Preoperative prediction for early recurrence in patients with pancreatic ductal adenocarcinoma: combining radiomics and abdominal fat analysis

Abstract

Background: The role of radiomics and abdominal fat analysis in the survival prediction of pancreatic ductal adenocarcinoma (PDAC) has attracted attention. This study aims to develop a preoperative model for predicting early recurrence (ER) in patients pathologically confirmed PDAC, combining radiomic and abdominal fat analysis.

Methods: A total of 177 patients (Hospital A) were retrospectively analyzed and allocated to the training cohort (n = 124) and internal validation cohort (n = 53). Another 71 patients (Hospital B) group formed the geographic external validation cohort. The threshold of ER was set at 6 months after surgery, and the primary endpoint was to determine the best model to predict ER of PDAC patients. A radiomics model for predicting ER was constructed by the least absolute shrinkage and selection operator Cox regression. Univariate and multivariate Cox regression analyses were used to build a combined model based on radiomics, fat quantitation, and clinical features. The combined model's performance was assessed using the Harrell concordance index (C-index). Based on the nomogram score, patients were stratified into high-risk and low-risk groups, and survival analysis of different risk groups was performed using the Kaplan-Meier (KM) method. All patients were divided into four subgroups according to recurrence patterns: local recurrence subgroup, distant recurrence subgroup, "local + distant" recurrence subgroup, and "multiple" recurrence subgroup. The predictive efficacy of the combined model was calculated in different subgroups.

Results: Radiomics scores (P < 0.001), CA19-9 (P = 0.009), and visceral to subcutaneous fat volume ratio(P = 0.009) were selected for the combined model. Compared to clinical and radiomics models, the combined model exhibited the best prediction performance. C indexes of the training cohort, internal validation cohort, and external validation cohort were 0.778 (0.711,0.845), 0.746 (0.632,0.860), and 0.712 (0.612,0.812) respectively, showing the improvement over the clinical model (without radiomics and fat quantitation features) in the internal validation and external validation sets (DeLong test: P = 0.027, P = 0.079). KM analysis showed significant differences between risk groups (all P < 0.05). The combined model also achieved robust performance in different subgroups of recurrence patterns.

Conclusion: The combined model effectively predicted the probability of ER in PDAC patients and may provide an emerging tool to preoperatively guide personalized treatment.

Clinical trial number: Not applicable.

Keywords: Abdominal fat; Computed tomography; Pancreatic cancer; Radiomics; Recurrence.

Fig. 1

Flowchart showing the patient inclusion criteria and exclusion criteria (PDAC = pancreatic ductal adenocarcinoma, ER = early recurrence, NER = nonearly recurrence)

Fig. 2

Overall workflow of our study

Fig. 3

Nomogram for preoperatively predicting early recurrence of PDAC patients (Rad-score = Radiomics score, VSR = Visceral to subcutaneous fat volume ratio, RFS = recurrence-free survival)

Fig. 4

Model assessment and comparison. a–c Time-dependent ROC curves comparing all different models for early recurrence(ER) prediction on the training, internal validation, and external validation sets. d–f Decision curve analysis of combined, clinical, and radiomics models for ER prediction on the training, internal validation, and external validation sets. g–i Kaplan-Meier curves of the combined model on the training, internal validation, and external validation sets

Fig. 5

Kaplan-Meier analysis in subgroups of radiomics score (Rad-score), visceral to subcutaneous fat volume ratio (VSR), and CA19-9. a KM analysis and comparison between high Rad-score group and low Rad-score group. b KM analysis and comparison between high VSR group and low VSR group. c KM analysis and comparison between high CA 19–9 group and low CA 19–9 group