Impact of Delineation and Partial Volume Effects Correction on PET Uptake Heterogeneity Quantification Through Textural Features Analysis for Therapy Response in Oncology

International audiencePurpose: characterization of intra tumoral tracer uptake heterogeneity in PET imaging for response to therapy assessment applications in oncology has been investigated in several recent studies. The use of textural features to quantify this heterogeneity has shown promising results for response to therapy prediction. However, there is no study available yet regarding the potential impact of pre-processing steps on the resulting heterogeneity quantification. The goal of this work was therefore to assess the dependency of heterogeneity parameters obtained through textural features on delineation and partial volume effects (PVE) correction (PVC). Methods: fifty patients with esophageal cancer were retrospectively analyzed. PVC of each FDG PET image was performed using iterative deconvolution with wavelet-based denoising. The tumors were delineated using fixed (FT) and adaptive thresholding (AT), and the fuzzy locally adaptive Bayesian (FLAB) algorithm. From the resulting delineations, uptake heterogeneity was quantified using the following features, selected for their reproducibility and robustness: entropy (E), homogeneity (H), dissimilarity (D), intensity variability (IV), size-zone variability (SZV) and high intensity emphasis (HIE). The results obtained with FLAB (the most accurate) on the non-corrected image were chosen as reference. Variability with respect to this reference depending on delineation or PVC was assessed using Bland-Altman analysis. Impact on the associated predictive value regarding the identification on non-responders was assessed by comparing areas under the receiver operating characteristic curves. Results: heterogeneity parameters were more dependent on delineation than PVC. The most sensitive parameters were IV and SZV (90-100% variability). The most independent were E, H and HIE (10-50%). The impact on the corresponding predictive value was not significant, except for SZV and HIE after deconvolution (p