Abstract A086: PancMAP: Pancreatic Morphology Analysis Pipeline for Early Detection of Pancreatic Ductal Adenocarcinoma

Abstract

Before pancreatic ductal adenocarcinoma (PDAC) becomes clinically apparent, diagnostic imaging may already show subtle pancreatic changes. These changes often include thinning of pancreas such as diffuse or focal atrophy, main pancreatic duct dilation, and ductal stenosis. Implementation of these features into clinical practice remains limited due to standardization (i.e., where and how to measure?) and reproducibility (i.e., how to measure and interpret consistently?). To overcome this challenge, we developed the Pancreatic Morphology Analysis Pipeline (PancMAP), a novel CT-based curved multiplanar reconstruction algorithm to capture early disease-associated changes in pancreatic morphology. PancMAP converts annotations of the pancreas and its internal structures, including main pancreatic duct and parenchyma, into normalized profiles along the medial axis of the pancreas. Then, it extracts 164 global or regional morphological features of the profiles using minimum, maximum, and derivative operations, as well as their products, to reflect spatial relationships among pancreatic structures. In this study, we used PancMAP to develop and validate a model for early detection of PDAC. We retrospectively identified prediagnostic contrast-enhanced CT scans of 41 patients (median age, 69.1±8.1 y.o.; 65.9% male) 6–36 months before diagnosis and 104 healthy controls (median age, 59±9.4 y.o.; 53.8% male) who did not develop PDAC within 60 months of follow-up. We applied a 5-fold nested cross-validation strategy to optimize hyperparameters, including the input feature set, while preserving model generalizability to unseen data. The final multivariable logistic regression was evaluated using sensitivity, specificity, accuracy, area under the curve (AUC), and F1 score to mitigate the imbalance between case and control. We also assessed how the final model could improve our current clinical practice using decision curve analysis. As a result, the final model selected 6 variables. Permutation importance and partial dependence curves showed that the predicted probability of PDAC was greater when PancMAP identified the combination of a larger main duct diameter, larger pancreatic head volume, and smaller total parenchymal volume. The final model showed balanced performance between false negatives and false positives, with a sensitivity of 0.90±0.09, a specificity of 0.81±0.09, an accuracy of 0.83±0.05, an AUC of 0.92±0.07, and an F1 score of 0.76±0.06. Decision curve analysis showed that, at a threshold of 0.28 corresponding to the study cohort prevalence, PancMAP may help avoid 48 unnecessary additional interventions per 100 patients while missing fewer than one cancer per 100 patients. These results indicate the potential of a standardized, quantitative method to characterize the pancreas on standard imaging for early detection of PDAC. Ongoing work includes blinded external validation of PancMAP, extension to magnetic resonance imaging, and investigation of its performance when combined with state-of-the-art blood-based biomarkers for early detection of PDAC.