Biomarkers for hypoxia, hpvness, and proliferation from imaging perspective

Recent advances in quantitative imaging with handcrafted radiomics and unsupervised deep learning have resulted in a plethora of validated imaging biomarkers in the field of head and neck oncology. Generally speaking, these algorithms are trained for one specific task, e.g. to classify between two or multiple types of underlying tumor biology (e.g. hypoxia, HPV status), predict overall survival (OS) or progression free survival (PFS), automatically segment a region of interest e.g. an organ at risk for radiotherapy dose or the gross tumor volume (GTV). Despite relatively good performances in external validation cohorts these algorithms still have not found their way into routine clinical practice. The reason this has not happened yet is complex, multifactorial, and can be usually divided into three categories: technical (a part of the algorithm or pre-processing step is not technically sound), statistical (mainly related to selection of subset of relevant biomarkers), and translational (not enough understanding by clinicians, not easily implementable within clinical workflow). We currently foresee that the next artificial intelligence (AI)-driven technique to find its way into clinical practice beside existing techniques (e.g. automatic organ at risk segmentation) will be the automatic segmentation of head and neck gross tumor volumes.