Investigation of the prognostic value of CT and PET-based radiomic image features in oropharyngeal squamous cell carcinoma

Sammanfattning: Background Medical data in the form of radiographic routine scans is steadily accumulating. The analysis of such data through automated quantitative methods is believed to produce new information which would allow for more personalization of therapy. The present thesis investigated the use of such methods in head and neck cancer. Material and methods Pretreatment positron emission tomography (PET) and computed tomography (CT) scans from 74 patients present with oropharyngeal squamous cell carcinoma were analyzed quantitatively and a total of 92 image-based features were calculated. These features attempt to describe the shape and size of the tumor, as well as the heterogeneity within. The prognostic value of these features, as well as common clinical variables, was investigated for tumor recurrence and disease-specific mortality, respectively. Additionally, prediction of treatment failure was attempted using an artificial neural network. All patients received intensity-modulated radiation therapy and there were thus treatment plans for each patient in addition to the PET/CT scans. The non-uniformity of the dose distribution was studied using custom features based on the gray-level size zone matrix. These custom features measured the number of disconnected regions receiving either too low or too high of radiation dose, and differences in the sizes of such regions. Results One PET- and two CT-based features were found to significantly differ between responders and non-responders. The PET-based feature was the correlation (p = 0.0011), which is a texture feature derived from the gray-level co-occurrence matrix. It described the irregularity in radiotracer uptake on a voxel-to-voxel basis and results suggest that non-responders have more irregular patterns of uptake. The CT-based features were the variance (p = 0.0012) and skewness (p = 0.0027), where the former was found to be significantly larger among responders and the skewness more negative. However, image-based features performed quite poorly in treatment failure prediction, as compared to clinical variables, which had an area under the receiver operating characteristic curve (AUC-ROC) of 0.87 (95% confidence interval, 0.73—0.96) for primary tumor recurrence and 0.73 (95% confidence interval, 0.52—0.87) for disease-specific mortality. Three image-based features did, however, contribute significantly when included to the model utilizing clinical variables, which suggests that they may contain additional information that is likely to be of value. Of the five custom features calculated on the dose distribution, the one emphasizing differences in the number of disconnected regions was observed to be significantly higher among non-responders. No statistical differences were found in the sizes of low-dose and high-dose regions, respectively, between the two groups. Conclusion Quantitative analysis of routine scans may provide additional information regarding tumor phenotype, which is likely to be of value when used in conjunction with clinical variables. Additionally, texture analysis of the dose distribution reveals differences between treatment plans that are not captured by dose-volume histogram metrics. These methods are, however, relatively new in use on medical data and there are certain details that require further investigation.