Radiomics for the prediction of the regression grade of colorectal cancer: a challenging classification problem

Prediction of the response of a cancer is a crucial research topic in the field of oncology. Over the years many different prediction methods were introduced where both non-ensemble and ensemble methods were proposed for the response prediction. These (non-)ensemble methods, although functioning well for numerous problems, are not always suitable. When considering a more challenging data set, they can fail for either both training and validation or validation alone. Since our interest lies in the use of the model for predicting the response of new patients, we needed to find a method that works good for both training and validation. We were interested in predicting the response grade of patients with colorectal cancer by examining their CT scans. From the medical images, we extracted radiomic features such that we obtained a quantitative features design. We predicted the Dworak response grade of the patients using the radiomics data. Our study showed that the classical prediction methods fail for this type of data. This is due to various problems we are facing in both the response parameter as well as the feature matrix. Even though radiomics are very promising, we showed that they are not yet ready to be given to doctors solely as a definitive prediction tool. We will need to construct a prediction method that is capable of learning from the radiomics data and adapting its algorithm to suit the characteristics of the data better.