PO.BCS02.04 · 生物信息与计算
CT影像组学在gBRCA1/2和PALB2突变的晚期胰腺癌患者中的预后价值
Prognostic value of CT radiomics in patients with gBRCA1/2 and PALB2 and advanced pancreas cancer
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
引言：纪念斯隆-凯特琳癌症中心（MSK）的一项II期试验（IRB #12-045）评估了顺铂/吉西他滨加或不加PARP抑制剂veliparib治疗gBRCA1/2和PALB2（核心同源修复缺陷；cHRD）突变的晚期胰腺癌（PDAC）患者。尽管该研究发现各治疗组之间的总生存期（OS）无显著差异，但接受veliparib治疗的患者具有更高的疾病稳定（SD）或部分缓解（PR）率，提示该人群内存在异质性的治疗获益（RR 74.1% vs. 65.2%；P = 0.55）。本研究利用该试验队列，旨在识别可预测OS的CT影像组学特征，并评估影像组学特征随时间的变化是否提供额外的预后价值。
方法：我们回顾性分析了入组该试验的cHRD和PDAC患者的对比增强CT扫描。一名训练有素的机器学习专家使用自动化胰腺分割工作流处理了基线、6周和3个月的扫描，并从每次扫描中提取了134个影像组学特征。去除了高度相关的特征（r > 0.9）。我们采用单变量Cox比例风险方法识别与OS相关的特征。将显著的前两个特征整合到多变量Cox模型中，使用重复3折交叉验证考察其预测能力以避免过拟合。Kaplan-Meier分析比较了影像组学定义的高风险组和低风险组之间的OS。
结果：在MSK入组的26例患者中，23例（88%）有基线（治疗前）CT扫描，25例有6周扫描，24例（92%）有3个月扫描。25例患者接受了顺铂和吉西他滨，其中16例还接受了veliparib。放射科医师评估的治疗反应为PR（n = 21）、SD（n = 2）和疾病进展（n = 3）。多个影像组学特征在所有三个时间点均与OS显著相关，选出前2个特征用于最佳预后分层。基线胰腺中较低的灰度非均匀性与较差的OS相关，表明胰腺纹理均质性增加与较差的结局相关。Delta影像组学分析表明胰腺纹理的纵向变化也与OS相关。
结论：CT影像组学特征在cHRD和晚期PDAC患者中显示出对总生存期的预后价值，支持影像组学作为未来试验中患者分层的一种有前景的工具。目前正在开展的工作评估影像组学特征与基因组图谱之间的相关性，以进一步完善预测模型。
查看英文原文 English abstract
Introduction: A phase II trial at Memorial Sloan Kettering Cancer Center (MSK), IRB #12-045 evaluated cisplatin/gemcitabine +/- PARP inhibitor veliparib in patients with gBRCA 1/2 and PALB 2 (core homologous repair deficiency; cHRD) and advanced pancreas cancer (PDAC). Although the study found no significant difference in overall survival (OS) between treatment arms, patients receiving veliparib had higher rates of stable disease (SD) or partial response (PR), suggesting heterogeneous treatment benefit within this population (RR 74.1% vs. 65.2%; P = 0.55). Leveraging this trial cohort, the present study aimed to identify CT radiomic signatures predictive of OS and to assess whether changes in radiomic features over time provide additional prognostic value.
Methods: We retrospectively analyzed contrast-enhanced CT scans of patients with cHRD and PDAC that were enrolled in the trial. A trained machine learning specialist processed baseline, 6-week, and 3-month scans using an automated pancreas segmentation workflow, and 134 radiomic features were extracted from each scan. Highly correlated features (r > 0.9) were removed. We performed univariate Cox proportional hazards method to identify features associated with OS. Significant top two features were integrated into multivariate Cox models to investigate their predictability using repeated 3-fold cross-validation to avoid overfitting. Kaplan-Meier analysis compared OS between radiomics-defined high- and low-risk groups.
Results: Among the 26 patients enrolled at MSK, baseline (pre-treatment) CT scans were available for 23 (88%), 6-week scans for 25, and 3-month scans for 24 (92%). Twenty-five patients received cisplatin and gemcitabine, and 16 also receiving veliparib. Radiologist-assessed treatment responses were PR (n = 21), SD (n = 2), and progressive disease (n = 3). Several radiomic features were significantly associated with OS across all three time points, with the top 2 features selected for optimal prognostic stratification. Lower gray-level non-uniformity in the baseline pancreas was associated with worse OS, indicating that increased pancreatic textural homogeneity correlates with poorer outcomes. Delta-radiomics analyses demonstrated that longitudinal changes in pancreatic texture also correlated with OS.
Conclusions: CT radiomic features demonstrate prognostic value for overall survival in patients with cHRD and advanced PDAC, supporting radiomics as a promising tool for patient stratification in future trials. Ongoing work is evaluating correlations between radiomic signatures and genomic profiles to further refine predictive models.
利益披露 Disclosure
S. Gandhi, None..
J. Chakraborty, None..
H. Ghahremannezhad, None..
N. Horvat, None.