Searching for biomarkers of disease-free survival in head and neck cancers using PET/CT radiomics

Abstract

The goals of this thesis were to (1) study methodologies for radiomics data analysis, and (2) apply such methods to identify biomarkers of disease-free survival in head and neck cancers.

Procedures for radiomics feature extraction and feature exploration in biomarker discovery were implemented with the Python(TM) programming language. The code is available at https://github.com/gsel9/biorad.

In a retrospective study of disease-free survival as response to radiotherapy, radiomics features were extracted from PET/CT images of 198 head and neck cancers patients. A total of 513 features were obtained by combining the radiomics features with clinical factors and PET parameters.

Combinations of seven feature selection and 10 classification algorithms were evaluated in terms of their ability to predict patient treatment response. By using a combination of MultiSURF feature selection and Extreme Gradient Boosting classification, subgroup analyses of HPV negative oropharyngeal (HPV unrelated) cancers gave 76.4 +/- 13.2 % area under the Receiver Operating Characteristic curve (AUC). This performance was superior to the baseline of 54 \% for disease-free survival outcomes in the patient subgroup.

Four features were identified as prognostic of disease-free survival in the HPV unrelated cohort. Among these were two CT features capturing intratumour heterogeneity. Another feature described tumour shape and was, contrary to the CT features, significantly correlated with the tumour volume. The fourth feature was the median CT intensity. Determining the prognostic value of these features in an independent cohort will elucidate the relevance of tumour volume and intratumour heterogeneity in treatment of HPV unrelated head and neck cancer.