文献速递腹部CT

光子计数探测器CT光谱重建用于影像组学肝脏病灶分类:多中心研究

Photon-counting Detector CT Spectral Reconstructions for Radiomics-based Liver Lesion Classification: A Multicenter Study.
PubMed 文献 · Invest Radiol · 2026-10-01

这项多中心研究纳入378例患者,利用光子计数探测器CT的9种光谱数据集提取影像组学特征,训练机器学习模型鉴别肝脏病灶良恶性。基于虚拟平扫图像的随机森林模型AUC达0.899,高能级和虚拟平扫测试重建结果更稳定,提示光谱影像组学可辅助无创风险分层。

要点

  • 基于虚拟平扫的随机森林模型分类性能最高,AUC为0.899。
  • 测试重建影响性能:高能级(110/140/180 keV)和虚拟平扫更稳定。
  • 低能级(40/50 keV)和碘密度图测试AUC较低且波动较大。

核心结论

  • 基于光子计数探测器CT光谱重建的影像组学机器学习可较可靠地鉴别肝脏病灶良恶性。
  • 测试重建的选择影响分类性能,高能级和虚拟平扫图像结果更稳定。
  • 将高能级重建和虚拟平扫纳入影像组学流程可能有助于无创风险分层并减少侵入性诊断需求。

研究设计与样本

  • 多中心研究,纳入2家机构共378例局灶性肝脏病灶患者,中位年龄62岁,女性占59.5%。
  • 良性190例(227个病灶),恶性188例(2681个病灶)。
  • 病灶由nnU-Net模型自动分割并经腹部放射科医师确认,采用体积分割从9个光谱数据集提取影像组学特征。
  • 9个光谱数据集包括40、50、70、90、110、140、180 keV虚拟单能图像、虚拟平扫和碘密度图。
  • 采用基于患者的分类分析,训练多种机器学习模型,病灶级预测按患者聚合,并用随机森林评估所有训练-测试重建组合。

诊断标准与关键数值

  • 基于虚拟平扫训练的随机森林模型性能最高:AUC 0.899(95% CI 0.840-0.951),准确率83.5%。
  • 交叉域分析中,所有训练重建的平均AUC相近(0.75至0.83)。
  • 在110、140、180 keV和虚拟平扫上测试的模型中位AUC≥0.83,稳定性最高。
  • 在40 keV、50 keV和碘密度图上测试的模型中位AUC≤0.71,较低且波动更大。

临床应用与阅片要点

  • 日常阅片或科研建模时,可优先考虑高能级虚拟单能和虚拟平扫重建用于肝脏病灶良恶性分类。
  • 低能级虚拟单能和碘密度图作为测试集时分类性能较低且不稳定,解读其结果需谨慎。
  • 影像组学模型为患者级聚合预测,报告时应明确训练和测试所用重建类型。

局限性

  • 原文未明确列出研究局限性。

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

OBJECTIVES: Accurate noninvasive classification of hepatic lesions remains a diagnostic challenge, particularly on conventional CT. Photon Counting Detector CT (PCD-CT) offers spectral imaging capabilities that may enhance tissue characterization. This study aimed to evaluate the performance of radiomics-based machine learning for differentiating benign and malignant liver lesions using multispectral virtual monochromatic and material decomposition images derived from contrast-enhanced PCD-CT.

MATERIALS AND METHODS: This multicenter study included patients with focal hepatic lesions, scanned on PCD-CT at 2 separate institutions. Lesions were automatically segmented by a nnU-Net model and confirmed by an abdominal radiologist. Volumetric segmentations were used to extract radiomic features from 9 spectral datasets: 40, 50, 70, 90, 110, 140, and 180 keV, virtual non-contrast (VNC), and iodine density maps (IDM). A patient-based classification analysis was conducted by training multiple machine-learning models, with lesion-level predictions aggregated per patient. A Random Forest model further evaluated classification performancef across all train-test reconstruction combinations.

RESULTS: The cohort comprised 378 patients (median age: 62; 59.5% female) from 2 centers, including 190 benign (n=227 lesions) and 188 with malignant (n=2681 lesions) hepatic findings. Random forest model trained on VNC achieved highest performance [AUC: 0.899; 95% CI (0.840-0.951); Accuracy 83.5%). In the cross-domain analysis, all training reconstructions achieved similar mean AUCs (0.75 to 0.83). Evaluation performance varied across test levels. Models tested on 110, 140, and 180 keV, and VNC showed the highest stability (median AUC ≥ 0.83), while 40 keV, 50 keV, and IDM resulted in lower and more variable AUCs (median AUC ≤ 0.71).

CONCLUSION: Radiomics-based machine learning enables reliable differentiation between malignant and benign hepatic lesions on PCD-CT. The choice of test reconstruction influenced performance: higher energy levels and VNC images yielded more stable results. Incorporating high-energy reconstructions and VNC images into radiomic workflows for liver lesion risk stratification may enable noninvasive triage and reduce the need for invasive diagnostics.

来源:Invest Radiol,查看原文。版权归原作者所有。

引用

  1. Pietsch FL, Vellala AK, Haag F, et al. Photon-counting Detector CT Spectral Reconstructions for Radiomics-based Liver Lesion Classification: A Multicenter Study. Invest Radiol. 2026. PMID: 41923596. doi:10.1097/RLI.0000000000001279

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。