Machine learning identifies multi-parametric functional PET/MR imaging cluster to predict radiation resistance in preclinical head and neck cancer models

Abstract

Abstract Purpose Tumor hypoxia and other microenvironmental factors are key determinants of treatment resistance. Hypoxia positron emission tomography (PET) and functional magnetic resonance imaging (MRI) are established prognostic imaging modalities to identify radiation resistance in head-and-neck cancer (HNC). The aim of this preclinical study was to develop a multi-parametric imaging parameter specifically for focal radiotherapy (RT) dose escalation using HNC xenografts of different radiation sensitivities. Methods A total of eight human HNC xenograft models were implanted into 68 immunodeficient mice. Combined PET/MRI using dynamic [18F]-fluoromisonidazole (FMISO) hypoxia PET, diffusion-weighted (DW), and dynamic contrast-enhanced MRI was carried out before and after fractionated RT (10 × 2 Gy). Imaging data were analyzed on voxel-basis using principal component (PC) analysis for dynamic data and apparent diffusion coefficients (ADCs) for DW-MRI. A data- and hypothesis-driven machine learning model was trained to identify clusters of high-risk subvolumes (HRSs) from multi-dimensional (1-5D) pre-clinical imaging data before and after RT. The stratification potential of each 1D to 5D model with respect to radiation sensitivity was evaluated using Cohen’s d-score and compared to classical features such as mean/peak/maximum standardized uptake values (SUVmean/peak/max) and tumor-to-muscle-ratios (TMRpeak/max) as well as minimum/valley/maximum/mean ADC. Results Complete 5D imaging data were available for 42 animals. The final preclinical model for HRS identification at baseline yielding the highest stratification potential was defined in 3D imaging space based on ADC and two FMISO PCs ($$p<0.001$$ p < 0.001 ). In 1D imaging space, only clusters of ADC revealed significant stratification potential ($$p=0.002$$ p = 0.002 ). Among all classical features, only ADCvalley showed significant correlation to radiation resistance ($$p=0.006$$ p = 0.006 ). After 2 weeks of RT, FMISO_c1 showed significant correlation to radiation resistance ($$p=0.04$$ p = 0.04 ). Conclusion A quantitative imaging metric was described in a preclinical study indicating that radiation-resistant subvolumes in HNC may be detected by clusters of ADC and FMISO using combined PET/MRI which are potential targets for future functional image-guided RT dose-painting approaches and require clinical validation.