to obtain a feature subset, and 238 patients were randomly divided into groups in a ratio of 7∶3. Finally, the ADC model screened out 5 omics features, and the MD model screened out 6 omics features. The MK model screened out 6 omics features, established alogistic regression model, calculated the threshold, accuracy, sensitivity, and specificity of the clinical models, radiology, and clinical-radiology models, and drew the receiver operating characteristic (ROC) curve. Calculate the area under the curve (AUC) and 95% confidence interval (CI), use the DeLong test to combine each model in pairs, compare whether the AUC values between the two groups are statistically significant, and further use decision curve analysis (DCA) to evaluate model performance.Results The AUC, specificity and sensitivity of the clinical model in the training set were 0.840 (95% CI: 0.778-0.901), 78.7% and 76.8%, and in the test set were 0.675 (95% CI: 0.539-0.812), 79.0% and 59.2%, respectively. The AUC, specificity and sensitivity of the ADC model in the training set were 0.927 (95% CI: 0.890-0.964), 81.9%, 86.9%, and in the test set were 0.909 (95% CI: 0.835-0.983), 90.6%, 84.1%, respectively; the AUC, specificity and sensitivity of the MD model in the trainingset were 0.934 (95% CI: 0.899-0.969), 85.1%, 84.0%, and in the test set were 0.960 (95% CI: 0.910-1.000), 93.0%, 85.1%, respectively; the AUC, specificity and sensitivity of the MK model in the training set were 0.935 (95% CI: 0.900-0.971), 90.4%, 84.0%, and in the test set were 0.856 (95% CI: 0.770-0.941), 81.3%, 66.6%, respectively. The AUC, specificity and sensitivity of the clinical-radiology model in the training set were 0.946 (95% CI: 0.912-0.980), 88.2% and 89.8%, and in the test set were 0.963 (95% CI: 0.925-1.000), 93.0% and 85.1%, respectively. DeLong test results showed that there was no significant difference between the radiology model and the clinical-radiology combined model (P>0.05). There was a significant difference in AUC value between the clinical model and the other two models (Z=2.836, P=0.004), and there was no significant difference between the other two groups of models (P>0.05). The decision curve shows that the threshold probability of each model is in the range of 0.1-1.0, which has a net benefit for clinical practice. Different models have a positive effect on the diagnosis of csPCa. The clinical-radiology model having the highest diagnostic performance.Conclusions The radiomics analysis technology of MRI mono-exponential and diffusion kurtosis model functional parameter map is an effective method for the detection of csPCa. The clinical-radiology combined model has high diagnostic value for csPCa, which can provide relevant technical support for early clinical diagnosis and treatment.
[关键词] 前列腺癌;磁共振成像;影像组学;扩散加权成像;诊断效能
[Keywords] prostate cancer;magnetic resonance imaging;radiomics;diffusion weighted imaging;diagnostic efficacy

杜兵 1, 2   戚轩 1   杨宏楷 1   齐东 1, 3   何永胜 1*  

1 马鞍山市人民医院影像科,马鞍山 243000

2 皖南医学院,芜湖 241002

3 安徽医科大学,合肥,230032

通信作者:何永胜,E-mail:heyongsheng881@163.com

作者贡献声明::何永胜设计本研究的方案,对稿件重要内容进行了修改;杜兵起草、撰写稿件,获取、解释本研究的数据,获得了安徽省重点研究与开发计划基金项目资助;杨宏楷、齐东扫描、整理患者资料,获取、分析或解释本研究的数据,对稿件重要内容进行了修改;戚轩统计、分析数据,对稿件重要内容进行了修改;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: 安徽省重点研究与开发计划项目 2022e07020065
收稿日期:2023-10-07
接受日期:2024-02-02
中图分类号:R445.2  R737.25 
文献标识码:A
DOI: 10.12015/issn.1674-8034.2024.02.012
本文引用格式杜兵, 戚轩, 杨宏楷, 等. 基于不同扩散模型参数图的影像组学分析磁共振早期诊断临床显著性前列腺癌的价值[J]. 磁共振成像, 2024, 15(2): 83-89. DOI:10.12015/issn.1674-8034.2024.02.012.

0 引言

       在全球范围内,男性第二大常见恶性肿瘤仍是前列腺癌(prostate cancer, PCa),在老年男性疾病中,PCa也是最常见的恶性肿瘤之一[1]。近年来,随着经济社会的发展,居民生活水平的提高和饮食习惯的不断改变,我国PCa的发病率及致死率也呈显著地快速上升趋势[2]。PCa的治疗和预后因其组织病理学的亚型不同而有所差异,容易出现转移的肿瘤一般预后都较差,因此临床上能够对临床显著性前列腺癌(clinically significant prostate cancer, csPCa)做出早期诊断具有重要意义。csPCa主要是根据PCa Gleason评分(Gleason Score)和浸润范围来区分的,csPCa是指Gleason评分≥7分及/或体积≥0.5 cm及/或前列腺包膜外侵犯,csPCa腺体结构多被破坏,细胞数量增多,细胞外间隙变小,恶性程度明显增高,侵袭性强,需要积极治疗。非临床显著性前列腺癌(non-clinically significant prostate cancer, ncsPca)是指Gleasno评分<7分的肿瘤(包含良性前列腺增生),ncsPca一般分化良好,仍有正常腺体结构和细胞间隙,恶性度低,侵袭性弱,疾病进展较缓慢,以主动监测和定期随访为主,过度诊疗不仅会给患者增加负担,还会降低患者的生存质量[3, 4, 5]。MRI作为前列腺首选的无创检查技术[6],具有多方向、多平面、多序列成像及没有电离辐射的优点,目前临床应用非常普遍。特别是扩散加权成像(diffusion-weighted imaging, DWI)在肿瘤病变的诊断中发挥着重要作用,近年来多项研究表明DWI技术在肿瘤的检测中也具有较高的特异性和敏感性[7, 8, 9]。DWI序列和表观扩散系数(apparent diffusion coefficient, ADC)值被认为是前列腺病灶检测和判断PCa侵袭性的基础,也有研究结果表明通过定量扩散MRI可以减少因假阳性前列腺病变而发生的不必要的活检[10, 11, 12]。而不同病理分化类型的肿瘤异质性不同,肿瘤异质性的评估是制订治疗计划和预测预后的关键步骤,影像组学能够提供一种量化图像信号异质性的方法,分析人眼无法感知的像素值空间分布的规律性和粗糙度[13, 14]。影像组学为PCa患者的诊断和治疗提供了个性化、精确的方法,但以往研究多是通过MRI定量参数值对前列腺病变进行分析[15, 16]。本研究引入了高阶扩散模型的扩散峰度模型,通过高b值、多方向、多维度分析,而以往研究中扩散峰度模型多是采用0、1 000、2 000 s/mm² 3个b值来进行模型的拟合,本研究中扩散峰度模型则加入了3个高b值参数(1 500、2 000和2 500 s/mm²),使模型更加稳定,能够更真实地反映组织的扩散特性。本研究主要分析MRI单指数模型和扩散峰度模型中不同功能参数图的影像组学特征,建立逻辑回归模型,分析各功能参数模型的差异,从而建立对csPCa和ncsPCa早期诊断的有效预测模型。

1 材料与方法

1.1 一般资料

       回顾性地分析2022年4月至2023年7月就诊于马鞍山市人民医院的前列腺疾患病例238例,经超声下引导穿刺或手术病理证实,其中csPCa 96例,非csPCa 142例,年龄56~84(62.34±7.62)岁。纳入标准:(1)患者在进行MRI检查之前,未进行内分泌治疗或前列腺手术;(2)患者在检查完成后,在两个月之内有超声引导下穿刺或手术病理结果证实;(3)患者前列腺病灶图像清晰,没有失真变形,病灶便于测量和勾画。排除标准:(1)临床资料不完整;(2)病灶体积过小(直径<5 mm),不利于病灶勾画;(3)病理结果显示为前列腺间叶肿瘤、鳞状上皮肿瘤或其他瘤样病变。本研究遵守《赫尔辛基宣言》,经马鞍山市人民医院伦理委员会批准,免除受试者知情同意,批准文号:2121006008。

1.2 方法

1.2.1 MRI检查

       采用Siemens MANGETO Prisma 3.0 T超导型MRI仪,受检者呈仰卧位,采用18通道体部线圈扫描。扫描包含横断面T2WI和多b值DWI序列。T2WI序列参数:FOV 200 mm×200 mm,TR 6 980 ms,TE 104 ms,层厚3 mm,层数23层。DWI序列参数:FOV 220 mm×220 mm,TR 4 000 ms,TE 65 ms,层厚3 mm,体素0.9 mm×0.9 mm×3.0 mm,层厚3 mm,层数23层,带宽1 750 Hz,b值0、50、100、200、500、1 000、1 500、2 000和2 500 s/mm²,在三个不同方向上采集,采用在线动态场校正技术以消除涡流导致的图像变形。

1.2.2 图像分析

       通过后处理生成标准扩散系数ADC伪彩图,并得到扩散峰度模型中的平均扩散峰度(mean kurtosis,MK)和平均扩散系数(mean diffusivty, MD)伪彩图。由两位分别具有5年和8年工作经验的主治医师和副主任医师独立分析图像资料,两人对病理结果均不知情,使用3D-slicer(http://www.3D-slicer.org)软件进行PCa病灶的分割,结合T2WI图像,选取病灶显示最大的层面,在T2WI上进行三维感兴趣区(volume of interest, VOI)的上下层面逐层手动勾画,在重复进行感兴趣区勾画时,尽可能保证勾画形态与大小一致,将病灶实质部位勾画进去,同时避开出血、囊变和坏死的位置。为保证观察者内和观察者间所获取特征的一致性,随机抽取60例患者的图像,在两周以后由另外两位具有5年和8年工作经验的主治医师和副主任医师重新逐层勾画患者病灶,计算观察者内和观察者间的组内相关系数(intra-class correlation coefficient, ICC),ICC>0.75代表一致性良好[17]。为保证基于T2WI图像获取的VOI与其ADC、MD和MK图像相匹配,采用SPM软件中的刚体变换,将图像与T2WI图像对齐。对齐完成后,将VOI和对齐图像导入3D-slicer软件中,肉眼评估VOI的位置是否准确,若存在不准确的情况,手动进行修改(图1)。

图1  男,72岁,病理诊断为前列腺癌,Gleason评分为3+5分。1A:中央腺体T2WI低信号结节;1B:红色ROI为病灶在T2WI的位置;1C~1E:经过配准后感兴趣区(ROI)在各个参数图的位置,1C为表观扩散系数(ADC)图,1D为平均扩散系数(MD)图,1E为平均扩散峰度(MK)图;1F:病灶逐层手动勾画形成的三维感兴趣区。
Fig. 1  Male, 72 years old, pathologically diagnosed with prostate cancer with Gleason score of 3+5. 1A: T2WI hypointense nodules in the central gland; 1B: Red region of interest (ROI) indicates the location of the lesion on T2WI; 1C-1E: The position of ROI in each parameter map after registration, 1C is the apparent diffusion coefficient (ADC) map, 1D is the mean diffusion coefficient (MD) map, 1E is the mean diffusion kurtosis (MK) map; 1F: 3D ROI formed by manual delineation of the lesion layer by layer.

1.2.3 临床数据收集

       收集纳入患者入院1周内的总前列腺特异性抗原(total prostate specific antigen, TPSA)、游离前列腺特异性抗(free prostate specific antigen, FPSA)、游离/总前列腺特异性抗原(free/total prostate specific antigen, f/tPSA)和前列腺体积(prostate volume, PV),其中PV=前列腺最大左右径×最大前后径×最大上下径×0.52,按照前列腺影像报告和数据系统(Prostate Imaging Reporting and Data System, PI-RADS)指南推荐测量方式在横断位T2WI上测量最大左右径及最大前后径,在矢状位T2WI上测量最大上下径。

1.2.4 影像组学特征提取

       使用开源Python软件包“pyradiomics”(version3.0.1)对上述图像进行图像归一化和影像组学特征计算处理。在特征提取之前,首先对图像进行预处理,包括图像重采样为1 mm×1 mm×1 mm的体素大小。采用binwidths=10进行图像灰度离散化。图像预处理后,提取1 056个影像组学特征,包括基于特征类:形状参数、一阶纹理参数、灰度游程长度矩阵(gray-level run length matrix, GLRLM)、灰度共生矩阵(gray-1evel co-occurrence matrix, GLCM)、灰度区域大小矩阵(gray-level size zone matrix, GLSZM)和灰度依赖矩阵(gray-level dependence matrix, GLDM)等;基于过滤器类:小波分析(wavelet analysis)、对数特征、拉普拉斯算子等。

1.2.5 特征选择及模型构建

       所有病例按照7∶3的比例被随机分组为训练集和测试集。训练集用于模型训练,测试集用于对模型独立验证。对ADC、MD和MK数据的训练集均采用最大相关最小冗余(maximum relevance minimum redundancy, MRMR)算法和最小绝对收缩和选择算子(least absolute shrinkage and selection operator, LASSO)消除冗余、进行特征降维,保留与标签高相关的特征。应用10倍交叉验证,得到特征子集。

1.3 统计学分析

       使用R语言(版本4.2.1,https://cran.r-project.org/)进行统计检验。符合正态分布的计量资料以均数±标准差(x¯±s)形式表示,两组间比较采用独立样本t检验;计数资料以n(%)表示,组间比较采用χ2检验。采用单因素和多因素逻辑回归向后法分析各临床特征指标。绘制受试者工作特征(receiver operating characteristic, ROC)曲线并计算曲线下面积(area under the curve, AUC)和95%置信区间(confidence interval, CI)、阈值、敏感度、特异度、准确度、阴性预测值和阳性预测值分析各个模型的诊断效能。采用DeLong检验两两比较各组间的差异,进一步使用决策曲线分析(decision curve analysis, D CA)评估模型的净获益。

2 结果

2.1 临床特征因素分析

       单、多因素logistic回归结果显示年龄、TPSA、PV、%fPSA是csPCa的独立危险因素(P<0.05),而FPSA的差异无统计学意义(P>0.05)(表1)。最终纳入年龄、TPSA、PV、%fPSA 4项临床指标作为临床模型构建的特征。

表1  临床指标单因素和多因素逻辑回归分析结果
Tab. 1  Results of univariate and multivariate logistic regression analysis of clinical indicators

2.2 影像特征的选择

       采用MRMR算法和LASSO消除冗余、进行特征降维,保留与标签高相关的特征,应用10倍交叉验证,得到特征子集,最终ADC模型共筛选出5个组学特征(图2A),其中小波变换纹理特征4个、灰度相关矩阵纹理特征1个;MD模型共筛选出6个组学特征(图2B),为小波变换纹理特征;MK模型共筛选出6个组学特征(图2C),其中小波变换纹理特征4个、灰度区域大小矩阵纹理特征2个。

图2  经过最终特征筛选ADC、MD和MK模型有关的影像组学特征及权重,横轴为每个特征对应的权重系数,纵轴为与每个模型相关的每个影像组学特征名称。wavelet:小波变换;firstorder:一阶纹理特征;gldm:灰度依赖矩阵;glszm:灰度区域大小矩阵。
Fig. 2  Radiomics features and weights related to ADC, MD and MK models after the final feature screening. The horizontal axis is the corresponding weight coefficient of each feature, and the vertical axis is the name of each radiomics feature related to each model. wavelet: wavelet transform; firstorder: first-order texture feature; gldm: grayscale dependence matrix; glszm: grayscale region size matrix.

2.3 各个模型预测效能的评估

       基于临床特征构建的临床模型,ADC、MD和MK图像组学特征构建的影像学模型以及临床-影像学联合构建的模型在训练集中(图3A)的AUC值分别为0.840(95% CI:0.778~0.901)、0.927(95% CI:0.890~0.964)、0.934(95% CI:0.899~0.969)和0.935(95% CI:0.900~0.971)、0.946(95% CI:0.912~0.980);在测试集中(图3B)的AUC值分别为0.675(95% CI:0.539~0.812)、0.909(95% CI:0.835~0.983)、0.960(95% CI:0.910~1.000)和0.856(95% CI:0.770~0.941)、0.963(95% CI:0.925~1.000)。各模型的阈值、准确度、敏感度、特异度、阳性预测值、阴性预测值详见表2。利用DeLong检验两两比较各组间AUC值的差异,结果显示影像学模型和临床-影像学联合模型间差异均无统计学意义(P>0.05),临床模型和其他两个模型的AUC值差异具有统计学意义(Z=2.836,P=0.004)。DCA中各模型的阈值概率在0.1~1.0范围内,对临床有净获益(图4A4B)。训练集与测试集中均以临床-影像学联合模型诊断效能较高。

图3  临床、ADC、MD、MK和临床-影像学联合模型在训练集(3A)和测试集(3B)中的ROC曲线。ROC曲线显示以临床-影像学联合模型的诊断效能最高。ADC:表观扩散系数;MD:平均扩散系数;MK:平均扩散峰度;ROC:受试者工作特征。
Fig. 3  ROC curves of clinical, ADC, MD, MK and clinical-radiology model in the training (3A) set and the test set (3B). ROC curves shows that the clinical-radiology combined model has the highest diagnostic efficiency. ADC: apparent diffusion coefficient; MD mean diffusivty; MK: mean kurtosis; ROC: receiver operating characteristic.
图4  临床、ADC、MD、MK和临床-影像学联合模型在训练集(4A)和测试集(4B)的决策曲线。横轴代表高风险阈值,纵轴代表校正净获益;决策曲线显示临床-影像学模型联合模型的临床净获益最高。CLI:临床;ADC:表观扩散系数;MD:平均扩散系数;MK:平均扩散峰度;COM:临床-影像学联合。
Fig. 4  The decision curves of clinical, ADC, MD, MK and clinical-radiology models in the training set (4A) and the test set (4B). The horizontal axis represents the high-risk threshold, and the vertical axis represents the net benefit of correction; the decision curves shows that the clinical-radiology combined model has the highest net clinical benefit. CLI: clinical; ADC: apparent diffusion coefficient; MD mean diffusivty; MK: mean kurtosis; COM: clinical-radiology combined.
表2  临床、ADC、MD、MK模型和临床-影像学联合模型在训练集和测试集中的效能
Tab. 2  The efficacy of clinical, ADC, MD, MK, clinical-radiology models in the training and testing sets

3 讨论

       本研究对临床特征进行单因素和多因素逻辑回归分析,从MRI扩散模型(ADC、MD、MK图)中提取并筛选出17个影像组学特征,建立逻辑回归模型诊断csPCa,结果显示在训练集和测试集中临床-影像学联合模型均有较好的诊断效能(AUC=0.946、0.963),可以为临床医生在患者的早期诊断和治疗中提供帮助。

3.1 单指数模型影像组学的诊断价值分析

       DWI可以通过水分子的扩散运动来反映组织内部的微观环境,能够以一种无创的方式反映肿瘤的异质性信息,ADC值能够反映水的微观流动性有关的信息,肿瘤病变一般表现为扩散受限[18]。XIONG等[19]研究结果表明通过ADC图提取的纹理特征,发现峰度、偏度和熵对csPCa具有一定的诊断能力,当对三个纹理参数组合分析时,ROC曲线下面积达到最大,AUC为0.846(95% CI:0.758~0.935)。SATIO等[20]研究结果表明直方图特征中第10和25百分位数对csPCa的诊断中具有显著优势。本研究结果中构建模型权重占比较高的主要是一阶特征,这与既往研究一致。而李梦娟等[21]研究结果显示在构建双参数影像组学模型中灰度相关矩阵和形态特征对诊断csPCa的贡献最大,与本研究存在差异,其研究结果显示一阶纹理特征特别是第10百分位数、熵、能量更能反映出肿瘤的异质性信息,一阶纹理参数较二阶和高阶纹理参数更容易理解[22]。本研究从一阶参数构建的模型在训练集和测试集中的AUC值为0.927和0.909,具有很好的诊断效能,有助于临床决策,更具有临床有效性。

3.2 扩散峰度模型影像组学的诊断价值分析

       单指数模型是一种高斯分布运动形式,水分子运动成单指数信号衰减,而真实的水分子在生物组织扩散中是非高斯分布的,研究结果表明非高斯扩散峰度模型对DWI信号的拟合优度要高于标准单指数模型的DWI数据[23]。本研究中引入了高阶扩散模型——扩散峰度模型,能够准确地反映出水分子在细胞内部环境及组织结构中扩散受限的真实状态[24, 25]。JIANG等[26]的研究通过对扩散峰度模型的直方图分析,发现MD的第10、25、50、75和90百分位数显著低于非癌性病灶(P<0.001),MK的平均值为第50、75和90百分位数在PCa患者中显著升高(P<0.05)。ZHOU等[27]的研究结果表明,MD和MK值在鉴别PI-RADS 3分的PCa病变中具有较好的临床价值(AUC值为0.919)。本研究结果中MD模型在训练集和测试集中的AUC分别为0.934、0.960,MK模型在训练集和测试集中的AUC分别为0.935、0.856。与单指数模型相比,扩散峰度模型MD模型的AUC要高于ADC模型,MD模型具有较好的诊断效能,这与既往的研究结果一致[28]。而本研究结果中MK模型的诊断效能在测试集中表现欠佳。MK模型中对模型构建占比权重较高的主要是灰度纹理特征,此纹理特征主要反映的是图像灰度的变化,主要体现在图像纹理的均匀度、粗糙度及平滑度等方面的差异[29]。MRI扩散模型的拟合与MRI设备、特定参数和b值设置有关[30],这些都可能导致MRI原始图像信号的差异,未来需进一步优化扫描参数,从而提高模型的预测效能。

3.3 临床因素结合影像组学模型的诊断价值分析

       先前研究结果表明临床因素也是构建PCa预测模型十分重要的预测因子[31, 32]。KRAUSS等[33]的研究分析将PSA密度和PI-RADS评分联合T2WI和ADC图的影像组学特征能有效区分外周区和过渡区的PCa(AUC为0.82)。QI等[34]的研究通过比较临床模型(纳入的临床特征包括年龄、PSA、前列腺体积、PSA密度)和临床-放射学模型在csPCa中的预测效能,发现临床-放射学模型在训练集和测试集中均具有较好的诊断效能(AUC为0.93)。本研究结果显示联合模型的在训练集和测试集中的AUC分别为0.946和0.963,DCA曲线显示具有较好的临床获益,表现出较好的预测效能。临床特征因素容易获得,结合影像组学特征能够提高对csPCa诊断的准确性,协助临床医生在手术或治疗前对患者的诊疗做出合理决策。

3.4 本研究的局限性

       本研究存在的不足:(1)本研究为回顾性的单中心研究,缺乏外部验证,未来会逐渐扩大样本量,保持数据均衡,引用多中心研究数据,采用深度学习或卷积神经网络技术进一步提高其诊断效能;(2)由于部分前列腺病灶体积较小,通过人为手动勾画病灶会不可避免地出现主观性的偏差,一些出血、坏死、囊变难以避开,最终会造成结果偏移,未来会采用人工和半自动化结合的勾画方式,避免出现主观视觉上的偏差造成结果的误差。

4 结论

       综上所述,基于临床因素及MRI扩散模型构建的临床、影像学、临床-影像学联合预测模型对鉴别csPCa和ncsPca均具有较好的应用价值,以临床-影像学联合模型诊断效能最高,未来能够在临床的前列腺疾病早期诊断中提供一定的参考价值,也对临床为患者的治疗和预后提供不同治疗计划提供依据。

[1]
BERGENGREN O, PEKALA K R, MATSOUKAS K, et al. 2022 update on prostate cancer epidemiology and risk factors-a systematic review[J]. Eur Urol, 2023, 84(2): 191-206. DOI: 10.1016/j.eururo.2023.04.021.
[2]
李星, 曾晓勇. 中国前列腺癌流行病学研究进展[J]. 肿瘤防治研究, 2021, 48(1): 98-102. DOI: 10.3971/j.issn.1000-8578.2021.20.0370.
LI X, ZENG X Y. Advances in epidemiology of prostate cancer in China[J]. Cancer Res Prev Treat, 2021, 48(1): 98-102. DOI: 10.3971/j.issn.1000-8578.2021.20.0370.
[3]
OERTHER B, ENGEL H, NEDELCU A, et al. Prediction of upgrade to clinically significant prostate cancer in patients under active surveillance: performance of a fully automated AI-algorithm for lesion detection and classification[J]. Prostate, 2023, 83(9): 871-878. DOI: 10.1002/pros.24528.
[4]
MATOSO A, EPSTEIN J I. Defining clinically significant prostate cancer on the basis of pathological findings[J]. Histopathology, 2019, 74(1): 135-145. DOI: 10.1111/his.13712.
[5]
RITCH C, COOKSON M. Recent trends in the management of advanced prostate cancer[J/OL]. F1000Res, 2018, 7: F1000FacultyRev-F1000Faculty1513 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/30345007/. DOI: 10.12688/f1000research.15382.1.
[6]
SANTORO A A, DI GIANFRANCESCO L, RACIOPPI M, et al. Multiparametric magnetic resonance imaging of the prostate: lights and shadows[J]. Urologia, 2021, 88(4): 280-286. DOI: 10.1177/03915603211019982.
[7]
FUSCO R, SANSONE M, GRANATA V, et al. A systematic review on multiparametric MR imaging in prostate cancer detection[J/OL]. Infect Agent Cancer, 2017, 12: 57 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/29093748/. DOI: 10.1186/s13027-017-0168-z.
[8]
WANG F, CHEN H G, ZHANG R Y, et al. Diffusion kurtosis imaging to assess correlations with clinicopathologic factors for bladder cancer: a comparison between the multi-b value method and the tensor method[J]. Eur Radiol, 2019, 29(8): 4447-4455. DOI: 10.1007/s00330-018-5977-y.
[9]
韩丽莹, 何冠勇, 喻晴, 等. MRI多模型扩散加权成像及PSA相关指标诊断前列腺癌的价值[J]. 中国CT和MRI杂志, 2023, 21(3): 121-123. DOI: 10.3969/j.issn.1672-5131.2023.03.043.
HAN L Y, HE G Y, YU Q, et al. Diagnostic Value of MRI Multi-model Diffusion-weighted Imaging and PSA-related Indicators in Prostate Cancer*[J]. CHINESE JOURNAL OF CT AND MR, 2023, 21(3): 121-123. DOI: 10.3969/j.issn.1672-5131.2023.03.043.
[10]
LUCARELLI N M, VILLANOVA I, MAGGIALETTI N, et al. Quantitative ADC: an additional tool in the evaluation of prostate cancer?[J/OL]. J Pers Med, 2023, 13(9): 1378 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/37763146/. DOI: 10.3390/jpm13091378.
[11]
YAO L, DING J, LI X Y, et al. Application of dynamic enhanced magnetic resonance imaging texture analysis combined with ADCs in predicting pelvic lymph metastasis of prostate cancer[J]. Arch Esp Urol, 2023, 76(6): 383-388. DOI: 10.56434/j.arch.esp.urol.20237606.46.
[12]
SEN S, VALINDRIA V, SLATOR P J, et al. Differentiating false positive lesions from clinically significant cancer and normal prostate tissue using VERDICT MRI and other diffusion models[J/OL]. Diagnostics, 2022, 12(7): 1631 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/35885536/. DOI: 10.3390/diagnostics12071631.
[13]
UENO Y, TAMADA T, SOFUE K, et al. Diffusion and quantification of diffusion of prostate cancer[J/OL]. Br J Radiol, 2022, 95(1131): 20210653 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/34538094/. DOI: 10.1259/bjr.20210653.
[14]
CAVINATO L, PEGORARO M, RAGNI A, et al. Imaging-based representation and stratification of intra-tumor heterogeneity via tree-edit distance[J/OL]. Sci Rep, 2022, 12(1): 19607 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/36380083/. DOI: 10.1038/s41598-022-23752-2.
[15]
HAMEED M, GANESHAN B, SHUR J, et al. The clinical utility of prostate cancer heterogeneity using texture analysis of multiparametric MRI[J]. Int Urol Nephrol, 2019, 51(5): 817-824. DOI: 10.1007/s11255-019-02134-0.
[16]
BANERJEE V, WANG S, DRESCHER M, et al. Radiogenomics influence on the future of prostate cancer risk stratification[J/OL]. Ther Adv Urol, 2022, 14: 17562872221125317 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/36160762/. DOI: 10.1177/17562872221125317.
[17]
JIANG L, YOU C, XIAO Y, et al. Radiogenomic analysis reveals tumor heterogeneity of triple-negative breast cancer[J/OL]. Cell Rep Med, 2022, 3(7): 100694 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/35858585/. DOI: 10.1016/j.xcrm.2022.100694.
[18]
GRANATA V, FUSCO R, BELLI A, et al. Diffusion weighted imaging and diffusion kurtosis imaging in abdominal oncological setting: why and when[J/OL]. Infect Agent Cancer, 2022, 17(1): 25 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/35681237/. DOI: 10.1186/s13027-022-00441-3.
[19]
XIONG H, HE X J, GUO D J. Value of MRI texture analysis for predicting high-grade prostate cancer[J]. Clin Imaging, 2021, 72: 168-174. DOI: 10.1016/j.clinimag.2020.10.028.
[20]
SAITO S, KOYAMA Y, UEDA J, et al. Relationship between apparent diffusion coefficient distribution and cancer grade in prostate cancer and benign prostatic hyperplasia[J/OL]. Diagnostics, 2022, 12(2): 525 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/35204614/. DOI: 10.3390/diagnostics12020525.
[21]
李梦娟, 张彩元, 赵文露, 等. 基于双参数MRI的影像组学模型在临床显著性前列腺癌中的诊断价值探讨[J]. 磁共振成像, 2022, 13(11): 76-81. DOI: 10.12015/issn.1674-8034.2022.11.014.
LI M J, ZHANG C Y, ZHAO W L, et al. Radiomics prediction model for the improved diagnosis of clinically significant prostate cancer on biparametric MRI[J]. Chin J Magn Reson Imag, 2022, 13(11): 76-81. DOI: 10.12015/issn.1674-8034.2022.11.014.
[22]
ÖZER H, KOPLAY M, BAYTOK A, et al. Texture analysis of multiparametric magnetic resonance imaging for differentiating clinically significant prostate cancer in the peripheral zone[J]. Turk J Med Sci, 2023, 53(3): 701-711. DOI: 10.55730/1300-0144.5633.
[23]
ŻUROWSKA A, PĘKSA R, GRZYWIŃSKA M, et al. Comparison of diffusion kurtosis imaging and standard mono-exponential apparent diffusion coefficient in diagnosis of significant prostate cancer-a correlation with gleason score assessed on whole-mount histopathology specimens[J/OL]. Diagnostics, 2023, 13(2): 173 [2023-01-10]. https://pubmed.ncbi.nlm.nih.gov/36672983/. DOI: 10.3390/diagnostics13020173.
[24]
张家慧, 郎宁, 袁慧书. 磁共振扩散峰度成像的临床研究进展[J]. 磁共振成像, 2018, 9(4): 316-320. DOI: 10.12015/issn.1674-8034.2018.04.016.
ZHANG J H, LANG N, YUAN H S. Research advances in diffusional kurtosis imaging[J]. Chin J Magn Reson Imag, 2018, 9(4): 316-320. DOI: 10.12015/issn.1674-8034.2018.04.016.
[25]
MALAGI A V, NETAJI A, KUMAR V, et al. IVIM–DKIfor differentiation between prostate cancer and benign prostatic hyperplasia: comparison of 1.5T vs. 3T MRI[J]. Magn Reson Mater Phys Biol Med, 2022, 35(4): 609-620. DOI: 10.1007/s10334-021-00932-1.
[26]
JIANG Y W, LI C M, LIU Y, et al. Histogram analysis in prostate cancer: a comparison of diffusion kurtosis imaging model versus monoexponential model[J]. Acta Radiol, 2020, 61(10): 1431-1440. DOI: 10.1177/0284185120901504.
[27]
ZHOU K P, HUANG H B, BU C, et al. Sub-differentiation of PI-RADS 3 lesions in TZ by advanced diffusion-weighted imaging to aid the biopsy decision process[J/OL]. Front Oncol, 2023, 13: 1092073 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/36845749/. DOI: 10.3389/fonc.2023.1092073.
[28]
LI C M, CHEN M, WAN B, et al. A comparative study of Gaussian and non-Gaussian diffusion models for differential diagnosis of prostate cancer with in-bore transrectal MR-guided biopsy as a pathological reference[J]. Acta Radiol, 2018, 59(11): 1395-1402. DOI: 10.1177/0284185118760961.
[29]
田士峰, 刘爱连, 郭妍, 等. 基于扩散峰度成像序列平均扩散峰度图的影像组学鉴别宫颈鳞癌与腺癌的应用[J]. 中国医学影像学杂志, 2021, 29(7): 716-720. DOI: 10.3969/j.issn.1005-5185.2021.07.015.
TIAN S F, LIU A L, GUO Y, et al. Application of radiomics methods based on DKI sequence MK map for differentiating squamous cell carcinoma from cervix adenocarcinoma[J]. Chin J Med Imag, 2021, 29(7): 716-720. DOI: 10.3969/j.issn.1005-5185.2021.07.015.
[30]
ZHOU B N, LIU X H, GAN H L, et al. Differentiation of prostate cancer and stromal hyperplasia in the transition zone with monoexponential, stretched-exponential diffusion-weighted imaging and diffusion kurtosis imaging in a reduced number of b values: correlation with whole-mount pathology[J]. J Comput Assist Tomogr, 2022, 46(4): 545-550. DOI: 10.1097/RCT.0000000000001314.
[31]
LIU T B, ZHANG X M, CHEN R, et al. Development, comparison, and validation of four intelligent, practical machine learning models for patients with prostate-specific antigen in the gray zone[J/OL]. Front Oncol, 2023, 13: 1157384 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/37361597/. DOI: 10.3389/fonc.2023.1157384.
[32]
CHEN Y C, XU D, RUAN M J, et al. A prospective study of the prostate health index density and multiparametric magnetic resonance imaging in diagnosing clinically significant prostate cancer[J]. Investig Clin Urol, 2023, 64(4): 363-372. DOI: 10.4111/icu.20230060.
[33]
KRAUSS W, FREY J, LAGERLÖF J H, et al. Radiomics from multisite MRI and clinical data to predict clinically significant prostate cancer[J/OL]. Acta Radiol, 2023: 2841851231216555 [2023-10-07]. https://pubmed.ncbi.nlm.nih.gov/38115809/. DOI: 10.1177/02841851231216555.
[34]
QI Y F, ZHANG S T, WEI J W, et al. Multiparametric MRI-based radiomics for prostate cancer screening with PSA in 4-10 ng/mL to reduce unnecessary biopsies[J]. J Magn Reson Imaging, 2020, 51(6): 1890-1899. DOI: 10.1002/jmri.27008.

上一篇 磁共振表观扩散系数鉴别诊断移行带高危前列腺癌及与病理分级分组的相关性
下一篇 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
  
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