Predicting the Response to Chemoradiotherapy in Rectal Cancer Patients Using Bayesian Evolutionary Random Forest and Three-Dimensional Discrete Fourier Transform

Rectal cancer remains a very deadly disease that often causes discomfort and decreases patients’ quality of life due to invasive surgeries. Therefore, it is crucial to develop a prediction method that can predict the tumor regression grade in advance, allowing us to tailor surgeries to the specific needs of each patient. In this study, we extracted quantitative data from planning CT images taken before the treatment and used them to predict the regression grade of rectal cancer after treatment. By making predictions in advance, a “wait-and-see” approach can be used for some patients, preserving their quality of life. We used the Discrete Fourier Transform to extract quantitative data from the images and created an Evolutionary Random Forest with this data. Additionally, we incorporated the prior distribution of the different regression grade groups obtained from our previous study into the Random Forest of this study. Our training results showed a normalized accuracy of 90.008%, with a total normalized accuracy of 74.968% for the Leave-One-Out cross-validation when accounting for the estimated priors. A Random Forest created without prior information yielded an unrealistic perfect classification of the training data and 71.483% in the Leave-One-Out cross-validation. The Random Forest with prior distribution information showed good results for both training and validation. However, without the prior distribution, the results were unrealistic as the regression grade has inherent variability.