Deep Texture Analysis Enhanced MRI Radiomics for Predicting Head and Neck Cancer Treatment Outcomes with Machine Learning Classifiers

Abstract

Background: Head and neck cancer treatment does not yield desired outcomes for all patients. This investigation aimed to explore the feasibility of predicting treatment outcomes from routine pre-treatment magnetic resonance images (MRIs). Radiomics features were “mined” and used to train machine learning (ML) classifiers to predict treatment outcomes. Moreover, iterative deep texture analysis (DTA) was explored to boost model performances. Methods: Radiomics features were determined from T1-weighted post-contrast MRIs of pathologically involved lymph node (LN) segmentations for n = 63 patients. SVM, k-NN, and FLD classifier models were trained, selecting for 1–10 features. The model with the top balanced accuracy was chosen for an iteration of DTA. New feature sets were used to retrain and test the ML. Radiomics features were explored for a total of three layers through two iterations of DTA. Results: Models proved useful in predicting treatment outcomes. The best model was a nine-feature multivariable k-NN model with a sensitivity (%Sn) of 93%, specificity (%Sp) of 74%, 86% accuracy (%Acc), and 86% precision (%Per). The best model for two of the three classifiers (k-NN and FLD) was trained using features from three layers. The performance of the average k-NN and FLD models trained with features was boosted significantly with the inclusion of deeper-layer features. Conclusions: Pre-treatment LN MRIs contain quantifiable texture information that can be used to train ML models to predict cancer treatment outcomes. Furthermore, DTA proved useful to boosting predictive models.