Image Analysis Enables Quantitation of Metastatic Risk in Pancreatic Intraductal Papillary Mucinous Neoplasms

Intraductal papillary mucinous neoplasms (IPMNs) are pre‐malignant neoplasms of the main and branch pancreatic ducts with a risk of malignancy that is dependent on the underlying grade of dysplasia. They are evaluated by noninvasive cross‐sectional imaging, based on subjective and objective imaging criteria including size, pancreatic duct size and mural nodularity. Surgical resection is recommended for IPMNs with high grade dysplasia or invasive cancer, as determined by high risk imaging features on cross sectional or endoscopic imaging. However, some of these criteria are subjective and without high accuracy. Additionally, due to the concern for progression to adenocarcinoma, “over‐treatment” of these lesions remains a major problem, since a large percentage of patients will not progress to invasive cancer. Thus, there is a critical need to develop more rigorous diagnostic criteria to distinguish the risk of high‐grade dysplasia or invasive IPMNs via imaging. To address this need, we developed a novel, semi‐automated approach to IPMN analysis from imaging data and used this approach to build a multivariate statistical model of risk classification. Briefly, IPMNs were manually segmented from patient computed tomography scans, and automated particle detection and measurement was performed to evaluate tumor volume and density. Slice‐wise image descriptors including heterogeneity in cross‐sectional area, eccentricity (major/minor axis ratio), density, and solidity (cross‐sectional area/convex hull) were also measured. We hypothesized that eccentricity and density heterogeneity, defined as the standard deviation of slice‐wise tumor density, would be significantly correlated with risk of metastasis and thus patient outcomes. Interestingly, we found that IPMN volume and eccentricity were not associated with patient outcomes, while low solidity and density heterogeneity were both significantly associated with patient mortality (p<0.05 by Fisher’s exact test). Additionally, we performed k‐means clustering, Ward hierarchical clustering, and principal component analysis to evaluate the efficacy of multivariate IPMN classification with respect to patient metastatic risk. We found that using both tumor solidity and density heterogeneity, all three clustering methodologies were concordant and naïve clustering separated IPMNs into two groups consistent with patient mortality. These results indicate the utility of our novel image processing approach, and suggest that similar strategies may be employed to rigorously evaluate IPMN metastatic risk in imaging data, and thus minimize potential over‐treatment. Support or Funding Information NIH K25HL136869 and University of South Alabama College of Medicine Internal Grant Program Intraductal papillary mucinous neoplasm treatment algorithms. A. Current treatment flow chart. B. Proposed treatment flow chart including new image‐based metastatic risk metrics. Figure 1