Prognostic models for classifying rectal tumour response to therapy using radiomics and CNN features

Purpose or Learning Objective To retrospectively evaluate the best radiomic features in predicting complete response to neoadjuvant therapy in patients affected by rectal cancer and to assess the possible correlation between them. Methods or Background A total of 109 handcrafted radiomic features and 4096 deep radiomic features were extracted from pre-treatment 3D-MRI of 43 patients, using transfer learning from a pre-trained convolutional neural network. The most widely explored 7 supervised machine learning-based classifiers and 6 different feature selection algorithms were validated and compared utilising all possible radiomic features, in order to examine their effectiveness in achieving an accurate predictive model. Cross-validation was performed in 100 rounds partitioning the data as 75% for training and 25% for testing. Results or Findings Using only handcrafted radiomic features, artificial neural network classifier and Fisher as feature selection algorithm delivered the best predictive performance on test data sets with the area under the curves (AUCs) [mean±SD] of 0.79±0.016 and 0.8±0.01, respectively. The best prognostic performance, using only deep radiomic features, was achieved by linear support vector machine (LSVM) classifier and Relief-based feature selection algorithm as 0.8±0.042 and 0.82±0.04, respectively. When using a combination of both handcrafted and deep radiomic features, almost all classifiers in combination with every feature selection algorithm generated better AUC than that obtained individually; the best AUCs were generated by the LSVM classifier and Relief-based feature selection as 0.84±0.025 and 0.87±0.013, respectively. Conclusion The best predicting models’ performance was achieved by integrating both handcrafted and deep radiomic features, i.e. LSVM classifier and Relief-based feature selection algorithms in combination with all classifiers.