深度学习肺癌自动分割流程助力CT影像组学研究
本研究开发并验证了一种基于nnU-Net的深度学习自动分割流程，用于接受手术切除的2A-3B期非小细胞肺癌患者的CT影像组学分析。自动分割时间约36秒/序列，体积和表面DSC均值分别为0.84和0.79；预训练中加入解剖先验可显著提升分割性能（p<0.001）。但多数影像组学特征在扰动和重采样后不可重复，加入影像组学特征并未显著改善2年无病生存预测（AUC 0.67 vs 0.63，p=0.28）。
要点
- 自动分割流程高效可行，体积DSC为0.84，但6.5%病例出现漏分割、6.5%需重大修改。
- 肿瘤胸壁粘连时分割性能下降，DSC与肿瘤形状（球形度、直径）及位置相关。
- 多数影像组学特征对分割扰动和重采样不稳健，传统影像组学重复性有限。
关键影像表现/征象
- 自动分割时间约36秒/序列，体积Dice-Sørensen系数（DSC）均值0.84（±0.28），表面DSC均值0.79（±0.34）。
- DSC评分与肿瘤形状（球形度、直径）和位置显著相关，胸壁粘连者分割性能更差；与重建核等批次效应无关。
- 总体6.5%病例出现分割遗漏，6.5%需要重大修改。
- 在解剖先验上预训练的分割效果优于仅用肿瘤标签训练（p<0.001）以及肿瘤加解剖标签训练（p<0.001）。
- 多数影像组学特征在扰动和重采样后不可重复。
诊断标准与关键数值
- 体积DSC均值0.84（±0.28），表面DSC均值0.79（±0.34）。
- 预测2年无病生存：临床模型AUC为0.63（95%CI 0.54-0.71），加入影像组学特征后AUC为0.67（95%CI 0.59-0.75），差异无统计学意义（p=0.28）。
- 预训练方式比较：解剖先验预训练 vs 仅肿瘤标签训练 p<0.001；解剖先验预训练 vs 肿瘤加解剖标签训练 p<0.001。
临床应用与阅片要点
- 自动分割可大幅减少手动勾画ROI的工作量，适合影像组学流程中的初筛，但仍需临床医生审核修改（约6.5%需重大修改）。
- 肿瘤靠近胸壁时分割准确性下降，阅片和审核时应重点关注胸壁粘连区域。
- 影像组学特征的可重复性有限，在报告影像组学结果时应谨慎解读，尤其涉及扰动或重采样后的特征稳定性。
局限性
- 单中心研究，样本量153例，外部泛化性未验证。
- 多数影像组学特征不可重复，传统影像组学建模的稳健性受限。
- 自动分割在6.5%病例中漏分割、6.5%需重大修改，仍需人工审核。
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
AIM: CT-based radio-biomarkers could provide non-invasive insights into tumour biology to risk-stratify patients. One of the limitations is the laborious manual segmentation of regions-of-interest (ROI). We present a deep learning auto-segmentation pipeline for radiomic analysis.
MATERIALS AND METHODS: 153 patients with resected stage 2A-3B non-small cell lung cancer (NSCLC) had tumours segmented using nnU-Net with review by two clinicians. The nnU-Net was pretrained with anatomical priors in non-cancerous lungs and finetuned on NSCLCs. Three ROIs were segmented: intra-tumoural, peri-tumoural, and whole lung. A total of 1967 features were extracted using PyRadiomics. Feature reproducibility was tested using segmentation perturbations. Features were selected using minimum-redundancy-maximum-relevance with Random Forest-recursive feature elimination nested in 500 bootstraps.
RESULTS: Auto-segmentation time was ∼36 seconds/series. Mean volumetric and surface Dice-Sørensen coefficient (DSC) scores were 0.84 (±0.28), and 0.79 (±0.34) respectively. DSC scores were significantly correlated with tumour shape (sphericity, diameter) and location (with worse performance associated with chest wall adherence), but not with batch effects (e.g. reconstruction kernel). Overall, 6.5% of cases had 'missed' segmentations; 6.5% required major changes. Pre-training on anatomical priors resulted in better segmentations compared to training on tumour-labels alone (p<0.001) and tumour plus anatomical labels (p<0.001). Most radiomic features were not reproducible following perturbations and resampling. Adding radiomic features did not significantly improve the clinical model in predicting 2-year disease-free survival: AUCs were 0.67 (95%CI 0.59-0.75) vs 0.63 (95%CI 0.54-0.71) respectively (p=0.28).
CONCLUSION: Our study demonstrates that integrating auto-segmentation into radio-biomarker discovery is feasible with high efficiency and accuracy. Whilst traditional radiomics show limited reproducibility, our auto-segmentation pipeline can facilitate future novel radio-biomarker development.
引用
- So ACP, Cheng D, Aslani S, et al. A deep learning lung cancer segmentation pipeline to facilitate CT-based radiomics. Clin Radiol. 2026. PMID: 42190286. doi:10.1016/j.crad.2026.107375
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。