Pancreatic cancer: Assessment of neoadjuvant chemotherapy outcome based on radiomics of pretreatment computed tomography.
e15767 Background: The objective response rate to neoadjuvant chemotherapy (NAC) was limited to around 35% in pancreatic cancer and as more as 30% patients show no benefit to NAC. In this instance, predicting the response to NAC may play an important role in individual treatment for pancreatic cancer patients. We aim to evaluate contrast enhanced-computed tomography (CE-CT) features in predicting treatment response and survival after neoadjuvant chemotherapy (NAC) for patients with borderline resectable and locally advanced pancreatic cancer. Methods: Sixty-one pancreatic cancer patients receiving NAC were enrolled and underwent abdominal CE-CT before treatment. All patients were divided into groups according to the changes of tumor size after treatment. 396 radiomics features were extracted from three-dimensional ROIs (region of interest) based on pretreatment CE-CT images of each patient. The optimal features were selected and three supervised machine learning classifiers were developed. Finally, univariate and multivariate analyses were performed to evaluate the capability of the selected features in predicting histopathologic response and outcomes. Results: Nine, seven and five radiomics features were selected as optimal features for three experiments respectively. Two features, Haralick Entropy and Histogram Entropy, were found consistent in experiments and were both higher in patients with tumor enlargement. Moreover, lower Histogram Entropy was significantly associated with a better histopathologic response (p = 0.008) and smaller tumor size (p = 0.041) in patients with tumor resection. In univariate Cox regression analysis, lower Histogram Entropy (P = 0.006) and lower Haralick Entropy (P = 0.001) predicted a better prognosis. Meanwhile, lower Haralick Entropy (p = 0.048) was independent predictor for longer survival time in multivariate Cox regression analysis. Conclusions: Radiomics features are strongly correlated with NAC treatment response and prognosis in pancreatic cancer, suggesting the great potential of imaging radiomics to help tailoring the treatment into the era of personalized medicine