Radiomics-based prediction of tumor phenotype from tumor microenvironment and medical imaging

Magnetic resonance imaging (MRI) and immunohistochemical tissue stainings are pivotal for radiotherapeutic workflows. Yet, recent efforts herald a paradigm shift: Radiomic methods are used to extract a large number of quantitative features from image data to detect high-dimensional patterns, which are correlated with relevant clinical endpoints. Preclinical experiments help to understand underlying mechanisms, yet require the backtranslation of clinically used methods and their application to a heterogeneous patient cohort. In the present preclinical experiment, we determine the tumor phenotype from MRI and tumor microenvironment (TME) features in a patient cohort of xenograft tumor models of the head and neck.
An artificial heterogeneous patient population was created by mixing two tumor models of different radiosensitivity (SAS & UT-SCC-14) in pooled cohort (N = 108) and exposure to one week of fractionated irradiation with photons and protons. After irradiation, contrast agent-enhanced T1-weighted 3D gradient-echo MRI scans were acquired, tumors were excised and characterized immunohistochemically regarding vascularity (CD31), hypoxia (Pimonidazole) and morphology (H&E). Approximately 200 quantitative features were extracted from MRI and light-microscopy image data with an automated medical image radiomics processor and trainable image segmentation, respectively. TME parameters were analyzed regarding effects of radiation with two-sided t-tests. A fully automated radiomic framework was used for feature selection, model generation using leave-one-out cross validation with the individual tumor model’s identity (i.e. its phenotype) as endpoint. Model performance was assessed through area under the curve (AUC).
The used image quantification methods allowed for robust feature extraction. No effects of radiation on the TME were detected except for changes in vessel-adjacent hypoxia. Radiomic analysis was able to predict the tumor model based on TME features (AUC = 0.86), MRI features (AUC = 0.90) or combined features (AUC = 0.86).
We demonstrated backtranslation of radiomic methods in a preclinical setting with multi-modal image data. Further analysis of automatically extracted MRI and TME features may allow for a more biologically informed interpretation of MRI data.