Enhancing Outcome Prediction for Pancreatic Cancer Following Stereotactic Body Radiotherapy Through MRI Radiomics Analysis

Abstract

Listen

Translate article

Purpose: Pancreatic cancer is an extremely aggressive and deadly cancer with a 5-year survival rate of less than 10%. Our study aims to establish an MRI radiomics-based model to predict survival for borderline resectable and locally advanced pancreatic ductal adenocarcinoma patients who have received radiation therapy. Methods: 71 borderline resectable and locally advanced pancreatic cancer patients (42 Male, 29 Female) were retrospectively selected for radiomics analysis with a median age of 63 years. The gross tumor of each patient was delineated on contrast-enhanced T1-weighted MRI images. Radiomics features were extracted using PyRadiomics and feature stability of the radiomics features was assessed under MRI intensity normalization and bin width variation. The 71 patients were randomly split into a training set (54 patients) and a testing set (17 patients). Using the training set, we trained three risk stratification models (clinical-only, radiomics-only, and a composite) through a penalized Cox model, which simultaneously established the predictive model and selects important features by incorporating L1 and/or L2 penalties to the Cox Proportional Hazards model. We also built a Random Forest classifier using the Synthetic Minority Over-sampling Technique (SMOTE) with the same set features selected in the penalized cox model to predict the 1-year survival of these pancreatic cancer patients. Results: Out of 924 extracted features, we identified 133 (14.4%) stable features with ICC > 0.75, against both intensity normalization and bin width variations. Survival models based on clinical endpoints alone, radiomics features alone, and a combination showed that including radiomics features can significantly improve survival prediction. Using the same number of features to construct survival models for clinical only, radiomics only, and a combination of clinical and radiomics features, we find that we are able to accurately distinguish low and high-risk groups and generate survival curves for the test group with a concordance index of 0.615, 0.654, and 0.716, respectively. The Random Forest classifier predicted the 1-year survival accuracies of 0.529, 0.824, and 0.765 for the clinical-only model, radiomics-only model, and the composite model, respectively.Conclusions: Magnetic resonance imaging (MRI) radiomics is promising in predicting the mortality of pancreatic cancer following SBRT and improving survival prediction capabilities. Intensity normalization is an essential preprocessing step to exclude unstable and/or redundant imaging features.