卵巢癌FOS免疫组化状态的CT影像组学与病理组学模型
本研究基于两个中心的289例卵巢癌患者,以训练队列FOS H-score中位数15为截断值,构建并验证了术前CT影像组学模型及术后病理组学模型。CT影像组学模型的内部和外部验证AUC分别为0.827和0.794,联合模型的AUC数值上更高但未建立增量判别价值。
要点
- 静脉期CT影像组学模型可无创预估FOS免疫组化状态,外部验证AUC为0.794。
- 影像组学特征选择后仅保留2个,模型简洁但与FOS H-score显著相关。
- 联合临床、影像和病理组学的模型数值上优于单纯CT模型,但缺乏推断性统计验证。
核心结论
- CT影像组学模型提供了卵巢癌FOS免疫组化状态的术前无创预估,外部验证AUC为0.794。
- 联合临床、影像组学和病理组学模型的外部验证AUC为0.871,数值上高于单纯CT模型,但未建立增量判别价值。
研究设计与样本
- 回顾性两中心研究,共纳入289例卵巢癌患者,分为训练集151例、内部验证集38例、外部验证集100例。
- 以训练队列FOS H-score中位数15为截断值,将患者分为H-score大于15和小于等于15两组。
- 从静脉期CT提取1834个影像组学特征,从H&E全玻片图像提取682个病理组学特征。
- 特征筛选采用可重复性过滤、冗余过滤、组间差异检验和五折交叉验证LASSO,最终影像组学保留2个特征,病理组学保留23个特征。
- 采用XGBoost构建分类模型,参考标准为FOS免疫组化H-score。
关键影像表现/征象
- 最终影像组学模型仅包含2个静脉期CT特征,特征数量极少但预测性能中等。
- Rad-score与连续FOS H-score显著相关(P < 0.001),提示影像特征可反映FOS表达水平。
- 高Rad-score与较差总生存期相关(P = 0.044),提示影像组学特征具有预后价值。
诊断标准与关键数值
- 二分类终点为FOS H-score大于15对比小于等于15,截断值锁定为训练队列中位数。
- CT影像组学模型外部验证AUC为0.794,临床模型为0.827,病理组学模型为0.842,联合模型为0.871。
- 联合模型的准确度为0.806,敏感度为0.846,特异度为0.750,F1分数为0.835。
- Rad-score与FOS H-score显著相关(P < 0.001),Path-score同样显著相关(P < 0.001)。
临床应用与阅片要点
- 术前静脉期CT影像组学模型可作为FOS免疫组化状态的无创预估工具,辅助术前决策。
- 仅2个影像组学特征即可达到外部验证AUC为0.794,提示模型简洁且具有泛化潜力。
- 组织依赖的病理组学模型无法替代直接FOS免疫组化检测,临床应用时应明确其辅助定位。
- 联合模型数值上优于CT模型,但缺乏推断性统计支持,不宜将其解释为确有增量受益。
局限性
- 联合模型与CT模型的增量比较因未保留患者级预测向量而无法进行推断性检验。
- 研究为回顾性设计,存在选择偏倚和混杂因素控制不足的风险。
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
RATIONALE AND OBJECTIVES: The primary objective was to develop and externally validate a CT-only model for preoperative classification of FOS immunohistochemical status dichotomized at the training-cohort median H-score. Secondary analyses evaluated tissue-derived H&E pathomics and a combined model after tissue acquisition. Survival stratification was exploratory.
MATERIALS AND METHODS: This retrospective two-center study included 289 patients in training (n = 151), internal validation (n = 38), and external validation (n = 100) cohorts. The binary endpoint was an H-score greater than 15 versus 15 or less; the training-median cutoff was locked for validation. Venous-phase CT and H&E whole-slide images yielded 1834 radiomic and 682 pathomic features. Feature reduction used reproducibility and redundancy filters, group-difference testing, and five-fold cross-validated LASSO. XGBoost modeled relative FOS status.
RESULTS: The final feature sets contained two radiomic and 23 pathomic features. External AUCs were 0.794, 0.827, 0.842, and 0.871 for clinical, radiomic, pathomic, and combined models. The combined-model AUC was numerically higher than the CT-only AUC. Inferential comparison was unavailable because aligned patient-level prediction vectors were not retained. Combined-model accuracy, sensitivity, specificity, and F1-score were 0.806, 0.846, 0.750, and 0.835. Rad-score and Path-score were associated with continuous FOS H-score (both P < 0.001) and worse OS (P = 0.044 and P = 0.012).
CONCLUSION: The primary CT-only model provided a noninvasive preoperative estimate of the data-derived FOS endpoint. Tissue-dependent models did not replace direct FOS immunohistochemistry. The combined model's numerically higher AUC did not establish incremental discrimination.
引用
- Qin, Lu, Liang, et al. Clinical, CT Radiomic, and Pathomic Models for FOS Immunohistochemical Status Defined by the Training-Cohort Median in Ovarian Cancer. Acad Radiol. 2026. PMID: 42731902. doi:10.1016/j.acra.2026.08.110
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。