Page 130 - Quantitative Imaging of Small Tumours with Positron Emission Tomography

Tumor Delineation An experienced nuclear medicine physician (DO) reviewed all [18F]PSMA PET- CT scans for intra-prostatic tumor localization. A mask was manually drawn around PET-avid intraprostatic tumor volumes to constrain region-growing and prevent inclusion of bladder activity. All masks were reviewed by a second observer. If needed, consensus was reached through joint revision. Next, tumors were delineated using a region-growing algorithm with a background-adapted peak threshold (19). The thresholds were varied incrementally from 50% to 70% (5% intervals). Delineation was performed on original and PVC scans separately to mimic clinical reality. Radiomics Extraction Radiomic features were extracted from the delineated tumors following descriptions of the Image Biomarker Standardization Initiative, as presented by Zwanenburg et al., using the RaCaT software (21,22). Voxel values were scaled to the net injected tracer dosage per kilogram bodyweight (Standardized Uptake 6 Value, SUV). Image voxels and volumes of interest were resampled to 2x2x2 mm isotropic voxels using tri-linear interpolation as recommended (23,24). Per tumor we extracted 480 radiomic features on intensity (n=50), morphology (n=22), and texture (n=408). Intensity features encompassed peak intensity, intensity-based statistics, intensity-volume histograms, and intensity histograms. 2D and 3D textural features based on grey-level co-occurrence matrices (GLCM), grey-level run length matrices (GLRLM), grey-level size zone matrices (GLSZM), grey-level distance zone matrices (GLDZM), neighborhood grey-tone difference matrices (NGTDM), and neighboring grey-level dependence matrices (NGLDM) were extracted. Before textural feature calculation, images were discretized using a fixed bin width of 0.25 SUV starting at SUVmin (23). To compare with radiomics, from the original and PVC PET images we also extracted standard PET features SUVmean, SUVpeak, SUVmax, PSMA-positive tumor volume, and PSMA-total lesion uptake (the product of SUVmean and volume) and used these data as input for the machine learning pipeline. Machine Learning ML algorithms may handle high-dimensional data and/or data with complex non- linear relations with clinical outcomes. We constructed a ML framework in Python 3.6 using Scikit-learn library 0.21 (pipeline in Fig. 6.1) (15,25). As ML model we used a Random Forest classifier (1000 decision trees), a commonly used non-parametric PSMA radiomics and machine learning 129