Theis Maike, Hong Wei, Lee Belinda, Nowak Sebastian, Luetkens Julian, Stuckey Stephen, Gibbs Peter, Thomson Benjamin, Michael Michael, Sprinkart Alois Martin, Ko Hyun Soo
Department of Diagnostic and Interventional Radiology, University Hospital Bonn, Bonn, Germany.
Personalised Oncology Division, The Walter and Eliza Hall Institute of Medical Research, Melbourne, VIC, Australia; Department of Cancer Imaging, Peter MacCallum Cancer Centre, Melbourne, Australia.
Surg Oncol. 2025 Aug;61:102251. doi: 10.1016/j.suronc.2025.102251. Epub 2025 Jun 20.
Utilizing a fully automated AI-generated body composition analysis (BCA) from PDAC staging computed tomography (CT) imaging to discover predictive imaging biomarkers for overall survival (OS).
Routine PDAC staging CTs (07/2012-12/2020) and clinicopathological data (Eastern Cooperative Oncology Group (ECOG) performance status, resection status, chemotherapy, age, CA19-9, Charlson Comorbidity Index, BMI) from four tertiary centers were collected retrospectively. Using a 3:1 split (training:holdout), we fitted Cox regression OS using every possible combination of 7 clinicopathological and 9 BCA variables: skeletal muscle index (SMI), area and density of total muscle compartment (TMC), skeletal muscle (SM), subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT) and selected the combination with the lowest information complexity (ICOMP). The added value of BCA was calculated by comparing the BCA model with the base model (without BCA variables).
Analysis included 472 PDAC patients (213 female, mean age 67.9 ± 11.5 years, resectable n = 170, unresectable n = 106, metastatic n = 196). Four clinicopathological (ECOG, resection status, chemotherapy, CA19-9) and 5 BCA variables (SMI, SM density, VAT density, TMC area, VAT area) were selected. Decreased SM density (myosteatosis) and increased VAT density showed strong association with OS (p = 0.0094 and 0.0019, respectively). The BCA model showed superior performance compared to the base model in all subgroups (AUC: resectable 0.76 versus 0.70, unresectable 0.76 versus 0.69, and metastatic 0.80 versus 0.75).
BCA-identified myosteatosis and increased VAT density to be predictive imaging biomarkers for OS in all PDAC subgroups, potentially adding value to upfront risk stratification.
利用来自胰腺癌（PDAC）分期计算机断层扫描（CT）成像的全自动人工智能生成的身体成分分析（BCA），发现总生存期（OS）的预测性影像生物标志物。
回顾性收集了四个三级中心的常规PDAC分期CT（2012年7月至2020年12月）及临床病理数据（东部肿瘤协作组（ECOG）体能状态、切除状态、化疗、年龄、CA19-9、Charlson合并症指数、体重指数）。采用3:1分割（训练集：验证集），我们使用7个临床病理变量和9个BCA变量（骨骼肌指数（SMI）、总肌肉腔室（TMC）的面积和密度、骨骼肌（SM）、皮下脂肪组织（SAT）、内脏脂肪组织（VAT））的每种可能组合拟合Cox回归OS模型，并选择信息复杂度（ICOMP）最低的组合。通过将BCA模型与基础模型（不包含BCA变量）进行比较，计算BCA的附加值。
分析纳入472例PDAC患者（女性213例，平均年龄67.9±11.5岁，可切除170例，不可切除106例，转移196例）。选择了4个临床病理变量（ECOG、切除状态、化疗、CA19-9）和5个BCA变量（SMI、SM密度、VAT密度、TMC面积、VAT面积）。SM密度降低（肌肉减少症）和VAT密度增加与OS密切相关（p分别为0.0094和0.0019）。在所有亚组中，BCA模型均显示出优于基础模型的性能（AUC：可切除组0.76对0.70，不可切除组0.76对0.69，转移组0.80对0.75）。
BCA识别出的肌肉减少症和VAT密度增加是所有PDAC亚组中OS的预测性影像生物标志物，可能为 upfront 风险分层增加价值。
