文献速递胸部CT

数学保真度与感知真实感:超分辨率影像组学预测肺腺癌侵袭性的多中心研究

Mathematical fidelity vs. perceptual realism: A multi-center study of super-resolution radiomics for lung adenocarcinoma invasiveness.
PubMed 文献 · Eur J Radiol · 2026-03-11

该多中心回顾性研究纳入373例肺腺癌患者,比较RCAN、Real-ESRGAN和HAT三种超分辨率模型对影像组学特征稳定性和诊断性能的影响。RCAN在特征稳定性(最高CCC 0.779,最低RFD 9.02)和诊断性能(平均AUC 0.911±0.008)上最优,显著优于原始分辨率基线(AUC 0.875±0.022,P=0.043)和临床模型(AUC 0.924 vs. 0.837,P=0.006),而Real-ESRGAN出现严重特征失真(RFD>300,000),HAT和Real-ESRGAN未显著提升诊断。

要点

  • RCAN超分辨率模型在肺腺癌侵袭性预测中平均AUC达0.911±0.008,显著优于原始分辨率基线(P=0.043)。
  • Real-ESRGAN虽感知质量高但特征失真严重(RFD>300,000),不适合定量影像组学。
  • RCAN模型AUC 0.924显著优于临床模型0.837(P=0.006),校准良好(Brier score 0.1766)。

核心结论

  • 基于CNN的RCAN在定量影像组学中优于生成式和Transformer架构,优先保证数学保真度而非感知真实感,可保留信号完整性以实现准确可靠的肺腺癌风险分层。

研究设计与样本

  • 多中心回顾性研究,纳入两个中心的373例肺腺癌患者,在内部数据集(81,545张图像)上微调三种超分辨率模型,提取107个影像组学特征,使用CCC和RFD评估特征保真度,采用五种机器学习分类器分析诊断性能并以Brier分数量化临床可靠性。

诊断标准与关键数值

  • RCAN特征稳定性最佳,一阶和纹理类别最高CCC为0.779、最低RFD为9.02;RCAN平均AUC 0.911±0.008,显著优于原始分辨率基线0.875±0.022(P=0.043),Brier分数0.1766;HAT(P=0.628)和Real-ESRGAN(P=0.923)未显示显著诊断改善;RCAN模型AUC 0.924显著优于临床模型0.837(P=0.006)。

临床应用与阅片要点

  • 作者结论认为,基于CNN的RCAN通过优先保证数学保真度而非感知真实感,可保留信号完整性以实现准确可靠的肺腺癌风险分层。

局限性

  • 原文未报告局限性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

OBJECTIVE: To identify the optimal super-resolution (SR) architecture for radiomics by comparing three models (Residual Channel Attention Network (RCAN), Real-Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN), and Hybrid Attention Transformer (HAT)) in predicting lung adenocarcinoma (LUAD) invasiveness while systematically evaluating feature stability and clinical reliability.

METHODS: We retrospectively enrolled 373 LUAD patients across two centers (Approval: Huadong Hospital No. 20230051; Nantong Third People's Hospital No. EL2025021). Three SR models were fine-tuned on an in-house dataset (81,545 images) to generate SR images from original-resolution (OR) scans. We extracted 107 radiomic features and assessed fidelity across feature categories using the Clinical Concordance Coefficient (CCC) and Radiomic Feature Distance (RFD). Diagnostic performance across five machine learning classifiers was analyzed, with clinical reliability quantified by Brier scores. The best signature was validated against a clinical-only model.

RESULTS: Despite high perceptual quality, Real-ESRGAN exhibited catastrophic feature distortion (RFD > 300,000). RCAN achieved optimal stability, especially in first-order and texture categories (highest CCC: 0.779; lowest RFD: 9.02). RCAN achieved the highest diagnostic performance and stability (Mean AUC: 0.911 ± 0.008), significantly outperforming the OR baseline (0.875 ± 0.022; P = 0.043) with superior calibration (Brier score: 0.1766). Conversely, neither HAT (P = 0.628) nor Real-ESRGAN (P = 0.923) showed significant diagnostic improvement. Furthermore, the RCAN model significantly outperformed the clinical-only baseline (AUC: 0.924 vs. 0.837, P = 0.006).

CONCLUSION: The CNN-based RCAN outperforms generative and transformer architectures for quantitative radiomics. By prioritizing mathematical fidelity over perceptual realism, RCAN preserves signal integrity to ensure accurate and reliable LUAD risk stratification.

来源:Eur J Radiol,查看原文。版权归原作者所有。

引用

  1. Liu MT, Jin L, Zhang T, et al. Mathematical fidelity vs. perceptual realism: A multi-center study of super-resolution radiomics for lung adenocarcinoma invasiveness. Eur J Radiol. 2026. PMID: 41833231. doi:10.1016/j.ejrad.2026.112794

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。