放射学实践2024，Vol.39Issue(8) ：1081-1088.DOI:10.13609/j.cnki.1000-0313.2024.08.015

深度学习重建算法对肾上腺肿瘤的检出及鉴别效能的影响

The effect of deep-learning image reconstruction algorithm on the visualization and classification of adrenal tumors

王诗耕 刘义军 童小雨 范勇 李贝贝 王旭 崔景景 陈安良
放射学实践2024，Vol.39Issue(8) ：1081-1088.DOI:10.13609/j.cnki.1000-0313.2024.08.015
✕
  • NETL
  • NSTL
  • 万方数据

深度学习重建算法对肾上腺肿瘤的检出及鉴别效能的影响

The effect of deep-learning image reconstruction algorithm on the visualization and classification of adrenal tumors

王诗耕 1刘义军 1童小雨 1范勇 1李贝贝 1王旭 1崔景景 2陈安良1
扫码查看

作者信息

  • 1. 116011 辽宁大连,大连医科大学附属第一医院放射科
  • 2. 100094 北京,联影智能医疗科技(北京)有限公司
  • 折叠

摘要

目的:探讨不同等级深度学习图像重建(DLIR)算法对肾上腺肿瘤的检出、组学特征可重复性和组学模型鉴别肿瘤类型效能的影响.方法:回顾性收集41例肾上腺功能性腺瘤(FAA)和46例肾上腺转移瘤(AM)患者的临床和影像资料.CT增强扫描完成后,对静脉期的原始数据采用4种强度等级(DL1、DL2、DL3、DL4)的DLIR算法进行重建.首先采用主、客观指标比较4种等级间图像质量的差异;然后使用Research Portal V1.1科研平台对各组重建图像上肾上腺肿瘤进行分割并提取450个影像组学特征,包括原始图像特征90个和拉普拉斯(LoG)滤波后的高阶特征(高斯核:0.5、1.0、1.5、2.0)360个.采用一致性相关系数(CCC)评估采用不同图像重建等级测量的FAA和AM组学特征的可重复性.最后,在各组重建图像中采用逐步特征选择策略,筛选出最优特征集并构建鉴别FAA和AM的组学模型.利用五折交叉验证法验证4个组学模型的鉴别效能,利用分层交叉验证法测评4个模型的泛化能力.结果:DL2和DL3在肾显示上腺肿瘤的清晰度方面最优,得分为4(4,5),优于DL1相应得分4(3,5)和DL4相应得分4(3,4),且差异具有统计学意义(F=139.045,P<0.05).随着DLIR降噪等级的提升,原始特征CCC值>0.85的个数逐渐减少,DL4中FAA和AM特征可重复的比例仅占39.3％(21/90)和50.9％(29/90).组学特征经过LoG滤波(高斯核2.0)处理后,CCC值>0.85的个数增加,DL4中FAA和AM特征可重复的比例占91.1％(82/90)和93.3％(84/90).4个组学模型在测试集中的曲线下面积(AUC)和符合率均>0.75,DeLong检验显示AUC的差异无统计学意义(Z=0.177～1.284,P=0.199～0.859).但分层交叉验证显示,DL4重建图像的泛化能力最弱,AUC和符合率均<0.75.结论:高降噪等级的DLIR算法会降低对肾上腺肿瘤显示的清晰度以及组学模型的泛化性.虽然LoG滤波器(高斯核:2.0)有助于提升组学特征测量的可重复性,但仍建议在肾上腺影像诊断和组学模型训练时,使用中低降噪等级的DLIR图像.

Abstract

Objective:To observe the impact of different levels of deep-learning image recon-struction (DLIR)algorithm on the conspicuity of adrenal tumors,the reproducibility of radiomics fea-tures,and the performance of radiomics models in differentiating tumor types.Methods:The clinical and CT data of forty-one patients with functioning adrenocortical adenoma (FAA)and forty-six pa-tients with adrenal metastases (AM)were collected retrospectively.After completion of CT enhanced scans,the raw data at venous phase was reconstructed using four levels of DLIR (DL1,DL2,DL3 and DL4)respectively.The image quality of different reconstruction levels was compared first using sub-jective and objective evaluation.Adrenal tumor segmentation and 450 radiomics features extraction in each group of reconstructed images were established by the Research Portal V1.1 .Radiomics features including 90 features from original image features as well as 360 features from higher-order features after Laplacian of Gaussian (LoG)filtering (with Gaussian kernels of 0.5,1.0,1.5 and 2.0,respec-tively).The concordance correlation coefficient (CCC)was used to evaluate the reproducibility of the radiomics features of FAA and AM after the DLIR level was increased.Finally,a step-wise feature se-lection strategy was employed in each group of reconstructed images to select the optimal feature set and construct a radiomics model to differentiate FAAs from Ams.The discrimination performance of the four models was verified using five-fold cross-validation,while their generalizability was assessed using stratified cross-validation.Results:Images of DL2 and DL3 were optimal in displaying adrenal le-sions with a score of 4 (4,5),which were better than those of DL1 with a corresponding score of 4 (3,5)and DL4 with a corresponding score of 4 (3,4),with statistical difference (F=139.045,P<0.05).With the increase in DLIR level,the number of original features with a CCC value>0.85 gradually decreased,and the reproducibility proportion of FAA and AM features in DL4 was only 39.3％ (21/90)and 50.9％ (29/90),respectively.After the radiomics features were processed with the LoG filter (Gaussian kernel:2.0),the number of features with CCC values>0.85 increased,with reproducibility proportions in DL4 being 91.1％ (82/90)for FAA and 93.3％ (84/90)for AM.The area under the curve (AUC)and accuracy of the four radiomics models were all>0.75 in the test sets,with no sta-tistically significant differences according to DeLong test (Z=0.177～1.284,P=0.199～0.859).However,stratified cross-validation showed that the generalization ability of images reconstructed with DL4 was the weakest,with both AUC values and accuracy<0.75.Conclusion:High-level noise reduc-tion in DLIR images reduces the conspicuity of adrenal tumors as well as the generalizability of the ra-diomics models.Although the LoG filter with Gaussian kernel of 2.0 can help for enhancing the repro-ducibility of radiomics features,it is still recommended to use medium-to low-noise-level DLIR images for adrenal imaging diagnosis and radiomics model training.

关键词

体层摄影术,X线计算机/影像组学/深度学习/重建算法/肾上腺肿瘤/可重复性

Key words

Tomography,X-ray computed/Radiomics/Deep learning/Reconstruction algorithm/Adrenal gland neoplasm/Reproducibility

引用本文复制引用

出版年

2024
放射学实践
华中科技大学同济医学院

放射学实践

CSTPCD北大核心
影响因子：1.08
ISSN：1000-0313
段落导航相关论文