Systematic multi-energy DECT radiomics integrating body composition and host inflammatory-metabolic environment for preoperative prediction of early recurrence in PDAC
Background Early recurrence (ER) after resection occurs in up to 80% of patients with pancreatic ductal adenocarcinoma (PDAC) yet remains difficult to predict preoperatively. We developed and externally validated a dual-energy CT (DECT)-based model integrating tumor radiomics and body composition to predict ER. Methods This retrospective bicentric study included 175 patients (training: n = 128; external validation: n = 47). Radiomic features were extracted from 40-keV, 70-keV, iodine map, and fusion series; features were selected using LASSO. Body composition (fat and skeletal muscle) was quantified at L3. A support vector machine with nested cross-validation was used for model development. The final model was frozen and applied to the external cohort without re-fitting. Performance was assessed using the area under the ROC curve (AUC), calibration (Brier score), and reclassification metrics. Results The optimal integrated model (70-keV tumor radiomics + body composition) achieved AUCs of 0.930 (training), 0.745 (cross-validation; 95% CI 0.654–0.848), and 0.836 (external validation; 95% CI 0.682–0.952). The external Brier score was 0.142. The model improved risk reclassification over the clinical baseline model (NRI = 0.31, 95% CI 0.05–0.57; IDI = 0.21, 95% CI 0.13–0.28). Skeletal muscle area contributed more to discrimination than fat-related metrics. Patients with ER had significantly lower muscle and fat indices ( p = 0.0052 and p = 0.0139). Negative correlations were found between subcutaneous fat and ALP ( p = 0.0023), skeletal muscle and ALP ( p = 0.0030). Conclusions A systematic multi-energy DECT model integrating 70-keV radiomics and body composition predicts early recurrence in PDAC with good external discrimination (AUC = 0.836), although external calibration was miscalibrated (slope 3.67) and requires recalibration. Skeletal muscle area emerges as a key imaging biomarker, providing a non-invasive tool for personalized preoperative risk stratification.
Authors
- Yongming Dai (ORCID: https://orcid.org/0000-0002-6084-9685)
- Hainan Ren (ORCID: https://orcid.org/0000-0002-3980-2807)
- Jieyu Yu (ORCID: https://orcid.org/0000-0002-8701-7440)
- Nandong Hu
- Shiteng Suo (ORCID: https://orcid.org/0000-0003-2528-2836)
- Yan Zhou (ORCID: https://orcid.org/0000-0001-9402-1109)
- Lijun Qian (ORCID: https://orcid.org/0000-0003-2329-8113)
- Chengwei Shao
- Zhenhui Xie (ORCID: https://orcid.org/0000-0002-5796-0876)
- Yuanzheng Wang (ORCID: https://orcid.org/0009-0008-9401-4011)
- Yun Bian
Institutions
- Shanghai Jiao Tong University (CN)
- Renji Hospital (CN)
- ShanghaiTech University (CN)
- Changhai Hospital (CN)
Publication Details
- Journal
- BMC Medical Informatics and Decision Making
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1186/s12911-026-03992-2
- Primary Topic
- Radiomics and Machine Learning in Medical Imaging
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Shanghai Municipal Health Commission
- National Natural Science Foundation of China
- Science and Technology Commission of Shanghai Municipality