Histogram analysis of magnetic resonance images: evaluation of intra-tumoral heterogeneity and correlation with pathological findings in solid pancreatic tumors.

Objectives To evaluate magnetic resonance (MR)-derived whole-tumor histogram analysis parameters in predicting aggressiveness of pancreatic ductal adenocarcinomas (PDACs) and neuroendocrine neoplasms (panNENs). Methods Pre-operative MR of 169 consecutive patients with PDAC or panNEN were retrospectively analyzed. T1-/T2-weighted images and apparent diffusion coefficient (ADC) maps were analyzed. Histogram-derived parameters were compared to several pathological features (grade, vascular infiltration, nodal and hepatic metastases) using Mann-Whitney U test. Diagnostic accuracy was assessed by receiver operating characteristic area under curve (ROC-AUC) analysis; sensitivity and specificity were assessed for each histogram parameter. Results No significant differences were found among histogram parameters for prediction of PDACs grade. ADCentropy was significantly higher in G2-3 panNENs with ROC-AUC 0.757; sensitivity was 83.3%. ADCentropy was significantly higher in PDACs with vascular involvement (p=.022; AUC=.641), with specificity of 92.2%. ADCskewness was significantly higher in PDACs with nodal metastases (p=.027; AUC=.642), with 72% specificity. ADCkurtosis was higher in panNENs with vascular involvement, nodal and hepatic metastases (p= .008, .021, and .008; ROC-AUC= 0.820, 0.709, and 0.820); sensitivity and specificity were: 85.7/74.3%; 36.8/96.5%; and 100/62.8%. No significant differences between groups were found for other histogram-derived parameters (p >.05). Conclusions Whole-tumors histogram analysis of ADC values is a valuable tool for predicting aggressiveness of PDACs and panNENs. Our results indicate that histogram metrics related to intra-tumor heterogeneity, as ADCentropy, ADCkurtosis and ADCskewness are the most accurate parameters for the identification of PDACs and panNENs with higher biological aggressiveness. Further and larger studies are needed to incorporate the results of the histogram analysis within decision support models and to mine these data to detect possible correlations with genomic patterns.