Room: Exhibit Hall | Forum 2
Purpose: To develop outcome prediction models by combining clinical data and radiomic features of the primary tumor for pancreatic cancer.
Methods: Seventy-eight pancreatic cancer patients previously receiving SBRT and chemotherapy were used for the study. The primary tumor was consistently contoured using contrast-enhanced CT and MRI images. Ninety-four CT-based radiomic features including image intensity, shape and texture were extracted from each tumor. The prognostic performances of the radiomic features for local recurrence, regional recurrence and overall survival were evaluated using the concordance index. We systematically studied the associations between the outcomes and the most prognostic radiomic features and other clinical parameters including TNM staging, SBRT prescription, chemotherapy, location of the tumor, age and sex of the patient etc. For model development, regression modeling, multinomial discriminant analysis, and machine learning (support vector machine, artificial neural networks, and convolution neural networks) were tested to classify the outcomes and correlate them with the features. Feature selection methods (LASSO with cox model, stepwise feature selection using AIC, and ANNOVA/MANOVA) were applied to identify the optimal feature set among all radiomic and clinical features. Model performances were evaluated using cross validation for a systematical comparison.
Results: Some radiomic features that described tumor shape and heterogeneity were found to be associated with prognostic performances. Both the regression and classification models including radiomic features outperformed those without them in predicting the outcomes. Regression model with stepwise feature selection and multinomial discriminant analysis with ANNOVA had better performance than other models.
Conclusion: By combining radiomic and clinical features, advanced statistical/computational models can be developed for pancreatic cancer prognosis prediction. Suitable feature selection methods and statistical models are important. Despite the preliminary success, the statistical variance is not ignorable. Increasing the sample size and including additional features, such as molecular-level information, may give rise to more powerful predictive models.