胃型宫颈腺癌术前PET/CT表型:形态、代谢、血清学与可解释机器学习整合分析
该回顾性研究纳入144例宫颈癌患者(GAS 22例、SCC 82例、UEA 40例),分析治疗前18F-FDG PET/CT形态、代谢参数、血清肿瘤标志物及播散特征。GAS表现为弥漫浸润性生长、囊性形态、宫腔积液、FDG摄取相对较低、CA19-9阳性及更频繁的远处和腹膜转移,呈现代谢-形态不匹配表型;可解释机器学习模型有助于特征整合与鉴别诊断。
要点
- GAS原发灶FDG摄取相对较低,但侵袭性形态和转移播散更显著,呈代谢-形态不匹配。
- GAS常见弥漫浸润生长、囊性形态、宫腔积液、CA19-9阳性及远处/腹膜转移。
- 整合PET/CT形态、代谢、肿瘤标志物和播散模式可提高术前GAS怀疑。
核心结论
- GAS具有可识别的PET/CT表型,表现为原发肿瘤FDG摄取相对较低但侵袭性浸润形态和转移播散显著。
- 整合PET/CT形态、代谢参数、肿瘤标志物和播散模式有助于术前怀疑GAS并指导进一步病理检查。
- 可解释机器学习可作为特征整合的补充工具,但临床实施前需外部验证。
研究设计与样本
- 回顾性研究纳入144例组织学确诊宫颈癌患者,均行治疗前18F-FDG PET/CT。
- 队列包括GAS 22例、鳞状细胞癌(SCC)82例和普通型宫颈腺癌(UEA)40例。
- 收集临床特征、血清肿瘤标志物、PET/CT形态特征、代谢参数和播散相关变量,并建立5种机器学习模型进行组织学鉴别。
主要发现
- GAS表现为弥漫浸润性生长、囊性形态、宫腔积液、FDG摄取相对较低、CA19-9阳性和更频繁的远处及腹膜转移。
- GAS宫颈病灶SUVmax中位数低于SCC和UEA,肝脏归一化和血池归一化SUV比值呈相似趋势。
- 尽管代谢活性相对较低,GAS表现出更具侵袭性的播散相关特征。
- 树基集成模型在探索性鉴别性能上优于Logistic回归和多层感知机。
- SHAP分析显示生长模式、囊性形态、宫腔积液、宫颈病灶SUVmax、肝脏SUV比值、血池比值、CA19-9和SCC抗原本是模型预测的主要贡献特征。
关键数值与推荐
- GAS宫颈病灶SUVmax中位数低于SCC和UEA,肝脏归一化和血池归一化SUV比值呈相似趋势。
- 树基集成模型在探索性鉴别性能上优于Logistic回归和多层感知机。
- SHAP分析主要贡献特征包括生长模式、囊性形态、宫腔积液、宫颈病灶SUVmax、肝脏SUV比值、血池比值、CA19-9和SCC抗原。
临床怎么用
- 术前PET/CT发现宫颈病灶FDG摄取相对较低但呈弥漫浸润性生长、囊性形态或宫腔积液时,应警惕GAS可能。
- 结合CA19-9、SCC抗原及远处/腹膜转移等播散特征综合评估,可提高GAS术前怀疑并指导进一步病理检查。
- 可解释机器学习可作为特征整合的补充工具,但临床实施前需外部验证。
局限性
- 本研究为回顾性研究。
- GAS样本量为22例。
- 机器学习模型需外部验证后方可临床实施。
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)。
原文
PURPOSE: Gastric-type endocervical adenocarcinoma (GAS) is an aggressive, non-HPV-associated cervical adenocarcinoma that is often difficult to recognize preoperatively. This study aimed to characterize the integrated PET/CT phenotype of GAS and evaluate whether morphological, metabolic, serological, and explainable machine-learning features could support its differentiation from squamous cell carcinoma (SCC) and usual-type endocervical adenocarcinoma (UEA). MATERIALS AND METHODS: This retrospective study included 144 patients with histologically confirmed cervical cancer who underwent pretreatment 18 F-FDG PET/CT, including 22 with GAS, 82 with SCC, and 40 with UEA. Clinical characteristics, serum tumor markers, PET/CT-derived morphological features, metabolic parameters, and dissemination-related variables were collected. Intergroup differences were assessed using appropriate statistical tests. Five machine-learning models were developed for histological differentiation, and model performance was evaluated using ROC analysis, classification metrics, calibration assessment, and decision curve analysis. SHAP analysis was used to interpret feature contributions. RESULTS: GAS demonstrated a distinctive PET/CT phenotype characterized by diffuse infiltrative growth, cystic morphology, intrauterine fluid accumulation, relatively lower FDG uptake, CA19-9 positivity, and more frequent distant and peritoneal metastasis. The median cervical lesion SUVmax was lower in GAS than in SCC and UEA, and similar trends were observed for liver-normalized and blood pool-normalized SUV ratios. Despite its relatively low metabolic activity, GAS showed more aggressive dissemination-related features. Among the machine-learning models, tree-based ensemble models showed better exploratory discriminative performance than Logistic Regression and multilayer perceptron. SHAP analysis indicated that growth pattern, cystic morphology, intrauterine fluid, cervical lesion SUVmax, liver SUV ratio, blood pool ratio, CA19-9, and SCC antigen were the main contributors to model prediction. CONCLUSION: GAS exhibits a recognizable PET/CT phenotype characterized by a descriptive metabolic-morphological mismatch, namely relatively low primary-tumor FDG uptake despite aggressive infiltrative morphology and metastatic dissemination. Integrated assessment of PET/CT morphology, metabolic parameters, tumor markers, and dissemination patterns may help raise preoperative suspicion of GAS and guide further pathological work-up. Explainable machine learning may serve as a complementary tool for feature integration, but external validation is required before clinical implementation.
引用
- Wang Y, Su Y, Zhang Y, et al. Preoperative PET/CT phenotype of gastric-type endocervical adenocarcinoma: integrated morphological, metabolic, serological, and explainable machine-learning analysis. Eur J Nucl Med Mol Imaging. 2026. PMID: 42484891. doi:10.1007/s00259-026-08088-7
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。