Fig. 1

a Workflow for the classification of volumes extracted from non-contrast CT images based on FDG uptake using radiomic signature (Experiments 1A and 1B). Tumour and adjacent soft tissue were segmented from the SUV map using threshold-based methods and the thyroid tissue was manually segmented on the paired/registered non-contrast CT image. Volumes were classified based on FDG uptake into three categories (High SUV, Low SUV, and negligible SUV). High/Low SUV were localised within the tumour volume (0.50 × Max SUV). Regions of negligible SUV included the soft tissue surrounding the tumour and thyroid tissue. Four sets of radiomic features were extracted within the segmented volumes from either the CT image or All Images (NCCT + Filtered Images). Filtered images included the CT image with applied Laplacian of Gaussian and Wavelet filters. Full details regarding image filtering can be found within the supplement. Feature reduction was performed in MATLAB using the minimum redundancy, maximum relevance (MRMR) algorithm. Tenfold cross-validation was performed (n = 100) and each of the validated models was applied on the testing cohort. b Pipeline for the Clinical Evaluation of Simulated SUV Maps. This pipeline is based on the work performed by Vallières et al. which focused on the application of SUV maps for the prediction of three clinical outcomes: (1) Locoregional tumour recurrence, (2) Distant Metastasis and (3) Death. The pipeline consists of radiomic feature extraction from the tumour regions within the SUV Map. Feature reduction, selection and model training were performed on the training cohort using an imbalance-adjustment strategy that was identical to Vallieres et al. Optimised models were evaluated on the testing cohort