Imaging habitat changes after preoperative radiotherapy for glioblastoma: utilisation of a bioinformatics and machine learning pipeline to characterise regional treatment response

Abstract AIMS Imaging of tumour habitats is a novel technique to assess regional treatment response by defining tumour ‘habitats’, which are regions with shared imaging characteristics. Here, we describe a novel pipeline to characterise imaging habitats in glioblastoma using bioinformatics and machine learning. We apply this pipeline to glioblastoma patients in the Preoperative Brain Irradiation in Glioblastoma (POBIG) trial that received preoperative radiotherapy for the first time. METHOD Glioblastoma patients underwent preoperative diffusion/perfusion MRI. A two- stage clustering pipeline was developed to define imaging habitats using relative cerebral blood volume (rCBV) and apparent diffusion co- effcient (ADC). 6 habitats were defined that occupied 10.6-21.5% of tumour volume. Treatment response was evaluated using Ktrans permeability. RESULTS In newly diagnosed cases the commonest habitats included those with a medium and high rCBV. In contrast, recurrent cases had significantly greater proportion of low rCBV habitats (p = 0.026). Two patients in the POBIG trial received 8 Gy preoperative radiotherapy. Within 24 hours of treatment there was a global decrease in vascularity with a decrease in high rCBV habitats (-3.7%) and increase in low rCBV habitats (+8.2%) in irradiated regions. Habitats demonstrated a heterogeneous response to preoperative radiotherapy, with a greater increase in Ktrans in high rCBV habitats (mean +83%) versus low rCBV habitats (mean +40%). CONCLUSIONS We have developed a unique pipeline to delineate imaging habitats in glioblastoma and applied it to a novel trial. Imaging habitats can allow regional monitoring of treatment response and could allow more targeted treatment in future.