PET/CT radiomics and machine learning for non-invasive molecular and prognostic characterization of oropharyngeal squamous cell carcinomas

Every day, extensive amounts of tomographic imagery including CT-, PET- and magnetic resonance-images are acquired in hospitals. Recent advances in radiomics suggest computer algorithms quantifying size, shape, intensity, and texture of imaging findings may provide added clinical value beyond physicians' qualitative visual evaluation. Radiomics may be a readily available, cost-efficient, and non-invasive means to expand the scope of standard-of-care imaging to include provision of quantifiable, objective biomarkers and prognostic models. The data-driven analysis of radiomic features is often accomplished by ML - a type of AI which is especially capable of exploiting the vast datasets generated in "omics" research. High-risk HPV strains are causally linked to a marked increase in OPSCC incidence around the world in recent decades. HPV-driven cancer affects younger, healthier patients, carries more favorable prognosis, and has recently been assigned a separate TNM classification - compared to frequently tobacco- and alcohol-related HPV-negative OPSCC. HPV testing is routinely performed on tissue samples of OPSCC. To devise non-invasive radiomic HPV biomarkers, we gathered a multicentric, multinational cohort and extracted radiomic features of 435 OPSCC primary tumors and 741 metastatic cervical lymph nodes on pretreatment FDG-PET and non-contrast CT. Combining different sources of radiomics input, feature dimensionality reduction techniques, and ML algorithms, we trained, optimized, and compared 360 candidate HPV classification models, reaching moderate to high performance in cross validation. Twelve select top performing models did satisfactorily generalize to an independent validation dataset, and the best PET-based models were additionally validated in an external set. A model combining radiomic PET and CT features of primary tumors as input and applying ridge regression feature selection and an extreme gradient boosting ML classifier achieved the highest overall performance. Locoregional treatment failure occurs in approximately 10% of HPV-associated cancers and entails worse outcome and morbid salvage therapies. Using a subset of 190 patients with HPV-attributable OPSCC and sufficient follow-up intervals, we pursued a similar analysis approach to develop ML models for prognostication of LRP after definitive therapy in curative intent. We generated and compared models relying on radiomic features or clinical variables including eighth edition UICC/AJCC TNM staging or a combination of both. A random survival forest ML model with radiomic PET and CT features of primary tumors as input was superior, and addition of clinical variables did not improve performance. A random forest classifier using the same radiomics input achieved significant stratification into high- and low-LRP-risk groups. All models were cross validated. Across HPV and LRP analyses, combining radiomic PET and CT features tended to yield higher performance than single-modality models, likely reflecting the complementarity of information gathered by metabolic PET and morphological CT imaging. Conversely, supplementing radiomic features of primary tumors with metastatic lymph node radiomics did not reliably improve model performance. Models relying on lymph node radiomics alone fared worse than primary tumor models in predicting HPV "status". The radiomic HPV biomarkers and LRP prognostication models devised in this work are promising. In the future, non-invasive PET/CT biomarkers may supplement or substitute tissue-based HPV testing, and LRP prediction models may guide therapy planning and alert physicians to increased risk of progression after treatment in certain patients who may especially benefit from tight surveillance. Radiomics and ML could become key enablers of personalized "precision medicine", which may help achieve the next major leap forward in cancer care. Before routine clinical application may be considered, higher model accuracy must be attained, and additional validation in large, prospective studies performed.