摘要
目的评估未采用深度学习重建(Deep Learning-based Reconstruction,DLR)与采用不同水平DLR的前列腺磁共振成像(Magnetic Resonance Imaging,MRI)图像中提取的影像组学特征的可重复性。方法纳入57例疑似前列腺癌患者的T2加权图像,分别采用常规重建方法及低、中、高3个水平DLR技术重建。从原始图像、高斯拉普拉斯(Laplacian of Gaussian,LoG)滤波图像和小波变换图像中提取影像组学特征,采用组内相关系数(Intra-Class Coefficient,ICC)量化影像组学特征的可重复性,并评估特征校正对减轻重建相关影响的作用。结果随着DLR水平升高,特征总体可重复性显著降低(P<0.05)。纹理特征中,灰度尺寸区域矩阵指标对重建水平最敏感,有5个特征的ICC<0.75;而一阶特征保持稳定,ICC均值为0.99。在滤波图像相关分析中,LoG滤波表现出最优的稳定性,所有LoG滤波影像组学特征的ICC>0.96;而通过小波变换高频图像提取的特征,随着高频成分占比增加,可重复性下降趋势显著。此外,开展特征校正能够有效提升影像组学特征的可重复性。结论DLR会影响前列腺MRI影像组学特征的可重复性,基于DLR的图像开展影像组学特征分析时,优先选择一阶特征或LoG滤波特征,并通过特征校正减小偏差。
Objective To evaluate the reproducibility of radiomics features extracted from prostate magnetic resonance imaging(MRI)images without deep learning-based reconstruction(DLR)and with different levels of DLR.Methods T2-weighted images of 57 patients suspected of prostate cancer were included,reconstructed using conventional reconstruction and three levels of DLR(low,medium,and high).Radiomics features were extracted from original images,Laplacian of Gaussian(LoG)-filtered images,and wavelet-transformed images.The intra-class coefficient(ICC)was used to quantify the reproducibility of radiomics features,and the effect of feature correction on mitigating reconstruction-related impacts was assessed.Results With the increase of DLR level,the overall reproducibility of features significantly decreased(P<0.05).Among texture features,gray-level size zone matrix indices were most sensitive to reconstruction levels,with 5 features showing ICC<0.75;while first-order features remained stable,with a mean ICC of 0.99.In filtered image correlation analysis,LoG filtering exhibited the optimal stability,with all LoG-filtered radiomics features having ICC>0.96;in contrast,features extracted from wavelet-transformed high-frequency images showed a significant downward trend in reproducibility as the proportion of high-frequency components increased.Additionally,feature correction effectively improved the reproducibility of radiomics features.Conclusion DLR affects the reproducibility of prostate MRI radiomics features.When performing radiomics feature analysis based on DLR images,first-order features or LoG-filtered features should be preferred,and feature correction should be applied to reduce deviations.
作者
陈静雯
方向明
万红艳
陈利华
段绍峰
汤然
朱宗明
CHEN Jingwen;FANG Xiangming;WAN Hongyan;CHEN Lihua;DUAN Shaofeng;TANG Ran;ZHU Zongming(Department of Radiology,The Affiliated Wuxi People’s Hospital of Nanjing Medical University,Wuxi Jiangsu 214023,China;Collaborative Innovation Department,United Imaging Health Care Group Co.,Ltd.,Shanghai 201800,China)
出处
《中国医疗设备》
2025年第12期151-158,共8页
China Medical Devices
基金
无锡太湖人才计划医学专家团队项目(THRC-TDYXYXK-2021)。
