Predicting treatment outcome of colorectal cancer from MRI images using machine learning

Abstract

In this thesis different machine learning algorithms have been utilised to predict treatment outcome for patients with colorectal cancer. The predicted treatment endpoint was overall survival. The patient cohort included 77 patients with histologically confirmed colorectal cancer who were recruited at Akershus University Hospital between 2013 and 2017. Radiomics was used to extract first-order statistics, shape and texture features from T2-weighed images and DWIs taken of the patients before starting treatment. These features, in addition to clinical data, were used to train machine learning models. The models were later combined into majority vote classifiers. Models and majority vote classifiers were built for three different patient subsets: all patients, patients who had received chemoradiotherapy, and patients who had not received chemoradiotherapy. Performance was estimated using k-fold cross validation with MCC as the the validation metric.

Repeated Elastic Net Technique (RENT) and PCA were used for feature reduction before training models and building the majority vote classifiers. RENT was also used to analyse feature importance for the radiomics data.

All the majority vote classifiers achieved mean MCC scores above 0, but had quite large mean standard deviations. The differences in the performances of the models between folds in the k-fold cross validation were severe, indicating that the data was susceptible to poor train-test splits. A handful of features with high selection frequency were singled out during the RENT analysis of the radiomics data.