| 郑志娟,李姝霖,马昆,向之明.深度学习重建算法对胸部CT图像质量及肺结节影像组学特征数据可重复性的影响[J].中国医学影像技术,2025,41(1):79~83 |
| 深度学习重建算法对胸部CT图像质量及肺结节影像组学特征数据可重复性的影响 |
| Impact of deep learning reconstruction algorithms on image quality of chest CT and reproducibility of lung nodule radiomics feature data |
| 投稿时间：2024-04-09 修订日期：2024-11-18 |
| DOI：10.13929/j.issn.1003-3289.2025.01.017 |
| 中文关键词: 深度学习 肺疾病 体层摄影术,X线计算机 影像组学 前瞻性研究 |
| 英文关键词:deep learning lung diseases tomography, X-ray computed radiomics prospective studies |
| 基金项目:国家自然科学基金(82171931)。 |
|
| 摘要点击次数: 413 |
| 全文下载次数: 128 |
| 中文摘要: |
| 目的 与自适应统计迭代重建(ASIR-V)算法比较,观察深度学习图像重建(DLIR)对于胸部CT图像质量、肺结节检出率及肺结节影像组学特征数据可重复性的影响。方法 前瞻性纳入75例同期接受胸部超低剂量CT(ULD-CT)及标准剂量CT(SDCT)检查的肺结节患者、共211个肺结节,分别以高强度DLIR(DLIR-H)、中强度DLIR(DLIR-M)及50%水平ASIR-V算法(50%ASIR-V)重建ULD-CT图像,以50% ASIR-V获得SDCT图像。针对ULD-CT和SDCT图像,以相同ROI内肺实质CT值标准差(SD)代表噪声并计算信噪比(SNR);计算肺结节检出率,并在不同图像间加以比较。基于自动分割法提取50%ASIR-V SDCT及各ULD-CT肺结节影像组学特征,分别计算各ULD-CT与50%ASIR-V SDCT影像组学特征的组内相关系数(ICC),并与不同ULD-CT算法进行比较。结果 相比50%ASIR-V SDCT图像,以不同算法重建的ULD-CT图像SD均较高而SNR均较低(P均＜0.05)。ULD-CT图像中,DLIR-H、DLIR-M及50%ASIR-V图像之间,SD依序增加而SNR渐次减低(校正P均＜0.05)。以50%ASIR-V SDCT图像为标准,于DLIR-H、DLIR-M及50%ASIR-V ULD-CT中均检出207个(207/211,98.10%)肺结节。ULD-CT图像中,50%ASIR-V重建图像肺结节纹理特征数据与50%ASIR-V SDCT的可重复性低于DLIR-H及DLIR-M重建图像(校正P均＜0.05),而后二者与50%ASIR-V SDCT的可重复性差异无统计学意义(校正P＞0.05)。基于3种算法重建的ULD-CT所获肺结节与50%ASIR-V SDCT一阶特征及形状特征数据的可重复性较好(中位ICC均＞0.75),不同算法间差异无统计学意义(P均＞0.05)。结论 以DLIR-H和DLIR-M算法重建胸部ULD-CT可在降低图像噪声的同时提高图像质量,并在一定程度上维持肺结节影像学特征的可重复性,尤以DLIR-H更佳。 |
| 英文摘要: |
| Objective To explore the impact of deep learning image reconstruction (DLIR) algorithms on image quality of chest CT, detection rate of lung nodule and reproducibility of lung nodule radiomics feature data compared with adaptive statistical iterative reconstruction V (ASIR-V) algorithms. Methods Seventy-five patients with 211 lung nodules who underwent both ultra-low-dose CT (ULD-CT) and standard-dose CT (SDCT) were prospectively enrolled. ULD-CT images were reconstructed using different algorithms, namely high-level DLIR (DLIR-H), medium-level DLIR (DLIR-M) and 50% ASIR-V (50%ASIR-V), while SDCT images were reconstructed by 50%ASIR-V. Image noise was represented by the standard deviation (SD) of lung parenchyma CT values within identical ROI in both ULD-CT and SDCT images, and signal-to-noise ratio (SNR) were calculated. The detection rate of lung nodule were obtained and compared among different images. Radiomics features of lung nodules in chest 50%ASIR-V SDCT and each ULD-CT were extracted based on automatic segmentation methods, and intra-class correlation coefficients (ICC) of each ULD-CT and 50%ASIR-V SDCT were calculated respectively, and then compared among different ULD-CT algorithms. Results Compared with SDCT images reconstructed with 50%ASIR-V algorithm, all ULD-CT images reconstructed by different algorithms showed higher SD and lower SNR (all P＜0.05). ULD-CT images reconstructed by DLIR-H, DLIR-M and 50%ASIR-V exhibited progressively increasing SD and decreasing SNR (all adjusted P＜0.05). Taken 50%ASIR-V SDCT images as standards, ULD-CT by DLIR-H, DLIR-M and 50%ASIR-V each detected 207 lung nodules (207/211, 98.10%), respectively. In chest ULD-CT images, the reproducibility with 50%ASIR-V SDCT for texture feature data of lung nodules on ULD-CT reconstructed by 50%ASIR-V algorithm was lower than that by DLIR-H and DLIR-M (both adjusted P＜0.05), while no significant difference was found between the latter two with 50%ASIR-V SDCT (adjusted P＞0.05). The first order and shape feature data of lung nodules on ULD-CT reconstructed by all 3 algorithms showed good reproducibility with 50%ASIR-V SDCT (median ICC＞0.75), and no significant difference was detected among them (all P＞0.05). Conclusion Compared with 50%ASIR-V ULD-CT, both DLIR-H and DLIR-M ULD-CT could significantly reduce image noise and improve image quality, as well as maintain reproducibility of radiomics features in lung nodules in a certain degree, especially DLIR-H. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
